LLM in Life Sciences News Tracker

AI Scientists & Autonomous Discovery Systems
June 2025 - October 2026
Dr Raminderpal Singh
0 / 0 items
Antibody Design Drug Discovery Platform

Xaira Breaks Two Years of Silence: X-Design Generative Antibody Platform Revealed Alongside First Two Pipeline Candidates, One Cancer Antibody Reported Designed From Scratch in Seven Weeks

Fierce Biotech reported on 2 October 2026 that Xaira Therapeutics, which launched in 2024 with $1 billion in funding and then went quiet, has disclosed its generative antibody design platform, X-Design, together with the first details of its pipeline. Chief Executive Marc Tessier-Lavigne told Fierce that developing a model able to design antibodies that are drug-ready from the start has been the company's major focus, and that X-Design sits alongside two further models, X-Cell and X-Patient; he described the company as tackling all three stages of discovery — identifying targets, making the drugs and identifying the patients most likely to respond. Two candidates were named without disclosing their targets: XA-1, a cancer candidate the chief executive said was created from scratch in seven weeks, and XA-4, aimed at a G-protein coupled receptor described as seemingly intractable. Two further preclinical candidates, XA-2 and XA-3, are also in the pipeline, and the company plans to advance XA-1 and XA-4 into human studies without disclosing timing. Co-founder and Nobel laureate David Baker said in a statement that X-Design shows AI beginning to design from scratch antibodies against proteins that traditional discovery approaches failed on. Contrary view: this is a company blog post and interview, not peer-reviewed data, and none of the central claims can be checked externally — the targets are undisclosed, no binding affinities, developability measurements or hit rates were released, and the seven-week figure covers design rather than a validated development candidate. The claim that most AI design work stops at hits while X-Design optimizes for drug readiness is a marketing differentiator asserted without comparative evidence, and competitors including Absci, Generate:Biomedicines, BigHat and Chugai make structurally similar claims. Tessier-Lavigne declined to disclose the company's remaining cash position, which for a two-year-old company that raised $1 billion and has nothing in the clinic is a material omission. Nothing here changes the tracker's recurring finding on AI antibody design: design speed is demonstrable, downstream clinical translation is not. Source
Drug Discovery Platform Autonomous AI

AstraZeneca and Turbine Authors Argue in Nature Machine Intelligence That the Binding Constraint in Biomedical Discovery Is No Longer Data but Evidence Integration — Commentary on the DeepEvidence Agent

Nature Machine Intelligence published a News & Views piece on 1 October 2026 (DOI: 10.1038/s42256-026-01313-w) by Shruti Shikhare and Jake Cohen-Setton of AstraZeneca Oncology R&D and Krishna C. Bulusu of Turbine Simulated Cell Systems, commenting on the DeepEvidence knowledge-graph research agent published in the same journal on 2 July 2026 (DOI: 10.1038/s42256-026-01266-0) and already recorded in this tracker. The authors' framing is the notable part: they argue that biomedical discovery has entered a period in which the limiting resource is no longer data but the ability to integrate and interpret evidence, and they position DeepEvidence as going beyond retrieving facts to constructing explicit representations of scientific evidence rather than answers. The commentary is indexed under target identification and data integration. Contrary view: this is a three-page commentary, not new evidence, and it is written by authors at a large pharma company and an AI-biology vendor, both of whom have a commercial interest in the claim that evidence-integration software is the field's rate-limiting layer. That claim is also contestable on its face — the Bender and colleagues Perspective recorded in this tracker in August argued that the field's problem is benchmarking against model validation rather than against improved decision making, which locates the constraint in how evidence is used rather than in how much of it can be assembled. An agent that organizes evidence more systematically still inherits whatever biases sit in the underlying knowledge graphs, and no result here demonstrates that a better-organized evidence base changed a target-selection decision or a downstream outcome. Source
Genomics & Proteomics Platform Regulatory

SynthID Bio Watermarks AI-Designed Proteins Without Losing Function — Nature Paper Reports Watermarked Binders Against VEGF-A, SARS-CoV-2 RBD and PD-L1 Matching Unwatermarked Hit Rates and Affinities in Wet-Lab Testing

Google DeepMind announced SynthID Bio on 30 September 2026, with the methods paper published in Nature (DOI: 10.1038/s41586-026-10965-y). The approach embeds a detectable signature into designed protein sequences by guiding amino acid choice, and into predicted structures by adjusting atomic coordinates, with the structure method implemented by fine-tuning part of AlphaFold 3's diffusion network so that predictions carry the signature regardless of who runs the model. Validation used AlphaProteo together with a SynthID Bio-enabled version of ProteinMPNN to design binders against three targets — VEGF-A, the SARS-CoV-2 spike receptor binding domain and PD-L1 — with in vitro work by Adaptyv Bio. The watermarked designs matched unwatermarked versions on hit rate, binding affinity and sequence diversity, which the company describes as the first watermarked and biologically functional protein binders. The stated rationale is DNA synthesis screening: because AI can generate sequences with little resemblance to known hazards, screeners can no longer treat an unfamiliar order as a presumed natural sequence, and a watermark offers an automated signal that an order came from a model with built-in safeguards. James Diggans of Twist Bioscience, who reviewed the work, described watermarking as a promising addition to the biosecurity toolbox. The code, in vitro data and model weights are being released. Contrary view: DeepMind states plainly that robustness against deliberate tampering remains an open problem, which is the case that matters most — a bad actor has every incentive to strip the mark, and a watermark that only survives honest use is a provenance aid rather than a safeguard. The scheme also only covers designs produced by participating models, so it does nothing about open-weight or unregistered generators, and uptake depends on synthesis providers and database curators adopting it, none of which is established. The bacteriophage genome extension with the Hie lab at Stanford and the Arc Institute is described as ongoing work with a manuscript still to come. For drug discovery specifically the relevance is indirect: this is infrastructure for provenance and screening, not a discovery method, though it bears directly on the compliance burden facing any group designing biologics with generative models. Source
Partnership Antibody Design Drug Discovery

Genentech Pays Earendil Labs $55M Upfront in Oncology Bispecific Deal Worth Over $1.5B — Fifth Major Agreement in 18 Months for an AI-Enabled Antibody Platform That Still Has No Disclosed Clinical Asset

Fierce Biotech reported on 25 September 2026 that Roche's Genentech is paying Earendil Labs $55 million upfront to discover and develop multiple bispecific antibodies in oncology, with development, regulatory and sales milestones that could take the total above $1.5 billion plus tiered royalties. Earendil leads antibody discovery and research against pre-agreed target combinations and advances programs through early clinical development before handing them to Genentech for global trials and commercialization. Targets were not disclosed. The deal continues a rapid sequence for Earendil: Sanofi paid $125 million for two bispecifics in April 2025 and a further $160 million in upfront and near-term milestones in January 2026, the company struck an $885 million agreement with WuXi XDC for antibody-drug conjugate linker technology in February and raised $787 million in March. Earendil describes its platform as designing antibody sequences at massive scale with tools for predicting antibody properties and optimizing molecules. Its pipeline is dominated by ADCs with bispecific targeting molecules, plus T-cell engagers against GUCY2C and DLL3 and a CEACAM5xCDH17 bispecific. Correction and update to a prior entry: the only dedicated entry this tracker holds for Earendil concerns a possible Hong Kong IPO; the company has since become one of the most heavily partnered AI biologics platforms in the field. Contrary view: the upfront-to-biobucks ratio here is roughly 27:1, which is less extreme than several deals previously flagged in this tracker but still means Genentech has bought an option rather than committed to the science. The AI contribution is asserted by the company and not independently characterized — no published hit rates, no developability data, no peer-reviewed description of the platform. Five large agreements in 18 months without a disclosed clinical-stage asset is a pattern this tracker has seen before in AI discovery platforms, where deal flow runs well ahead of demonstrated molecules in patients. Bispecific oncology is also a crowded modality where the binding constraint is tumor biology rather than sequence design. Source
Funding Platform Drug Discovery

Basecamp Research Raises $140M Series C at an $800M Valuation With Anthropic and Nvidia Participating — EDEN Biological Foundation Models Redirected Toward In Vivo Cell Therapy, No Clinical Timeline Given

Tech.eu reported on 23 September 2026 that London-based Basecamp Research raised $140 million in an oversubscribed Series C, bringing total funding to roughly $225 million since 2019; Reuters reported the round valued the company at $800 million. The round was led by S32, the firm founded by former GV head Bill Maris, with participation from Nvidia, Anthropic's Anthology Fund, Catalio Capital Management, the NATO Innovation Fund, the Rockefeller Foundation, Singular and others; the company's Series B was $60 million in October 2024. Basecamp said the money will train a new generation of its EDEN biological foundation models and move a pipeline of AI-designed therapies toward clinical development, initially focused on in vivo cell therapy — reprogramming a patient's cells inside the body rather than extracting, engineering and reinfusing them. The stated mechanism is pairing the models' ability to design long and complex DNA sequences with specialized delivery enzymes, with the aim of making cell therapies more customizable and simpler to administer than current treatments the company says cost hundreds of thousands of dollars per patient. EDEN is trained on the Trillion Gene Atlas, a proprietary genomic dataset assembled through partnerships in more than 30 countries. Former Biogen executive Richard Pearce was appointed chief business officer. Contrary view: Basecamp said it has strong preclinical results across several therapy types and diseases but gave no timeline for starting clinical trials and released no data, so the claim cannot be assessed. In vivo cell therapy is among the hardest delivery problems in the field and the step from designing a long DNA sequence to getting it into the right cells in a patient at a therapeutic dose is where the difficulty concentrates; sequence design capability does not address it. An $800 million valuation on a company with no clinical asset continues the pattern of capital concentrating in AI platforms ahead of clinical evidence that this tracker has recorded repeatedly in 2026. Note also that Anthropic now appears in this tracker as model vendor, drug developer, wet-lab operator, acquirer and investor; the concentration of a single AI company across every layer of the stack is itself worth watching. Source
Drug Discovery Platform

SynCraft in Nature Machine Intelligence: Reframing Synthesizability as Atom-Level Editing Rather Than Sequence Generation — LLM Reasoning Used to Rescue Shelved RIPK1 Candidates and Repair Generated Molecules While Preserving Pharmacophores

Junren Li and Luhua Lai of Peking University published SynCraft in Nature Machine Intelligence on 23 September 2026 (DOI: 10.1038/s42256-026-01304-x), addressing a bottleneck this tracker has returned to repeatedly: a substantial fraction of molecules produced by generative models cannot be synthesized. Existing remedies — filtering generated output after the fact, or projecting molecules into predefined synthetic templates — either discard structural novelty or disrupt the pharmacophore. SynCraft instead treats the problem as structural editing, predicting executable sequences of atom-level edits rather than emitting SMILES strings directly, which the authors argue sidesteps the syntactic fragility of language models while using their chemical reasoning. The framework targets what the paper calls the synthesis cliff, where small structural changes produce large gains in synthetic feasibility. Reported results include outperforming state-of-the-art baselines on synthesizable analogue generation with high structural fidelity, comparable performance across proprietary and open-weight model backends, replication of expert medicinal chemistry intuition in editing PLK1 inhibitors, and rescue of high-scoring RIPK1 candidates previously discarded in the generative chemistry literature. The framework is packaged as an agent skill with an end-to-end rescue workflow demonstrated on candidates generated for the SARS-CoV-2 main protease. Code is MIT-licensed on GitHub and datasets are on Figshare. Contrary view: every reported result is computational. Synthetic accessibility is measured by scoring functions and retrosynthesis software, not by anyone making the molecules, and binding-mode preservation is assessed by docking and interaction profiling rather than by assay. The RIPK1 rescue is described as prospective but no synthesis or biochemical data accompanies it. Retrospective replication of expert intuition on a published series is a weaker test than it appears, since the expert answer exists in the literature the backend models were trained on. The work is nonetheless unusually well scoped: it addresses a concrete, acknowledged failure mode rather than claiming end-to-end discovery, and open code with open-weight backend support makes it independently testable. Source
Partnership Drug Discovery Platform

Boehringer Ingelheim and Envisagenics Sign Potential $1B Oncology Pact Built on RNA Splicing — SpliceCore Screens Over 14 Million Splicing Events to Find Tumor-Specific Isoforms for ADCs and T-Cell Engagers

Fierce Biotech reported on 22 September 2026 that Boehringer Ingelheim and Envisagenics agreed a multi-year collaboration with potential value of $1 billion, under which Boehringer will validate tumor-specific targets identified by Envisagenics' SpliceCore platform. SpliceCore combines machine learning with large-scale transcriptomics to screen more than 14 million splicing events and prioritize targets while predicting disease specificity, patient prevalence and therapeutic suitability. The scientific premise is a therapeutic-window argument rather than a potency argument: chief executive Maria Pineda told Fierce that an isoform present only in tumor tissue has a unique region that is clean in normal tissue, which widens the window and reduces off-target effects compared with less selective approaches. Boehringer gains the option to exclusively license selected targets and will assess their potential for antibody-drug conjugates, T-cell engagers and multi-specific antibodies. Upfront and milestone payments were not itemized; Pineda said they will fund taking the company's lead internal CNS candidate into the clinic, and highlighted ENV-375, a preclinical ALS asset. The agreement builds on prior SpliceCore collaborations with Bristol Myers Squibb, Johnson & Johnson and Biogen, and follows Boehringer's $15 million deal with Immunai in June 2025 and an earlier agreement with OpenProtein.AI. Contrary view: no upfront figure was disclosed, so the $1 billion headline carries less information than usual even by biobucks standards, and the structure — Boehringer validating targets, with an option to license — means the pharma partner has bought cheap optionality on computational predictions. Four pharma collaborations on the same platform without a disclosed clinical-stage asset from any of them is the relevant signal: SpliceCore generates target hypotheses, and target hypotheses are the least scarce commodity in oncology. The differential-expression logic that underpins isoform specificity has a long record of producing targets that look clean in bulk transcriptomic data and less clean in patients. Source
Funding Drug Discovery Clinical Trials

Iambic Therapeutics Files to Go Public With $207.9M in the Bank — AI-Discovered HER2 Inhibitor IAM1363 Named Top Spending Priority, Registrational Trial Planned for Early 2027

Fierce Biotech reported on 22 September 2026 that Iambic Therapeutics filed with the Securities and Exchange Commission to go public, without yet disclosing share count or price. The filing names IAM1363 as the top spending priority: an oral, brain-penetrant small-molecule HER2 inhibitor currently in a Phase 1 study in advanced HER2-altered solid tumors, with a registrational trial planned for early 2027. Iambic states in the filing that it believes IAM1363 is the only known HER2 tyrosine kinase inhibitor binding the inactive DFG-out conformation of the HER2 kinase domain, and that this Type II binding mode may account for its selectivity for HER2 over EGFR and its coverage of activating HER2 variants. Early-phase data presented at ESMO in 2025 showed anti-tumor activity in HER2-mutant non-small cell lung cancer and HER2-positive breast and gastric cancers. Two further candidates are named: IAM217, a brain-penetrant allosteric KIF18A inhibitor, and IAM-C1, a dual CDK2/CDK4 inhibitor. The company had 171 employees and $207.9 million in cash at the end of June, and has partnerships with Revolution Medicines, Jazz, Lundbeck, Nvidia, Bayer, AbbVie and Takeda, the last worth up to $1.7 billion in biobucks. This matters to the tracker's central question because an IPO prospectus forces a disclosure standard that AI discovery companies rarely meet voluntarily. Contrary view: the filing is a selling document and the specifics that would settle whether AI did the work — how IAM1363's binding mode was found, what the platform contributed versus conventional medicinal chemistry, and the attrition behind the three named candidates — are not in the public reporting. A Phase 1 asset with a registrational trial planned is real progress relative to platform announcements, but it remains Phase 1. Iambic is one of several companies filing into a crowded 2026 biotech IPO window, and the timing suggests market conditions as much as pipeline readiness. Source
Drug Discovery Platform

Stanford Transfers Antimicrobial Peptide Knowledge to Polymers and Synthesizes Ten Potent Candidates — Active Learning Trained on Model Disagreement Narrows 1.7 Million Molecules, All Ten Beat Expectations Against E. coli

Stanford announced on 21 September 2026, with the paper published in Matter (DOI: 10.1016/j.matt.2026.103026), a machine learning workflow that identified antimicrobial polymers by transferring knowledge from antimicrobial peptide data. Senior author Eric Appel and first author Shoshana Williams targeted polymers rather than peptides because peptides are short-lived and expensive to synthesize while polymers are cheap, stable, safe and distributable to low-resource settings. Both classes kill bacteria physically — permeabilizing the membrane rather than inhibiting a specific protein or pathway — which is the proposed basis for evading resistance, since altering membrane lipid structure or surface charge is a harder evolutionary step than rejecting a chemical drug. The data problem was that antimicrobial polymer datasets are small while peptide datasets are large, so the team featurized thousands of evaluated peptides and connected them to standardized polymer outputs. Their active-learning strategy is the methodologically interesting part: rather than synthesizing the candidates models agreed on, they synthesized the 20 candidates on which models disagreed most, tested them, and fed the results back. The refined model narrowed a library of 1.7 million candidates to 10, all of which were synthesized and all of which inhibited E. coli above expectation, with one particularly effective against biofilms; activity against Staphylococcus aureus was also reported. Contrary view: this is in vitro work against laboratory strains, with no animal data, no pharmacokinetics and no toxicity profile, and membrane-disrupting agents have a long history of being potent in a dish and unacceptably toxic to mammalian cells in vivo — selectivity for bacterial over human membranes is the question that decides whether any of this becomes a drug, and it is not answered here. The claim that these are among the most potent antimicrobial polymers reported is attributed to the first author's own awareness rather than a systematic comparison. The resistance-evasion argument is mechanistic reasoning, not an experimental result: no serial-passage resistance selection is reported. The deliberate-disagreement sampling strategy is the transferable contribution and is directly applicable wherever labeled data are scarce. Source
Autonomous AI Drug Discovery Clinical Trials

Stanford's Virtual Biotech Published in Science — Up to 37,000 Agents Re-Annotate 37,075 Phase II/III Trials in Six Hours and Find Drugs Against Cell-Type-Specific Switch-Like Targets 48% More Likely to Reach Market

Zhang, Eckmann, Miao, Mahon and Zou published the Virtual Biotech in Science on 17 September 2026 (DOI: 10.1126/science.aeg6779), scaling the group's earlier Virtual Lab from a single research question to a simulated drug-development organization. A chief scientific officer agent receives a query and delegates to specialized scientist agents organized into four divisions — target discovery and validation, safety assessment, modality selection and clinical trial review — each with its own databases and tools, including Open Targets and ClinicalTrials.gov, with independent reviewer agents checking findings. Asked to extend prior work showing that genetic evidence predicts trial success, the system first identified data quality as the obstacle, then assigned one agent per trial to re-annotate outcomes for 37,075 Phase II and III trials from registries, papers and press releases, completing in about six hours. A second agent built a two-part gene scoring system measuring cell-type specificity and whether expression is switch-like or graded. Combining the two produced the headline association: drugs aimed at switch-like genes active in few cell types were 48% more likely to reach market, 40% more likely to advance from Phase 1 to Phase 2, and showed 32% fewer adverse events. Applied to B7-H3 in lung cancer using only data available before January 2025, the system proposed an antibody-drug conjugate strategy; in August 2025 ifinatamab deruxtecan received FDA breakthrough therapy designation on the same strategy, which Zou characterized as independent third-party validation. Contrary view: the B7-H3 result is retrospective validation against an outcome the authors knew before running the analysis, and B7-H3 ADCs were an active, heavily pursued area well before January 2025 — more than 60 programs were in development by early 2026 — so arriving at that strategy is less discriminating than the framing suggests. The 48% figure is an association across heterogeneous historical trials, with confounding by indication, sponsor and era that re-annotation does not remove, and the annotation itself was performed by the same class of model whose output is being evaluated. Target selection is one step; the paper produces no molecule, no experiment and no prospective test. The six-hour runtime is a genuine result about throughput, not about scientific validity. Source
Autonomous AI Genomics & Proteomics Platform

Paper2Agent in Nature Converts Published Methods Into Executable Agents — 74 of 100 Computational Biology Papers Agentified, 593 of 599 Tools Passed Validation, and Three Paper Agents Collaborated to Prioritize GPR137 for Psoriasis

Miao, Davis, Zhang, Pritchard and Zou published Paper2Agent in Nature on 16 September 2026 (DOI: 10.1038/s41586-026-11044-y, open access), an automated framework that converts a research paper and its codebase into a Model Context Protocol server and connects it to a chat agent, so that a published method can be invoked in natural language rather than cloned, configured and debugged. The pipeline uses specialized sub-agents for environment setup, tutorial discovery, execution, tool extraction and testing, with tools that fail validation excluded rather than shipped. Scale results: of 100 computational biology papers sampled from bioRxiv without filtering for code quality, 74 were agentified, producing 599 tools of which 593 passed automated validation; failures were dominated by missing code, missing data artifacts and unresolvable environments. The AlphaGenome agent was built in about 45 minutes for roughly $14 and reached 98.7% accuracy on tutorial-derived queries and 100% on novel ones, against 82.7% and 78.7% for Claude Code with direct repository access. In the discovery case study, three paper agents — AlphaGenome, an MPRA-coupled scCRISPRi screen and a CD4+ T cell Perturb-seq dataset — were combined to prioritize GPR137 at the psoriasis-associated rs887314 locus, with concordance between the regulatory-element perturbation signature and GPR137 knockdown reaching significance only under stimulation. The authors propose an "agent availability" section in papers alongside data and code availability. Contrary view: the validation standard is reproduction of the paper's own tutorial outputs, which measures faithful execution rather than scientific correctness — a tool that reliably reproduces a flawed analysis passes. The 26% agentification failure rate reflects the state of published research code and will not improve because this framework exists. The psoriasis result is a computational prioritization built on three published datasets, not new experimental evidence, and the authors say explicitly that hypothesis generation and mechanistic interpretation remain human-in-the-loop; the human researcher selected the validation strategy from ten the agent proposed. Exposing research code as callable agents also raises security and attribution questions the authors acknowledge but do not resolve. For pre-clinical groups the practical significance is lower friction in reusing published computational methods, which is a real constraint, rather than any new discovery capability. Source
Autonomous AI Genomics & Proteomics Platform

Anthropic Launches Biology Wet Lab and Reports First Find — Roughly 950 AI Agents Spent Over 21 Hours Mining Sequence Databases and Surfaced CRISPR-Like Repeat Arrays in Giant Viruses, Function Still Unknown

Nature reported on 25 September 2026 (Heidi Ledford, DOI: 10.1038/d41586-026-03039-6) that Anthropic has launched a biology wet lab in which human scientists and AI agents design and run experiments together, and has announced one of the group's first findings: a repeated short DNA sequence in the genomes of several giant viruses. The company announced the lab and released the associated preprint on 23 September; the work is posted on alphaXiv and bioRxiv and has not been peer reviewed. Methodologically, researchers directed roughly 950 AI agents to survey a database of DNA sequences encoding billions of proteins, looking for proteins that might act together with reverse transcriptases. The agents spent more than 21 hours exploring in a self-directed fashion, discussing preliminary results and deciding next steps among themselves. While examining DNA surrounding a reverse transcriptase gene in a giant virus, they noticed the same short sequence recurring, judged it worth closer inspection, and went on to find similar patterns in other viral genomes. The repeat arrays resemble those in bacterial CRISPR systems, where repeated DNA chunks separated by invader-derived segments guide a partner enzyme to cut intruding DNA. Eric Kauderer-Abrams, Anthropic's head of life sciences, framed the goal as systematizing and scaling the kind of research that comes from finding unusual things in microbes, and described the result as a promising lead. Contrary view: the function of these repeats is unknown, there is scant evidence they behave like CRISPR arrays, and no partner DNA-slicing enzyme has been identified — the distance between spotting a repeat pattern and characterizing a usable genome-editing tool is where essentially all the difficulty lies, as Kauderer-Abrams himself notes. The work is a preprint from a company with a direct commercial interest in demonstrating agent capability, and the headline framing of 950 agents over 21 hours is a compute-scale claim rather than a scientific result. Database mining for recurring motifs is a task unusually well suited to parallel scale and pattern matching; it is search rather than hypothesis-driven discovery closed by experiment. Despite the wet lab being the announcement's centerpiece, no experimental validation of this finding has been reported. Relevance to therapeutics is also indirect: genome-editing tools sit upstream of drug programs, and this tracker has recorded several AI-discovered editors that have yet to reach the clinic. Source
Partnership Platform Drug Discovery

Federated OpenFold3 Results: Five Pharma Companies Fine-Tune a Co-Folding Model on 20,000+ Proprietary Structures Without Pooling Data — AISB-1-Fed Reported to Beat Every Public Model on Held-Out Private Structures

Bio-IT World reported on 22 September 2026 on results from the Federated OpenFold3 Initiative, run by Apheris through its AI Structural Biology network in collaboration with the AlQuraishi Lab at Columbia University, which develops OpenFold3. AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson and Takeda each fine-tuned OpenFold3 Preview 2 locally inside their own environments on structures from active drug-discovery programs, with only model parameters periodically aggregated centrally; no structures were pooled. The combined training data spanned more than 20,000 experimentally determined protein-ligand structures, which the organizers state roughly triples the drug-relevant co-folding data available for training. The resulting model, AISB-1-Fed, is reported to outperform the foundation model it started from, every public model tested, and every model fine-tuned on a single partner's data alone, evaluated on 1,056 held-out private structures using the fraction of predictions meeting PL-lDDT of at least 0.8 and ligand RMSD within 2 Å. Nature covered the initiative on 14 September (Ewen Callaway, DOI: 10.1038/d41586-026-02882-x). The organizers' framing of why private data matters is itself notable: since the AlphaFold3 training cutoff the Protein Data Bank has grown by roughly 70,000 structures, but almost all are cofactors, metabolites, ions and crystallographic additives, with only around 10,600 public structures containing an approved or investigational drug and roughly 3,000 of those new since the cutoff. A follow-on effort, AISB Bind, extends the approach from structure to binding prediction with AbbVie, AstraZeneca, Bristol Myers Squibb and Johnson & Johnson. Contrary view: the results paper is not peer reviewed and was published on the vendor's own website, and Apheris is a commercial federated-computing provider with a direct interest in the finding that federation works. The benchmark consists of held-out private structures from the same five partners, so it measures fit to those companies' own medicinal-chemistry series rather than generalization to genuinely novel chemical space — the local-model problem that the Bender and colleagues Perspective covered in the prior cycle identifies as endemic. Dense analog series around targets a company already works on are exactly the data most likely to flatter a held-out split. Because the underlying structures and the trained weights remain private to participants, no external group can verify the result. And this remains structure prediction: the tracker has repeatedly noted that co-folding accuracy is an imperfect proxy for binding affinity, developability or in vivo behavior, and no evidence has been presented that the improvement changed a discovery decision. Source
Drug Discovery Platform

Nature Reviews Bioengineering: Machine Learning in Nanoparticle Drug Delivery — Review Maps the Shift From Trial-and-Error Formulation to Computationally Guided Design Across Discovery and Preclinical Nanomedicine

Nature Reviews Bioengineering published a review (DOI: 10.1038/s44222-026-00495-7) on the role of artificial intelligence and machine learning in nanoparticle-based drug delivery systems. Its central claim is that ML is shifting drug development pipelines away from traditional trial-and-error discovery toward computationally guided architectures that accelerate identification of optimal formulations, and it surveys applications across the discovery and preclinical phases of the nanomedicine pipeline. The review draws together a body of work this tracker has followed in pieces, including the LUMI-lab self-driving laboratory that discovered brominated-tail ionizable lipids for mRNA delivery, the AGILE platform for accelerating lipid nanoparticle development, transformer-based neural networks for lipid nanoparticle design, and the combination of machine learning with combinatorial chemistry to screen virtual ionizable-lipid space. Delivery is a genuine rate-limiting step for RNA therapeutics, so formulation optimization is one of the few areas where computational gains translate fairly directly into a developable product. Notably, this is also a domain that partly escapes the central criticism made in the Bender and colleagues Perspective covered alongside it in this cycle: lipid nanoparticle datasets are typically generated by deliberate experimental design across combinatorial libraries with direct functional readouts, rather than assembled opportunistically from heterogeneous public sources. Contrary view: this is a review rather than primary research, so it consolidates rather than extends the evidence base. Formulation optimization is a materially easier problem than target identification or molecular design — the search space is constrained, the readout is direct, and the feedback loop is short — so success here should not be read as evidence that AI works equally well across drug discovery. Nanomedicine also carries a long history of preclinical promise that failed to translate, most prominently the enhanced permeability and retention effect, whose generality in human tumors has been contested for over a decade. ML-optimized formulations still face the same in vivo translation gap as everything else: potency in a cell assay or a mouse model is not delivery in a patient, and none of the platforms surveyed has yet produced an approved product. Source
Drug Discovery Platform Clinical Trials

Nature Reviews Drug Discovery Perspective: Evidence of AI's Clinically Relevant Impact Is "Disappointingly Limited" — Bender, Scannell, Grisoni and 13 Co-Authors Argue Benchmarking Must Shift From Model Validation to Improved Decision Making

Andreas Bender, Morgan C. Thomas, Jack W. Scannell, David A. Shaywitz, Francesca Grisoni, Isidro Cortes-Ciriano and ten further co-authors published a 22-page Perspective in Nature Reviews Drug Discovery on 7 August 2026 (DOI: 10.1038/s41573-026-01496-2) offering a critical assessment of a decade of AI in drug discovery. Their conclusion is blunt: although a wide variety of AI methods have been developed, applied and benchmarked, evidence of clinically relevant impact is so far disappointingly limited. The diagnosis has several parts. Effort concentrates on preclinical stages where labeled data happen to exist, which the authors compare to looking for keys where the light is rather than where the keys are; their cost modeling indicates that reducing failure rates in phase II has by far the greatest effect on capitalized cost per successful launch, and that projects using biomarkers for patient stratification cost roughly half as much per launch as those that do not. Biological data are highly conditional and epistemically opaque, so labels are often unreliable in ways that propagate into every supervised model built on them. Chemical space is effectively unsamplable at roughly 10^60 small molecules, and the authors show that four commonly used ADME benchmark datasets share only 0.1% of compounds and 0.6% of scaffolds, leaving models with incompatible applicability domains. A dedicated box argues that ligand discovery and drug discovery are routinely conflated, noting there are more than a million bioactive ligands in ChEMBL and PubChem against roughly a thousand marketed drugs. The core recommendation is that benchmarking studies must move on from model validation and instead measure whether a model improves decision making, formalized as model value being the product of improvement in clinical success rate and project applicability domain. The Perspective also addresses items this tracker carries directly, noting that attributing phase IIa successes such as the generative-AI-discovered TNIK inhibitor to AI alone remains difficult because the lead-optimization steps linking AI-generated hits to final candidates are often insufficiently described or undisclosed, and that Isomorphic Labs' IsoDDE claims superior performance while offering too few technical details to permit independent verification. Note: this entry falls outside the strict cycle window (14-19 September 2026) and is included as deliberate backfill; it is the most directly relevant peer-reviewed assessment of this tracker's own subject matter published in 2026, and its omission until now was an oversight. Contrary view: the authors are careful to distinguish absence of evidence from evidence of absence, and readers should hold that distinction too — the earliest genuinely AI-originated programs entered the clinic around 2020 and 2021, so the cohort remains too young for approvals. Several co-authors have long-standing public positions as critics of AI-in-drug-discovery claims, and the Perspective is argumentative rather than a systematic review, with no quantitative meta-analysis of AI versus non-AI program outcomes. The recommendation to benchmark on decision quality rather than model accuracy is correct in principle but difficult to operationalize, since decision quality in drug discovery is typically only legible years later and is confounded by every other choice in the project. Source
Genomics & Proteomics Drug Discovery Platform

Nature Machine Intelligence: NucleicBERT Interprets RNA Sequence Space Through Self-Supervised Language Modelling — Deriving Biologically Meaningful Structure and Function Signals From Sequence Correlations Where Annotations Are Scarce

Upadhyay, Herold and Schug published in Nature Machine Intelligence on 3 September 2026 (DOI: 10.1038/s42256-026-01295-9, open access) NucleicBERT, a self-supervised language model trained on large-scale RNA sequence data. The problem it addresses is a structural asymmetry in RNA biology: sequence data is abundant while structural and functional annotation is scarce, so supervised approaches are constrained by the small labelled fraction. NucleicBERT instead learns from sequence correlations alone, with the authors reporting that biologically meaningful patterns relating to RNA structure and function emerge from that self-supervised signal. The work belongs to the RNA-therapeutics thread this tracker has been building — alongside the IRES-LM framework for programmable cap-independent translation and the Alnylam-Inceptive siRNA foundation-model partnership — where sequence-level design capability is the rate-limiting step for mRNA, circRNA and oligonucleotide payload engineering. Note: this entry falls marginally outside the strict cycle window (6-13 September 2026) and is included as catch-up backfill; it was published three days before the window opened and was not captured by the previous sweep. Contrary view: emergent biological signal from self-supervised sequence modelling is a claim that has been made repeatedly across protein and DNA language models, and the recurring difficulty is that the evaluations which demonstrate it are often the same annotation-scarce benchmarks the model was meant to circumvent. RNA structure prediction in particular has a history of models performing well on curated families while failing on novel folds, since RNA secondary and tertiary structure is strongly context-, ion- and temperature-dependent in ways sequence alone cannot capture. Without prospective experimental validation — which does not accompany this work — the distance between a biologically meaningful representation and a therapeutically actionable design tool remains substantial. The tracker has applied this same caution to every foundation model it carries, and nothing here warrants an exception. Source
Drug Discovery Platform

Nature Machine Intelligence: Uncertainty-Calibrated Language Models as Optimizers for Experimental Discovery — Training Molecular-Design LLMs to Account for Data Uncertainty Rather Than Emit Unqualified Predictions

Ranković, Griffiths and Schwaller published in Nature Machine Intelligence on 28 August 2026 (DOI: 10.1038/s42256-026-01283-z, open access) a method for training language models as uncertainty-calibrated optimizers for experimental discovery. The premise is a specific and consequential gap: language models have become genuinely useful in molecular design, but they are not typically calibrated for uncertainty, so they return suggestions without a defensible statement of how much confidence those suggestions warrant. In an experimental-design loop — where each proposed molecule consumes synthesis and assay capacity — uncalibrated confidence is not a cosmetic problem but a direct driver of wasted wet-lab effort. The authors present a training approach that takes the uncertainty of the underlying data into account, positioning the model as an optimizer within a discovery loop rather than a generator of unranked candidates. This connects directly to two items already tracked here: van Tilborg and Grisoni's "unfamiliarity" metric for quantifying when a molecular model has left its training chemical space, and the RXNGraphormer reusability report's finding that retrosynthetic prediction degrades under distribution shift while forward prediction holds. Together these mark a shift in the field's attention from raw predictive performance toward knowing when a prediction should be trusted. Note: this entry falls outside the strict cycle window (6-13 September 2026) and is included as catch-up backfill from late August. Contrary view: calibration methods are evaluated on how well predicted confidence matches observed error on held-out data, which is itself an in-distribution property — a model can be well calibrated on familiar chemistry and confidently wrong on novel chemistry, which is precisely the regime that matters for discovery. Bayesian-optimisation-style framing of molecular design has a long history in cheminformatics and repeated demonstrations that it works on retrospective benchmarks; the shortage has always been prospective campaigns where the method chose what to synthesise and the outcome was recorded either way. No such prospective validation accompanies this work. Calibrated uncertainty improves triage and honesty; it does not extend a model's reach into chemistry it has never seen. Source
Drug Discovery Antibody Design Platform

Nature Machine Intelligence: VITAL Dual-Channel Framework for Quantitative, Interface-Aware Peptide–Protein Interaction Prediction — Co-Learning Sequence and Structural Context to Map Binding Interfaces and Estimate Affinity

Chen, Wang, Li and Ji published in Nature Machine Intelligence on 19 August 2026 (DOI: 10.1038/s42256-026-01291-z) VITAL, a dual-channel deep learning framework that co-learns sequence and structural contexts to predict peptide–protein interactions quantitatively, map the residues forming the binding interface, and estimate binding affinity. The dual-channel design targets a known limitation in the peptide-interaction literature: sequence-only methods are broadly applicable but coarse, while structure-dependent methods are more precise yet constrained by the availability of high-resolution complex structures, which are especially scarce for peptides given their conformational flexibility. Producing interface maps and affinity estimates alongside binary interaction calls is the practically useful part, since peptide therapeutic programmes need to know not merely whether a peptide binds but where and how strongly. Peptide therapeutics remain an under-represented modality in this tracker relative to small molecules and antibodies, despite sustained commercial interest. Note: this entry falls outside the strict cycle window (6-13 September 2026) and is included as catch-up backfill from mid-August. Contrary view: peptide–protein interaction prediction is evaluated against curated benchmark sets drawn from the same structural databases that constrain the field, so reported gains partly measure fit to an existing, biased sample of solved complexes rather than generalisation to the flexible, transient interactions that characterise much peptide biology. Affinity estimation from sequence and structure remains among the least reliable tasks in computational biology — errors of an order of magnitude are common across methods — and quantitative claims warrant particular scrutiny. Interface mapping is a prediction about geometry, not a measurement, and no wet-lab validation accompanies the work. As with the structure-prediction lineage this tracker follows through AlphaFold3, Chai and HelixFold, accurate interface geometry has repeatedly proven an imperfect proxy for developability or in-vivo behaviour. Source
Drug Discovery Platform

Nature Machine Intelligence Reusability Report: Independent Team Reproduces RXNGraphormer and Tests It on External Data — All Major Results Replicate, but Retrosynthesis Degrades Under Distribution Shift While Forward Prediction Holds

Zhang, Zeng, Lin and Liao (Guangzhou National Laboratory and AIChemEco) published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01257-1; Nat Mach Intell 8, 1066-1074) an independent reusability report on RXNGraphormer, the unified graph-transformer framework for cross-task reaction performance prediction and synthesis planning published by Xu and colleagues in the same journal in August 2025 and trained on more than 13 million reactions. Using only the released implementation, pretrained checkpoint and benchmark datasets, the team reproduced all major regression and sequence-generation results, including the relative difficulty patterns in out-of-sample evaluations — an outcome the authors read as evidence of the stability and transparency of the original published workflow. They then extended the assessment beyond the original setting: on several standardised high-throughput experimentation datasets (sulfoxonium-ylide, hindered meta-C-H arylation and amide-coupling), the pretrained encoder adapted efficiently and performed strongly with minimal fine-tuning; on an external sequence-prediction benchmark beyond the original USPTO data, forward prediction held up well while retrosynthetic prediction proved markedly more sensitive to distributional shift. Reproduction code is released via Zenodo and GitHub. Note: this entry falls outside the strict cycle window (26 August - 5 September 2026) and is included as deliberate backfill of a peer-reviewed item the tracker had flagged but not previously dated. Its inclusion is editorially deliberate: independent reproduction studies are rare in AI-for-chemistry, and this tracker has attached generalisation caveats to nearly every molecular-ML entry it carries without often being able to point to a formal replication attempt. Contrary view: a reusability report is not an adversarial audit — the authors used the original team's own checkpoint and benchmarks, so this establishes computational reproducibility and transfer, not independent validation of the underlying scientific claims from scratch. Successful reproduction from a released checkpoint is a considerably weaker test than retraining from raw data, which was not attempted. The finding that retrosynthesis degrades under distribution shift while forward prediction persists is the report's most useful result and reinforces a pattern seen across the field, but the external benchmarks remain literature-derived and curated rather than prospective. Both the original work and this report originate from Chinese chemical-AI groups working within a shared benchmarking culture, so genuinely adversarial replication from an unrelated methodological tradition would be more informative still. Source
Genomics & Proteomics Drug Discovery Platform

Nature Machine Intelligence: CrossDNA Models Explicit Double-Strand Interactions in DNA Language Modelling — A Few Million Parameters Match or Beat Models Hundreds of Times Larger on Regulatory-Region and Non-Coding Variant Tasks

Yang, Liu and colleagues published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01249-1) CrossDNA, a DNA sequence language model that models cross-strand interactions explicitly and dynamically rather than treating DNA as unidirectional text or enforcing reverse-complement symmetry through data augmentation or architectural equivariance. The authors argue that both prevailing approaches approximate double-strand interactions implicitly and statically, which limits the capture of context-dependent information exchange between strands — whereas in biological systems duplex coordination is shaped by physical coupling, functional synergy and information transfer. CrossDNA uses dual branches with alternating exposure to forward and reverse-complement segments as a duplex-inspired training strategy, establishes interstrand communication through a lightweight cross-strand communication module, and couples a recurrent long-context backbone with sliding-window attention to support long sequences. The headline efficiency claim is that CrossDNA, at only a few million parameters, matches or surpasses models with hundreds of millions of parameters across diverse genomics benchmarks, substantially reducing training and inference cost. Relevance to drug discovery is via regulatory-region interpretation and non-coding variant prioritisation — the same translational bottleneck covered in this tracker through Sanjana's CRISPR-perturbation work on connecting GWAS variants to causal genes. Note: this entry falls outside the strict cycle window (26 August - 5 September 2026) and is included as deliberate backfill of a peer-reviewed item previously flagged but undated. Contrary view: parameter efficiency is a real and underrated contribution, but matching much larger models on existing genomics benchmarks partly reflects that those benchmarks are saturating and may not discriminate well between architectures — a criticism levelled at the DNA-language-model field generally. Non-coding variant prioritisation is a ranking task evaluated against curated variant sets whose ground truth is itself derived from prior computational and statistical annotation, so strong benchmark performance does not establish that the model identifies causal variants that experimental perturbation would confirm. The biological motivation for explicit cross-strand modelling is elegant but the paper demonstrates benchmark gains rather than a mechanistic result, and no wet-lab validation accompanies the work. Whether duplex-aware representations improve real target-identification decisions remains untested. Source
Drug Discovery Genomics & Proteomics Platform

Nature Machine Intelligence: MAP Knowledge-Driven Framework Predicts Single-Cell Responses for Unprofiled Drugs — MAP-KG Unifies 14 Resources Across 187,089 Drugs and 694,246 Mechanistic Relationships, Enabling Zero-Shot Perturbation Prediction

Feng, Zhao, Zhang, Xie and colleagues published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01286-w) MAP, a framework that integrates structured biological knowledge into cellular perturbation modelling and supports zero-shot prediction for small molecules with scarce or absent experimental profiles. The problem it targets is a real constraint on the virtual-cell programme: experimentally profiled compounds cover only a small fraction of chemical space, and existing perturbation models generalise poorly to unprofiled drugs because they treat compounds as isolated identifiers rather than encoding mechanistic relationships. MAP has three components: MAP-KG, a knowledge graph unifying 14 public resources spanning 187,089 drugs, 22,924 genes and 694,246 mechanistic relationships; a knowledge-driven pretraining strategy that aligns molecular structures, protein sequences and textual mechanistic descriptions into a shared embedding space to produce mechanism-aware, transferable gene and compound embeddings; and coupling of those representations to a pretrained single-cell foundation model that conditions response prediction. Evaluated under two zero-shot regimes — unseen cell-type/drug combinations, and the stricter unprofiled-drug setting — MAP reports improvements in top-50 differentially-expressed-gene Pearson delta correlation of up to 13.3% and 12.2% respectively over the strongest baselines across three benchmarks. Pathway-level GSEA analysis for in-silico screening is reported to yield mechanism-consistent programmes for unprofiled candidates, prioritising four of five approved anti-cancer drugs in A-549 non-small-cell lung cancer. The work extends the perturbation-modelling thread this tracker has followed through CMonge and related conditional optimal-transport methods, but attacks generalisation via mechanistic priors rather than transport geometry. Contrary view: knowledge-graph grounding imports the coverage biases of its 14 source databases, so mechanism-aware embeddings will be richest exactly where the literature is already dense and thinnest for the novel targets where prediction would be most valuable — the same structural limitation flagged for DeepEvidence and MIDAS. Recovering four of five approved anti-cancer drugs in a single well-characterised cell line is a retrospective enrichment result on compounds whose activity is already documented in the underlying knowledge graph, which risks circularity, and it is a modest experimental bar relative to prospective identification of an unknown active. Improvements of 12-13% in DEG correlation are meaningful but incremental, and the field still lacks agreement on whether DEG correlation predicts anything a medicinal chemist would act on. No prospective wet-lab validation accompanies the paper. Source
Regulatory Clinical Trials Drug Discovery

Insilico Receives First-Ever Fast Track Designation in Its Pipeline — FDA Grants FTD to ISM6331, an AI-Designed Pan-TEAD Inhibitor for Advanced Mesothelioma, Establishing a Second Clinical Asset Beyond Rentosertib

The FDA granted Fast Track Designation to ISM6331, Insilico Medicine's pan-TEAD inhibitor for advanced mesothelioma, in an announcement dated 28 July 2026. This is the first FTD granted to any programme in Insilico's AI-driven pipeline. ISM6331 is a potential best-in-class pan-TEAD inhibitor engineered via Chemistry42 to restore Hippo pathway balance; it received Orphan Drug Designation for mesothelioma in June 2024, dosed its first patient in a global multicentre Phase I trial in January 2025, and its first-in-human Phase I data has been accepted for a brief oral presentation at ESMO 2026. Fast Track confers more frequent FDA meetings and written feedback on trial design and biomarker strategy, plus eligibility for Rolling Review, Accelerated Approval and Priority Review if criteria are met. Note: this entry falls outside the strict cycle window (24-30 August 2026) and is included as deliberate backfill. This tracker treats Insilico as its live source-concentration risk (16+ dedicated entries, the most of any organisation covered) and has excluded or concentration-framed its recent partnership announcements; this item qualifies on the stated exception for regulatory actions. It is also analytically significant for a reason separate from the designation itself: the tracker has repeatedly noted that Insilico's entire clinical thesis rested on rentosertib as a single validated asset, and ISM6331 reaching FTD with Phase I data heading to ESMO establishes a genuine second programme — narrowing, though not closing, that criticism. Contrary view: Fast Track is a process designation, not evidence of efficacy — it reflects the seriousness of the indication and unmet need rather than any judgement on ISM6331's clinical performance, and the FDA grants it liberally (hundreds annually) with no implication that approval is more likely. Mesothelioma is a small, poor-prognosis indication where the unmet-need threshold is easily met, so the designation says more about the disease than the drug. Phase I first-in-human data presented as a brief oral at a congress is early, unblinded and typically dose-escalation safety rather than efficacy. Two clinical assets is still a thin base from which to claim platform reproducibility, and TEAD inhibitors as a class have a mixed record, with several programmes at other sponsors encountering tolerability ceilings. The Hippo-pathway rationale is well-founded biology but not AI-specific; what AI contributed here was molecular design, not target discovery. Source
Drug Discovery Clinical Trials

Pipeline Accounting at the Eight-Year Mark: No AI-Discovered Medicine Approved in Any Major Market — ASCO Analysis Counts 117 AI-Enabled Assets Across 63 Companies in Human Trials, With 60 Past Phase I and Only 8 Past Phase II

The Pharmaletter published an assessment of the AI drug discovery field's cumulative clinical record, framing August 2026 as roughly eight years into the boom with no medicine discovered or designed using AI yet approved by a regulator in any major market. What the field has accumulated, on this accounting, is capital, trial starts and a thin first layer of human data. The piece notes that counting the pipeline is harder than it appears: one analysis presented at the American Society of Clinical Oncology this year identified 117 AI-enabled assets across 63 companies that had entered interventional human trials, of which 60 had completed Phase I and only eight had completed Phase II — an attrition profile that, on its face, does not yet distinguish AI-originated programmes from conventional ones. Other trackers put the programme count nearer 175, with roughly $60 billion invested since 2019 and an estimated 15 to 20 candidates potentially entering pivotal Phase III trials during 2026. The divergence between the 117 and 175 figures is itself informative, reflecting unresolved definitional questions about what counts as "AI-enabled" versus "AI-discovered" — a distinction this tracker has repeatedly flagged, since a molecule whose target was nominated by AI, whose structure was generated de novo, and whose development was AI-guided (as with Insilico's rentosertib, now in Phase III) sits in a materially different category from one where machine learning assisted a conventional medicinal-chemistry campaign. Note: this is a trade-press analytical assessment rather than a peer-reviewed disclosure, and the underlying article is partially gated. Contrary view: "no approvals after eight years" is a rhetorically powerful framing but a weak analytical one, because typical discovery-to-approval timelines run ten to fifteen years and the earliest genuinely AI-originated programmes only entered the clinic around 2020-2021 — the cohort is simply too young for approvals, so their absence is close to uninformative. Conversely, the eight-past-Phase-II figure is the number that should worry advocates: Phase II is where AI's claimed advantage in target selection ought to show up as improved translational success, and eight completions is far too small a denominator to demonstrate anything either way. Both the optimistic and pessimistic readings are currently underdetermined by the data, and will remain so until the pivotal readouts arrive around 2029. Readers should also note that ASCO-presented pipeline analyses are conference-tier disclosures, not peer-reviewed epidemiology, and company self-reporting of AI involvement is unaudited. Source
Partnership Platform Infrastructure

Novo Nordisk Names AWS Preferred Cloud and Strategic AI Partner — Agentic AI for Target Identification and Candidate Ranking, London Co-Innovation Hub, Four Months After the OpenAI Deal

Novo Nordisk and Amazon Web Services announced on 10 August 2026 (Bagsværd, Denmark) a strategic partnership establishing AWS as Novo Nordisk's preferred cloud provider and strategic AI partner, with the stated aim of compressing the path from drug target to first human dose using agentic AI and cloud technologies. The centrepiece is a co-innovation hub in London where AWS engineers and Novo Nordisk scientists work jointly on AWS AI technologies applied to Novo Nordisk data; reported use cases include identifying potential drug targets, designing therapies, analysing biological information, and generating and ranking candidate molecules before selection. AWS is deploying its recently launched Forward Deployed Engineering organisation, embedding technical staff directly with the customer to co-develop agentic applications. Novo Nordisk reports existing AI work has already reduced clinical documentation time, with more than 25,000 employees using AWS-based AI tools; the collaboration spans research, clinical development, commercial operations and IT. Financial terms were not disclosed. The deal sits alongside Novo Nordisk's April 2026 OpenAI partnership (covered separately here) and extends a broader Amazon relationship spanning Amazon Pharmacy, Amazon Ads and One Medical; AWS has also announced 2026 life-sciences AI collaborations with Flagship Pioneering and Johnson & Johnson. Contrary view: this is an infrastructure and enterprise-IT agreement rather than a therapeutic-discovery result, and the tracker's recurring caution applies with unusual force — no financial terms, no disclosed targets, no timelines for when AI-assisted programmes might yield candidates, and no mechanism by which an outside observer could later verify success. The headline productivity metrics offered (reduced documentation time, 25,000 employees made more productive) measure administrative throughput, not discovery output, and conflating the two is precisely the category error this tracker exists to flag. Stacking a preferred-cloud AI partner on top of an April OpenAI partnership within four months suggests portfolio hedging across vendors rather than settled conviction in any one stack, and raises integration and data-governance questions the announcement does not address. Agentic candidate generation and ranking is a capability claim, not evidence; the binding constraint remains wet-lab validation throughput and clinical attrition. Source
Regulatory Clinical Trials

HHS Launches Operation TrialBlazer Spanning FDA, NIH and ARPA-H — Umbrella Initiative Behind the FDA Real-Time Clinical Trials Programme; Pilot Selection Criteria and Participants Still Unpublished as of Mid-August 2026

The US Department of Health and Human Services launched Operation TrialBlazer on 22 June 2026, a department-wide initiative involving the FDA, NIH, ARPA-H and other HHS components, framed by HHS as restoring American leadership in clinical trials and shortening the path from drug discovery to first-in-human testing. The launch provides the institutional context for the FDA Real-Time Clinical Trials (RTCT) activity this tracker has followed across several cycles: the two proof-of-concept trials reporting data to the agency as it accrues (AstraZeneca's TRAVERSE in mantle cell lymphoma and Amgen's STREAM-SCLC in small cell lung cancer, the latter via Paradigm Health's platform), and the Request for Information on a proposed RTCT pilot (Docket FDA-2026-N-4390, comment period extended to 29 June 2026). Note: this entry falls outside the strict cycle window (10-16 August 2026) and is included as deliberate backfill, because the tracker recorded the constituent FDA actions without capturing the HHS umbrella programme that frames them. It also records a negative finding worth logging: the FDA stated it intended to publish final pilot selection criteria in July 2026 and complete pilot selections in August 2026, and as of mid-August 2026 neither the selection criteria nor any additional named participants had been published. Contrary view: departmental initiatives of this kind are organisational and rhetorical framings rather than regulatory instruments — RTCT is explicitly not a new regulatory pathway and does not change IND requirements under 21 CFR Part 312 or Phase 1/2 requirements, so the practical effect on sponsors remains modest until concrete criteria appear. The two proof-of-concept trials predate the pilot and were not selected through any published competitive process, which limits what they demonstrate about how a broader programme would operate. The slipping July/August timeline is the more informative signal: it echoes the EU's 16-month deferral of AI Act high-risk obligations (covered separately) and suggests that on both sides of the Atlantic, regulatory ambition around AI in trials is currently outpacing implementation capacity. Separately, secondary sources have described an FDA "Accelerated AI Pathway Pilot" selecting ten AI-focused sponsors; this tracker has been unable to corroborate that against any primary FDA source and therefore does not record it. Source
Funding Platform Antibody Design

Chai Discovery Closes $400M Series C at $3.8B Valuation — Index Ventures Leads; Valuation Nearly Triples in Seven Months as Chai-3 Reaches Eli Lilly and Pfizer Pipelines

Chai Discovery announced on 14 July 2026 a $400 million Series C at a $3.8 billion post-money valuation, led by Index Ventures alongside Kleiner Perkins, Sequoia Capital and Dimension, with new investors including Bain Capital Ventures, Battery Ventures, Baillie Gifford, BDT & MSD, Sapphire Ventures and Avra Capital, and existing backers Thrive Capital, OpenAI, Oak HC/FT, Menlo Ventures and General Catalyst participating. Cumulative funding now stands above $600 million across a $30M seed, $70M Series A, $130M Series B (December 2025, $1.3B valuation) and this round — a near-tripling of valuation in roughly seven months. The company, founded in 2024 by Joshua Meier, Jack Dent, Matthew McPartlon and Jacques Boitreaud, builds generative models for pre-clinical molecular design; Chai-2 was the first zero-shot de novo antibody design platform to reach double-digit experimental success rates, and Chai-3 is now deployed with Eli Lilly and Pfizer against hard-to-drug and historically undruggable targets. Note: this entry falls outside the strict cycle window (2-9 August 2026) and is included as backfill. It also corrects the tracker's earlier Chai entry, which recorded pre-close reporting of a prospective raise at a $3.4 billion valuation; the round in fact closed at $3.8 billion. Contrary view: a valuation tripling in seven months on a pre-revenue-scale platform is a capital-markets signal, not a scientific one, and the tracker has repeatedly noted that funding milestones are the weakest form of validation in this field. Despite roughly $20 billion invested across generative-AI drug discovery to date, no AI-discovered drug has been approved, and Chai has no clinical-stage asset of its own — its value rests on partner pipelines whose outcomes will not be legible for years. Chai-3's headline claim of doubling Chai-2's success rate remains undisclosed in any peer-reviewed venue, so the central technical assertion underpinning this valuation is vendor-reported. OpenAI's presence on the cap table alongside Lilly and Pfizer deployments also concentrates strategic dependencies. The round is best read as evidence of investor conviction in foundation-model approaches to biology, which is a claim about sentiment rather than about molecules. Source
Drug Discovery Platform

Nature Machine Intelligence: "Unfamiliarity" Metric Quantifies When Molecular Deep Learning Models Leave Their Training Chemical Space — Joint Bioactivity-Reconstruction Modelling Across 33 Datasets, Wet-Lab Validated With Low-Micromolar Actives on Two Kinase Targets

Van Tilborg, Rossen and Grisoni (Eindhoven University of Technology) published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01216-w; Nat Mach Intell 8, 575-587) a joint modelling approach that trains a network to simultaneously predict bioactivity and reconstruct molecules, yielding "unfamiliarity" — a reconstruction-based metric estimating how far a query molecule sits outside a model's training chemical space, and therefore how much its prediction should be trusted. Across a systematic analysis spanning 33 experimentally labelled bioactivity datasets, unfamiliarity is reported both to identify out-of-distribution molecules and to correlate strongly with classifier performance, remaining informative under strong distribution shift on large-scale libraries. Critically for a field dominated by retrospective benchmarks, the authors carried the metric into the wet lab and identified several compounds with low-micromolar activity against two clinically relevant kinase targets. An accompanying News & Views by Tomberg and Carson (AstraZeneca; DOI: 10.1038/s42256-026-01251-7, Nat Mach Intell 8, 862-863) frames the contribution for medicinal chemists. This work speaks directly to the caveat this tracker has attached to nearly every molecular-ML entry — that models excel in-distribution and degrade precisely where novel chemistry would deliver most value — by offering a principled way to measure that degradation rather than merely assert it. Note: this entry falls well outside the cycle window (2-9 August 2026) and is included as deliberate backfill of a peer-reviewed primary that fills a conceptual gap in the tracker's coverage. Contrary view: unfamiliarity is a reconstruction-based proxy, so it measures distance in the model's learned representation rather than distance in any chemically meaningful sense, and a molecule can be reconstructible yet mechanistically novel (or vice versa). Low-micromolar kinase hits are a modest experimental bar — kinases are among the most tractable and best-precedented target classes, and micromolar potency is a starting point rather than a lead. The 33 datasets are ChEMBL-lineage bioactivity sets that carry known assay-heterogeneity and activity-cliff problems, which flatter distribution-shift analyses. Most importantly, knowing that a model is unreliable outside its training space does not make it reliable there; the metric improves triage and honesty about uncertainty, which is valuable, but it does not extend genuine predictive reach into novel chemistry. Source
Regulatory Platform

EU AI Act Amended: Digital Omnibus Regulation (EU) 2026/1744 Enters Into Force, Deferring High-Risk Obligations From 2 August 2026 to 2 December 2027 (Annex III) and 2 August 2028 (Annex I Regulated Products, Including Medical Devices and IVDs)

Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal of the European Union on 24 July 2026 and entered into force on 27 July 2026 — three days after publication, an accelerated timetable the regulation itself justifies as a matter of urgency because the general application date it amends fell on 2 August 2026. This is the first substantive amendment to the AI Act and it supersedes the tracker's earlier entry covering the Commission's November 2025 Omnibus proposal. The central change rewrites the third paragraph of Article 113: application of obligations for stand-alone high-risk systems under Annex III moves from 2 August 2026 to 2 December 2027, while high-risk AI embedded in products regulated under Annex I harmonisation legislation — the category covering medical devices under the MDR and in-vitro diagnostics under the IVDR — moves to 2 August 2028. The package passed the European Parliament on 16 June 2026 (423-57, 174 abstentions) and received the Council's final approval on 29 June 2026. Other provisions: national authorities now have until 2 August 2027 to establish at least one AI regulatory sandbox; the legal basis for processing special-category personal data for bias detection is extended beyond high-risk systems; and a new Article 5 prohibition on AI-generated non-consensual intimate imagery and child sexual abuse material applies from December 2026. Relevance to pre-clinical R&D: the deferral changes compliance sequencing for pharma AI touching diagnostics, clinical decision support and any SaMD-adjacent tooling, and buys time for the delegated acts the Commission must adopt by 2 August 2027. Note: this entry corrects a forward-looking assumption carried in earlier tracker planning notes that Annex III high-risk obligations would take effect on 2 August 2026; in law they will not. Contrary view: a deferral is not a dismantling — the Act's risk-based architecture, conformity-assessment regime, GPAI track and AI Office oversight all remain intact, and firms treating the Omnibus as a licence to pause AI governance work are misreading it. Article 50 transparency obligations were largely unaffected and remain live. The deferral itself is best read as an admission that the supporting regulatory infrastructure (harmonised standards, notified-body capacity, delegated acts) did not materialise on schedule, which is a signal about implementation capacity rather than a relaxation of intent. Much AI used purely in early discovery may fall outside the high-risk category altogether depending on intended use, so the practical impact on pre-clinical work is narrower than headline coverage implies — the classification guidance pharmaceutical companies have been waiting for is still outstanding. Source
Antibody Design Drug Discovery Platform

Nature Machine Intelligence: HelixFold-S1 Contact-Guided Conformational Sampling Improves Biomolecular Complex Structure Prediction — Order-of-Magnitude Reduction in Sampling Requirements, Notable Gains on Antibody-Antigen Interfaces Where Co-Evolutionary Signal Is Weak

Liu, Liu, Ye and colleagues at Baidu's PaddleHelix team published in Nature Machine Intelligence on 2 July 2026 (DOI: 10.1038/s42256-026-01264-2; Nat Mach Intell 8, 1115-1126) HelixFold-S1, a guided planning approach for biomolecular complex structure prediction. The premise is that generating large ensembles of candidate conformations — standard practice for improving structure prediction — is inefficient and costly, producing many redundant conformations with limited diversity, so additional compute yields diminishing returns. HelixFold-S1 instead predicts interchain contact probabilities that serve as a blueprint of the conformational space, directing computational effort towards higher-probability, low-redundancy contacts that constrain structure generation. Across diverse complex benchmarks the method reports markedly higher structural accuracy than unguided sampling while reducing sampling requirements by an order of magnitude, with gains reported across protein-antibody, protein-protein, protein-ligand, protein-RNA and protein-DNA interfaces. The predicted contact probabilities additionally function as a rough indicator of prediction difficulty, allowing compute to be allocated adaptively. The antibody-antigen result is the most directly relevant to therapeutic discovery, since antibody diversity and the absence of co-evolutionary signal make these among the hardest complexes for AlphaFold-lineage methods. Source code and weights are publicly released via PaddleHelix. Note: this entry falls outside the strict cycle window (26 July - 1 August 2026) and is included as deliberate backfill of a peer-reviewed primary from earlier in July. Contrary view: this is an inference-time sampling strategy layered on a HelixFold3 variant rather than a new predictive architecture, so the gains are efficiency and search-quality improvements rather than a step change in underlying structural understanding. Benchmarks are PDB-derived and retrospective; accuracy on complexes structurally unlike anything in the training distribution — precisely the novel antibody-antigen pairs a discovery programme cares about — is not established by these results. The order-of-magnitude sampling reduction is a compute claim, and compute was rarely the binding constraint in antibody engineering; experimental validation throughput was. No prospective wet-lab validation accompanies the paper, and structure prediction accuracy remains an imperfect proxy for binding affinity or developability, as the tracker has noted repeatedly across the AlphaFold3, Chai and Boltz lineages. Source
Autonomous AI Drug Discovery Platform

Science: Biomni General-Purpose Biomedical AI Agent Autonomously Executes Research Tasks Across 25 Domains — Action-Discovery Agent Mines Tools, Databases and Protocols From Thousands of Publications; Generalises to Drug Repurposing and Causal Gene Prioritisation Without Task-Specific Tuning

Huang, Zhang, Wang, Qu and colleagues (Stanford, Genentech, Arc Institute; senior authors Le Cong, Aviv Regev and Jure Leskovec) published in Science on 9 July 2026 (DOI: 10.1126/science.adz4351) Biomni, a general-purpose biomedical artificial intelligence agent designed to autonomously execute diverse research tasks rather than serve a single predefined workflow. To map the biomedical action space, Biomni's action-discovery agent mines tools, databases and protocols from thousands of publications spanning 25 biomedical domains, assembling a unified agentic environment. Its generalist architecture integrates large language model reasoning with retrieval-augmented planning and code-based execution, dynamically composing workflows without predefined templates. Systematic benchmarking reports strong generalisation across heterogeneous tasks — causal gene prioritisation, drug repurposing, rare-disease diagnosis, microbiome analysis and molecular cloning — without task-specific tuning, and real-world case studies describe Biomni interpreting multimodal datasets, optimising protein stability and orchestrating wet-lab instruments. The system is publicly accessible via Stanford. Biomni joins the agentic-science cluster the tracker has documented across 2026 (FutureHouse Robin and Google Co-Scientist, both Nature; DeepEvidence, covered separately this cycle), and is notable for breadth of task coverage rather than depth on a single discovery. Note: this entry falls outside the strict cycle window (20-25 July 2026) and is included as deliberate backfill of a major peer-reviewed primary from earlier in July. Contrary view: "autonomously executes" requires careful reading — the benchmark tasks are retrospective problems with known ground truth (drug repurposing and gene prioritisation are precisely the settings where the answer already exists in the literature the agent mines), which is a materially easier regime than generating and validating genuinely novel biology. The real-world case studies are demonstrations rather than controlled comparisons against expert baselines, so claims that performance approaches human experts (prominent in secondary coverage) are not straightforwardly supported by the published benchmarks. General-purpose LLM agents inherit well-documented failure modes — hallucinated tool invocations, fluent but incorrect syntheses, and silent errors in code execution — that are especially consequential when the output is an experimental protocol. Two senior authors are Genentech employees and one holds Roche equity, a commercial interest the reader should weigh. The bioRxiv preprint appeared in June 2025 and the Science version in July 2026, a roughly 13-month review interval suggesting substantial revision. As with every agentic system tracked here, no Biomni-originated therapeutic candidate has yet cleared prospective wet-lab or clinical validation. Source
Autonomous AI Drug Discovery Clinical Trials

Nature Machine Intelligence: DeepEvidence Multi-Agent Deep Research System for Biomedical Evidence Synthesis — Breadth-First and Depth-First Strategies Over Knowledge Graphs Plus an Auditable Evidence Graph, Benchmarked Across Target Identification, Mechanistic Reasoning and Trial Development

Wang, Chen, Yang, Jin, Peng, Lu and Sun (Keiji AI, Osaka University, Kyoto University, NIH National Library of Medicine, Weill Cornell Medicine and the University of Illinois Urbana-Champaign) published in Nature Machine Intelligence on 2 July 2026 (DOI: 10.1038/s42256-026-01266-0; Nat Mach Intell 8, 1142-1156) DeepEvidence, a deep research agent for evidence exploration and synthesis across heterogeneous biomedical knowledge sources. The system coordinates multi-agent collaboration that combines breadth-first and depth-first research strategies to search, explore and aggregate evidence from multiple biomedical knowledge bases and the literature, and incrementally constructs an evidence graph of key entities and observations to support transparent tracking, attribution and validation of the research process — an auditability feature that distinguishes it from general-purpose deep-research tools. DeepEvidence is reported to outperform generic AI agents across four open benchmarks, and the authors establish a further seven benchmark tasks spanning major stages of biomedical discovery including target identification, mechanistic reasoning, preclinical experimentation, clinical trial development and evidence-based medicine. Code and test datasets are publicly released. Note: this entry falls outside the strict cycle window (20-25 July 2026) and is included as deliberate backfill of a peer-reviewed primary from earlier in July; it pairs directly with the Biomni Science paper covered separately this cycle. Contrary view: evidence synthesis is a materially different and easier problem than discovery — retrieving, ranking and aggregating what is already recorded does not by itself generate novel biology, and the tracker's recurring caution applies that literature-derived target nomination reproduces the field's existing attention biases. Seven of the eleven benchmark tasks were established by the authors themselves, and self-defined benchmarks structurally favour the proposing system; independent evaluation on externally curated tasks is the diagnostic test. Knowledge-graph-grounded agents inherit the coverage gaps and curation biases of their underlying sources (KEGG, Open Targets, Reactome, UniProt and similar), so under-studied targets — exactly those where AI would add most value — remain under-represented. The corresponding-author affiliation includes a commercial AI company, and there is no prospective validation demonstrating that DeepEvidence-derived syntheses changed a real discovery decision. Deep research agents in medicine have also been shown to produce fluent, well-structured outputs that non-experts struggle to distinguish from expert work, which is a risk as much as a capability. Source
Clinical Trials Drug Discovery Platform

Insilico's Rentosertib Enters Phase III for Idiopathic Pulmonary Fibrosis — First Drug With AI-Identified Target, AI-Generated Structure and AI-Guided Clinical Prediction to Reach Late-Stage Pivotal Testing (320 Patients, NCT07687459)

Insilico Medicine (HKEX: 3696) announced on 7 July 2026 the initiation of a Phase III clinical trial for rentosertib (formerly ISM001-055 / INS018_055), its potentially first-in-class oral small-molecule TNIK inhibitor for idiopathic pulmonary fibrosis (IPF). This is a genuine field milestone that supersedes the tracker's existing Phase 2a entry: rentosertib is the first molecule whose target was prioritised by AI (Biology42/PandaOmics), whose structure was generated by generative AI (Chemistry42), and whose clinical development was informed by AI outcome prediction (Medicine42/inClinico) to reach late-stage pivotal testing. The GENESIS-IPF Phase 2a study (published Nature Medicine 2025) showed a mean forced-vital-capacity improvement of +98.4 mL at 12 weeks in the 60 mg once-daily arm with manageable safety; the discovery-to-clinic journey was documented in Nature Biotechnology. The Phase III programme (identifiers NCT07687459 / CTR20262475) is described as a prospective, randomised, double-blind, placebo-controlled, parallel-group study enrolling approximately 320 IPF patients over 40 years of age across roughly 47 centres, running 52 weeks, with company-sketched timelines pointing to initial results around December 2029, potential NDA around March 2030 and possible approval around September 2030. Note: this milestone lands amid an intense concentration of Insilico news this cycle — a $2.5B CDMO alliance with Bora Pharmaceuticals (15 July), a second China Medical System pact (~$165M, ~13 July) and the ~$600M Takeda collaboration (covered separately). This tracker treats Insilico as its live source-concentration risk (15+ entries) and has deliberately excluded or concentration-framed the partnership announcements; this Phase III entry is included on distinct clinical-milestone merit, not deal-flow grounds. Contrary view: initiating a Phase III trial is not the same as succeeding in one — the 12-week +98.4 mL FVC signal in 71 patients must now hold across 320 patients over a full 52 weeks, and IPF is a notoriously difficult indication where promising mid-stage antifibrotic signals have repeatedly failed to translate to durable Phase III benefit. The AI-origin framing, while legitimate and historically significant, does not de-risk the biology: TNIK is a novel mechanism without prior clinical validation in IPF, and novel-mechanism Phase III trials carry elevated failure risk precisely because the target has not been vetted by prior drugs. Company-sketched regulatory timelines (NDA 2030, approval 2030) are best-case projections that assume clean enrolment, no interim safety holds, and a positive readout — none of which is guaranteed. The result, positive or negative, around 2029-2030 will be one of the most consequential single data points for the entire AI-drug-discovery thesis, which is precisely why it should not be over-weighted before the data exist. Source
Partnership Platform Drug Discovery

Takeda Signs ~$600M AI Drug Discovery Collaboration With Insilico Medicine — Pharma.AI Platform Access for Early-Stage Discovery; Insilico's Deal Velocity Becomes a Source-Concentration Signal in Its Own Right

Takeda has entered a strategic collaboration with Hong Kong-based Insilico Medicine (HKEX: 3696) to apply AI in early-stage drug discovery across Takeda's therapeutic areas, in a deal reported around 6 July 2026 with a headline value of approximately $600 million (upfront terms undisclosed; therapeutic areas and specific targets not disclosed). The agreement gives Takeda access to Insilico's Pharma.AI platform — PandaOmics for target discovery, Chemistry42 for de novo small-molecule generation, and InClinico for clinical-trial transition prediction — with Insilico leading AI-driven discovery and Takeda taking selected candidates through clinical development under exclusive worldwide development, manufacturing and commercialisation rights. Takeda R&D head Chris Arendt framed the deal as combining Takeda's disease-biology depth with Insilico's AI-enabled discovery, adding that Takeda is also integrating automation, robotics and generative AI into its own discovery work. Editorial note on source concentration: this tracker has flagged Insilico Medicine as a live source-concentration risk (14+ dedicated entries — the most of any single organisation tracked), triggering the same weighting discipline previously applied to Tempus AI. This item is included not as another Insilico promotional milestone but because Insilico's 2026 deal cadence has itself become the story: Lilly ($2.75B, March), the Servier ($888M, 2025) partnership, SK Biopharmaceuticals ($2.5B, BIO 2026, deliberately omitted from the prior cycle on concentration grounds), and now Takeda (~$600M) inside roughly four months. Contrary view: a ~$600M headline with undisclosed upfront tells readers very little about genuine risk transfer — the tracker has repeatedly noted that biobucks-to-upfront ratios, not headline totals, reveal how much a pharma partner is actually committing, and Insilico's serial-partnership model means each individual deal represents a fractional, milestone-gated bet rather than deep conviction. Insilico's underlying differentiator remains rentosertib (the one clinically validated AI-discovered asset), and no amount of partnership volume substitutes for a second independent clinical readout. The concentration of pharma willingness to sign non-exclusive-feeling platform deals with a single AI vendor also raises the question of whether these agreements reflect genuine capability differentiation or diversified option-buying across an unproven category. Readers should weight this disclosure accordingly. Source
Genomics & Proteomics Drug Discovery Platform

Nature Machine Intelligence: Programmable RNA Translation via Deep-Learning IRES Discovery and De Novo Generation — IRES-LM Language-Model Ensemble Trained on 46,774 Sequences Enables Cap-Independent Translation Control for RNA Therapeutics

Chu, Yin, Yu and colleagues published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01213-z; Nat Mach Intell 8, 559-574, 2026) an end-to-end AI framework for programmable RNA translation built around internal ribosome entry sites (IRES) — structured RNA elements that enable cap-independent translation initiation, a key tool for controlling therapeutic payload expression in RNA medicines. The framework unifies IRES identification, evolutionary optimisation and de novo generation. Its core, IRES-LM, is an ensemble of two language models trained on 46,774 sequences that predicts linear mRNA IRESs with a reported 15% improvement in AUC and F1 over existing methods, and demonstrates cross-applicability to circular RNA IRESs — correctly identifying all 21 experimentally validated circRNA IRESs and outperforming benchmark methods. The work is directly relevant to the RNA-therapeutics wave the tracker has documented (Alnylam-Inceptive siRNA foundation-model deal; broader mRNA/circRNA vaccine and payload engineering): precise, designable translation control is a recognised bottleneck, and cap-independent IRES-driven expression is especially valuable for circular RNA constructs prized for nuclease resistance. Note: this entry falls outside the strict cycle window (6-12 July 2026); it is a peer-reviewed Nature Machine Intelligence primary from earlier in 2026 included here as deliberate backfill because RNA-therapeutics payload design is an under-represented category in the tracker and the paper is a substantive methodological contribution. Contrary view: IRES element design has a long history of context-dependence — an IRES that performs well in a reporter assay or one cell type frequently fails to transfer to the therapeutic mRNA, cell type, or in-vivo delivery context of interest, and the paper's validation is via high-throughput functional assays rather than in-vivo therapeutic expression. A 15% AUC/F1 improvement over prior IRES predictors is meaningful but incremental, and the de novo generation claims need independent wet-lab replication across diverse target payloads before being treated as a general design capability. As with essentially all sequence-design language models tracked here, the binding constraint is whether in-silico-designed elements retain function under real delivery and disease conditions. Source
Platform Drug Discovery Partnership

Anthropic Launches Claude Science and Announces Intent to Develop Its Own Drugs — AI Vendor Moves Downstream Into Therapeutic Discovery to Gain First-Hand Scientific Experience

Anthropic announced on 30 June 2026 (San Francisco) that it will begin developing drugs of its own, disclosed alongside the launch of Claude Science, a new application aimed at researchers and the pharmaceutical industry (reported by STAT, Brittany Trang and Matthew Herper). Eric Kauderer-Abrams, Anthropic's head of life sciences, framed the move as the company's answer to what it should be doing beyond training models and building products; multiple executives emphasised the value of gaining first-hand experience using Anthropic's own products to solve real scientific problems. The announcement extends Anthropic's steady push into life sciences previously tracked here: Claude for Life Sciences (launched October 2025), the Coefficient Bio acquisition (April 2026, drawing talent from Genentech's Prescient Design), the Gates Foundation partnership (May 2026), and the Vas Narasimhan board appointment. It positions Anthropic more directly against Google's Isomorphic Labs and OpenAI's life-sciences efforts, moving an AI-model vendor downstream from tool-provider toward therapeutic originator. Note: this entry is an industry / corporate-strategy disclosure rather than a peer-reviewed scientific result; the underlying STAT reporting is partly paywalled and the operational specifics (targets, modalities, wet-lab strategy, whether Anthropic intends to commercialise candidates) were not disclosed. Contrary view: an announcement of intent to develop drugs is not a drug-discovery capability - Anthropic has no disclosed internal wet-lab, no clinical pipeline, and no therapeutic track record, and STAT noted it is unclear whether the company even intends to bring candidates to commercialisation. The "use our own products on real problems" rationale is as much about model-improvement and marketing for Claude Science as about a genuine therapeutic programme, and the history of technology firms entering drug R&D (for example, prior Big Tech health ventures) is littered with retrenchments once the biological attrition rate becomes apparent. The move also raises an evident conflict-of-interest question: Anthropic simultaneously sells AI tools to pharma customers and now proposes to compete with them. As with all in-house AI-designer efforts, the binding constraint remains wet-lab validation and the ~90% clinical failure rate that no amount of model capability has yet been shown to move. Source
Drug Discovery Platform Partnership

Edison Scientific (FutureHouse Commercial Spinout) Partners With Population Health Partners to Launch AI-Agent-Founded Biotechs — Team Behind Metsera Backs Robin-Lineage Discovery Engine

Edison Scientific, the commercial spinout of non-profit AI-research lab FutureHouse, announced on 29 June 2026 a strategic partnership with biotech investment firm Population Health Partners to leverage AI agents in drug discovery and development, with the stated intent of creating new biotech companies (reported by STAT). Population Health Partners is associated with the team behind Metsera, the obesity-drug developer. The partnership operationalises the technology underlying FutureHouse's Robin multi-agent system - the platform that, in its Nature publication (Ghareeb et al., previously tracked here), autonomously identified ripasudil as a repurposing candidate for dry age-related macular degeneration. FutureHouse CEO Sam Rodriques (formerly of the Francis Crick Institute) and co-founder Andrew White (formerly University of Rochester) founded FutureHouse in 2023; STAT reported that a top-10 drugmaker approached FutureHouse with a $30 million offer to use its agents, and another big-pharma AI head characterised FutureHouse's agents as the "Ferrari" of agents relative to building in-house on OpenAI or Anthropic. Note: this is a corporate / commercialisation disclosure rather than a peer-reviewed result, and financial terms of the Edison-Population Health partnership were not disclosed. Contrary view: spinning a non-profit research lab's agent technology into a company that founds biotechs stacks two unproven bets - that AI agents can originate viable drug programmes, and that a repeatable company-creation engine can be built on top of them. Robin's single published success (ripasudil, a repurposed clinically approved ROCK inhibitor) is the easiest class of validation, and no Edison/FutureHouse-originated novel chemical entity has yet cleared clinical proof. The "Ferrari of agents" framing is vendor positioning; the durability of any agent-quality moat is questionable as foundation-model providers (including Anthropic and OpenAI, both entering this space directly) improve their own scientific-agent offerings. Company-creation partnerships of this type also concentrate incentives toward volume of programmes launched rather than clinical success, a metric that will only be testable over multi-year horizons. Source
Drug Discovery Platform Clinical Trials

BIO 2026 (San Diego) Marks Tonal Shift in AI Drug Discovery — From "Can It Work?" to "What Is Working?" as Generative-Protein Platforms Move From Pipeline-Filling to Clinical Candidates; Chinese Originator IP Competition Emerges as Strategic Theme

Industry coverage of the BIO International Convention (22-25 June 2026, San Diego; ~20,000 attendees, theme "Driven by Purpose", second-annual AI Summit with seven marquee panels across 20 AI sessions) reported a measurable shift in how investors and operators discuss AI drug discovery: the dominant question moved from whether AI drug design can work to which approaches are working, which are not, and how fast incumbents must respond. The reframing was attributed to the accumulating clinical evidence base — most prominently Insilico Medicine's rentosertib (TNIK inhibitor) Phase 2a idiopathic-pulmonary-fibrosis readout published in Nature Medicine in 2025, and the migration of generative-protein platforms (Xaira Therapeutics, Generate Biomedicines, Isomorphic Labs) from pipeline-filling toward clinical candidates. A second, harder theme dominated strategic discussion: Chinese biopharma has moved from contract manufacturing and fast-following to originating pharmaceutical IP, with Chinese companies reportedly securing more than $30 billion in oncology licensing deals across the 2024-2025 cycle and Western pharma signing over $10 billion for Chinese cardiovascular and oncology assets in late 2025 — exposure the BIOSECURE Act (focused on manufacturing) was not designed to address. Coverage also flagged regulatory ambiguity (the evolving FDA AI posture) as a planning variable rather than a settled framework. Note: this entry is an industry-synthesis item rather than a primary peer-reviewed disclosure, included because BIO functions as the field's annual barometer and the tonal shift it documents has direct implications for AI-drug-discovery capital allocation and pre-clinical R&D strategy. Contrary view: convention-derived "tonal shifts" are sentiment signals, not evidence; the underlying clinical proof points (rentosertib Phase 2a) are unchanged from 2025 and remain a single mid-stage readout, not a Phase 3 validation or approval. "Moving toward clinical candidates" is not the same as demonstrating improved clinical success rates over traditional discovery, which remains unproven against the industry's persistent ~90% attrition. The China-IP framing, while substantive, is partly shaped by US policy advocacy interests and should be read alongside the genuine scientific contributions of Chinese AI-biotech rather than purely as a competitive threat narrative. Source
Drug Discovery Genomics & Proteomics Platform

Nature Machine Intelligence: Data-Driven Surrogates of Rational Design Enable Antimicrobial Peptide Optimization — Closed-Loop Generative Approach Refines Biologically Complex Activity Scaffolds Under Practical Design Constraints

Mauša, Kalafatovic and colleagues published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01258-0) a study presenting data-driven surrogate models of rational design for antimicrobial peptide (AMP) optimization, framed around a closed-loop generative-optimisation workflow for antibiotic design. The work targets a sharpening question in the AMP field: rising pathogen drug resistance has made next-generation antimicrobial peptides a global priority, and while generative AI has accelerated the broad proposal of candidate peptides with high predicted therapeutic potential, the harder and less-solved challenge is whether such approaches can refine biologically complex activity scaffolds under practical, real-world design constraints rather than simply exploring sequence space broadly. The study positions data-driven surrogate models — trained to approximate the outcomes of rational, mechanism-informed design — as a route to constraint-aware optimisation, complementing the broader generative AMP literature (including the ApexGO transformer-VAE Bayesian-optimisation approach previously tracked here, and the Wong/de la Fuente-Nunez/Collins closed-loop antibiotic-design line of work). Direct relevance to drug discovery: antibiotic lead optimisation against drug-resistant bacteria, a therapeutic area where the economic and scientific pipeline has been notably stalled. Contrary view: surrogate models of rational design inherit the limitations of the rational-design heuristics they approximate, and the "practical design constraints" framing is only as good as the constraint set encoded; AMP optimisation studies have a recurring pattern of strong in-silico and in-vitro results that do not translate to in-vivo efficacy, stability, or acceptable toxicity/haemolysis profiles. The antimicrobial-peptide field as a whole has produced very few clinically approved agents despite a decade of computational enthusiasm, so methodological advances should be weighed against that sober translational track record. As with the broader single-cell and molecular-design methods tracked across recent cycles, the absence of prospective in-vivo validation in the core result limits direct read-across to a real antibiotic-development campaign. Source
Genomics & Proteomics Drug Discovery Platform

Nature Machine Intelligence: Conditional Monge Gap (CMonge) Enables Generalizable Single-Cell Perturbation Modelling — Conditional Optimal Transport Predicts Drug, Dose and Combination Responses Across Unseen Contexts

Driessen, Rajwade, Harsanyi, Rapsomaniki and Born published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01242-8) the Conditional Monge Gap (CMonge), a neural optimal transport method that learns transport maps conditioned on arbitrary covariates — drug, dose, cell type, or combinations thereof — to predict single-cell perturbation responses. The method addresses a recognised limitation: most prior neural optimal transport approaches to single-cell perturbation modelling cannot condition on treatment context, requiring a separate model per condition and failing to generalise to unseen treatments. CMonge instead trains a single global estimator that can be conditioned at inference time on a potentially unseen condition. On the SciPlex3 dataset (three cell lines, 187 compounds, four doses), a single CMonge model reportedly matches the performance of separate condition-specific models and generalises to previously unseen drug–dose combinations, outperforming the conditional baseline chemCPA in out-of-sample prediction using only drug structure (SMILES) as input. Direct relevance to drug discovery: in silico perturbation screening, dose-response simulation, and combination-effect prediction — capabilities that bear on target validation and lead prioritisation. Code released as the open-source cmonge package. Contrary view: single-cell perturbation modelling is a crowded methodological field with competing approaches (chemCPA, scGen, CellOT, diffusion-based methods such as LCD and scPPDM, and flow-matching methods such as PRiMeFlow); CMonge's reported state-of-the-art is benchmark-specific, evaluated principally on SciPlex3 and a 40-plex 4i melanoma imaging dataset, and the out-of-distribution gains are stronger for effect-based (mode-of-action) embeddings than for structure-based ones. The authors themselves note that fine-tuning on giga-scale datasets (e.g. Tahoe-100M) or on patient-derived organoids would be required to unlock the method's full translational potential — meaning the present results are a methodological advance rather than a validated drug-discovery tool. Single-cell perturbation prediction in general remains constrained by the inherent heterogeneity and dropout noise of scRNA-seq, and the absence of prospective wet-lab validation in this study limits direct read-across to real screening campaigns. Source
Drug Discovery Platform

Nature Chemical Biology: PBCNet2.0 Cartesian-Tensor Siamese Network for Protein–Ligand Affinity Prediction — Trained on 8.6M Complex Pairs, 7.18-Fold Lead Optimization Speed-Up, Emergent Mutation-Effect Prediction

Yu and colleagues published in Nature Chemical Biology (DOI: 10.1038/s41589-026-02241-x) PBCNet2.0, a Cartesian tensor-based Siamese neural network for predicting protein–ligand relative binding affinity at scale. Trained on 8.6 million protein–ligand complex pairs (vs ~20,000 unique molecules and 960 targets in the original PBCNet), PBCNet2.0 reports zero-shot accuracy comparable to physics-based simulations (typically Schrödinger FEP+) while remaining substantially more computationally efficient. Retrospective prioritisation experiments report 7.18-fold improvement in lead optimisation efficiency with 41% reduced compute resource use. The authors highlight an emergent capability: although the model was not explicitly trained on mutation data, PBCNet2.0 exhibits the ability to predict affinity changes induced by binding-pocket residue mutations — directly relevant to resistance analysis and patient-stratification for precision oncology programmes. The work bridges two trends previously tracked here: (a) the scale-data thesis for AI drug discovery (more diverse, more contexts), and (b) Cartesian-tensor / geometric-deep-learning architectures that preserve SE(3) symmetry without explicit data augmentation. Contrary view: zero-shot performance "comparable to physics-based simulations" benchmarks against a specific FEP+ implementation and a curated dataset; head-to-head against the most current Schrödinger FEP+, Cresset Spark, or OpenFE workflows on prospective campaigns is the diagnostic test. The 7.18-fold optimisation efficiency improvement is retrospective, measured on internal benchmarks; prospective validation on a real drug discovery programme (with all the standard confounders of medicinal chemistry decision-making) has not yet been disclosed. Emergent mutation-effect prediction is a strong claim that needs independent replication across more diverse target classes before being treated as a general capability. Source
Drug Discovery Platform

Nature Machine Intelligence: ConfSeq Conformation Description Language Bridges 3D Molecular Structures and Transformer Architectures — SMILES-Compatible Tokenisation with SE(3) Invariance for Sequence-Based 3D Modelling

Xiong and colleagues published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01250-8) ConfSeq, a conformation description language designed to encode three-dimensional molecular structures into discrete token sequences suitable for standard transformer architectures. ConfSeq integrates SMILES with internal coordinates — dihedral angles, bond angles, and a novel pseudo-chirality descriptor — ensuring SE(3) invariance while preserving the conciseness and human readability inherent to SMILES. By reformulating 3D molecular modelling tasks (conformation prediction, de novo 3D molecular generation, and 3D molecular representation learning) as sequence modelling problems, ConfSeq enables standard transformer architectures to achieve reported state-of-the-art performance across diverse benchmarks, with the authors noting unique advantages over diffusion-based approaches in inference efficiency, generation controllability, and scoring of generated molecules. The work addresses a long-standing limitation in chemical language modelling: while 2D SMILES-based language models have been broadly successful, extending these to 3D conformation modelling has been constrained by the absence of effective 1D token-based 3D representations. Direct relevance to drug discovery: structure-based generative chemistry, 3D-aware lead optimisation, and conformer-conditioned property prediction. Contrary view: language-model-based 3D molecular modelling is an active area with multiple competing approaches (graph neural networks with explicit equivariance, diffusion models with geometric priors, point-cloud methods); ConfSeq's reported state-of-the-art is benchmark-specific and the choice of comparators matters. The pseudo-chirality descriptor is a clever workaround for handling stereochemistry but introduces a discretisation step whose granularity-vs-accuracy trade-off needs careful prospective evaluation. The "unique advantages in inference efficiency" claim is likely accurate against autoregressive diffusion but may not hold against modern flow-matching methods. Source
Partnership Platform Antibody Design

Chai Discovery Licenses Chai-3 AI Antibody Design Platform to Pfizer — Doubles Chai-2 Hit Rate, Plus Custom Model Trained on Pfizer Proprietary Data Tailored to Internal Workflows

Chai Discovery announced on 4 June 2026 a license agreement with Pfizer that grants Pfizer scientists access to Chai's AI drug discovery platform, including the previously-undisclosed Chai-3 model and a custom model trained on Pfizer's proprietary data and tailored to Pfizer's internal R&D workflows. Chai-3 reportedly delivers step-change improvements in AI-driven antibody design, doubling the experimental success rate of its predecessor Chai-2 and producing antibodies that meet required therapeutic standards. The model advances progress on therapeutic binding capabilities, multi-specifics (bispecific and tetraspecific antibodies), molecules that bind hard-to-drug targets, and improved generalisation across target classes. Chai-2 — published in 2025 — was the first zero-shot antibody design platform to achieve double-digit experimental hit rates (~20%) and de novo CDR generation from target-plus-epitope prompts, a reported 100-fold improvement over sub-0.1% rates typical of legacy methods. The Pfizer deal positions Chai alongside Eli Lilly (existing Chai partnership) and follows Chai's December 2025 $130M Series B at a $1.3B valuation; subsequent reporting indicates Chai is in talks for a further $400M raise at a $3.4B valuation. Note: this entry falls within the cycle 6 window (30 May - 7 June 2026) and is included here as substantive backfill missed in the prior cycle. Contrary view: "doubling the success rate" benchmarks Chai-3 against Chai-2's already-impressive ~20% hit rate, but absolute hit rates and therapeutic-developability rates are the binding constraints for clinical translation, not relative improvement against the prior model. Financial terms of the Pfizer deal are undisclosed, which makes it difficult to calibrate the strategic weight relative to Chai's December $130M Series B or the recent Alnylam-Inceptive $2B partnership covered in cycle 6. Custom model training on proprietary pharma data introduces vendor-lock-in considerations and raises questions about Chai's competitive positioning vs Anthropic Coefficient Bio (acquired April 2026), Isomorphic Labs ($2.1B Series B May 2026), and Inceptive Nucleics (Alnylam partnership June 2026) — all racing to be the preferred AI biology partner for big pharma. The Chai-3 model has not been publicly disclosed in a peer-reviewed venue, which limits independent validation. Source
Drug Discovery Platform Funding

Nature Biotechnology Feature: "Programming Biology" — Next-Gen AI Firms Raise Billions to Design Better Medicines, but Data and Development Bottlenecks Remain

Nature Biotechnology published a feature analysis by Melanie Senior (DOI: 10.1038/s41587-026-03170-9) surveying the rapidly expanding AI-drug-design fundraising landscape and the persistent technical bottlenecks that capital alone cannot solve. The piece examines the wave of "next-generation" AI biology firms raising billion-dollar rounds — including Isomorphic Labs ($2.1B Series B May 2026, covered separately), Chai Discovery, Generate Biomedicines, Xaira, Iambic, and the Flagship-backed cluster — and frames the central thesis as "programmable therapeutics": the move from AI as virtual-screening accelerator toward AI as molecule designer that produces drugs by specification rather than by serendipity. Senior identifies four persistent bottlenecks: (a) the data-quality gap — most pharma proprietary datasets are not structured for foundation-model training; (b) the developability gap — AI-designed molecules pass binding and selectivity tests but fail on PK, immunogenicity, and manufacturability; (c) the clinical-validation gap — AI-discovered drugs face the same ~90% Phase 1/2 failure rates as traditional approaches; (d) the integration gap — most pharma still treat AI as an external service rather than an integrated R&D platform. The piece cross-references the Chai Discovery preprint (bioRxiv 10.1101/2025.11.29.691346), Fanton et al on AI-designed therapeutics (Nat Biotechnol s41587-025-02895-3), Durrant et al on protein design (Nat Biotechnol s41587-022-01494-w), and Madani et al on language models for protein engineering (Nat Biotechnol s41587-022-01618-2), positioning the May 2026 cluster as a generational shift in capability rather than incremental advance. Note: this entry falls outside the strict cycle 7 window (8-13 June 2026) and is included here as substantive Nature Biotechnology analytical backfill from late May 2026 that was missed in the prior cycle; it provides essential synthesis context for the rapidly-accumulating Robin/Co-Scientist/Chai-3/PBCNet2.0/ConfSeq cluster covered separately in cycles 5-7. Contrary view: Nature feature pieces are journalism rather than peer-reviewed research, and Melanie Senior's framing — while well-sourced — necessarily privileges the venture-backed company narrative over academic and open-source AI biology efforts. The "programmable therapeutics" framing is appealing but risks under-stating the role of empirical biological discovery (CRISPR-perturbation screens, organoid models, patient-derived datasets) that the AI design phase ultimately depends on. The four-bottleneck framework is useful but does not adequately address the regulatory bottleneck (FDA/EMA AI guidance still maturing, EU AI Act high-risk provisions effective 2 August 2026) which will increasingly shape what "programmable therapeutics" can actually mean in clinical practice. Source
Drug Discovery Platform

FutureHouse Robin Multi-Agent System Formally Published in Nature — Autonomous AI Identified Ripasudil (ROCK Inhibitor) as Novel Repurposing Candidate for Dry Age-Related Macular Degeneration via Lab-in-the-Loop Iteration

FutureHouse re-announced on 4 June 2026 that its Robin multi-agent system has been formally published in Nature (Ghareeb et al., DOI: 10.1038/s41586-026-10652-y) as the first AI system to autonomously orchestrate the full intellectual cycle of biomedical discovery: literature search, hypothesis generation, experimental design, data analysis, and follow-up insight generation. Robin combines three specialised agents (Crow for literature search and synthesis, Falcon for experimental design and scientific evaluation, Finch for experimental data analysis) and was applied to dry age-related macular degeneration (dAMD), a leading cause of irreversible blindness with limited treatment options. Robin proposed enhancing retinal pigment epithelium (RPE) phagocytosis as a therapeutic strategy and, through iterative drug-candidate testing, identified ripasudil — a clinically used Rho-kinase (ROCK) inhibitor approved for glaucoma — as a novel repurposing candidate, with a reported 7.5-fold increase in RPE phagocytosis activity. Robin then proposed and analysed a follow-up RNA-seq experiment that implicated ABCA1 as a candidate downstream target. FutureHouse reports Robin analysed 551 papers in approximately 30 minutes versus an estimated 540 hours of manual processing, and the entire discovery cycle spanned approximately 2.5 months. Validation was conducted in primary human RPE stem cells. The work has been operationalised via FutureHouse's commercial spinout Edison Scientific. Contrary view: "first to autonomously orchestrate" claims need calibration against Google Co-Scientist's parallel Nature publication (covered separately this cycle), which uses a similar multi-agent architecture for AML drug repurposing and liver fibrosis target discovery; both teams' positioning as "first" requires careful definitional review. Physical experiments were executed by human researchers, so Robin's autonomy is "intellectual" rather than full lab autonomy — the closed-loop self-driving lab pattern remains aspirational. The 200-fold time-reduction headline conflates literature-review time savings (where AI excels) with end-to-end discovery time, where wet-lab cycles dominate. Repurposing a clinically approved drug is the easiest validation pathway (known safety, known PK, known manufacturing) — generalisability to novel chemical entity discovery remains untested. Source
Partnership Platform Drug Discovery

Alnylam–Inceptive Strike Up-To-$2B Three-Year AI Collaboration to Apply Foundation Models to RNAi Therapeutic Discovery — $30M Upfront, Milestone-Heavy Structure

Alnylam Pharmaceuticals (NASDAQ: ALNY) and Inceptive Nucleics announced on 3 June 2026 a three-year strategic collaboration valued at up to $2 billion to apply Inceptive's foundation models for sequence-based medicines to Alnylam's RNAi therapeutic discovery. The upfront consideration is $30 million in cash plus the purchase of Inceptive equity, with the remaining $1.97 billion tied to preclinical, regulatory, and commercial milestones. The collaboration pairs Alnylam's 20+ years of proprietary siRNA data and six approved RNAi medicines with Inceptive's biology foundation models, focused on modelling target mRNAs, exploring novel chemical modifications to enhance potency and efficacy, and predicting top-performing therapeutic candidates in preclinical models. Inceptive was founded in 2021 by Jakob Uszkoreit, co-author of the seminal "Attention is all you need" paper that established the transformer architecture underpinning modern large language models. In joint exploratory work, Inceptive reports its foundation model achieved "exceptional performance within weeks" characterising siRNA molecules from "relatively small datasets". Contrary view: the $2B/$30M biobucks-to-upfront ratio (67:1) is even more extreme than the industry-typical 50:1 ratio Drug Target Review flagged in its April 2026 AI drug discovery predictions piece, indicating substantial discounting of probabilistic future milestones. Alnylam's six approved RNAi medicines were all discovered without AI foundation models, so the marginal value of AI augmentation over Alnylam's existing rational-design platform is the binding empirical question — and it is unaddressed by the deal terms. Inceptive's "foundation model of life" framing has a strong technology-leader pedigree (Uszkoreit's transformer credentials) but limited publicly-documented siRNA-specific validation. The pace of similar pharma-AI platform deals through 2026 (Lilly-Insilico $2.75B, Takeda-Iambic $1.7B, Roche-Manifold, BMS-Anthropic enterprise) raises the question of whether these are diversification bets across an unproven category. Source
Drug Discovery Platform

ITIF Policy Analysis: AI Drug Discovery Systems Could Strengthen Biopharmaceutical Innovation — Robin Highlighted as Proof-of-Concept for Combinatorial Synthesis Across Scientific Literature

The Information Technology and Innovation Foundation (ITIF) published a policy analysis on 2 June 2026 framing the FutureHouse Robin Nature paper as a proof-of-concept for AI-enabled drug discovery that warrants policy attention. The ITIF piece anchors three policy claims: (a) AI systems like Robin demonstrate "combinatorial synthesis" — the ability to identify non-obvious connections across disparate areas of scientific literature that human expert knowledge, typically more specialised, can miss; (b) a Boston Consulting Group study of 20 AI-focused pharmaceutical companies found 5 of 15 AI-assisted drug candidates that advanced to clinical trials did so in under four years versus the historical 5-6 year average, and BCG/Wellcome Trust projects AI-enabled efforts could reduce discovery and preclinical time and cost by 25-50%; (c) realising this potential at scale requires policy support for data infrastructure, regulation, and incentives. ITIF singles out Robin's literature-processing speed (551 papers in 30 minutes versus an estimated 540 hours of manual processing) as illustrative of AI's breadth advantage, complementing rather than replacing human depth expertise. Contrary view: the BCG 5-of-15 statistic is a small-sample observational comparison that does not control for indication mix, target tractability, or sponsor selection effects (well-funded AI-focused companies tend to select higher-tractability programmes). ITIF is a centre-right policy think tank with explicit pro-innovation framing; its policy claims should be balanced against the rising pattern of AI-discovered-drug clinical failures (Verge VRG50635 ALS Phase 1 failure, Recursion REC-994 disappointment, ProRenata trial discontinuation). The "combinatorial synthesis" framing is appealing but not yet quantitatively benchmarked against human expert teams with equivalent literature access. Source
Drug Discovery Genomics & Proteomics Platform

Google Co-Scientist Formally Published in Nature — Multi-Agent Hypothesis Tournament Validated in AML Drug Repurposing and Liver Fibrosis Target Discovery (Vorinostat 91% Anti-Fibrotic Activity in Hepatic Organoids)

Google's Co-Scientist multi-agent AI system, originally announced as a Google Research blog post in February 2025 and previously tracked here, was formally published in Nature in May 2026 (Gottweis et al., DOI: 10.1038/s41586-026-10644-y), with TechTimes and other outlets covering the peer-reviewed validation around 30 May 2026. The published work documents two preclinical biomedical applications: (a) Co-Scientist proposed drug repurposing candidates for acute myeloid leukaemia (AML) that were validated in multiple AML cell lines in vitro at clinically applicable concentrations; (b) for novel target discovery, Co-Scientist proposed Vorinostat — an FDA-approved anti-cancer drug — as a candidate for liver fibrosis treatment, reducing a key TGFβ-induced chromatin structural change by 91% in human hepatic organoid tests (Stanford collaboration with Gary Peltz; follow-on publication in Advanced Science, DOI 10.1002/advs.202508751). The system is built on Gemini 2.0 and operates a "scientist-in-the-loop" paradigm with specialised hypothesis-generation, ranking-tournament, and experimental-planning agents. Contrary view: Co-Scientist's Nature paper represents peer-reviewed validation of the system, not clinical validation of the drug candidates — all results remain in vitro / organoid, and none have entered human trials. Vorinostat's 91% TGFβ-response reduction in hepatic organoids is a single endpoint measured in an in vitro system that has not been validated against fibrosis regression in patients. The "scientist-in-the-loop" framing concedes that human researchers provided substantial input throughout the process, which complicates clean attribution of "AI-discovered" claims. The arxiv preprint (2502.18864) was posted in February 2025; Nature peer review took approximately 15 months, suggesting substantial revisions were required. Together with FutureHouse Robin (covered separately this cycle), these two parallel Nature publications mark the formal peer-reviewed entry of autonomous-AI-scientist systems into the biomedical literature. Source
Drug Discovery Platform Genomics & Proteomics

Verge Genomics Rebrands as Verge Labs Following VRG50635 ALS Phase 1 Failure — ~90% Workforce Reduction, Pivot to AI Foundation Models of Human Brain Biology (12,000+ Brain Transcriptome Dataset)

Verge Genomics announced on 27 May 2026 a rebrand as Verge Labs and a fundamental strategic pivot, repositioning as an "AI lab for human disease biology" providing pharma and biotech partners with data and AI models for CNS target identification and patient stratification, rather than developing its own therapeutic pipeline. The rebrand follows the December 2025 readout of VRG50635, the company's AI-designed PIKfyve inhibitor for ALS, which failed to benefit patients in its Phase 1 trial. CEO Alice Zhang confirmed Verge laid off approximately 90% of its workforce as part of the restructuring. The new Verge Labs is built around VergeDB, the company's proprietary database of more than 12,000 brain transcriptomes from approximately 6,000 patients — described as one of the largest proprietary multimodal CNS patient datasets ever assembled. Four senior hires join from Altos Labs, Calico, PostEra, and Flatiron Health across AI, business, product, and computational biology functions. Note: this entry falls within the cycle 5 window (25-31 May 2026) and is included here as substantive backfill missed in the prior cycle; it is one of the most consequential AI-drug-discovery failure cases of 2026 and deserves prominent tracker visibility alongside the optimistic Robin/Co-Scientist Nature publications. Contrary view: Verge's pivot is presented as a positive strategic re-positioning, but the underlying economics are that an AI-discovered ALS drug failed in the clinic, the company lost 90% of its workforce, and Verge has been forced to monetise its dataset rather than its drug pipeline. The "frontier AI lab" framing emulates patterns from generalist AI labs (OpenAI, Anthropic) that may not translate to the pharma services market, which is dominated by established CROs (IQVIA, ICON) with deeper enterprise relationships. The 90% layoff figure is a strong indicator that prior $100M+ venture financing and the Eli Lilly ALS partnership did not produce sufficient pipeline value to sustain operations. Verge's outcome is a real-world test of the thesis (covered repeatedly in this tracker) that AI-discovered drugs face the same ~90% clinical failure rate as traditional approaches. Source
Drug Discovery Platform

Nature Editorial Companion: "Why AI Cannot Do Good Science Without Humans" — Editorial Frames Robin and Co-Scientist Nature Papers as Demonstrations Requiring Sustained Human Wisdom, Empathy and Domain Expertise

Nature published an editorial companion piece around 17 May 2026 (d41586-026-01551-3, "Why AI cannot do good science without humans"), accompanied by a second commentary "Teams of AI agents boost speed of research" (d41586-026-01596-4), framing both the FutureHouse Robin paper (Ghareeb et al., s41586-026-10652-y) and the Google Co-Scientist paper (Gottweis et al., s41586-026-10644-y) within an editorial argument that human wisdom, empathy, and domain expertise remain indispensable to good science even as multi-agent AI systems demonstrate substantial speed advantages. The editorial highlights Robin's claimed 200-fold time reduction on its dAMD project but emphasises that the underlying experimental design choices, data interpretation, and ethical framing remained anchored in human researcher input. The commentary surveys parallel AI-agent systems in materials science (Aygün et al., s41586-026-10658-6) and biomedical research, positioning the May 2026 cluster as an inflection point for "AI scientist" claims in peer-reviewed venues. Note: this entry falls outside the strict cycle 6 window (30 May - 7 June 2026) and is included as substantive backfill from mid-May 2026 that was missed in the prior cycle; it provides essential editorial context for the Robin and Co-Scientist primary publications covered separately in this cycle. Contrary view: Nature editorial commentaries are positional advocacy, not original research, and should not be confused with empirical evaluation of AI-scientist systems' actual contribution to discovery outcomes. The "human in the loop" framing is genuinely correct for the current generation of multi-agent systems but may underweight the rate at which agentic capabilities are expanding — a 200-fold speedup today could become a 2,000-fold speedup within 18 months under continued model and infrastructure improvements. Nature has a commercial interest in defending the editorial value of peer-reviewed scientific publication, which the rise of fully-autonomous AI scientists could disrupt; the editorial framing should be read with that institutional context in mind. Source
Antibody Design Genomics & Proteomics Drug Discovery

Nature Machine Intelligence: UniAIR Unified Multimodal Framework for Generalizable Mutation-Effect Prediction Across Antibody, Antigen and T-Cell Receptor Interactions

Han and colleagues published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01243-7) UniAIR, a modular multimodal framework for predicting mutation effects across diverse adaptive immune recognition settings. The framework integrates a standardised data pipeline, an interface-centric sequence-structure fusion transformer that combines evolutionary information with geometric representations, and a suite of extensions for multi-expert consensus and adaptation to predicted-structure inputs. The work directly addresses a long-standing limitation: most prior mutation-effect prediction approaches are designed for specific tasks or modalities (antibody-antigen binding alone, or peptide-MHC alone, or TCR-peptide alone) and struggle to generalise across the heterogeneous, multimodal landscape of immune recognition. UniAIR's value to pre-clinical antibody engineering and TCR therapeutic design is in enabling a single framework for affinity-maturation predictions, escape-mutation modelling, and developability scoring across all three modalities. Code released as UniAIR v1.0.0 (Zenodo DOI 10.5281/zenodo.19471285). Contrary view: unified multimodal frameworks consistently show strong in-distribution performance but degrade on truly novel target classes; the paper's generalisability claims will need independent replication across antibody libraries and TCR repertoires from groups outside the authors' funding consortium. The reliance on predicted structures (rather than experimentally determined complexes) introduces a known failure mode for low-confidence regions of antibody CDR loops. Source
Clinical Trials Platform Genomics & Proteomics

Tempus AI Announces 37 ASCO 2026 Abstracts — Largest Accepted Collection To Date Spanning Multimodal Data and AI-Driven Precision-Medicine Insights Across Oncology

Tempus AI (NASDAQ: TEM) announced on 28 May 2026 that 37 of its abstracts have been accepted for presentation at the 2026 ASCO Annual Meeting (29 May - 3 June, Chicago) — the company's largest accepted ASCO collection to date. Chief Scientific Officer Kate Sasser framed the disclosure as evidence that multimodal data and AI-driven insights are now central to precision oncology, with the abstract breadth covering target discovery, biomarker development, treatment-response prediction, and clinical-evidence generation from real-world data. The abstract count is a leading indicator of platform breadth: a comparable peer pre-AACR/ASCO press release from Insilico Medicine carried four abstracts in April 2026, while Tempus's 37 reflects both the company's data-network advantage and ASCO's editorial appetite for AI-supported clinical-evidence work. Contrary view: abstract acceptance is not peer-reviewed publication; ASCO accepts a high volume of poster-tier work, and the 37 abstracts are heavily dominated by retrospective real-world-evidence studies rather than prospective AI-designed-therapeutic readouts. Tempus's commercial model relies on persuading payers and providers to consume its multimodal data and AI products; abstract count is a marketing-aligned KPI as much as a scientific one. The tracker has flagged source-concentration risk for Tempus AI as a recurring pattern — readers should weight this disclosure accordingly. Source
Clinical Trials Platform Partnership

Massive Bio Unveils Reticulum Nexus AI Operating System for Oncology Trial Access at ASCO 2026 — OpenAI Collaboration Converts Trial Criteria to Machine-Readable Parameters

Massive Bio announced at ASCO 2026 (29 May - 3 June, Chicago) the launch of Reticulum Nexus, framed as an "AI operating system for oncology access", designed to move cancer-trial matching from point-solution search to enterprise infrastructure. The announcement consolidates several prior Massive Bio milestones: an OpenAI collaboration to transform complex trial-eligibility criteria into structured, machine-readable parameters for AI-enabled pre-screening; partnership with the American Cancer Society ACS ACTS programme for equitable trial access; DiMe Seal designation; CMS Medicare App Library listing; and publication of peer-reviewed prospective evidence of AI-driven trial matching at scale. The framing positions trial-access infrastructure as the operational layer that translates AI-discovered targets and biomarker-stratified populations into actual patient enrolment — a category that has historically been a clinical-operations bottleneck on AI-discovered drug programmes. Contrary view: this item sits at the borderline of the tracker's pre-clinical R&D scope (Tempus Active Follow-Up category previously flagged as similarly borderline) but is included here because Reticulum Nexus's OpenAI-translated eligibility-criteria capability has direct implications for AI-discovered-drug trial design. Trial-matching platforms have historically struggled to demonstrate that improved patient-criteria parsing translates to enrolment-rate improvements at portfolio level. Massive Bio's commercial model includes hospital-network licensing that introduces vendor-lock-in considerations. Source
Clinical Trials Drug Discovery

Mayo Clinic 30+ ASCO 2026 Studies Spanning Precision Oncology and AI — Sinicrope Deep-Learning Tumour Microenvironment Quantification Linked to Postoperative ctDNA in Phase 3 FOLFOX Adjuvant Colon Cancer Trial (N0147 Alliance, Abstract 3525)

Mayo Clinic Comprehensive Cancer Center announced on 28 May 2026 that more than 30 of its studies will be presented at the 2026 ASCO Annual Meeting (29 May - 2 June, McCormick Place Chicago), highlighting advances in precision oncology, early cancer detection, AI, and personalised cancer care. The headline AI-driven study (Abstract 3525, poster session 30 May): Frank Sinicrope and colleagues describe a deep-learning approach to quantify tumour-microenvironment features associated with postoperative ctDNA status and outcomes in the Phase 3 FOLFOX-based adjuvant colon cancer trial N0147 (Alliance). The work links spatial computational pathology to a clinically validated minimal-residual-disease biomarker (ctDNA) within an already-completed, statistically-powered randomised Phase 3 cohort — providing an unusually strong AI-image-analysis-meets-clinical-outcome validation framing. Contrary view: ASCO poster format limits the depth of methodological disclosure; deep-learning histopathology features-associated-with-ctDNA studies have a track record of in-cohort overfitting that fails to replicate in independent validation sets. The N0147 cohort is well characterised but ethnically narrow; generalisability to community-oncology populations remains untested. The tumour-microenvironment-to-ctDNA correlative claim does not by itself prove causal mechanism. Source
Genomics & Proteomics Drug Discovery

Nature Machine Intelligence: pUniFind Large-Scale Unified Foundation Model for Peptide Mass Spectrum Interpretation Trained on 100M+ Spectra — 42.6% Increase in Identified Peptides in Immunopeptidomics

Zhao, Mao, Wang, Chi and colleagues published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01234-8) pUniFind, a large-scale multimodal foundation model for proteomics that unifies end-to-end peptide-spectrum scoring with open zero-shot de novo sequencing. Trained on over 100 million open-search-derived spectra, pUniFind aligns spectral and peptide modalities through cross-modality prediction alongside carefully designed pre-training tasks. Reported performance: 42.6% increase in identified peptides in immunopeptidomics applications; supports over 1,300 modifications; identifies 60% more peptide-spectrum matches than existing de novo methods despite a 300-fold larger search space; an integrated deep-learning quality-control module recovers 38.5% additional peptides, including 1,891 mapped to the genome but absent from reference proteomes. The model addresses a long-standing limitation: most prior mass-spectrometry deep-learning systems are feature extractors rather than unified scoring frameworks, which has constrained their applicability to modification-rich and de novo workflows. Direct relevance to drug discovery: neoantigen identification (immunopeptidomics), proteomics target validation, and post-translational modification mapping. Contrary view: large-scale pretrained proteomics models show strong benchmark numbers but have historically struggled with the long tail of biologically interesting low-abundance peptides; the 42.6% immunopeptidomics improvement is measured against curated benchmarks rather than prospective patient samples. The 300-fold larger search space risks higher false-discovery rates that the QC module's claims need to be audited against prospectively. Source
Clinical Trials Platform

FDA Extends AI-Enabled Early-Phase Clinical Trials RFI Comment Period to 29 June 2026 — Initial 30-Day Window Generated Sufficient Industry Engagement to Justify Extension

The U.S. Food and Drug Administration published a Federal Register notice (FR Doc. 2026-10602, public inspection 27 May 2026) extending the comment period on the AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program Request for Information (Docket FDA-2026-N-4390) by 30 days, with comments now due by 29 June 2026. The original RFI was published in the Federal Register 29 April 2026 (91 FR 23100) and sought industry input on a proposed pilot programme to evaluate how AI-enabled technologies can improve efficiency, speed and quality of decision-making in early-phase clinical trials, with the pilot guided by principles aligned with the NIST AI Risk Management Framework. The extension was prompted by a request for additional response time, indicating substantive industry engagement and likely reflecting the operational complexity of the input the RFI requested (trial scenarios, technology readiness, collaboration models, evaluation metrics). Contrary view: comment-period extensions are common procedural events and do not by themselves indicate the agency intends to act on the responses; FDA's previous AI RFIs (including the January 2025 draft AI guidance) attracted significant industry engagement without producing rapid binding rules. The extension delays the pilot-selection timeline disclosed in the original RFI, which had targeted July 2026 final selection criteria and August 2026 pilot selections. Industry resistance to exposing proprietary AI model architectures and training data remains the binding constraint on pilot uptake. Source
Drug Discovery Platform

Drug Target Review (Clozel/Owkin): Why AI Drug Discovery Models Need Large-Scale Patient-Derived Data to Translate Computational Predictions to Clinical Outcomes

Drug Target Review published a contributor article from Thomas Clozel MD (Owkin Co-founder and CEO) framing the central limitation of most current AI drug-discovery systems: they are trained primarily on laboratory datasets or controlled experimental systems that do not fully reflect the biological complexity seen in patients, and as a result, predictions that perform well in computational or pre-clinical settings often fail to translate to clinical outcomes. Clozel argues patient-derived data is essential for improving the biological relevance of drug-discovery models, and describes Owkin's accumulated multimodal patient datasets from a network of more than 800 hospitals, combined with experimental and clinical validation integrated directly into model development. The piece situates Owkin's agentic K Pro AI Scientist and its NVIDIA and Anthropic partnerships within this thesis, alongside an INVOKE Phase 1a/1b trial as the company's own validation pathway. Contrary view: the piece is a contributor article rather than peer-reviewed work, framing Owkin's commercial differentiation as a methodological principle; competing AI-drug-discovery approaches built on phenomic screening (Recursion), structural prediction (Isomorphic), or generative chemistry (Insilico, Iambic) would frame the "clinical translation gap" differently. Owkin's 800-hospital data network is a commercial moat that introduces vendor-lock-in considerations the piece does not engage. The argument that patient-derived data is necessary for clinical translation is plausible but not yet proven in head-to-head benchmarking against laboratory-data-trained alternatives. Source
Clinical Trials Platform

Median Technologies Showcases eyonis SaMD Suite and iCRO AI Imaging at ASCO 2026 — Cachexia Body-Composition Analysis and Radiopharmaceutical Image Processing

Median Technologies (Euronext: ALMDT) announced on 26 May 2026 that the company will participate in ASCO 2026 (29 May - 2 June, McCormick Place Chicago, booth 36102) to showcase its eyonis Software-as-Medical-Device suite for AI-powered early cancer diagnosis and its iCRO central imaging services for oncology drug developers. Two technical sessions are scheduled: a 31 May radiopharmaceutical image-processing demonstration, and Antoine Iannessi (VP Medical Affairs, iCRO) presenting body-composition AI analysis for cachexia assessment in oncology trials. The cachexia angle is notable because body-composition deep learning of CT cross-sections has emerged as a tractable AI imaging endpoint for oncology drug developers, with applications spanning patient stratification, sarcopenia prognostication, and chemotherapy dose individualisation. The iCRO offering covers central imaging and AI analysis for oncology trials. Contrary view: SaMD AI imaging for cachexia/body composition has a credible methodological basis but remains a relatively narrow regulatory pathway with FDA/EMA acceptance still developing for AI-derived endpoints in oncology trials. The company's stock-listed status (Euronext: ALMDT) means corporate disclosure incentives are tilted toward maximising ASCO visibility relative to scientific specificity. The framing of body-composition analysis as a primary AI capability rather than an adjunct sits in tension with the wider AI-oncology landscape, where target-discovery and biomarker-prediction are higher-impact applications. Source
Genomics & Proteomics Drug Discovery

Drug Target Review (Sanjana, NYU/New York Genome Center): Scalable CRISPR Screens Link Noncoding GWAS Variants to Causal Genes and Therapeutic Targets

Drug Target Review published a contributor article from Professor Neville Sanjana (NYU, Core Faculty New York Genome Center) addressing one of the most stubborn translational gaps in drug discovery: genome-wide association studies have linked thousands of genetic variants to disease, but the great majority remain disconnected from drug-relevant biology because GWAS measures correlation, not causal mechanism. Sanjana describes how scalable CRISPR perturbation screens combined with single-cell sequencing — including the STING-seq workflow developed by his lab — can systematically link noncoding variants to their causal target genes, producing the mechanistic substrate that enables therapeutic target nomination. The work sits at the intersection of two trends the tracker has documented through 2026: (a) the move from associational genomics to functional perturbation as the basis for AI target-identification, and (b) the integration of CRISPR-screen perturbation data with multimodal AI models (covered separately under Recursion/Owkin/Insilico approaches). Contrary view: STING-seq and similar functional-genomics workflows are powerful but scale-limited; they typically require pre-specification of candidate loci and produce mechanistic insight only for variants the screen is designed to probe. The article is contributor-tier framing rather than primary disclosure of new data; Sanjana's underlying STING-seq paper was published in Science in 2023 (covered in the tracker), so this piece is best read as analytical synthesis. The argument that CRISPR-perturbation-derived target nomination is now ready for routine AI integration is plausible but not yet demonstrated at portfolio scale. Source
Drug Discovery Genomics & Proteomics

Nature Machine Intelligence: MIDAS Multimodal Graph Neural Network for Immuno-Oncology Target Discovery — Outperforms OpenTargets Baseline, Validated on Patient-Derived Platform

Augustine and colleagues published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01201-3) MIDAS (Mining Immunotherapy Drug tArgetS), a multimodal graph neural network system for immuno-oncology target discovery. MIDAS leverages gene interactions, multi-omic patient profiles, immune-cell biology, antigen processing, disease associations, and phenotypic consequences of genetic perturbations within a single integrated graph-learning architecture. Reported performance: generalises to time-sliced data; outcompetes state-of-the-art baselines including OpenTargets; ranks approved targets above those in clinical development; recovers immunotherapy-response-associated genes in unseen patients, thereby capturing immunotherapy response determinants. Promising MIDAS-nominated candidates were validated on a clinically relevant patient-derived platform — addressing the wet-lab validation gap that the Owkin/Clozel contributor piece (also covered this cycle) flagged as the central limitation of contemporary AI drug-discovery models. Note: this entry falls outside the strict cycle window (25-31 May 2026) but is included as substantive peer-reviewed methodology coverage from earlier in May that was missed by the prior cycle's USF/PanPep-focused Nature Machine Intelligence selection. Contrary view: graph-neural-network target-discovery systems have a well-established track record of producing strong in-distribution rankings but degrading on truly novel biology where the underlying knowledge graphs are sparsest — i.e. exactly the targets that would deliver the most value. The OpenTargets baseline is well-curated but is not the strongest available competitor; head-to-head benchmarking against Insilico PandaOmics, Recursion phenomic-screen-derived targets, and Open Targets Genetics Portal would be more diagnostic. The patient-derived-platform validation is encouraging but unblinded. Source
Drug Discovery Platform

Drug Target Review AACR 2026 Recap: AI-Driven Drug Discovery Maturing From Faster Design Cycles to Improved Clinical Probability of Success — KRAS Inhibitors, TF Degraders, LabGenius Closed-Loop Antibody Systems

Sam Wightwick of Drug Target Review published an analytical recap ("AACR 2026 part 1: AI design, precision biology and the next wave of oncology innovation") summarising on-floor discussions with drug discovery leaders at the AACR Annual Meeting (San Diego, 17-22 April 2026). The piece characterises the dominant shift across the meeting as one from isolated technological breakthroughs to integrated discovery systems connecting AI, biology, and translational execution. Specific examples cited include: AI-generated KRAS inhibitors using non-obvious transient binding sites; transcription-factor degraders advanced via generative chemistry; LabGenius Therapeutics' closed-loop antibody design operating across roughly 2 million unique antibody architectures with primary human cell-based assays as the readout; and an interview with Insilico's Dr Halle Zhang on AI-driven KRAS work. The framing reflects what AACR's first dedicated AI plenary signals — that the conference programme committee views AI as warranting the same programmatic weight as clinical trials and discovery science. Contrary view: aggregator-tier conference coverage, not peer-reviewed primary data; many of the cited "AI-generated" molecules remain pre-clinical with no head-to-head benchmark against traditional discovery. The piece's repeated framing of AI as "decision support, not automation" (echoed across multiple AACR 2026 analytical pieces) is a useful reality check against the "compute halves the time" claim pattern Anthropic and Isomorphic-style messaging often produces. A promised part 2 (spatial biology, cell therapy manufacturing, ADC design) had not yet published as of 24 May. Source
Drug Discovery Platform

Collaborative Drug Discovery Industry Roundup: PubMed "Rethinking Nature's Pharmacy" Frames AI-Era Renaissance for Natural-Product Drug Discovery After 1990s Pharma Pullback

The Drug Discovery Industry Roundup by Barry Bunin (Collaborative Drug Discovery, 18 May 2026) flags "Rethinking Nature's Pharmacy: AI Era and Natural Product Drug Discovery" (PubMed) as the analytical paper of the week. The paper argues that natural products — bioactive small molecules from plants, animals, fungi and microorganisms — constituted the cornerstone of medical practice for thousands of years before the pharmaceutical sector experienced a pronounced decline in NP-driven research beginning in the 1990s, and that AI techniques (foundation models trained on natural-product chemistries, scaffold-aware contrastive learning, biosynthetic gene cluster mining) now make NP discovery tractable again at industrial scale. The roundup pairs this with separate coverage of AWS Amazon Bio Discovery's bioFM catalogue. Cross-referencing with the NaFM Nature Machine Intelligence paper (Ding et al., DOI 10.1038/s42256-026-01226-8, covered in the prior cycle) shows a coherent thesis emerging: that AI-driven natural-product re-exploration is the under-invested category for the next 24 months. Contrary view: aggregator-tier source citing other coverage rather than primary disclosure; Bunin's CDD has a commercial interest in CDD Vault adoption that the roundup format does not flag explicitly. The "1990s pharma pullback" framing is broadly accurate but does not engage with why pharma withdrew — supply chain unpredictability, IP weakness on natural scaffolds, and metabolic-stability failures in clinical translation — which AI tools do not address directly. Source
Drug Discovery Platform

Goethe University Frankfurt + Philipps Marburg + Fraunhofer ITMP: genESOM Generative AI Reduces Lab Animal Use in Pre-Clinical Drug Testing by 30-50%

Drug Target Review (12 May 2026) reported that researchers at Goethe University Frankfurt, Philipps University Marburg, and the Fraunhofer Institute for Translational Medicine and Pharmacology (ITMP) have developed genESOM, a generative artificial intelligence platform that could reduce the number of laboratory animals needed in early-stage drug testing by between 30 and 50%. The system is a network of thousands of artificial neurons that learn the internal structure of small experimental datasets and then generate additional synthetic data points whose statistical properties reproduce those of the original real experiments — effectively simulating a larger animal cohort than was actually used. Developed by data scientist and clinical pharmacologist Jorn Lotsch (Goethe University) with computer scientist Alfred Ultsch (Philipps University Marburg); the researchers explicitly note they do not conduct animal experiments themselves, framing the work as a methodological contribution to the 3Rs (replacement, reduction, refinement) of pre-clinical research. Contrary view: synthetic-data augmentation introduces a category of statistical risk distinct from sampling noise — if the generative model amplifies dataset-specific biases, the "additional animals" are not independent observations and can produce false positives that survive into in vivo confirmation studies. The 30-50% reduction range is presented as projected rather than independently audited across a portfolio of pre-clinical programmes. Regulatory acceptance under FDA/EMA AI principles for replacement of mandated animal study sample sizes remains unestablished. Source
Genomics & Proteomics Drug Discovery

Nature Machine Intelligence: USF Reusability Report on PanPep Meta-Learning for T-Cell Receptor Antigen Recognition — Evaluation Framework Tests AI Reliability for Cancer Immunotherapy and Vaccine Design

He, Wang and Xu (University of South Florida Health Morsani College of Medicine) published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01236-6) a Reusability Report on PanPep (Pan-peptide meta-learning), an AI model designed to predict how T-cell receptors recognise and bind antigens. The USF team developed a systematic evaluation framework that can be applied to a broad class of immunology prediction problems including peptide-HLA binding, peptide-TCR interaction, antigen presentation, and other peptide- or antigen-driven interactions. The framework tests AI predictions under realistic conditions rather than relying solely on curated controlled-data benchmarks. The work is directly relevant to pre-clinical immunotherapy and vaccine design programmes where TCR-antigen prediction informs neoantigen identification and patient-stratification decisions. Contrary view: a Reusability Report is a validation/replication contribution rather than a methodological breakthrough; the authors themselves emphasise that meta-learning approaches still require careful testing and refinement before they can be safely used to guide personalised care, and that real-world applications often involve entirely new immune targets where it remains unclear how well these models handle truly unseen cases. The piece raises but does not resolve generalisation-from-curated-benchmarks as the binding constraint on AI immunology models. Source
Drug Discovery Partnership

Aarvik Therapeutics + ArriVent BioPharma: MUTTA and AQUALINK Tetravalent ADC Platforms at AACR 2026 — AV-P138-ADC Dual-Target MUC16/NaPi2b for Ovarian and Endometrial Cancer

Aarvik Therapeutics presented two posters and a minisymposium talk at AACR 2026 (San Diego, 17-22 April 2026) showcasing its MUTTA (Multi-epitope Targeting Tetravalent Antibody) and AQUALINK platforms for next-generation antibody-drug conjugate (ADC) design. The headline collaboration is AV-P138-ADC (also designated ARR-002), a site-specifically conjugated dual-target tetravalent ADC against MUC16/NaPi2b for ovarian and endometrial cancers, discovered by Aarvik as part of a research collaboration with ArriVent BioPharma (NASDAQ: AVBP); ArriVent has since exclusively licensed the molecule and plans global development. Aarvik's framing argues that tetravalent dual-target ADCs may overcome the safety and efficacy limitations of conventional single-target or bivalent bispecific ADCs. Note: this entry falls outside the strict 14-day cycle window (17-24 May 2026) but is included as substantive ADC platform disclosure missed by the prior cycle. Contrary view: pre-clinical platform disclosures of tetravalent ADCs have a history of underperforming bispecific simpler designs in clinical translation; the safety case for tetravalent ADCs depends on linker-payload stability across multiple conjugation sites, which the AACR poster format does not stress-test. The Aarvik-ArriVent commercial structure means downstream clinical disclosure will be ArriVent-controlled. Source
Platform Clinical Trials Drug Discovery

Debiopharm at AACR 2026: First Phase I Disclosure of WEE1+PKMYT1 Combination (Zedoresertib + Lunresertib MYTHIC NCT04855656) Across CCNE1/FBXW7/PPP2R1A-Altered Solid Tumours; Dual-Payload ADC Technology and AI-Driven Biomarkers

Debiopharm International presented at AACR 2026 (San Diego, 17-22 April 2026) the first clinical data disclosure from the Phase I MYTHIC study (NCT04855656) — a first-in-class synthetic-lethal combination of WEE1 inhibitor zedoresertib (Debio 0123) and PKMYT1 inhibitor lunresertib (Debio 2513) in patients with advanced solid tumours harbouring CCNE1, FBXW7, or PPP2R1A genomic alterations, presented by Timothy A. Yap (MD Anderson). Pre-clinical work supporting the trial showed dual WEE1/PKMYT1 inhibition produces deeper tumour regressions than either agent alone via G2/M checkpoint collapse in cancers addicted to replication stress. Debiopharm also presented dual-payload ADC technology designed for therapeutic synergy and resistance management, and a 3D virtual-immunohistochemistry tool that integrates multi-omics and spatial profiling for sample-specific bispecific antibody/ADC treatment selection. Note: this entry falls outside the strict 14-day cycle window but is included as a substantive clinical milestone in AI-supported biomarker-driven oncology missed by the prior cycle. Contrary view: synthetic-lethal combinations have a track record of strong pre-clinical signals failing to replicate in Phase I-Ib biomarker-selected populations once heterogeneity is properly accounted for. The 3D virtual-IHC tool is framed as "AI-driven" but the methods sit closer to spatial-statistics dashboarding than to deep learning, raising the usual labelling-inflation question. Source
Drug Discovery Platform

Owkin K Pro at AACR 2026: Agentic Research Co-Pilot Now Integrates Multimodal Patient Data from 100+ Partner Hospitals With Expert-Defined Reproducible Analytical "Skills"

Owkin presented its K Pro agentic research co-pilot at AACR 2026 (San Diego, 17-22 April 2026), one year after the system was first introduced at AACR 2025. K Pro now integrates multimodal patient data from Owkin's network of 100+ partner hospitals with a growing library of expert-defined, reproducible analytical workflows (which Owkin calls "skills") built from a decade of biopharma experience — including multimodal target characterisation, spatial differential expression analysis, and outcome prediction. The system's framing is that an agentic AI for translational oncology should not just generate plausible answers but should run the analysis, show its work, and produce results auditable against expert-defined methodology. The library grows through ongoing collaboration with translational scientists. Contrary view: agentic AI in oncology research has a short track record (Genentech CLADD, AstraZeneca ChatInvent, and now K Pro all launched within ~6 months) and there is no independent benchmarking framework comparing "skills"-based architectures against unstructured LLM-driven workflows on real translational questions. The 100+ partner-hospital data network is Owkin's commercial moat, which introduces vendor-lock-in considerations that the conference-tier framing does not engage. Source
Drug Discovery Platform

Drug Target Review: AI Drug Discovery Platform Models Protein Flexibility During Molecular Binding to Improve Hit Rates and Reduce Late-Stage Failures

Drug Target Review (15 April 2026) reported on a new AI-driven drug discovery platform that explicitly models protein dynamics — the conformational flexibility a protein exhibits during ligand binding — rather than treating the target as a static structure. The platform's framing addresses one of the longest-running structural-biology critiques of pure AlphaFold-style static predictions: a substantial fraction of clinically relevant binding events involve cryptic pockets, induced-fit conformational changes, or transient states that single-structure prediction misses. By incorporating dynamics, the developers argue, the platform should improve binding-affinity predictions and reduce the high failure rates associated with conventional in silico drug-discovery campaigns where computational hits do not translate to wet-lab activity. Note: this entry falls outside the strict 14-day cycle window but is included as substantive methodology coverage missed by the prior cycle. Contrary view: dynamics-aware structure prediction has been a research thesis for over a decade (e.g. ensemble docking, molecular-dynamics-augmented scoring) and the practical gains over static-structure approaches have been narrow and target-specific in published benchmarks. The DTR piece is news coverage rather than peer-reviewed validation; the underlying platform's training data, benchmark performance, and prospective wet-lab validation are not addressed in the reporting. Source
Antibody Design Drug Discovery

Drug Target Review: "AI to Antibody in Days" — High-Throughput Integration Compresses Wet-Lab Bottleneck (Sino Biological: 10,000+ Antibodies/Month, 10 Days Gene to Antibody)

Drug Target Review (10 April 2026) published an analytical feature on the wet-lab bottleneck that has historically constrained AI-designed-antibody timelines. AI sequence design can now generate thousands of candidate antibodies per day, but downstream expression, purification, and characterisation typically required weeks per candidate — a bottleneck that has muted the productivity gains AI-design vendors report. The piece highlights Sino Biological's high-throughput integrated platform combining deep expertise in HTP gene synthesis, vector construction, and optimised transient antibody expression, delivering industry-leading throughput (10,000+ antibodies/month) and 10-day gene-to-antibody timelines, paired with immediate access to a catalogue of 10,000+ premium recombinant proteins for binding validation. For projects requiring even faster turnaround or involving difficult-to-express proteins, the article notes cell-free protein synthesis (CFPS) systems as a rapid alternative. Note: this entry falls outside the strict 14-day cycle window but is included as substantive workflow infrastructure coverage missed by the prior cycle. Contrary view: throughput is necessary but not sufficient — the binding validation step against a recombinant-protein library does not test developability (aggregation, off-target binding, manufacturability) that determine clinical translation. "Days to antibody" is a manufacturing-stage metric, not a drug-discovery-success metric. Source
Antibody Design Drug Discovery

Drug Target Review: Promatix Biosciences Bispecific ADCs Built on Membrane Proteomics — Targeting Previously Unexplored Tumour Biology for Selectivity and Reduced Toxicity

Drug Target Review (6 April 2026) reported on Promatix Biosciences' approach to bispecific antibody-drug conjugates (ADCs) leveraging membrane proteomics to identify novel tumour-selective antigen pairs. CEO and co-founder Dr Michael Hunter framed the company's thesis around the persistent challenge of ADC development: how to deliver potent cytotoxic payloads to tumour cells without damaging healthy tissue. Promatix's bispecific design targets tumour-restricted antigen combinations identified via deep membrane proteomic surveying — an approach intended to improve selectivity beyond what single-target ADCs achieve and to enable lower-toxicity payload delivery. The article positions the work within the broader 2026 ADC field push toward dual-target and multispecific designs (also seen at AACR 2026 from Aarvik, Debiopharm, NEOK Bio, Whitehawk and others). Note: this entry falls outside the strict 14-day cycle window but is included as substantive ADC methodology coverage missed by the prior cycle. Contrary view: membrane proteomics surveys produce candidate antigen pairs but do not by themselves prove tumour-restricted expression versus normal-tissue expression at the levels relevant for ADC therapeutic index. The bispecific-ADC field has multiple credible architectures competing simultaneously; differentiation will depend on clinical data, not pre-clinical mechanistic framing. Source
Partnership Platform Drug Discovery

Anthropic and Gates Foundation Launch $200M Four-Year AI Partnership Targeting Vaccines for Polio, HPV, Preeclampsia and Disease Forecasting

Anthropic and the Bill & Melinda Gates Foundation announced (14 May 2026) a four-year, $200 million partnership covering global health, life sciences, education, and economic mobility. Roughly half the value comes as Claude usage credits and Anthropic engineering support; the remainder is Gates Foundation grant funding. The health and life-sciences component focuses on using Claude to screen vaccine candidates computationally before pre-clinical development for overlooked diseases, starting with polio, HPV, and eclampsia/preeclampsia (HPV is cited as causing approximately 350,000 deaths annually, 90% in low- and middle-income countries). A separate integration with the Foundation's Institute for Disease Modeling (IDM) aims to make malaria and tuberculosis transmission forecasts more accessible and more predictive. The deal extends Anthropic's recent push into healthcare (Vas Narasimhan board appointment; Coefficient Bio acquisition in April 2026; Claude for Life Sciences launched October 2025). Contrary view: a $200M commitment over four years (~$50M/year) is small relative to large pharma R&D budgets or to single deals such as the recent Isomorphic Labs Series B. The work funds vaccine-candidate screening and disease-forecasting tools, not novel target discovery or clinical trials. Public-goods commitments at this scale also function as regulatory and reputational positioning for Anthropic ahead of evolving healthcare AI rules. Concrete pre-clinical R&D output attributable to the partnership will be measurable only over the four-year horizon. Source
Partnership Clinical Trials Platform

Tempus AI Expands Strategic Collaboration with Bristol Myers Squibb to Optimise Five Oncology and Neuroscience Clinical Trial Programmes Using AI on Real-World Data

Tempus AI (NASDAQ: TEM) and Bristol Myers Squibb (NYSE: BMY) announced a new initiative applying Tempus's Lens AI-enabled analytical platform and its multimodal real-world data library to optimise clinical trial design and improve Probability of Technical & Regulatory Success (PTRS) across five initial development programmes. Initial scope covers solid-tumour oncology assets in lung, colon, and prostate cancers, with extension into Alzheimer's disease neuroscience programmes planned. The collaboration uses de-identified longitudinal patient records to pressure-test trial assumptions, validate control group selection, characterise patient heterogeneity, and identify subgroups most likely to benefit from investigational therapies. The initiative builds on existing Tempus-BMS work, including the Next Pathways programme across 13 community-based health systems for advanced non-small-cell lung cancer patient identification. Contrary view: AI-driven PTRS optimisation is a process improvement, not a pipeline-acceleration claim; the partnership does not change the underlying drug candidates being tested. Tempus reported Q1 2026 revenue of $348.1M (up 36.1% year-over-year) but analyst price targets diverge sharply ($35 from Jefferies to $80 from BTIG), reflecting unresolved questions about how durable real-world-data multi-modal AI is as a defensible commercial moat versus other clinical-data platforms (e.g. Flatiron, Komodo Health, Truveta). Source
Drug Discovery Platform

Nature Machine Intelligence: Penn ApexGO Generative AI Optimises Peptide Antibiotics — 85% of AI-Designed Candidates Active In Vitro, Mouse Efficacy Matches Polymyxin B

Torres, Zeng, Wan, Maus, Gardner and de la Fuente-Nunez published in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01237-5) ApexGO — APEX generative optimisation — a generative-AI framework that takes a promising-but-imperfect antimicrobial peptide template and iteratively redesigns it for improved activity. The method combines a transformer-based variational autoencoder, which embeds peptide sequences in a continuous latent space, with Bayesian optimisation that proposes sequence edits to boost predicted antimicrobial potency. Starting from 10 peptide templates, 85% of AI-generated derivatives showed antimicrobial activity in laboratory tests against disease-causing bacteria, and 72% outperformed their starting peptides; two of the ApexGO-designed peptides reduced bacterial counts in mice at levels comparable to polymyxin B, an FDA-approved last-resort antibiotic. The work builds on the group's prior APEX discriminative model and addresses a key limitation: APEX could rank candidates but could not propose improvements. Contrary view: ApexGO optimises against another computer model (APEX), introducing a feedback-loop risk that the generative model exploits APEX's blind spots rather than producing genuinely better antimicrobials. The 85% in vitro hit rate is impressive but is measured against template-derivative candidates rather than first-in-class novel scaffolds, and only two compounds advanced to mouse infection models — far from the validation depth required for clinical candidate nomination. Generalisability to non-peptide modalities is asserted but unproven. Source
Funding Platform Drug Discovery

Isomorphic Labs Closes $2.1B Series B Led by Thrive Capital with UK Sovereign AI Fund Participation; First-in-Human Oncology Trials Targeted for End-2026

Isomorphic Labs (the London-based Alphabet/Google DeepMind spinout founded in 2021) closed a $2.1 billion Series B on 12 May 2026 — one of the largest single funding rounds for an AI-first drug discovery company. Thrive Capital led the round; new investors include Temasek (Singapore), MGX (Abu Dhabi), and the UK Sovereign AI Fund, with continued participation from Alphabet, GV, and CapitalG. Total capital raised reaches approximately $2.6 billion. Stated use of proceeds: scaling the IsoDDE (Isomorphic Drug Design Engine) platform; expansion across London, Cambridge (Massachusetts), and Lausanne sites; and advancing both partnered and internal programmes toward first-in-human clinical studies. The company has multi-billion-dollar collaborations with Eli Lilly, Novartis, and J&J, and reaffirms an end-2026 target for its first oncology Phase I — a date previously slipped from end-2025. Contrary view: zero human patients have been dosed to date with any Isomorphic-designed candidate. President Demis Hassabis's earlier "first clinical trials by end of 2025" statement later clarified as referring to pre-clinical work; the distinction is non-trivial, and analysts characterise the revised end-2026 target as reflecting the difficulty of converting AI-generated predictions into clinically validated therapeutics. MedCity News reported the raise without any disclosed compound names, target diseases, or specific Phase I timelines beyond "by end of 2026." The UK Sovereign AI Fund participation marks one of the first direct sovereign capital deployments into AI drug design specifically. Source
Drug Discovery Platform Clinical Trials

Moderna Confirms AI-Led Pre-Clinical Hantavirus Vaccine Research Pre-Dated MV Hondius Outbreak; 750+ Internal AI Models Deployed Since OpenAI Partnership

Following a hantavirus outbreak aboard the MV Hondius cruise ship in April 2026, Moderna confirmed (Bloomberg, 8 May 2026; broader analytical coverage 12 May) that it had been conducting pre-clinical hantavirus vaccine research in collaboration with the U.S. Army Medical Research Institute of Infectious Diseases (USAMRIID) and the Vaccine Innovation Center at Korea University College of Medicine before the outbreak. The work, under Moderna's mRNA Access Program, targets hantaviruses causing Haemorrhagic Fever with Renal Syndrome (HFRS) — designated by WHO as a potential "Disease X" pathogen. The company says it has deployed more than 750 internal AI models across scientific, regulatory, and operational workflows since its early-2023 OpenAI partnership, supporting a strategic plan to bring up to 15 new mRNA products to market in five years with a ~6,000-person workforce — Moderna CEO Stéphane Bancel argues that traditional biopharma methods would have required "a hundred thousand people". Contrary view: the research is pre-clinical, has not entered human testing, and is not part of any approved or planned vaccination programme; the AI tool stack is primarily operational and workflow-acceleration rather than de novo antigen design, and most disclosed mRNA-design steps remain conventional. The disclosure created a near-6% share-price bump on the day of broader coverage, raising the usual question about whether outbreak-adjacent timing inflates short-term valuation without changing underlying clinical-stage milestones. Source
Genomics & Proteomics Platform

Cell: RegVelo (Stowers Institute + Helmholtz Munich) Connects Gene-Regulatory Inference with Single-Cell Velocity to Predict Cell Fate Decisions

Researchers at the Stowers Institute for Medical Research and Helmholtz Munich published RegVelo in Cell (DOI: 10.1016/j.cell.2026.04.022), a deep-learning framework that links two previously separate areas of single-cell biology — RNA velocity (which infers a cell's near-future transcriptional trajectory) and gene-regulatory network inference (which identifies the transcription factors steering that trajectory). RegVelo allows researchers not only to predict where a cell is headed in developmental space but also to identify which regulators are most likely steering it to its final fate. The team demonstrated the framework on developmental cell-fate decisions including zebrafish embryos. RegVelo is methodologically relevant to pre-clinical drug discovery because cell-fate-decision regulators are a class of therapeutic target (e.g. in oncology, regenerative medicine, and fibrosis programmes) historically difficult to identify from omics data alone. Contrary view: the published demonstrations focus on developmental biology rather than therapeutic-target validation; integrating gene-regulatory-network inference with velocity inherits the failure modes of both component methods (velocity estimates assume steady-state kinetics, GRN inference is noisy for low-abundance regulators). Independent benchmarking against alternative GRN-aware velocity frameworks remains limited, and translation from regulator identification to drug-discoverable therapeutic targets requires substantial downstream wet-lab validation. Source
Platform Genomics & Proteomics

Nature Machine Intelligence Perspective: CRG Hunklinger and Ferruz Propose Roadmap for Explainable AI in Protein Language Models — Five Roles Beyond Today's "Evaluator" Use

Andrea Hunklinger and Noelia Ferruz (Centre for Genomic Regulation, Barcelona) published a Perspective in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01232-w) surveying the application of explainable artificial intelligence (XAI) to protein language models (pLMs). The authors review existing literature and propose organising XAI applications around four points in the modelling pipeline — training data, user-provided inputs, internal model architecture, and input-output relationships — and identify five potential roles for XAI in protein research: Evaluator, Multitasker, Engineer, Coach, and Teacher. The authors note that almost all current XAI work in pLMs is restricted to the Evaluator role (checking whether the model has learned patterns biologists already know, such as binding sites or structural motifs). The Perspective frames the absence of explainability as both a scientific and a safety concern as pLMs increasingly shape real-world biotechnology decisions, including for therapeutic enzyme and antibody design. Contrary view: a Perspective paper is a position piece, not a methodological contribution; the authors do not introduce new XAI techniques or benchmark existing ones. The roadmap is qualitative — operationalising it for industrial protein design requires substantial follow-up methodology development. The piece does, however, name a real gap: current pLM-based design pipelines (used in commercial settings including biotech-pharma partnerships) operate largely as black boxes, with implications for regulatory acceptance under emerging FDA/EMA AI principles. Source
Platform Drug Discovery

Nature Machine Intelligence: NaFM Foundation Model Pre-Trained for Small-Molecule Natural Products — Scaffold-Aware Contrastive Learning for Taxonomy, Genome Mining and Virtual Screening

Ding, Qiang, Zhou and colleagues (Peking University, Xi'an Jiaotong University) published NaFM in Nature Machine Intelligence (DOI: 10.1038/s42256-026-01226-8) — a foundation model pre-trained specifically for small-molecule natural products. NaFM departs from the prevailing "one-model-for-each-task" pattern in natural-product deep learning, instead using a single pre-trained representation supporting natural-product taxonomy classification, genome mining (linking biosynthetic gene clusters to chemical scaffolds), and virtual screening for drug discovery applications. The model's pre-training strategy combines contrastive learning (weighted by scaffold similarity) and masked-graph learning, emphasising the evolutional information carried in molecular scaffolds while capturing side-chain modifications. Reported state-of-the-art results across multiple downstream natural-product benchmarks, with virtual-screening experiments suggesting the learned representations support more effective identification of potential drug candidates. Natural products account for roughly 60% of FDA-approved drugs since 1981, but their chemical complexity has historically limited the applicability of deep-learning methods trained on synthetic-chemistry datasets. Contrary view: foundation-model-style approaches in chemistry have a history of benchmark-suite overperformance that does not translate to wet-lab hit rates; the NaFM evaluation is largely retrospective on curated databases rather than prospective on novel biosynthetic gene clusters. The contribution is best characterised as architectural and data-pipeline rather than a validated end-to-end natural-product discovery system. Source
Platform Drug Discovery

AstraZeneca Q1 Earnings: Open-Source Reinvent Generative AI Framework Reportedly Halves Time to Identify Candidate Molecules; QCS Algorithm Used for Patient Selection

On its 29 April 2026 Q1 earnings call, AstraZeneca CEO Pascal Soriot disclosed that the company's open-source generative AI framework, Reinvent, has approximately halved the time required to identify molecular structures suitable as candidate medicines, characterising the system as "fundamentally changing the speed at which we can move from concept to lead molecules." AstraZeneca also continues to use its Quantitative Continuous Scoring (QCS) algorithm to identify patients most likely to respond to a given investigational therapy. The disclosure forms part of a broader Q1 industry pattern in which large pharmaceutical companies are moving from generic "AI hype" messaging to specific operational metrics — Soriot's framing parallels GSK CEO Luke Miels's same-day characterisation of AI as "the number one priority for the usage of AI is the innovation dimension" with emphasis on translational early-stage R&D. Note: this entry falls just outside the strict 14-day window but is included as a substantive disclosure missed by the prior cycle, reflecting genuine measurable claims rather than partnership-announcement noise. Contrary view: "halved time to identify lead molecules" is unaudited and not validated against pre-Reinvent baseline data in any peer-reviewed publication; comparable claims from competitors (Insilico, Exscientia, BenevolentAI) have historically not survived independent replication at face value. Open-source release improves community trust but introduces competitive-erosion risk; the disclosure of QCS for patient selection raises FDA AI guidance questions yet to be tested in a submission context. Source
Drug Discovery Platform

Drug Target Review Q1 2026 Analysis: AI-Designed Molecules Move From Computational Benchmarks to Validated Preclinical Outputs Across Antibiotics, Enzymes and Protein Binders

A Drug Target Review analytical feature documented a cluster of Q1 2026 peer-reviewed publications in which AI-designed molecules moved from computational-only benchmarks to experimental validation in pre-clinical settings — described as a qualitative shift from leaderboard performance to lab-notebook results. Reported examples include: a designed antimicrobial peptide active in a mouse infection model; engineered enzymes outperforming both nature and directed evolution; and AI-designed protein binders progressed to physical testing. The article also flags NVIDIA's Proteina-Complexa generative protein-binder model — launched at GTC 2026 as part of the BioNeMo platform — being used by Novo Nordisk, Viva Biotech, and Manifold Bio for therapeutic-target binder design with experimental validation, alongside the AlphaFold Protein Structure Database's expansion by ~30 million AI-predicted protein-complex structures. The piece also covers CLADD (Genentech Research's multi-agent LLM system, published in AAAI 2026 proceedings) and AstraZeneca's ChatInvent agentic decision-support tool. Note: this entry falls just outside the strict 14-day window but is included as a substantive trend-tracking analysis missed by the prior cycle. Contrary view: the article aggregates multiple sources of varying peer-review quality; conference-tier results (AAAI) sit alongside Nature-tier studies. "Validated in pre-clinical settings" spans a wide range of stringency from cell-line activity to live-animal efficacy. The shift from benchmark to wet-lab matters but does not by itself address the Phase II/III translation gap that remains the field's binding constraint. Source
Partnership Platform Drug Discovery

AstraZeneca Extends Immunai AI Immune-Modelling Collaboration Through 2027 with up to $37.5M Commitment

AstraZeneca expanded its multi-year AI-driven collaboration with Immunai, extending the partnership through 2027 with up to US$37.5 million in additional payments to the biotech. The deal continues integration of Immunai's AMICA-OS platform — an AI operating system that models the human immune system at single-cell resolution from large-scale clinical immunology datasets — into AstraZeneca's oncology research programmes. Stated applications include cancer biomarker discovery, patient stratification, dose optimisation, and treatment-response analysis. The expansion is the latest in a series of biopharma–AI infrastructure deals in 2026 (e.g. Lilly–Profluent, Bayer–Cradle, GSK–Noetik) that frame AI not as a single drug-target tool but as a longitudinal data and modelling platform across pipelines. Contrary view: announced milestone-and-extension figures are not equivalent to committed cash; the $37.5M ceiling is small relative to AstraZeneca's R&D budget, and disclosed metrics on translational impact (clinical candidates progressed because of AMICA-OS) remain absent. Single-cell immune-state predictions also remain difficult to validate prospectively in heterogeneous patient populations. Source
Platform Drug Discovery

OpenBind UK Releases First Open Structure-Affinity Dataset and OpenBind v1 Predictive Model for Drug Discovery

The UK-led OpenBind consortium released its first publicly available structure-affinity dataset and a newly trained predictive model, OpenBind v1, both freely accessible to researchers worldwide. The release contains 925 crystallographic binding events from 699 compounds against EV-A71 2A protease (using closely related Coxsackievirus A16 2A protease as a surrogate), with K_D affinity measurements (Creoptix WAVEsystem) for 601 compounds. The pipeline produced these 800+ high-quality measurements in approximately seven months — historically a multi-year process — by combining automated chemistry, high-throughput crystallography at Diamond Light Source's XChem fragment-screening facility, and AI training on the UK's Isambard-AI cluster. The initiative was supported in early stages by the UK Department for Science, Innovation and Technology (DSIT) and aims to address the lack of standardised, AI-ready experimental protein-drug binding data — analogous to what the PDB enabled for AlphaFold2. Future tranches will target COVID-19, malaria, dengue, Zika, and oncology. Contrary view: a 925-event single-target dataset, while methodologically rigorous, is small relative to commercial pharma datasets (which typically run to millions of measurements). The benchmark utility for assessing pose-prediction, complex-prediction, and affinity-prediction methods is acknowledged by the consortium as "an initial reference point" — not a definitive evaluation. Generalisability to non-fragment-screen contexts and to PPI/membrane-protein targets remains untested. Source
Funding Platform

Sanofi Commits CAD$294M to Expand Toronto AI Centre of Excellence, Adding 50 ML/AI Roles

Sanofi announced a CAD$294 million expansion of its Global Artificial Intelligence Centre of Excellence in downtown Toronto, creating 50 additional high-skilled jobs in artificial intelligence and machine learning. The expansion builds on more than 150 roles created since the COE was established in 2022, and is supported by a conditional grant of up to CAD$5 million from the Invest Ontario Fund. The new roles will design and deploy AI tools across Sanofi's global research, manufacturing and business operations. The investment is consistent with Sanofi's stated "all-in" AI strategy under CEO Paul Hudson and aligns with the company's Modulus modular vaccine/biologics manufacturing facilities (France, Singapore) coming online in 2026. Contrary view: corporate AI-hub investments are weakly correlated with downstream pre-clinical R&D output. Sanofi's prior AI partnerships (Owkin, Formation Bio + OpenAI, Earendil, Aqemia) have produced infrastructure and process automation gains but no disclosed clinical candidate attributable specifically to AI-led discovery. The Toronto expansion is positioned as a talent-and-tools play, not a pipeline-acceleration claim. Source
Platform Drug Discovery

Insilico Medicine ChemCensor Retrosynthesis Benchmark Accepted for Presentation at ICML 2026

Insilico Medicine announced that its research paper, "When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs," has been accepted for presentation at the 43rd International Conference on Machine Learning (ICML 2026, Seoul, July 6–11). The paper introduces ChemCensor, a chemistry-aware metric that evaluates retrosynthesis model performance using reaction centres and functional groups rather than rigid Top-K accuracy against a single "ground-truth" answer. Additional contributions include CREED, a curated dataset of 6.4 million validated reactions; benchmark results from the C3LM model (which Insilico states surpasses LocalRetro, R-SMILES and Retro-KNN); and URSA-expert-2026, a 100-target expert-annotated out-of-domain benchmark designed to reduce data leakage. Supporting materials are to be released via Zenodo, Hugging Face and GitHub. The work was led by Insilico's Generative AI and Quantum Computing R&D Centre in Abu Dhabi. Contrary view: ICML acceptance and open-data release are credibility signals, but a single-vendor benchmark designed by the platform's own team is not a substitute for independent inter-laboratory evaluation; the field's history of LLM-for-chemistry benchmarks (e.g. USPTO-50K) shows headline accuracy gains often degrade significantly on out-of-distribution chemistries. The pre-print arXiv ID cited (2602.03554) lies in the future-numbering range and should be verified at https://arxiv.org/abs/2602.03554 before quoting. Source
Clinical Trials Drug Discovery Partnership

XtalPi-Discovered PEP08 PRMT5 Inhibitor Hits Phase I Solid-Tumour Enrolment Milestone; Second Synthetic-Lethality Programme Launched with PharmaEngine

XtalPi (HKEX: 2228) and PharmaEngine (TWO: 4162) announced that PEP08 — an MTA-cooperative PRMT5 inhibitor discovered using XtalPi's AI-and-robotics drug discovery platform — has begun enrolment in a Phase I solid-tumour clinical trial in Australia and Taiwan, having previously cleared regulatory review in June 2025. The two companies also initiated a second oncology programme targeting an undisclosed synthetic-lethality mechanism, building on the PEP08 collaboration. PEP08 belongs to the same MTA-cooperative PRMT5 inhibitor class targeted by Insilico's UAE-based ISM0387 and addresses the synthetic-lethal interaction between PRMT5 and MTAP-deleted cancers. Synthetic-lethality oncology has attracted multiple early-stage licensing deals exceeding $1 billion in recent years. Contrary view: Phase I enrolment is a process milestone, not an efficacy or safety readout; the broader PRMT5 inhibitor class has historically faced selectivity, on-target tolerability and patient-selection challenges. Multi-target synthetic lethality has been a fashionable thesis for several years without yet producing approved AI-attributable assets. PR-newswire-grade disclosure (rather than peer-reviewed clinical data) is the source. Source
Platform Drug Discovery Genomics & Proteomics

Scientific Reports: Insilico's TargetPro–TargetBench Disease-Specific AI Framework Validated for Drug Target Identification

Leung et al. (Insilico Medicine) published in Scientific Reports a peer-reviewed validation of TargetPro, a disease-specific machine-learning workflow trained on clinical-stage targets across 38 diseases (oncology, neurological, immune, fibrotic, and metabolic), and TargetBench 1.0, a standardised benchmarking system for evaluating target-identification systems including general-purpose LLMs. TargetPro integrates 22 distinct omics and text-based scores from Insilico's PandaOmics platform; the paper reports an overall precision-at-top-K of 71.6% in retrieving known clinical targets — a 1.7 to 5.5-fold improvement over leading LLM-based approaches — alongside 95.7% protein-structure resolvability for novel targets, 86.5% druggability classification, and 46% repurposing-potential overlap. DOI: 10.1038/s41598-026-47765-3. The paper is foundational to Insilico's subsequently announced TargetPro 2.0 / TargetBench 2.0 expansion to 100 indications across 10 therapeutic areas (announced 20 April 2026). Note: this entry sits just outside the strict 14-day window and is included as a peer-reviewed item missed by the prior cycle. Limitation: precision-at-top-K against known clinical targets rewards systems that recapitulate the existing target landscape; novel target validation requires prospective experimental work, and the 46% repurposing-overlap metric is a feature/limitation depending on whether genuinely first-in-class targets are the goal. Independent benchmarking by non-Insilico teams remains limited. Source
Genomics & Proteomics Platform

AACR 2026: Kindai University DNA-Methylation Machine Learning Model Identifies Origin of Cancers of Unknown Primary

At the AACR Annual Meeting 2026 (San Diego, 17–22 April), Marco A. De Velasco, PhD, and colleagues at Kindai University presented a machine-learning model that classifies the tissue of origin of cancers of unknown primary (CUP) using CpG-based DNA methylation patterns. The model achieved approximately 95% accuracy on a held-out test cohort and 87% accuracy on an independent validation cohort comprising 31 cases across 17 cancer types. CUP — metastatic malignancies with no identifiable primary tumour — accounts for roughly 3–5% of all cancers and historically carries among the worst prognoses precisely because tissue-of-origin uncertainty prevents site-directed therapy. The work supplements existing methylation-array-based commercial assays (e.g. EpiSeq, EPICUP) by demonstrating that a CpG-feature ML approach generalises across ancestrally diverse cohorts and can leverage liquid-biopsy methylomes. Note: this entry sits just outside the strict 14-day window and is included as an AACR 2026 highlight missed by the prior cycle. Limitation: AACR-presented data is conference-tier and not yet peer-reviewed in a journal; 87% validation accuracy on 31 cases is a small sample by clinical-deployment standards, and prospective utility (does the methylation call change clinical management) was not addressed. Source
Drug Discovery Genomics & Proteomics Platform

Nature Communications: AlphaFold-Multimer Virtual Screen Identifies Functional Nanobody Binders to MRGPRX2 GPCR — First Fully In Silico Antibody Discovery for a Druggable Receptor

A Nature Communications paper (Smith et al., DOI: 10.1038/s41467-026-72093-5) reports a prospective in silico screen using AlphaFold-Multimer (AF-M) that identified nanobody binders to MRGPRX2, a G-protein coupled receptor implicated as a therapeutic target for itch and IgE-independent pseudo-allergic drug reactions. The team co-folded approximately 10,000 nanobody candidates against MRGPRX2, ranked them by a learned LCF metric, and experimentally validated the top-ranked clones in flow-cytometry binding assays on ROSA mast cells expressing MRGPRX2. Nanobody Sim8619 (rank 1), Sim9877 (rank 5), and Sim4784 were among those showing receptor binding and signalling-assay activity. The work is significant for two reasons: (a) GPCRs are notoriously difficult antibody targets due to small extracellular surface area and conformational flexibility — only ~3 GPCR-targeting antibodies are clinically approved; and (b) it demonstrates that publicly available structural-prediction tools (AF-M, training cutoff 30 Sep 2021) can support fully in silico antibody discovery without commercial-platform retraining. Note: this entry sits just outside the strict 14-day window and is included as a peer-reviewed item missed by the prior cycle. Limitation: the manuscript is an unedited early-access version pending final editing; the result is for one GPCR target with a relatively well-populated nanobody-GPCR PDB training prior (over 150 deposited nanobody-GPCR structures by 2023), and generalisability to under-studied GPCRs and to soluble or non-GPCR targets remains an open question. Source
Platform Drug Discovery

PNAS: University of Virginia YuelDesign Diffusion-Model Suite Designs Drug Molecules Against Flexible, Induced-Fit Protein Targets

Researchers at the University of Virginia School of Medicine, led by Nikolay Dokholyan, PhD, published in PNAS (DOI: 10.1073/pnas.2524913123) an open-access AI suite — YuelDesign, YuelPocket, and YuelBond — that simultaneously designs both drug molecules and the protein binding pockets they target, treating proteins as conformationally flexible rather than as static structures. YuelDesign uses diffusion models to co-generate the protein pocket geometry and a candidate small molecule, capturing "induced fit" — the conformational rearrangement proteins undergo on ligand binding — which most existing structure-based design pipelines (including those built on rigid AlphaFold predictions) ignore. YuelPocket uses graph neural networks to identify drug-binding sites including on AF-predicted structures; YuelBond addresses the binding interaction step. The team demonstrated the approach on CDK2, a clinically validated kinase target in oncology. Note: this entry sits just outside the strict 14-day window and is included as a peer-reviewed PNAS item missed by the prior cycle. Limitation: induced-fit modelling improves theoretical accuracy but raises computational cost; benchmark comparisons against rigid-structure baselines are presented mostly on retrospective targets. The CDK2 demonstration is one case study, and prospective wet-lab validation against multi-target panels remains the field's standard for robustness. Source
Drug Discovery Platform

Insilico Medicine Nominates ISM6200 — Generative-AI-Designed NR3C1 Antagonist — for Ovarian Cancer, Cushing's Syndrome, and Cortisol-Driven Obesity

Insilico Medicine announced the nomination of ISM6200, a small-molecule NR3C1 (glucocorticoid receptor) antagonist, as its 29th preclinical/developmental candidate. Discovery and optimisation were driven by Chemistry42, the generative chemistry engine within Insilico's Pharma.AI platform. The candidate is positioned as potentially best-in-class with a stated lower drug-drug-interaction (DDI) risk than incumbent NR3C1 inhibitors, and is being progressed for platinum-resistant ovarian cancer (in combination with paclitaxel — the FDA has already cleared a separate clinical proof-of-concept study of NR3C1 antagonist relacorilant plus nab-paclitaxel), Cushing's syndrome (hypercortisolism), and glucocorticoid-related obesity. ISM6200 reportedly demonstrated dose-dependent anti-tumour efficacy in a CDX model and efficacy in preclinical Cushing's, hypercortisolism-obesity, and glaucoma models. Insilico cites a cumulative 12 IND clearances and 3 Phase II programmes since 2021 across its AI-discovered portfolio. Note: this entry sits just outside the strict 14-day window and is included as a substantive AI-derived preclinical-candidate disclosure missed by the prior cycle. Contrary view: PCC nomination is an internal milestone, not regulatory or clinical validation; NR3C1 modulators have a long history of selectivity and tolerability challenges, and combining a glucocorticoid receptor antagonist with paclitaxel raises composite safety considerations that will be tested only in human studies. The cardiometabolic-and-oncology indication breadth is strategically broad, and disclosed pharmacology is from company press materials rather than peer-reviewed publication. Source
Drug Discovery Platform

Recursion Founder Chris Gibson to Step Off Board as Najat Khan Consolidates Leadership of AI-Native Pipeline

Recursion (Nasdaq: RXRX) announced that founder Chris Gibson, PhD, will complete his current term through June 2026 and will not seek re-election to the company's Board of Directors, remaining as an advisor. The transition consolidates leadership under CEO Najat Khan, PhD, who succeeded Gibson on 1 January 2026 and will deliver Q1 2026 financial results on 6 May 2026. Recursion's stated platform priorities for 2026 include scaling its full-stack AI Operating System following the first AI-enabled clinical proof-of-concept (REC-4881 in familial adenomatous polyposis), advancing five differentiated clinical programmes with defined milestones, and managing cash operating expense below $400 million (approximately 10% lower than 2025 guidance). Contrary view: founder departures from publicly traded AI drug discovery companies have historically coincided with strategic resets that markets price as risk; Recursion's share price (~$3.19 at recent close) reflects continued scepticism that any TechBio platform can convert clinical proof-of-concept in a single rare-disease indication into a generalisable pipeline economic advantage. Nvidia's complete exit from its RXRX position in February 2026 underscores the unresolved investor debate. Source
Genomics & Proteomics Drug Discovery

Nature Aging: Multimodal Deep Learning Reveals Asynchronous Aging Across Female Reproductive Organs

Researchers at the Barcelona Supercomputing Center (BSC-CNS), using the MareNostrum 5 supercomputer, integrated deep-learning analysis of 1,112 histological images with RNA-sequencing data from 659 samples across seven reproductive organs in 304 female donors aged 20–70 years. The study, led by Marta Melé and published in Nature Aging, reports asynchronous aging dynamics: ovary and vagina age gradually, while the uterus undergoes an abrupt transition around menopause. Tissue segmentation identifies the myometrium as the most age-affected structure (extracellular matrix remodelling and immune activation); vaginal epithelial layers also show sharp menopausal shifts mirroring transcriptional changes in developmental pathways. Critically, organ-specific aging signatures were detectable in independent plasma proteomics data, suggesting non-invasive biomarker potential. The paper identifies candidate genes and molecular processes that constitute new target hypotheses for menopause-associated chronic disease (cardiovascular, osteoporosis, neurodegeneration). DOI: 10.1038/s43587-026-01098-y. Limitation: cross-sectional GTEx-derived design cannot establish causality between menopausal transitions and downstream chronic-disease risk; prospective longitudinal validation and tissue-resident drug-target tractability work remain. Source
Clinical Trials Platform

FDA Opens Formal Consultation on AI-Enabled Optimisation of Early-Phase Clinical Trials

The U.S. Food and Drug Administration published a Request for Information (Federal Register Docket FDA-2026-N-4390) seeking input on a proposed pilot programme to evaluate how AI-enabled technologies can improve efficiency, speed and quality of decision-making in early-phase clinical trials. The pilot's stated remit covers safety monitoring, dose selection, and early go/no-go regulatory decisions while maintaining FDA's scientific and regulatory standards and aligning with NIST trustworthy-AI principles. The agency seeks comment on scope and focus, participant selection, collaboration models, operational structure, timeline and milestones, and knowledge sharing. Comments are due by 29 May 2026, with final selection criteria targeted for July and pilot selections for August. Contrary view: an RFI is consultation, not policy; previous FDA AI-related RFIs (including the January 2025 draft AI guidance) attracted hundreds of comments without producing rapid binding rules. The pilot's success will hinge on whether sponsors are willing to expose proprietary AI model architectures and training data to regulator review, which industry has historically resisted on intellectual-property grounds. Source
Clinical Trials Platform

FDA Launches First Real-Time Clinical Trials, Targeting AI- and Cloud-Enabled Continuous Drug Development

The FDA announced the successful initiation of two proof-of-concept Real-Time Clinical Trials (RTCTs) that will report endpoints and safety signals to the agency in a continuous data stream rather than at discrete protocol-defined intervals. FDA Commissioner Marty Makary, MD, MPH, framed the announcement as a challenge to the long-held assumption that drug approval requires 10–15 years; FDA Chief AI Officer Jeremy Walsh stated the approach could ultimately reduce 20–40% of overall clinical-trial time. The proof-of-concept builds on AI and cloud-computing infrastructure for continuous safety monitoring and may eliminate inter-phase hiatuses by enabling FDA to view trial data as it is collected. Contrary view: real-time data sharing creates new regulatory and operational risks, including premature signal interpretation, data-privacy concerns, and asymmetric burden on sponsors who must build interoperable integration layers and continuous-exchange pipelines. The proof-of-concept involves only two trials, and operational scaling depends on stakeholder collaboration that has not historically produced rapid consensus on data standards in clinical research. Source
Partnership Genomics & Proteomics Drug Discovery

Eli Lilly and Profluent Sign $2.25 Billion Pact for AI-Designed Recombinases Targeting Kilobase-Scale DNA Editing

Profluent and Eli Lilly announced a multi-programme strategic research collaboration in which Profluent will apply its protein-design foundation models (built on an atlas of over 115 billion unique protein sequences) to design site-specific recombinases — enzymes that cut and rejoin DNA — for multiple genomic targets. Lilly receives an exclusive licence to advance selected recombinases through preclinical, clinical and commercialisation stages. Profluent receives an undisclosed upfront payment plus committed R&D funding and is eligible for up to $2.25 billion in development and commercial milestone payments and tiered royalties. The deal targets kilobase-scale DNA editing — the ability to insert long stretches of DNA at precise genomic locations — which conventional CRISPR-Cas tools cannot reliably achieve due to payload-size and specificity constraints. The collaboration follows Lilly's January 2026 $1.12bn Seamless Therapeutics recombinase deal and underscores big pharma's increasing willingness to pay platform-level premiums for AI-native protein design. Contrary view: the headline figure is potential, not committed. The reported "biobucks" structure — undisclosed upfront against $2.25bn in milestones — fits the industry pattern in which announced totals exceed actual paid value by 50:1 or more. Profluent's OpenCRISPR-1 publication established platform credibility, but no clinical asset, safety dataset, or in vivo durability data was disclosed with the announcement; off-target recombination, immunogenicity, and delivery-tissue specificity remain unresolved technical risks. Source
Genomics & Proteomics Platform

Nature Communications: DDA-BERT End-to-End Transformer Improves Peptide Identification Across Species and Immunopeptidomics

Researchers introduced DDA-BERT, a transformer-based end-to-end deep-learning model for peptide-spectrum match (PSM) rescoring in data-dependent acquisition (DDA) mass-spectrometry-based proteomics. The model was pre-trained on approximately 271 million PSMs from 11 species and replaces the traditional pipeline of separately trained shallow classifiers used for the final scoring stage. DDA-BERT consistently outperformed existing tools across species-specific benchmarks: 2.24%–269.35% increases in peptide identifications on human, 3.73%–141.46% on yeast, 5.53%–45.64% on Drosophila, and 3.68%–62.77% on Arabidopsis datasets. Authors also report improved performance on HLA immunopeptidomics, which is directly relevant to neoantigen identification for cancer vaccine and TCR-engineered therapy design. DOI: 10.1038/s41467-026-72246-6. Limitation: rescoring improvements are computational benchmarks against existing engines; the upstream limits remain physical (instrument duty cycle, dynamic range, fragmentation efficiency), and no orthogonal wet-lab validation of additionally identified peptides was reported. The model also requires substantial GPU compute compared with existing shallow-classifier rescoring approaches. Source
Genomics & Proteomics Platform Drug Discovery

Nature Communications: ECG-LFM Self-Supervised Foundation Model Pre-trained on 10 Million ECGs Surfaces Novel Cardiovascular Genetic Targets

Researchers introduced ECG-LFM (Electrocardiogram Large-scale Foundation Model), a self-supervised foundation model pre-trained on more than ten million 12-lead ECGs across multiple datasets. The model integrates contrastive learning with masked language modelling to capture both global and fine-grained ECG representations. Beyond cardiovascular disease prediction, ECG-LFM was used to perform genome-wide association studies on derived ECG-feature embeddings, surfacing novel genetic markers associated with heart-rhythm and conduction phenotypes. The work directly bridges self-supervised biosignal foundation models and downstream drug-target identification for cardiovascular disease. DOI: 10.1038/s41467-026-72436-2. Limitation: this is an unedited early-access manuscript; AI-derived GWAS associations require independent functional validation and replication in ancestrally diverse cohorts before any can be considered druggable targets. The model's interpretability claims will need scrutiny before clinical or target-discovery deployment. Source
Genomics & Proteomics Platform

Nature Computational Science: MethylVI Deep Generative Model Improves Single-Cell Bisulfite Sequencing Analysis

Nature Computational Science published a News & Views and accompanying paper introducing MethylVI, a deep generative model for analysing single-cell bisulfite sequencing methylomic data. The model explicitly accounts for the unique technical and biological sources of variability in single-cell methylation data — including coverage sparsity, cell-to-cell technical heterogeneity, and the bimodal nature of methylation signal — that have historically limited the utility of conventional integration approaches. MethylVI enables more reliable cell-type identification, batch correction, and downstream regulatory inference at single-cell resolution, which is directly relevant to epigenetic target identification in oncology, neuroscience and immunology. Limitation: methylVI is one of several generative-model approaches to single-cell epigenomics; comparative benchmarking against scBS-VAE and methylSCAN remains limited, and biological interpretability of latent embeddings is not addressed in depth. Single-cell bisulfite datasets at scale remain expensive to generate, constraining real-world deployment. Source
Platform Genomics & Proteomics Drug Discovery

Frontiers in AI: AlphaFold Series Now Accounts for Approximately 40% of New PDB Depositions, with AF3 Driving Structure-Based Drug Discovery

A peer-reviewed review in Frontiers in Artificial Intelligence (DOI: 10.3389/frai.2026.1739303) synthesises the current status and trajectory of the AlphaFold (AF) series following its 2024 Nobel Prize. The review reports that approximately 40% of structures newly deposited in the Protein Data Bank from 2024 to 2025 incorporated AlphaFold predictions, and that AF3 — DeepMind's diffusion-based model for protein-ligand, protein-nucleic acid and protein-protein complex prediction — has materially advanced structure-based drug discovery (SBDD). The article specifically cites a 60% hit rate when AF predicted protein structures were used to identify TAAR1 agonists, outperforming conventional methods. The review traces architectural progression (AF1 deep neural networks → AF2 Evoformer → AF3 Pairformer) and accessibility (AlphaFold Database, AF Server) as a combined driver of the structure-prediction-to-drug-discovery pipeline. Note: this entry falls outside the strict 14-day window but is included to capture a substantive review missed by the prior cycle. Contrary view: AF3 weights remain closed-source with commercial use gated through Isomorphic Labs partnerships, and the AF Server caps researchers at 10 predictions per day. AF3 confidence scores fall below 50% for intrinsically disordered regions, affecting roughly 30% of human proteome regions involved in signalling and regulation. The 60% TAAR1 hit-rate result is one targeted case study, not generalised performance. Source
Genomics & Proteomics Platform

Nature: Bacterial Antiviral Defence Atlases Expand Dramatically as DefensePredictor ML Tool Identifies Hundreds of Novel Immune Systems

Two research teams independently mined genomic data from bacteria to create databases containing thousands of antiviral defence proteins, with results reported in Nature. The Laub group at MIT designed DefensePredictor, a machine-learning tool that predicts bacterial immune proteins using gene and protein data from 17,000 bacterial genomes. Tested on 69 diverse Escherichia coli strains, DefensePredictor identified 624 candidate defence-related proteins, more than 100 of which were previously unknown; the team confirmed defence activity for 42 of these in laboratory experiments. The work expands the catalogue of microbial defence systems analogous to those that gave rise to CRISPR-Cas, restriction-modification, and abortive-infection tools — the historical source pool for genome-editing biotechnologies and antiviral therapeutics. Note: this entry falls outside the strict 14-day window but is included as a substantive Nature-tier publication missed by the prior cycle. Contrary view: candidate-defence-system catalogues have historically converted to deployable biotechnologies at low rates; only a small fraction of the originally catalogued CRISPR-Cas variants have produced therapeutically relevant tools, and structural/biochemical characterisation of newly predicted defence systems is laborious. The 100+ "novel" predictions still require functional validation in mammalian cells before drug-discovery utility can be assessed. Source
Drug Discovery Platform

Insilico Medicine Nominates First UAE-Based AI-Designed Preclinical Candidate for Glioblastoma

Insilico Medicine announced ISM0387, an MTA-cooperative PRMT5 inhibitor for glioblastoma multiforme (GBM), as the UAE's first AI-designed developmental candidate. The molecule was discovered locally at Insilico's Abu Dhabi R&D centre using the Pharma.AI platform, with the entire workflow from molecular design to lead optimisation completed in under 12 months. Insilico's team screened 90 AI-generated candidates using Chemistry42 and its 40+ models, focusing on CNS-penetrant traits, completing lead discovery in six months. ISM0387 targets the synthetic-lethal interaction between PRMT5 and MTAP-deleted cancer cells. The nomination represents Insilico's 30th AI-supported preclinical candidate. The milestone was witnessed by the Emirates Drug Establishment (EDE), Department of Health Abu Dhabi, and Abu Dhabi Investment Office. Limitation: preclinical candidacy is an early milestone; PRMT5 inhibitors as a class face historical selectivity and tolerability challenges, and ISM0387 has not yet entered IND-enabling studies or human testing. Source
Drug Discovery Genomics & Proteomics

McMaster University Generative AI Designs Novel Antibiotic Synthecin, Validated In Vivo Against Drug-Resistant Staph

Researchers at McMaster University led by Jonathan Stokes published in Molecular Systems Biology (selected as the cover of the June issue) a generative AI model that designed novel water-soluble antibiotics targeting Staphylococcus aureus. From a batch of 79 model-proposed antibacterials, the team identified a lead compound named synthecin, formulated it as a topical cream, and demonstrated efficacy controlling otherwise drug-resistant wound infections in mouse models. The work extends Stokes lab's prior generative-AI antibiotic discovery efforts, including the high-profile abaucin and halicin programmes, by adding water-solubility constraints during model training. Limitation: efficacy was demonstrated in topical cream form on mouse wound infections only, with no systemic dosing, no pharmacokinetic profile, no resistance-emergence study, and no toxicology beyond the early in vivo work; antibiotic translation from mouse infection models to human clinical use has historically had a high attrition rate. Source
Partnership Clinical Trials Genomics & Proteomics

Tempus AI and Keck Medicine of USC Form Multi-Faceted AI-Enabled Precision Medicine Partnership

Tempus AI announced a multi-faceted collaboration with Keck Medicine of USC and the Keck School of Medicine to embed Tempus's AI-enabled molecular diagnostics, clinical trial matching, and care-gap analytics across more than 1.5 million annual patient visits in USC-affiliated hospitals and clinics. The partnership couples Tempus's multimodal data and AI infrastructure with USC's high-volume patient base to create a large-scale testbed for new diagnostics, AI tools, and research insights, with the USC Norris Comprehensive Cancer Center as a primary integration site. The collaboration spans clinical genomic profiling, real-time AI-assisted clinical trial matching, and joint R&D initiatives. Contrary view: financial terms were not disclosed; while large academic-health-system testbeds are valuable for model validation, conversion of clinical-network adoption into recurring high-margin software and data revenue has historically lagged for AI-precision-medicine companies, and key risks include reimbursement uncertainty, ongoing GAAP losses, and the difficulty of demonstrating attributable clinical-outcome improvements at scale. Source
Drug Discovery Platform

Insilico Medicine Publishes Comprehensive AI Target Identification Review in Nature Reviews Drug Discovery

Insilico Medicine published a peer-reviewed review titled "Target identification and assessment in the era of AI" in Nature Reviews Drug Discovery, the highest-impact journal in pharmaceutical research (impact factor 101.8). The review surveys multimodal AI approaches to therapeutic target selection, including foundation models trained on tens of millions of single-cell transcriptomes (Geneformer, scGPT), domain-specific LLMs and AI agent frameworks (BioGPT, OriGene), and knowledge-graph-based graph neural networks for predicting synthetic lethality. The article uses Insilico's TNIK inhibitor rentosertib (ISM001-055) for idiopathic pulmonary fibrosis as a worked example of AI-driven target prioritisation moving from project initiation to preclinical candidate nomination in approximately 18 months. The review concludes the field still requires progress on data quality, explainability, standardised benchmarking, and digital-twin closed-loop platforms. Contrary view: as a single-company-led review, the framing prioritises the author's own platform; competing approaches such as Recursion's phenomic screening and Isomorphic Labs' physics-informed structural prediction are mentioned without comparable critical analysis. Source
Drug Discovery Platform Funding

10x Science Raises $4.8M Seed to Address AI-Generated Drug Candidate Validation Bottleneck

10x Science, founded December 2025 by chemical biologist David Roberts, biologist Andrew Reiter, and serial founder Vishnu Tejus, announced a $4.8 million seed round led by Initialized Capital with backing from Y Combinator, Civilization Ventures, and Founder Factor. The company is building a SaaS platform that uses AI to characterise and validate candidate molecules generated by upstream drug-design models, addressing a growing bottleneck as generative AI produces more drug candidates than experimental teams can characterise via traditional spectrometry. Early adopters include Rilas Technologies, a contract chemical analyses provider; an external scientist using the platform reported it correctly identified a protein from filename context alone and adapted to multiple molecule classes. Investors including Initialized partner Zoe Perret framed the platform as a recurring-revenue picks-and-shovels play independent of any single drug candidate's success. Contrary view: the seed round size is small relative to incumbent contract analytical-services providers, and the platform's claimed cross-modality generalisation has not been independently benchmarked or peer reviewed; "AI for AI validation" tools face the same trust and reproducibility challenges as the upstream models they evaluate. Source
Genomics & Proteomics Drug Discovery Clinical Trials

Path-IO Biology-Guided Pathomics Model Outperforms PD-L1 for Immunotherapy Selection at AACR 2026

At the 117th AACR Annual Meeting in San Diego (April 17–22), researchers presented Path-IO, a biology-guided pathomics machine-learning model that consistently outperformed PD-L1 — the FDA-validated standard-of-care biomarker for guiding immunotherapy in non-small-cell lung cancer (NSCLC) — across both discovery and test cohorts. Presented by Sarit Bandyopadhyay, the model differs from prior pathomics approaches by grounding predictions in tissue structures familiar to clinicians, rather than relying on opaque whole-slide image embeddings. Bandyopadhyay noted that further investigation is required to extend Path-IO beyond identifying which patients benefit from immunotherapy to predicting which class of immunotherapy is most appropriate. The abstract was published in the AACR Cancer Research Proceedings supplement on April 17, 2026. Limitation: results are from retrospective discovery and test cohorts; prospective validation, multi-site reproducibility, and demonstration of clinical-decision impact remain outstanding before regulatory or guideline adoption. Source
Drug Discovery Platform

Insilico Medicine Presents Four Generative-AI-Designed Oncology Candidates at AACR 2026

Insilico Medicine presented four poster abstracts at the AACR Annual Meeting 2026 (April 17–22, San Diego), each detailing a distinct AI-designed oncology candidate produced by its end-to-end Pharma.AI platform. Programmes disclosed include ISM6166, an oral pan-KRAS (ON/OFF) inhibitor with anti-tumour activity in solid tumours bearing KRAS alterations (~17% of solid tumours); ISM3830, a Cbl-b small-molecule inhibitor that restored function of suppressed T cells and increased tumour infiltration of T and NK cells in vivo; ISM6210, an AI-discovered selective CDK4 inhibitor for HR+/HER2- breast cancer; and a fourth candidate from the company's UAE-based R&D centre. Two of the four programmes originated from the UAE site. Contrary view: AACR poster presentations are early-stage data disclosures; KRAS, Cbl-b, and CDK4 are crowded fields with multiple competing programmes from established oncology developers, and the abstracts disclose preclinical data only — none of these candidates has a publicly disclosed IND filing date or human pharmacokinetic data. Source
Genomics & Proteomics Partnership

Tempus AI and Predicta Biosciences Expand Co-Branded Whole-Genome Sequencing Assay for Hematologic MRD Monitoring

Tempus AI and Predicta Biosciences announced commercial expansion of GenoPredicta, a co-branded ultrasensitive whole-genome sequencing (WGS) assay for genomic characterisation of haematologic malignancies and measurable residual disease (MRD) monitoring. The assay integrates flow cytometry with WGS, detects alterations from as few as 50 tumour cells (sensitivity approximately 1-in-1,000,000), is clinically validated in multiple myeloma, and is now available to Tempus life-sciences partners for research and clinical development. The platform is positioned for biopharma trial design, patient stratification, and clinical biomarker development in haematologic oncology. Contrary view: the announcement is a commercial expansion of an existing assay rather than a novel clinical or analytical validation; MRD-detection thresholds in the 1-in-10⁶ range face well-established analytical-noise challenges, and clinical utility for treatment decisions in myeloma remains an active area of guideline debate. Source
Platform Drug Discovery

Insilico Medicine Pharma.AI Spring Kickoff Showcases PandaClaw Agent and ICLR 2026-Accepted MMAI Gym Framework

Insilico Medicine hosted its Pharma.AI Spring Kickoff webinar showcasing updates across the Pharma.AI ecosystem. Highlights included PandaClaw, an agentic AI tool that enables natural-language multi-omics analyses, hypothesis generation, and target evaluation against PandaOmics; the MMAI Gym for Science framework, with the underlying training-and-benchmarking paper accepted at ICLR 2026; and Chemistry42 updates including multi-target molecule generation, Nach01-MMAI, and Absolute Binding Free Energy (ABFE) calculations in Alchemistry. Insilico reported MMAI-trained foundation models achieved up to 10x performance gains on key drug-discovery benchmarks compared with general-purpose foundation models, which fell short on approximately 75–95% of tasks. The previously disclosed LFM2-2.6B-MMAI (v0.2.1) model from the Liquid AI collaboration was reiterated as the framework's first deliverable. Contrary view: the 10x performance claim is internally benchmarked; independent evaluation, head-to-head comparisons against domain models from competing labs (Recursion's Phenom-2, Owkin's agentic infrastructure, Iambic's NeuralPLexer), and reproducibility outside Insilico-curated datasets are not publicly available. Source
Platform Clinical Trials

Tempus AI Launches Active Follow-Up Service Embedding Continuous AI Updates Into Genomic Reports

Tempus AI launched an automated Active Follow-Up clinical update service designed to keep genomic reports current with evolving clinical guidelines and biomarker-therapy relationships. The integrated workflow places patients on a continuous follow-up track and delivers context-aware notifications to clinicians via Hub, Tempus's secure AI-enabled physician portal. The service represents a shift from one-off genomic reports toward continuous, AI-informed decision support inside oncology workflows, and is positioned to extend Tempus's "dry lab" software-and-data revenue line beyond initial test reimbursement. Contrary view: continuous-update services depend on consistent ingestion of clinical-guideline changes and validated AI summarisation of new evidence; reimbursement pathways for ongoing decision-support services are unclear, the launch overlaps competitively with comparable offerings from Foundation Medicine and Caris Life Sciences, and adoption metrics inside oncology workflows have not yet been disclosed. Source
Platform Drug Discovery

OpenAI Launches GPT-Rosalind, Its First Domain-Specific Reasoning Model for Life Sciences

OpenAI introduced GPT-Rosalind, a frontier reasoning model purpose-built for biology, drug discovery, and translational medicine. The model is available through ChatGPT, Codex, and the API as a research preview to qualified enterprise customers in the US. Launch partners include Amgen, Moderna, the Allen Institute, Thermo Fisher Scientific, and Los Alamos National Laboratory for AI-guided protein and catalyst design. A free Life Sciences research plug-in for Codex connects scientists to over 50 tools and data sources across human genetics, functional genomics, protein structure, biochemistry, and clinical evidence. OpenAI states the system includes "high-precision flags" that monitor for bioweapons concerns and access is gated to organisations conducting legitimate research with public benefit. Following the announcement, shares of Recursion Pharmaceuticals and Schrödinger each fell more than 5%, IQVIA fell up to 3.2%, and Charles River dropped 2.6%. Contrary view: access is restricted, no peer-reviewed benchmarks have been released, and the model has not yet been independently validated against domain-specialised tools. Source
Platform Drug Discovery

DualGPT-AB Published in Nature Computational Science Demonstrates Multi-Property Antibody Optimisation

Researchers published DualGPT-AB in Nature Computational Science, a dual-stage conditional generative pre-trained transformer framework for therapeutic antibody design that simultaneously optimises antigen-binding specificity, viscosity, clearance, and immunogenicity. The framework represents desired properties as learnable embeddings and incorporates a reinforcement-learning strategy to improve sequence exploration efficiency. Computational experiments showed DualGPT-AB generated CDRH3 sequences fulfilling multiple property constraints, with experimentally validated designed antibodies showing favourable HER2-binding affinity and tumoricidal activity. The work addresses a long-standing limitation of single-property antibody optimisers that struggle to balance developability with potency. DOI: 10.1038/s43588-026-00976-0. Limitation: validation focused on a single, well-characterised target (HER2) with a known therapeutic benchmark (trastuzumab); generalisation to novel or low-data targets remains untested. Source
Partnership Platform Drug Discovery

Novo Nordisk and OpenAI Form Strategic Partnership Spanning Discovery to Commercial Operations

Novo Nordisk announced a strategic partnership with OpenAI to deploy advanced AI capabilities globally across drug discovery, manufacturing, supply chain, and commercial functions, with structured workforce upskilling treated as a core deliverable. Stated goals include analysing complex datasets to identify drug candidates and reducing the time from research to patient access. The partnership is structured with strict data governance, human-in-the-loop oversight, and pilots intended to scale to deployment by late 2026. CEO Mike Doustdar framed AI as enabling analysis at "previously impossible" scale. Contrary view: the announcement is high-level with no disclosed financial terms, milestones, target areas, or specific model endpoints; comparable enterprise AI roll-outs have shown variable productivity gains, and Novo's competitive position in obesity and metabolic disease will not be determined by general-purpose AI access alone. Source
Platform Drug Discovery

Amazon Web Services Launches Amazon Bio Discovery for Computational Drug Discovery Workflows

AWS launched Amazon Bio Discovery, an AI application that combines computational methods with wet-lab workflows in a single interface, allowing scientists to run complex computational pipelines without writing code. The product targets early-stage drug discovery and is positioned as enabling "lab-in-the-loop" workflows that reduce the gap between computational designers and experimentalists. The launch represents AWS's first fully integrated life-sciences-specific AI application, joining established competitors Microsoft Azure and Google Cloud in the pharma cloud-AI segment. Contrary view: cloud platform launches in pharma have historically generated more press than measurable adoption among research organisations with deeply embedded internal tooling; pricing, model specifics, validation data, and named launch customers were not disclosed in the launch announcement. Source
Partnership Drug Discovery Platform

Daiichi Sankyo Partners with Imagene AI for Multimodal Biomarker Discovery in ADC Oncology Programmes

Imagene AI announced a collaboration with Daiichi Sankyo to apply its OI Suite multimodal platform, powered by the CanvOI foundation model, to support biomarker discovery and response prediction in select antibody-drug conjugate (ADC) development programmes. The platform integrates haematoxylin and eosin (H&E) and immunohistochemistry (IHC) whole-slide images with molecular profiles and longitudinal clinical outcomes, applying Composite Continuous Scoring to inform patient stratification and companion diagnostic strategy. The deal is the latest in a series of multimodal-AI partnerships in oncology, building on Daiichi Sankyo's heavy ADC pipeline. Contrary view: financial terms were not disclosed, the collaboration's outputs are gated to internal Daiichi Sankyo decision-making, and pathology foundation models still face open questions on cross-site reproducibility and stain-normalisation robustness. Source
Platform Drug Discovery

Chemify Publishes Three Peer-Reviewed Papers Validating LLM-Augmented Chemputation for Autonomous Drug Discovery

Chemify, the deep-tech spin-out led by Lee Cronin, published three peer-reviewed papers in close succession in PNAS, Nature Communications Chemistry, and Nature Communications Biology, collectively validating Chemputation as an integrated paradigm for AI- and ML-driven molecular discovery coupled with automated robotic synthesis. The Nature Communications Chemistry paper (3 April 2026), "Verification and execution of the scientific literature via chemputation augmented by large language models," describes a multi-agent system that reads and corrects chemical literature so procedures can be safely executed by robots. The Nature Communications Biology paper (27 March 2026) demonstrates iterative kinase-inhibitor design in a chemputer-based system targeting KRAS-mutant colorectal cancer, providing a practical pre-clinical example. Limitation: closed-loop autonomous chemistry remains constrained by reagent availability, hardware throughput, and the narrow scope of validated reaction classes; generalisation beyond demonstrated chemistry types is not yet established. Source
Platform Genomics & Proteomics Drug Discovery

Peer-Reviewed Review in International Journal of Molecular Sciences Maps AI Methods for B-Cell Receptor Repertoire Modelling

A peer-reviewed review published in the International Journal of Molecular Sciences synthesises advanced deep-learning architectures applied to B-cell receptor (BCR) repertoire modelling, including antibody-specific language models, graph neural networks, and generative frameworks. The review addresses persistent challenges in interpreting AIRR-seq data: strong inter-individual heterogeneity, non-linear sequence-structure-function relationships, dynamic clonal evolution, and the rarity of functionally relevant clones. The work is positioned as a framework for translational immunology applications including therapeutic antibody discovery, vaccine response characterisation, and patient stratification. DOI: 10.3390/ijms27073296. Limitation: as a synthesis review, the article does not introduce new benchmarks or models; the field still lacks standardised evaluation protocols that would allow head-to-head comparison of competing antibody language models on therapeutic-design tasks. Source
Partnership Funding Drug Discovery Platform

Anthropic Acquires Stealth AI Drug Discovery Startup Coefficient Bio for Approximately $400 Million

Anthropic acquired Coefficient Bio, an eight-month-old stealth biotech startup with fewer than 10 employees, in an all-stock deal valued at approximately $400 million. The team, drawn largely from Genentech's Prescient Design computational drug discovery unit, will join Anthropic's healthcare and life-sciences group to extend Claude for Life Sciences, the dedicated research tool launched in October 2025. Co-founders Samuel Stanton and Nathan C. Frey contributed to Cortex (a modular deep-learning architecture for drug discovery) and Beignet (an open-source standard for molecular representation). Coefficient Bio's stated platform enabled AI to draft drug R&D plans, manage regulatory strategies, and identify candidates across the pipeline. This is Anthropic's first major acquisition. Contrary view: a sub-10-person, pre-product startup priced at $400 million is overwhelmingly a talent-and-IP transaction; Anthropic has no disclosed pharma partnerships, no internal wet-lab capability, and competes with OpenAI's GPT-Rosalind and Google's Isomorphic Labs from a much smaller life-sciences footprint. Source
Platform Genomics & Proteomics Drug Discovery

Frontiers in Microbiology Review Examines Dual-Use Risks of Generative AI in Protein Design for Biosecurity

A peer-reviewed review published in Frontiers in Microbiology examines how AI-driven protein design enables both pandemic-preparedness applications (engineered binders against viral surface proteins, biosensor capture proteins, novel medical countermeasures) and creates new biosecurity risks. The authors warn that AI-generated proteins can be functionally equivalent to known toxins while sharing little sequence similarity with them, rendering current homology-based screening approaches blind to such designs. The review surveys emerging mitigation approaches including DNA-language-model watermarking and distributional/evolutionary watermarks for protein structure generative models (FoldMark). DOI: 10.3389/fmicb.2026.1817535. Contrary view: proposed watermarking schemes are early-stage, easily stripped by adversarial fine-tuning, and currently lack adoption commitments from the major model developers; without coordinated governance, technical mitigations alone are unlikely to close the screening gap. Source
Partnership Funding Drug Discovery Platform

Merck Signs $838 Million AI-Driven Antibody Discovery Collaboration with Infinimmune

Merck (MSD outside the US and Canada) entered a multi-target antibody discovery collaboration with California-based Infinimmune valued at up to $838 million in undisclosed upfront and milestone payments tied to the progression of multiple antibody candidates. Infinimmune will deploy its Anthrobody discovery platform combined with the GLIMPSE antibody language model, screening millions of single memory B cells from human immune repertoires to identify natively paired antibodies against undisclosed Merck-selected targets. Merck holds exclusive development and commercialisation rights to candidates arising from the collaboration. Founded by former 10x Genomics employees, Infinimmune has raised approximately $22 million since launch and is also advancing internal IL-22 and IL-13 programmes for atopic dermatitis, IgA nephropathy, and ulcerative colitis. The deal supports Merck's pre-Keytruda-LoE biologics build-out. Contrary view: the headline $838 million is biobucks (milestone-contingent), the upfront payment was not disclosed, and the targets remain unnamed — limiting external visibility into scientific differentiation. Source
Drug Discovery Platform

UVA School of Medicine: YuelDesign AI Diffusion Suite Designs Drug Molecules for Flexible Protein Targets

UVA researchers published a suite of AI tools — YuelDesign, YuelPocket, and YuelBond — across PNAS, JCIM, and Science Advances. The centrepiece, YuelDesign, uses diffusion models to design drug molecules for protein pockets that flex and shift shape during binding — unlike most existing methods that treat targets as rigid structures. YuelPocket uses graph neural networks to identify binding sites (including on AlphaFold-predicted structures), while YuelBond validates chemical bond accuracy. Tested on CDK2 (cancer), YuelDesign uniquely captured conformational changes upon drug binding. All tools released as open-source. DOI: 10.1126/sciadv.aeb7045, 10.1073/pnas.2524913123 Science Advances | UVA Health
Drug Discovery Platform Partnership

Pharmaphorum Roundup: Verily $300M, BMS/insitro ALS Expansion, Latent Labs Agent, Zealand AI Hub

Key AI drug discovery developments compiled: Verily completed a $300M private placement (led by Series X Capital), transitioning from Alphabet majority control to independent Verily Health Inc. — funds accelerate precision health AI platform. BMS expanded insitro ALS collaboration with two new TDP-43 targets (ALS-2, ALS-3) via insitro's Virtual Human platform, triggering $10M payment; insitro advancing its own antisense oligonucleotide programme for ALS-1. Latent Labs (ex-DeepMind) launched Latent-Y, an autonomous antibody design agent completing campaigns 56x faster than expert estimates. Zealand Pharma establishing Cambridge, MA AI hub for US headquarters. Source
Clinical Trials Drug Discovery Platform

173 AI-Discovered Drug Programs Now in Clinical Development — Phase III Results Will Define 2026

As of early 2026, over 173 AI-discovered drug programs are in clinical development globally: approximately 94 in Phase I, 56 in Phase II, and 15 in Phase III. Between 15 and 20 programs are expected to enter pivotal trials this year. Phase III remains the critical test — AI-discovered compounds are only now reaching this stage, and results later in 2026 will determine whether the technology delivers clinically validated drugs at scale. Gero signed a research agreement with Chugai Pharmaceutical for AI-discovered therapies targeting age-related diseases. Scripps/Gero AI-identified anti-aging candidates extended lifespan in animal models, with 70%+ compounds showing significant results. Source
Drug Discovery Platform

Drug Target Review: AI Drug Discovery Enters Pivotal Year — FDA Guidance, EU AI Act, Market Consolidation

Drug Target Review analysis examines where AI is delivering measurable gains in early discovery and where hype outpaces reality. Market projected to grow from $5–7B (2025) to $8–10B (2026). FDA draft AI guidance expected to be finalised in 2026, requiring credibility assessment plans for high-risk AI applications. EU AI Act high-risk provisions take effect 2 August 2026, potentially classifying some drug development AI as high-risk. Smaller AI drug discovery companies face existential pressures — multiple shutdowns, 20%+ workforce reductions, and delisting despite substantial backing. 50:1 ratio between announced biobucks and actual upfront payments. Fundamental challenge remains data availability, not algorithmic sophistication. Source
Partnership Funding Drug Discovery

Eli Lilly and Insilico Medicine Strike $2.75 Billion AI Drug Discovery Deal

Eli Lilly signed a $2.75B deal with Insilico Medicine for exclusive global rights to develop, manufacture, and commercialise preclinical oral drug candidates discovered using Insilico's Pharma.AI generative platform. Insilico receives $115M upfront; remainder tied to development, regulatory, and commercial milestones plus tiered royalties. This is a tenfold escalation from their $100M November 2025 partnership. Insilico pipeline now includes 28 drugs, nearly half at clinical stage. Targets span oncology, metabolic disease, immunology — including a GLP-1 candidate and pan-KRAS inhibitor. Insilico shares jumped 15% on announcement. Sceptics note persistent gaps between in-silico promise and clinical reality, and concerns about opaque training data limiting reproducibility. BioPharma Dive | Bloomberg
Platform Drug Discovery Genomics & Proteomics

Latent Labs Launches Latent-Y: First Lab-Validated Autonomous AI Agent for De Novo Antibody Design

Latent Labs (founded by ex-DeepMind AlphaFold2 co-lead Dr Simon Kohl; $50M total funding from Radical Ventures, Sofinnova, Google chief scientist Jeff Dean) launched Latent-Y, an AI agent that autonomously designs therapeutic antibodies from a text prompt — compressing weeks of expert work into hours. Powered by Latent-X2, the agent handles target analysis, epitope selection, candidate design, computational validation, and iteration. In user studies, PhD-level experts completed campaigns 56x faster. Demonstrated across diverse campaign types including cross-species binder design and blood-brain barrier crossing. All design decisions logged with reasoning for scientist review. Latent Labs | BusinessWire
Partnership Drug Discovery Platform

BMS and insitro Expand ALS Collaboration — Two Additional TDP-43 Targets via Virtual Human AI Platform

Bristol Myers Squibb nominated two additional AI-identified ALS targets (ALS-2 and ALS-3) through insitro's Virtual Human platform, expanding the six-year collaboration with a $10M payment. These join ALS-1 (nominated December 2024). insitro is advancing its own antisense oligonucleotide for ALS-1 while simultaneously progressing a small molecule candidate for BMS. The collaboration focuses on processes modulating TDP-43 mislocalisation — a central disease mechanism in nearly 97% of ALS patients. Virtual Human platform applies machine learning, human genetics, and functional genomics to generate predictive in vitro models. Source
Platform Genomics & Proteomics Drug Discovery

Nature Biotechnology Review: Generalist Biological AI (GBAI) — Modelling the Language of Life

A major review in Nature Biotechnology synthesises rapid advances in biological AI to interpret and generate DNA, RNA, proteins, and cellular systems. The paper charts a course toward comprehensive generalist systems performing multiple critical biological tasks simultaneously. Key opportunities: synergising language and structural AI, leveraging specialised models, and improving AI agents for autonomous discovery. GBAI could deepen understanding of disease pathways and biomarkers, advance automated therapeutic design, and integrate within virtual cells to simulate biological activity. Significant barriers remain in data availability, biological complexity, scalability, and experimental validation. DOI: 10.1038/s41587-026-03064-w Source
Drug Discovery Genomics & Proteomics Platform

Nature Biotechnology: AI Turbocharges Antibody Discovery — AIntibody Challenge Benchmarks 29 Organisations

Nature Biotechnology feature reports on AI-assisted antibody discovery, including results from the AIntibody Challenge with 29 participating organisations (pharma, startups, academics). David Baker's lab published the first peer-reviewed description of de novo antibody design from scratch (Nature, November 2025). Xaira ($1B founding, six Baker lab co-founders) developing AI-predicted antibody binders with drug-like properties for human studies. Key challenge: most industry AI models remain hidden, limiting collective progress. Experts call for open benchmarking and shared failure data. DOI: 10.1038/s41587-026-03048-w Source
Genomics & Proteomics Platform

Nature Biotechnology: Genome Editing's Third Act — Mutation-Agnostic Therapies Enter the Clinic

Nature Biotechnology reports the newest gene editors are shifting from tackling mutations one by one toward universal, mutation-agnostic therapies. Tessera Therapeutics received US and Australian regulatory clearance for its gene writer therapy clinical trials, backed by a $150M Regeneron partnership. Alltrna plans IND filing in 2026 for engineered transfer RNA correcting shared premature stop codons across multiple diseases. ARPA-H launched Rare Disease AI/ML for Precision Integrated Diagnosis programme. California Institute for Regenerative Medicine committed $100M to RAPID (Rare-disease Acceleration through Platform Innovation and Delivery). David Liu (Broad Institute): strategy is to develop genome-based therapeutics applicable beyond single specific mutations. DOI: 10.1038/s41587-026-03058-8 Source
Platform Genomics/Proteomics
Mar 26, 2026

Nature Communications: RNA Base Pairing Rules Learned with Only 21 Parameters

Researchers demonstrated that the fundamental biological rules of RNA base pairing can be learned from sequences alone — with no structural training data — using a model with only 21 parameters. This challenges the dominant assumption that deep learning in biology requires massive parameter spaces and extensive training data. The work shows that biologically grounded inductive biases can dramatically reduce model complexity while maintaining predictive accuracy, with implications for RNA structure prediction and therapeutic RNA design where training data remains scarce. Source
Drug Discovery Platform
Mar 26, 2026

Bio-IT World: AI Running 10-20 Million Predictions/Day in Drug Discovery — but What's a Prediction Worth?

Expert Systems CEO Tudor Oprea reports generative AI is now running an estimated 10-20 million predictions per day across the pharmaceutical industry to explore new molecules and reactions. However, Oprea argues a critical question is being ignored: the economic value of each prediction. His position is that a trustworthy prediction should be worth at least 5% of the actual experiment it replaces. LLM outputs still hallucinate and cannot be fully trusted, though reasoning-specialised LLMs perform better on complex multi-step problems. AstraZeneca's PIP platform alone makes ~1 million predictions daily. Active learning coupled with agentic AI gives GPU-rich companies the highest probability of success. Source
Clinical Trials Platform
Mar 25, 2026

Nature Communications: TrialMatchAI — Open-Source LLM System for Clinical Trial Patient Matching

Researchers published TrialMatchAI, an end-to-end AI recommendation system that automates patient-to-trial matching using fine-tuned open-source LLMs within a retrieval-augmented generation (RAG) framework. The system normalises biomedical entities, retrieves trials using hybrid lexical/semantic search, and performs criterion-level eligibility assessment via medical chain-of-thought reasoning. In real-world oncology validation, 92% of patients had at least one relevant trial in the top 20 results; expert assessment validated >90% accuracy in eligibility classification, particularly for biomarker-driven matches. Open-source and deployable locally for privacy. DOI: 10.1038/s41467-026-70509-w. Source
Platform Genomics/Proteomics
Mar 2026

Nature Communications: CellScope — Tree-Structured Framework Reveals Multi-Level Single-Cell Hierarchies

Li and colleagues introduced CellScope, a tree-structured computational framework that reveals multi-level cellular hierarchies and gene functions from single-cell RNA-seq data. Unlike flat clustering approaches, CellScope provides hierarchical organisation that mirrors biological tissue structure, with intuitive visualisation and deep biological views into cell types and their functional relationships. The approach addresses a key limitation of existing single-cell analysis: most methods produce flat clusterings that obscure the nested, hierarchical nature of cell populations in tissues. Source
Drug Discovery Platform
Mar 2026

Nature Communications: 3D Molecular Foundation Model Predicts Protein and Small Molecule Properties

Researchers published a 3D molecular foundation model trained across diverse biological domains that accurately predicts properties of both proteins and small molecules from a unified architecture. The model aids in the discovery of potential antiviral compounds, demonstrating cross-domain transfer learning from general molecular representation to specific therapeutic applications. This represents a shift from separate protein and small-molecule models toward unified molecular representations — a prerequisite for modelling the full drug-target interaction landscape computationally. Source
Drug Discovery Platform
Mar 2026

Nature Communications: CAMPER — Mechanistic AI Platform Designs Antimicrobial Peptides Effective Against MRSA

Researchers introduced CAMPER, a mechanistic artificial intelligence platform for designing antimicrobial peptides targeting methicillin-resistant Staphylococcus aureus (MRSA). The platform identified a stable peptide that eradicates MRSA biofilms and persister cells — two of the most treatment-resistant bacterial states — and demonstrated activity in mouse infection models. The work addresses the critical antimicrobial resistance (AMR) crisis by demonstrating that AI-designed peptides can overcome resistance mechanisms that defeat conventional antibiotics. Source
Platform Genomics/Proteomics Drug Discovery
Mar 2026

Nature Communications: GenSLM Protein Language Model Designs Functional Enzymes Outperforming Natural Variants

Researchers demonstrated that using the GenSLM protein language model to design TrpB enzyme variants yields stable, active enzymes with broad substrate promiscuity, outperforming both natural and laboratory-evolved counterparts. The study addresses a fundamental bottleneck in enzyme engineering: identifying functional starting points for optimisation. By generating diverse, functional enzyme variants computationally, the approach accelerates biocatalyst development for pharmaceutical synthesis and industrial biotechnology applications. Source
Drug Discovery Platform
Mar 2026

Nature Communications: M2UMol — Multimodal Knowledge Transfer for Drug Discovery with Missing Data

Researchers introduced M2UMol, a framework that transfers multimodal molecular knowledge (3D structure, spectral data, bioactivity profiles) into 2D molecular representations, enabling accurate property predictions even when experimental modalities are missing. Drug discovery often suffers from incomplete multimodal data — a compound may have structure data but lack spectral or binding data. M2UMol addresses this by distilling cross-modal knowledge during training, then making predictions from whatever data is available at inference time. Source
Platform Genomics/Proteomics
Mar 20, 2026

Nature Communications: UCASpatial — Ultra-Precision Deconvolution for Spatial Transcriptomics

Xu and colleagues published UCASpatial, an entropy-weighted algorithm that enables ultra-precision deconvolution of spatial transcriptomics data. The method resolves fine-grained immune cell landscapes and explores intercellular communication mechanisms at a resolution not achievable with existing approaches. Accurately mapping diverse cell types in complex tissues remains a major challenge for spatial transcriptomics; UCASpatial's entropy-weighting approach addresses this by adaptively handling varying confidence levels across spatial spots. Source
Platform Partnership Drug Discovery
Mar 19, 2026

GEN Edge: GTC 2026 Signals Agentic AI Inflection in Healthcare — 70% Adoption, 5,000+ Startups

GEN Edge analysis of NVIDIA GTC 2026 reports that healthcare AI has reached an inflection point: 70% of healthcare organisations now actively adopt AI (up from 63% in 2024), 69% use generative AI/LLMs (up from 54%), and NVIDIA's Inception programme has grown to 5,000+ healthcare/life sciences startups. NVIDIA VP Kimberly Powell declared "the transformer moment is now for biology." Healthcare AI startup ecosystem captured >85% of sector AI spending last year. The $4.9 trillion healthcare industry is deploying AI at more than twice the rate of the broader economy. Pharmaceutical leaders Roche and Lilly are investing in AI infrastructure at unprecedented levels. Source
Drug Discovery Platform
Mar 20, 2026

Frontiers in Bioinformatics: HAMGNN Graph Neural Network for LLM-Enhanced Drug Repurposing

Researchers published HAMGNN, a heterogeneous attention-based meta-learning graph neural network that integrates LLM-extracted therapeutic knowledge into a biomedical knowledge graph (DrugBank, DisGeNET, Hetionet; 2.2M+ edges). The model uses relation-sensitive multi-head attention and disease-focused meta-learning for rapid adaptation to unseen diseases. HAMGNN achieved ROC-AUC of 0.98 and precision of 0.95 on cold-start disease generalisation, representing 10-15% improvement over TxGNN and GAT-GNN baselines. DOI: 10.3389/fbinf.2026.1755412. Source
Platform Partnership Drug Discovery
Mar 18, 2026

NVIDIA GTC 2026: Healthcare AI Stack Expands with nvQSP, BioNeMo Genomics Blueprints

At GTC 2026, NVIDIA detailed expanded healthcare AI capabilities. nvQSP, a GPU-accelerated quantitative systems pharmacology simulation engine, achieves 77x speedup over traditional methods for testing treatment scenarios before clinical trials. New BioNeMo Blueprints for single-cell analysis and genomics were released. Basecamp Research is using BioNeMo/Parabricks for its Trillion Gene Atlas; Tahoe Therapeutics trains virtual cell models on single-cell data at scale; PerturbAI applies the platform to a large in-vivo CRISPR functional genomics atlas. 70% of healthcare organisations now actively adopt AI, up from 63% in 2024. Source
Platform Partnership
Mar 16, 2026

Roche Deploys Pharma's Largest Hybrid-Cloud AI Factory with 3,500+ NVIDIA GPUs

Roche announced expansion of its AI infrastructure with 2,176 NVIDIA Blackwell GPUs on-premises across the US and Europe, bringing its total to 3,500+ GPUs — the largest announced hybrid-cloud AI factory in the pharmaceutical industry. The infrastructure supports Genentech's "Lab-in-the-Loop" strategy using NVIDIA BioNeMo for drug discovery, Omniverse digital twins for manufacturing (including a new GLP-1 facility in North Carolina), and Parabricks for diagnostics. Builds on a 2023 Genentech-NVIDIA research collaboration. Announcement made at NVIDIA GTC 2026. Source
Platform Drug Discovery Genomics/Proteomics
Mar 16, 2026

NVIDIA GTC: Proteina-Complexa Model + AlphaFold DB Expansion with 30M Protein Complexes

At GTC 2026, NVIDIA launched Proteina-Complexa, a generative model for protein binder design as part of the BioNeMo platform. Novo Nordisk, Viva Biotech and Manifold Bio are early adopters, designing and experimentally validating generated protein binders. Separately, NVIDIA, Google DeepMind, EMBL-EBI and Seoul National University expanded the AlphaFold Protein Structure Database with ~30 million new AI-predicted protein complex structures and 1.7 million high-confidence predictions. This significantly broadens the resource for structure-based drug discovery. Source
Platform Genomics/Proteomics
Mar 16, 2026

Nature Communications: CellVQ Foundation Model for Single-Cell Transcriptomics

Researchers introduced CellVQ, a 500-million parameter single-cell foundation model trained on 68 million cells. Key innovation: a Single-Cell Discretisation (SCD) module that transforms high-dimensional sparse single-cell data into a compact "cell code," addressing data heterogeneity and improving interpretability. CellVQ-Graph extends the model by integrating cell features with multimodal data (genes, cell communication, annotations) into a knowledge graph for biological discovery. Addresses key limitations of existing single-cell foundation models around data sparsity and interpretability. DOI: 10.1038/s41467-026-70071-5. Source
Platform Clinical Trials
Mar 16, 2026

npj Digital Medicine: Comprehensive Review of AI-Driven Virtual Cell Models for Preclinical Research

A comprehensive review in npj Digital Medicine maps the landscape of AI-driven virtual cell models — computational models simulating cellular functional states, signalling networks, and dynamics under diverse perturbations. The review covers technical pathways (single-cell RNA-seq, spatial transcriptomics, proteomics), validation mechanisms, and clinical translation potential. Highlights terminological fragmentation across "virtual cell," "digital cell," and "digital twin" as an obstacle to cross-disciplinary communication and regulation. Calls for ISO-style standardisation and explainable-AI integration to improve clinical acceptability. DOI: 10.1038/s41746-025-02198-6. Source
Drug Discovery Platform
Mar 2026

AAAI 2026: CLADD — Multi-Agent LLM Framework for Drug Discovery Outperforms Domain-Specific Models

Published in AAAI 2026 proceedings, CLADD (Collaborative LLM Agents for Drug Discovery) from Genentech Research introduces a RAG-empowered multi-agent system using general-purpose LLMs for drug discovery. The framework includes specialised Planning, Knowledge Graph and Molecule Understanding teams that dynamically retrieve from biomedical knowledge bases without domain-specific fine-tuning. CLADD outperforms both general-purpose and domain-specific LLMs, as well as traditional deep learning approaches, on tasks including drug-target prediction, toxicity classification and molecular captioning. Code publicly available. Source
Clinical Trials Platform
Mar 14, 2026

Nature Communications: Federated Learning Enables Precision Oncology for Rare Cancers (atomCAT)

The international atomCAT consortium published federated multivariable Cox models trained across 14 centres (1,428 patients) and externally validated at 2 additional centres (277 patients) for anal cancer prognosis. The approach achieves consistent calibration and discrimination (c-indices 0.68-0.79) without centralising sensitive patient data. Identifies T stage, nodal involvement, tumour volume, sex and chemotherapy regimen as key prognostic factors. Demonstrates federated learning as a viable path for precision oncology in rare cancers where centralised data collection is impractical. DOI: 10.1038/s41467-026-70297-3. Source
Platform Genomics/Proteomics
Mar 14, 2026

Nature Communications: EPInformer — Scalable Deep Learning for Enhancer-Gene Expression Prediction

Researchers from Harvard/MIT published EPInformer, a deep learning framework that integrates promoter-enhancer sequences, epigenomic signals and chromatin contacts to predict gene expression. With only 0.4 million parameters, EPInformer outperforms existing methods while providing interpretable enhancer-gene interactions. The lightweight architecture makes it practical for genome-wide screening of regulatory variants, addressing a key bottleneck in connecting non-coding genetic variation to disease mechanisms. Complements larger models like AlphaGenome with a more targeted, interpretable approach. Source
Clinical Trials Platform Genomics/Proteomics
Mar 10, 2026

Nature Communications: TRIM — TCR-RNA Model Predicts Immunotherapy Response from Pre-Treatment Blood

Researchers developed TRIM (TCR-RNA Integrating Model), which jointly analyses T cell receptor sequences and single-cell RNA profiles to predict intra-tumour T cell signatures following checkpoint inhibitor treatment. Critically, the model works from pre-treatment blood or tissue samples, enabling prospective patient stratification before immunotherapy begins. This addresses a major unmet need in precision immuno-oncology: identifying which patients will respond to expensive checkpoint inhibitors before treatment starts, rather than after. Source
Funding Drug Discovery
Mar 13, 2026

Earendil Labs Considers Hong Kong IPO for AI Drug Discovery

Bloomberg reports AI drug discovery startup Earendil Labs is considering a Hong Kong listing that could raise up to $500 million. The company is working with China International Capital Corp. (CICC) and Morgan Stanley on the potential share sale. The move follows Insilico Medicine's successful December 2025 HK IPO, which raised $293M and was massively oversubscribed. Earendil's consideration signals continued investor appetite for AI-first drug discovery platforms in Asian capital markets. Source
Drug Discovery Platform
Mar 12, 2026

Nature: Machine Learning Predicts Drug Enantioselectivity from Sparse Data

UCLA and University of Utah researchers published in Nature a machine learning system that predicts how molecules form during drug synthesis. The method, trained on sparse data, cuts months of lab work to days. Lead author Prof. Abigail Doyle notes the tool is "highly applicable" for optimizing reactions in drug development phases. The transferable enantioselectivity model addresses a key challenge in pharmaceutical chemistry where small molecular changes can dramatically affect drug efficacy and safety. DOI: 10.1038/s41586-026-10239-7. Source
Platform Genomics/Proteomics
Mar 10, 2026

Nature Communications: MINT Protein Language Model for Protein-Protein Interactions

Researchers introduced MINT, the first protein language model (PLM) specifically trained on protein-protein interactions using the STRING database of 96 million interactions. MINT outperforms existing PLMs (including ESM-2) in binding affinity prediction and mutational effect estimation. The model excels at modeling complex protein assemblies, antibody-antigen interactions, and T-cell receptor-epitope binding. Key innovation: adapting model architecture to handle multiple protein sequences simultaneously while maintaining scalability. Source
Platform Genomics/Proteomics Drug Discovery
Mar 10, 2026

Nature Communications: PPLM Achieves State-of-the-Art Protein Interaction Prediction

Zhang Lab published the Protein Pair Language Model (PPLM), which jointly encodes paired protein sequences to learn interaction-aware representations. PPLM achieves state-of-the-art performance on cross-species protein-protein interaction prediction (human, mouse, fly, worm, E. coli, yeast). PPLM-Affinity outperforms both ESM2 and structure-based methods on binding affinity modeling, including challenging antibody-antigen and TCR-pMHC complexes. The approach advances therapeutic discovery by improving target-drug interaction predictions. Source
Clinical Trials Drug Discovery Regulatory
Mar 10, 2026

AI Drug Discovery Status: 200+ Programs in Clinical Trials, Phase III Readouts Imminent

Comprehensive analysis confirms over 200 AI-discovered drugs now in clinical development, with 15-20 entering pivotal Phase III trials in 2026. Insilico's rentosertib Phase IIa results (published Nature Medicine): 98.4mL FVC improvement (60mg) vs -62.3mL placebo decline over 12 weeks. FDA final AI guidance expected Q2 2026. First AI-discovered drug approval probability: ~60% by late 2026 or early 2027. Key watchpoints: Takeda's zasocitinib (TYK2 inhibitor) Phase III in psoriasis, Recursion's REC-4881 registrational path, and Generate Biomedicines' GB-0895 Phase III in asthma. Source
Platform Partnership Drug Discovery
Mar 8, 2026

Pharma AI Platform Licensing Emerges as New Business Model

GEN analysis: 2026 marks shift from single-asset AI deals to platform licensing. Noetik-GSK: $50M upfront, 5-year subscription for foundation models predicting cancer clinical outcomes—described as "first true foundation model licensing deal in biotech." Chai-Lilly: Platform deployment for biologics design across multiple targets. Boltz-Pfizer: Small molecule drug discovery. Generate Biomedicines CEO Mike Nally: typical 10-15 year discovery-to-approval journey could compress to 8 years with AI platforms. Pattern suggests pharma investing in AI infrastructure, not just molecules. Source
Clinical Trials Platform
Mar 2026

Medable Launches Agent Studio: First Agentic AI for Clinical Development

BioMed Nexus reports Medable launched Agent Studio, described as the industry's first agentic AI platform for clinical development. The platform automates routine clinical trial processes and reduces unproductive "white space" caused by manual workflows. Medable's end-to-end system integrates eCOA, eConsent, remote data collection, telemedicine, and connected device management. Partners include retail pharmacy networks, home health providers, and connected sensor companies. Part of broader trend of AI moving into clinical operations beyond drug discovery. Source
Genomics/Proteomics Platform
Mar 2026

Nature Biotechnology: GRAPE-LM Enables One-Round RNA Aptamer Evolution

Nature Biotechnology published GRAPE-LM (Generator of RNA Aptamers Powered by activity-guided Evolution and Language Model), a generative AI framework enabling one-round generation of short RNA binders. When guided by CRISPR-Cas-based intracellular screening, GRAPE-LM outperforms traditional multi-round aptamer evolution methods. The approach combines generative AI with single wet lab evolution round to generate high-affinity RNA aptamers, significantly accelerating therapeutic RNA development. DOI: 10.1038/s41587-026-03008-4. Source
Funding Partnership Drug Discovery
Jan 4, 2026

Insilico Medicine Signs $888M Servier Cancer Partnership Post-IPO

Less than a week after its Hong Kong IPO, Insilico Medicine announced an oncology discovery and development partnership with French pharma Servier valued at up to $888 million. Deal includes $32 million in upfront and R&D-related payments. Focused on "challenging targets" in cancer research using Insilico's Pharma.AI platform. Insilico leads AI-driven discovery; Servier shares R&D expenses and leads clinical validation. Part of ongoing validation of AI drug discovery platforms attracting major pharma investment. Source
Funding Drug Discovery Platform
Dec 30, 2025

Insilico Medicine Completes $293M Hong Kong IPO—Largest AI Biotech Listing

Insilico Medicine listed on Hong Kong Stock Exchange, becoming first AI-driven biotech on HKEX Main Board under Chapter 8.05 rules. Raised HKD 2.277 billion ($293M)—largest biotech IPO in Hong Kong 2025. Public offering oversubscribed 1,427x, locking HKD 328B+ in subscription funds. Cornerstone investors: Eli Lilly, Tencent, Temasek, Schroders, UBS AM, Oaktree. Lead candidate rentosertib (TNIK inhibitor) in Phase IIb/III for IPF. Post-IPO allocation: 48% clinical R&D, 20% early discovery, 15% AI model development, 12% automated lab expansion. Source
Clinical Trials Platform Drug Discovery
Mar 6, 2026

Recursion Pharmaceuticals Achieves First AI Clinical Proof-of-Concept

Recursion Pharmaceuticals reported Q4 earnings marking a key milestone: first AI-enabled clinical proof-of-concept. Phase 2 data for REC-4881 in familial adenomatous polyposis showed 43% median polyp reduction at 4mg once-daily dose, with 75% of patients responding. Responses remained durable after three-month treatment pause. Company engaging FDA on registrational path targeting 2026. Over $500M in cumulative partnership milestones. Platform now houses 50+ petabytes of multimodal data across phenomics, transcriptomics, proteomics, ADME, and 300M+ real-world patient lives. AI-driven discovery engine synthesizes only 300-330 compounds per program vs ~2,500 for industry peers. Source
Platform Drug Discovery Partnership
Mar 3, 2026

Insilico Medicine + Liquid AI Release LFM2-2.6B-MMAI Foundation Model

Insilico Medicine and Liquid AI announced partnership creating lightweight scientific foundation models for pharmaceutical research. LFM2-2.6B-MMAI (v0.2.1) is a single 2.6B-parameter model performing at state-of-the-art levels across drug discovery subdomains. Outperformed TxGemma-27B (10x larger) on 13/22 pharmacokinetics and toxicology tasks. Achieved 98.8% success rate on molecular optimization benchmarks. Trained on ~120B tokens across 200+ tasks. Operates entirely on private infrastructure, addressing pharma's data security concerns. Part of Insilico's Pharmaceutical Superintelligence (PSI) roadmap. Source
Drug Discovery Regulatory
Mar 1, 2026

Wiley Review Confirms No AI-Only Drug Has Achieved FDA Approval

Comprehensive peer-reviewed analysis published in Drug Development Research confirms that despite more than a decade of intensive research, no AI-only originated drug has achieved full regulatory approval. AI-guided molecules have progressed into clinical trials with encouraging early-phase success rates. Key limitations identified: poor data quality and accessibility, lack of model interpretability, gaps between computational predictions and chemical feasibility, and inherent complexity of biological systems. AI remains a supportive tool rather than standalone solution, underscoring continued need for human expertise and experimental validation. Source
Platform Drug Discovery Infrastructure
Feb 23, 2026

2026 AI Power Shift: Biotech Industry Enters "Builder Phase"

Drug Discovery News analysis based on Benchling 2026 Biotech AI Report: industry has entered "builder" phase where successful organizations are reshaping data environments and organizational structures to make AI a default operating model. 80% of organizations plan to increase AI budgets in next 12 months; 23% expect to double spend. 50% report faster time-to-target; 42% see accuracy uplift. Predictive models lead adoption: protein structure prediction (73%), docking models (52%). Poor data quality cited as #1 reason AI pilots fail (55% of organizations). Capital moving into data infrastructure and scientific modeling capabilities. Source
Infrastructure Platform
Feb 23, 2026

Ardigen Report: 95% of Enterprise GenAI Pilots Failed

Ardigen's AI in Biotech 2026 trends report cites MIT 2025 study finding nearly 95% of enterprise generative AI pilots failed to deliver measurable business impact. Failures occurred because systems remained disconnected from real workflows, data foundations, and organizational ownership. Shift from algorithms to data infrastructure: next phase of AI in biotech will be defined less by new algorithms and more by whether organizations can move from experimentation to dependable infrastructure. US AI biotech market approximately $2.1B in 2025, with growth driven by drug discovery, genomics, and precision medicine. Platform-oriented strategies dominating (e.g., Lilly TuneLab, Ginkgo Datapoints). Source
Infrastructure
Feb 18, 2026

AI Slashing Jobs Across Industries—But Pharma May Be Spared

PharmaVoice analysis indicates that despite a wave of AI-fueled layoffs across industries, pharma and biotech could be spared from massive job losses for now. AI integration requires talent to run and understand the technology. Companies that demonstrate willingness to adopt AI tools depends on risk appetite and having budget and skills to create AI solutions in-house or source externally. AI particularly helps early-stage companies where AI-derived trial improvements are more prominent. Integration can be tougher for bigger companies with legacy systems. Source
Platform Drug Discovery
Feb 17, 2026

University of Missouri Releases PSBench: 1.4 Million Protein Models

University of Missouri researchers released PSBench, the world's largest collection of protein models with quality assessment—1.4 million annotated protein structure models verified by independent experts. Database gives scientists reliable information to build more accurate AI systems for assessing protein structure model quality, critical for developing medical treatments for Alzheimer's, cancer, and other diseases. Led by Prof. Jianlin "Jack" Cheng, a Curators' Distinguished Professor in Bioinformatics. Even AlphaFold has limitations—no single AI tool is consistently accurate for every protein type, making quality benchmarks essential. Source
Regulatory Clinical Trials
Feb 2026

Clinical Trials 2026: AI Fluency Becomes #1 Organizational Differentiator

Applied Clinical Trials Online outlines sharp divide emerging in 2026: organizations building AI fluency into every layer of clinical process vs. legacy operators still piloting standalone "AI use cases." AI fluency—measured in talent, governance, and operational agility—will dictate survival. Key trends: living protocols with machine-readable design auto-created from biomedical concept libraries; secondary data utilization unlocking billions in latent insights; trial workforce transformation with demand for clinical data product managers, digital trial architects, and AI governance leads. FDA wants traceable, explainable logic and robust data provenance for clinical decisions. Source
Regulatory Platform
Feb 2026

FDA Deploys Agentic AI Capabilities Across Agency

FDA announced deployment of agentic AI capabilities for all agency employees (December 2025), launching two-month Agentic AI Challenge for staff to build and demonstrate solutions. Follows June 2025 launch of Elsa, an agency-wide generative AI assistant for scientific reviewers and investigators. January 2026: FDA published "Guiding Principles of Good AI Practice in Drug Development." CDER has seen significant increase in drug applications using AI components across nonclinical, clinical, postmarketing, and manufacturing phases. CDER AI Council established in 2024 to coordinate AI activities and ensure consistent external communications. Source
Clinical Trials Drug Discovery
Feb 2026

PitchBook: AI-Native Biotechs Show 80-90% Phase I Success Rate

PitchBook analysis (via BioSpace) indicates AI-native biotechs—companies using AI as foundational technology—have achieved approximately 80-90% Phase I success rate, compared to industry average of 40-65%. Phase II success rate dropped to 40%, still above industry average of 29%. Caveat: AI-focused biotechs have completed only ~10 clinical trials so far, making dataset nascent. Higher Phase I success rates may reflect improved target selection. AI enables biotechs to produce more "shots on goal" without escalating costs—structural de-risking. Smaller companies can now compete with big pharma on trial efficiency. Source
Funding Drug Discovery Platform
Feb 26, 2026

Generate Biomedicines Raises $400M in Largest Biotech IPO of 2026

Flagship Pioneering-backed Generate Biomedicines raised $400M in the year's largest biotech IPO, trading on Nasdaq as GENB. Lead candidate GB-0895, an AI-designed anti-TSLP monoclonal antibody, is in Phase 3 for severe asthma with twice-yearly dosing (vs. monthly competitors). The company previously demonstrated computer-to-clinic speed of 17 months for an earlier COVID candidate. Board includes Nobel laureate Frances Arnold and Moderna CEO Stéphane Bancel. Fifth biotech IPO in February 2026, bringing total monthly proceeds to ~$1.4B. Source
Infrastructure Platform Drug Discovery
Feb 26, 2026

Eli Lilly AI Factory Goes Live with Industry's Most Powerful Pharma Supercomputer

Eli Lilly activated its AI factory, the pharmaceutical industry's most powerful AI supercomputer built on NVIDIA DGX SuperPOD with DGX B300 systems. The "LillyPod" platform enables scientists to build chatbots, agentic workflows, and research lab agents. Select models will be available through Lilly TuneLab via federated learning infrastructure built on NVIDIA FLARE. Supports the $1B co-innovation lab with NVIDIA announced in January 2026. Chief AI Officer Thomas Fuchs: "This machine is exactly how AI should be used—for science, to lessen suffering and improve the human condition." Source
Platform Drug Discovery
Feb 24, 2026

Benchling 2026 Biotech AI Report: Industry Enters "Builder Phase"

Benchling's 2026 Biotech AI Report based on 100 biotech organizations reveals AI "killer apps": literature review (76% adoption), protein structure prediction (71%), scientific reporting (66%), and target identification (58%). Half of organizations report faster time-to-target; 80% plan to increase AI budgets in next 12 months. Data quality and availability cited as #1 reason AI pilots fail. 66% of scientists report increased LLM confidence but maintain "trust but verify" approach. Industry shifting from task copilots to integrated discovery systems. Source
Autonomous AI Drug Discovery Platform
Feb 24, 2026

LUMI-Lab Self-Driving Laboratory Discovers Novel mRNA Delivery Materials

University of Toronto researchers published LUMI-lab (Large-scale Unsupervised Modeling followed by Iterative experiments) in Cell, integrating molecular foundation models pretrained on 28M+ molecular structures with robotic systems. The SDL screened 1,700+ lipid nanoparticles across ten active-learning cycles, unexpectedly discovering brominated-tail ionizable lipids that enhance mRNA delivery to human lung cells—a class not previously linked to mRNA delivery. Results outperformed FDA-approved benchmarks with favorable safety profiles. Source
Infrastructure Clinical Trials
Feb 24, 2026

Jeeva Clinical Trials: Infrastructure, Not AI, Is the Bottleneck

Maryland-based Jeeva Clinical Trials amplified a key message from JPMorgan Healthcare Conference and BIO International: AI won't transform drug development without infrastructure evolution. CEO Harsha Rajasimha: "AI is not the constraint. The constraint is infrastructure. If you deploy advanced intelligence on siloed, outdated systems, you amplify inefficiency. If you deploy AI on a unified, cloud-native architecture, you amplify speed, compliance, and patient impact." Companies modernizing digital infrastructure today will define the next decade of clinical research. Source
Platform Drug Discovery
Feb 19, 2026

Isomorphic Labs Launches IsoDDE "AlphaFold 4" Drug Design Engine

DeepMind spinoff Isomorphic Labs released IsoDDE (Isomorphic Labs Drug Design Engine), described by Columbia's Mohammed AlQuraishi as "a major advance on the scale of an AlphaFold 4." IsoDDE more than doubles AlphaFold 3 accuracy on protein-ligand predictions, predicts binding affinities faster than gold-standard physics-based methods (FEP+), and identifies cryptic binding pockets from sequence alone—including sites that took researchers 15+ years to discover experimentally (cereblon). Unlike AlphaFold, IsoDDE is proprietary with limited technical disclosure. Already deployed in Novartis, Eli Lilly, and Johnson & Johnson partnerships. Source
Partnership Drug Discovery Platform
Feb 18, 2026

Merck and Mayo Clinic Announce AI Drug Discovery Partnership

Merck and Mayo Clinic announced an R&D agreement applying AI, advanced analytics, and multimodal clinical data to drug discovery—Mayo's first collaboration of this scale with a global biopharma company. Merck gains access to Mayo Clinic Platform_Orchestrate, enabling AI models to run inside Mayo's secure environment with de-identified clinical, genomic, imaging, and molecular data. Merck SVP Greg Hersch emphasized Mayo's "unique wealth of de-identified clinical, molecular multimodal data sets" not readily available elsewhere in clean, curated form. Initial focus: inflammatory bowel disease, atopic dermatitis, multiple sclerosis. Source
Platform Drug Discovery
Feb 16, 2026

MIT LLM Predicts Optimal Codons for Protein Drug Manufacturing

MIT researchers developed an LLM that analyzes the genetic code of industrial yeast Komagataella phaffii—specifically the codons it uses—to predict which codons work best for manufacturing therapeutic proteins. The model boosted production efficiency of six proteins including human growth hormone and monoclonal antibodies used to treat cancer. "Having predictive tools that consistently work well is really important to help shorten the time from having an idea to getting it into production. Taking away uncertainty ultimately saves time and money." Source
Drug Discovery Regulatory
Feb 16, 2026

AI in Drug Discovery 2026 Predictions: First FDA Approval Expected 2027-2028

Drug Target Review analysis notes no AI-discovered drug has achieved FDA approval as of December 2025, with first approval expected 2027-2028. Chinese AI drug discovery companies increased share of global biotech licensing deals from 21% (2023-2024) to 32% (Q1 2025). Market consolidation accelerating: multiple companies shut down despite substantial backing, others announced 20%+ workforce reductions, several pursued delisting. Valuations collapsed since 2021-2022 IPO peaks. 50:1 ratio between announced "biobucks" and actual upfront payments reveals industry caution. Source
Platform Drug Discovery
Feb 10, 2026

Benchling AI Now Generally Available with 500+ Biotech Companies

Benchling AI reached general availability after deployment across 500+ biotech companies from AI-native startups to top-20 pharma. Features AI agents for scientific tasks including Ask, Compose, Deep Research, and Data Entry. Integrates scientific models (AlphaFold 2, Chai-1, Boltz-2) directly in workflows. At Beam Therapeutics, regulatory writers use Deep Research for document preparation. In one large organization, hundreds of scientists adopted it within weeks to unlock years of captured data. Compose is being used to migrate 20,000 legacy ELN entries into structured, searchable formats. Source
Platform Autonomous AI
Nov 5, 2025

Google Launches Deep Research with Gmail and Drive Integration

Google announced that its Deep Research feature, available to all Gemini users, can now access emails and private documents from Gmail, Drive, and Chat to provide better answers. The autonomous "deep browse" capability creates comprehensive reports by pulling information directly from users' documents, email threads, and project plans alongside web sources. Source
Platform Drug Discovery
Nov 4, 2025

Recursion Pharmaceuticals Achieves $30M Milestone from Roche/Genentech

Recursion received a $30 million milestone from Roche/Genentech for delivering a whole-genome map of microglial immune cells, bringing cumulative partner payments to over $500 million. The company announced leadership transition with CEO Chris Gibson stepping down January 2026, replaced by Najat Khan, while reporting $785M cash runway through 2027. Source
Platform Autonomous AI
Nov 4, 2025

Cambridge Researchers Launch Denario: AI Assistant for Complete Scientific Process

Trinity Hall Cambridge announced the development of Denario, an AI-powered scientific assistant designed to accelerate the scientific process by helping researchers identify new research questions, analyze and interpret data, and produce scientific documents covering every step from hypothesis generation to final publication. Source
Platform Autonomous AI
Oct 30, 2025

Google Unveils Gemini 2.0 AI Co-Scientist for Scientific Discovery

Google Research announced a multi-agent AI system built with Gemini 2.0 as a virtual scientific collaborator. The system uses test-time compute scaling to iteratively reason and improve outputs through self-play scientific debates for hypothesis generation and ranking tournaments. Domain experts testing 15 open research goals found the AI co-scientist outperformed other state-of-the-art agentic models, with some AI-generated hypotheses already validated experimentally. Source
Infrastructure Drug Discovery
Oct 29, 2025

Lilly Deploys World's Largest AI Factory for Drug Discovery

Eli Lilly unveiled the world's largest AI factory wholly owned by a pharmaceutical company, featuring 1,016 NVIDIA Blackwell Ultra GPUs in a DGX SuperPOD system. This AI factory will train large-scale biomedical foundation models for drug discovery and development, with select models available through Lilly TuneLab platform for the broader biotech ecosystem. Source
Partnership Infrastructure
Oct 28, 2025

Microsoft-OpenAI Partnership Restructured with $250B Azure Commitment

Microsoft and OpenAI reached a new deal allowing OpenAI to restructure into a public benefit corporation while securing a $250 billion Azure cloud services contract. The agreement maintains Microsoft's frontier model partnership while allowing OpenAI more independence, including ability to release open-weight models and develop products with third parties. Source
Platform Drug Discovery
Oct 28, 2025

Harbour BioMed Launches World's First Fully Human Generative AI Antibody Model

Unveiled at Global R&D Day in Shanghai, Harbour BioMed introduced its first fully human Generative AI HCAb Model powered by the Hu-mAtrIx™ AI platform. The closed-loop system achieved 78.5% success rate with 107 de novo generated binder sequences, with validated molecules showing nanomolar-level binding affinity and average yields exceeding 700 mg/L. Source
Platform Autonomous AI
Oct 27, 2025

OpenAI Targets Autonomous AI Researcher by March 2028

OpenAI CEO Sam Altman revealed ambitious internal timelines to create an AI research intern by September 2026 and a fully autonomous AI researcher by March 2028. Chief Scientist Jakub Pachocki described this as a "system capable of autonomously delivering on larger research projects," with current models already matching top human performers. The announcement coincided with OpenAI's restructuring to a public benefit corporation, with the non-profit foundation committing $25 billion to use AI for curing diseases. Source
Partnership
Oct 27, 2025

Thermo Fisher Scientific Partners with OpenAI for Drug Development

Thermo Fisher Scientific revealed a strategic collaboration embedding OpenAI APIs into its Accelerator Drug Development platform and PPD clinical research business. Modeling shows 25-60% time savings across trial documentation (10,000-15,000 documents per trial) and site activation processes. Source
Platform Multi-Omics
Oct 27, 2025

World's First Multi-Omics LLM Showcased in Riyadh

BioAro Inc. announced the world's first large language model constructed from multi-omics data, called "The BioIntelligence™," powering their PanOmiQ™ platform. This breakthrough was showcased at the Global Health Exhibition in Riyadh and promises to decode the complete vocabulary of human biology by integrating genomics, transcriptomics, metabolomics, and proteomics data. Source
Platform Autonomous AI
Oct 27, 2025

"A Survey of AI Scientists" Published on arXiv

A comprehensive survey paper on AI scientists was published by researchers from multiple institutions. The paper charts the field's evolution from early Foundational Modules (2022-2023) to integrated Closed-Loop Systems (2024), and finally to the current frontier of Scalability, Impact, and Human-AI Collaboration (2025-present). The survey analyzes how artificial intelligence is transitioning from a computational instrument to an autonomous originator of scientific knowledge. Source
Platform Infrastructure Autonomous AI
Oct 26, 2025

Duke Engineers Build AI-Powered Autonomous Research Microscope

Duke University published research in ACS Nano describing ATOMIC, an AI optical microscope that analyzes 2D materials autonomously. The system doesn't just follow instructions—it understands them, can assess samples, make decisions independently, and produce results as well as a human expert. This represents a significant advancement toward autonomous research where AI systems work alongside humans. Source
Platform Autonomous AI
Oct 22, 2025

Sakana AI's AI Scientist v2 Produces First Peer-Reviewed AI Paper

Sakana AI announced that The AI Scientist v2 successfully produced the first fully AI-generated paper to pass peer review at the ICLR 2025 workshop level. The updated system adds agentic tree search for open-ended idea exploration, vision-language model reviewer capabilities, and parallel execution. The accepted paper documented a "negative result," sparking debate about the quality and role of AI-generated research. Source
Regulatory Drug Discovery
Oct 22, 2025

MHRA Launches AI-Powered Drug Interaction Prediction System

The UK's Medicines and Healthcare products Regulatory Agency announced three AI projects, including a groundbreaking study using artificial intelligence and NHS data to predict side effects from drug combinations before they reach patients. The flagship project, backed by £859,650 in government funding, will analyze anonymized NHS data focusing on cardiovascular medicines. Source
Platform
Oct 20, 2025

Anthropic Launches Claude for Life Sciences Platform

Anthropic unveiled Claude for Life Sciences, marking its formal entry into the life sciences sector. Built on Claude Sonnet 4.5, the platform integrates with Benchling, PubMed, 10x Genomics, and Synapse.org. Early adopters like Sanofi report daily usage among the majority of employees, while Novo Nordisk continues to reduce clinical study report production from 10+ weeks to 10 minutes. Source
Platform Drug Discovery
Oct 17, 2025

Google DeepMind's AI Model Discovers Novel Cancer Immunotherapy Pathway

Google DeepMind and Yale University's Cell2Sentence-Scale 27B (C2S-Scale) foundation model generated and experimentally validated a novel hypothesis for cancer treatment. The 27-billion-parameter AI identified that combining silmitasertib with low-dose interferon increases antigen presentation by 50%, potentially making "cold" tumors visible to the immune system. Source
Partnership Drug Discovery
Oct 15, 2025

Absci and Owkin Partner for Generative AI Drug Discovery

Absci and Owkin formed a strategic partnership to co-develop therapeutic candidates in immuno-oncology, immunology, and inflammation. Owkin's predictive AI models will optimize target selection using extensive biomedical datasets and patient-derived organoids, while Absci's generative AI Drug Creation platform can go from AI-designed antibodies to wet lab-validated candidates in as little as six weeks. Source
Platform Drug Discovery
Oct 9, 2025

AlphaFold 3 Database Major Update and Widespread Adoption

The AlphaFold Database received significant updates. Published analysis in Precision Clinical Medicine documented AlphaFold 3's unprecedented accuracy in predicting drug-like interactions, achieving 50% improvement over traditional methods and 76% accuracy in protein-ligand interactions, accelerating target identification and drug optimization. Source
Partnership Clinical Trials
Oct 2, 2025

Sanofi Ventures Invests in QuantHealth for AI Clinical Trial Simulation

QuantHealth secured strategic investment from Sanofi Ventures to accelerate AI-driven clinical trial simulation and digital twin technologies. The platform enables pharmaceutical companies to virtually simulate trials using over 350 million patient records, predicting trial outcomes and optimizing protocol design to improve success rates and reduce costs. Source
Platform Infrastructure Drug Discovery
Oct 1, 2025

BioNTech AI Day 2025 Showcases Kyber Supercomputer and BFN Protein Models

Held in London, BioNTech and InstaDeep showcased their near-exascale Kyber supercomputer and Bayesian Flow Network (BFN) models for protein sequence generation. The event highlighted AbBFN2, achieving 90% success rate in antibody humanization completed in under 20 minutes while preserving binding affinity, and new InstaNovo model delivering 10-15% accuracy increase in peptide sequencing. Source
Platform Drug Discovery
Aug 14, 2025

MIT Researchers Design Novel Antibiotics Using Generative AI for Superbugs

MIT's Antibiotics-AI Project announced that generative AI designed over 36 million novel compounds, identifying effective candidates against drug-resistant gonorrhoea and MRSA. The compounds work through novel membrane-disruption mechanisms and showed effectiveness in both laboratory and animal studies. Source
Drug Discovery Clinical Trials
Jun 3, 2025

Insilico Medicine's AI-Designed Drug Rentosertib Shows Efficacy in Phase 2a Trial

Published in Nature Medicine, Rentosertib became the first fully AI-discovered and AI-designed drug to demonstrate both safety and preliminary efficacy in human trials. The TNIK inhibitor for idiopathic pulmonary fibrosis showed a mean lung function improvement of +98.4 mL in the 60mg dosage group versus -20.3 mL decline in placebo, compressing target discovery to clinical candidate selection into just 18 months. Source
Platform Multi-Omics
Nov 23, 2025

Protein Set Transformer: Genome Language Model for Viromics

Researchers published Protein Set Transformer (PST), a protein-based genome language model that models complete genomes as sets of proteins without requiring functional labels. The transformer architecture enables high-diversity viral genomics analysis, addressing exponential increases in microbial and viral genomic data. PST helps overcome limitations of homology-based analyses that struggle with rapid viral genome divergence. Source
Infrastructure Regulatory
Nov 22, 2025

UK Launches £600M Health Data Research Service for AI Drug Discovery

The UK Government announced the Health Data Research Service (HDRS) backed by up to £600M from government and Wellcome Trust. HDRS will provide a single access point to large-scale, AI-ready pathology, radiology, and genomic datasets from multiple sources. The infrastructure supports AI model development and validation for biomarker identification, disease modelling, and pre-clinical models to improve drug response prediction. Source
Regulatory Clinical Trials
Nov 21, 2025

UK Plans Radical Clinical Trial Acceleration Using AI

The UK Government unveiled plans to reduce clinical trial set-up times from nine months to ten weeks, part of broader efforts to strengthen the life sciences sector. Chris Meier from Boston Consulting Group emphasized that AI, data, and analytics are critical to shortening timelines and enabling new drug discovery modes. The initiative aims to make UK hospitals more competitive globally for pharmaceutical company trials. Source
Partnership Platform Multi-Omics
Nov 20, 2025

NVIDIA and Sheba Medical Center Build 'ChatGPT for Genomics'

NVIDIA partnered with Sheba Medical Center (Israel) and Mount Sinai Hospital (New York) on a three-year, multi-million dollar project to create large language models trained on biological language. The ambitious goal is to decode the majority of the poorly understood human genome, enabling users to input whole genome sequences and receive personalized health risk assessments and treatment recommendations. Success depends on data quality, interpretability, and validation, with regulatory approval and ethical frameworks remaining uncertain. Source
Platform
Nov 20, 2025

NVIDIA Releases BioCLIP 2 Foundation Model for Biodiversity

NVIDIA launched BioCLIP 2, an NVIDIA-accelerated biology foundation model trained on the largest and most diverse dataset of organisms to date. The model identifies over a million species, demonstrating the potential of foundation models to organize and analyze massive biological datasets. This represents advancement in using AI for biodiversity research and ecological monitoring. Source
Platform Drug Discovery
Nov 20, 2025

MarkerPredict: Machine Learning for Clinically Relevant Biomarkers

Researchers published MarkerPredict, a bio-primed machine learning approach designed to enhance discovery of clinically relevant biomarkers. The methodology addresses challenges in biomarker identification by integrating biological priors with machine learning algorithms, improving the reliability and clinical utility of discovered biomarkers across multiple disease contexts. Source
Platform Drug Discovery Multi-Omics
Nov 18, 2025

Multimodal AI Framework for Drug-Target Interaction Prediction

Scientists introduced the Unified Multimodal Molecule Encoder (UMME) with Adaptive Curriculum-guided Modality Optimization (ACMO) for drug discovery. The framework integrates molecular graphs, protein sequences, and omics profiles while handling missing modality scenarios. The system uses confidence-based ranking and curriculum learning to prioritize reliable data during training, maintaining strong performance even with modality absence or noise, which mimics realistic drug screening conditions. Source
Platform Drug Discovery
Nov 14, 2025

Pistoia Alliance: Multi-Agent LLMs Excel in Database Queries

The Pistoia Alliance completed Phase 1 of its LLMs in Life Sciences project, focusing on target discovery and validation. Key finding: multi-agent LLM systems that challenge each other's outputs and interact with users significantly outperform single models in querying biological databases like Open Targets. The approach offers flexibility without requiring prior knowledge of database structure or query templates. Phase 2 will focus on creating proper benchmarks for natural language data mining systems. Source
Platform Drug Discovery
Nov 10, 2025

MAGE: AI-Designed Antibodies Without Existing Templates

Vanderbilt researchers published in Cell their development of MAGE (Monoclonal Antibody Generator), using protein language models to design functional human antibodies from scratch without needing existing antibody templates. The system was trained on H5N1 avian influenza antibodies and successfully generated antibodies recognizing related but previously unseen influenza strains. Traditional antibody discovery takes months; MAGE could reduce this to hours, dramatically accelerating pandemic response capabilities. In silico predictions require extensive experimental validation. Source
Platform Drug Discovery
Nov 15, 2025

Foundation Models Survey Published in Biology and Chemistry

A comprehensive survey on large language models in biology and chemistry was published, reviewing molecular representation strategies for biological macromolecules and small organic compounds. The review confirms that approximately 30% of published foundation models are now multimodal, with 20% trained specifically on protein sequences. The survey highlights the rapid evolution of protein language models that enable faster structure prediction, drug screening, and protein design without traditional multiple sequence alignment requirements. Source
Partnership Platform
Nov 13, 2025

OpenAI Exploring Consumer Health Applications

OpenAI is considering a significant push into consumer health products, including a personal health assistant and health data aggregator. The company has hired Nate Gross (Doximity co-founder) to lead healthcare strategy and Ashley Alexander (formerly Instagram) as VP of health products. With ChatGPT attracting 800 million active weekly users—many seeking medical advice—investors believe OpenAI could address longstanding challenges in personal health record management that have stymied previous Big Tech attempts. Source
Drug Discovery
Nov 13, 2025

AI in Biotechnology Market Projected to Reach $22.7 Billion by 2035

The AI in Biotechnology Market is projected to reach $22.7 billion by 2035, with rapid growth driven by machine learning integration in drug discovery, personalized medicine, and genomics research. Multiple reports confirm AI will drive 30% of new drug discoveries by 2025, with companies like Pfizer achieving 30-day drug discovery timelines (reduced from years) and saving 16,000 research hours annually. Source
Partnership Drug Discovery
Nov 9, 2025

Insilico Medicine and Eli Lilly Research Collaboration

Insilico Medicine announced a research collaboration with Eli Lilly, combining Insilico's Pharma.AI platform with Lilly's development expertise to discover innovative therapies. The agreement includes over $100 million in potential payments (upfront, milestones, and tiered royalties). This expands their 2023 AI software licensing partnership and validates Insilico's AI-driven drug discovery capabilities, which have nominated 22 preclinical candidates at an average pace of 12-18 months per program—significantly faster than the traditional 2.5-4 year timeline. Source
Platform Drug Discovery
Nov 9, 2025

DeepMind's AlphaFold 3 AI-Designed Drugs to Enter Human Trials

Google DeepMind CEO Demis Hassabis announced that AI-designed drugs using AlphaFold technology will enter human trials this year. AlphaFold 2 has predicted structures for virtually all 200 million identified proteins—work that would have taken a billion years of PhD time at traditional rates. AlphaFold 3's ability to model protein-ligand interactions with 50% greater accuracy than traditional physics-based methods is accelerating drug discovery timelines from years to potentially weeks or months. Source
Platform Drug Discovery
Nov 5, 2025

David Baker Lab: AI Achieves Atomic Precision in Antibody Design

Nobel Laureate David Baker's lab at the University of Washington announced a breakthrough using RFdiffusion AI to design full-length antibodies from scratch with atomic precision (RMSD values as low as 0.3 Å). The technology can design all six complementarity-determining regions de novo, compressing discovery timelines from years to weeks without requiring animal immunization or extensive screening. This could revolutionize the $200 billion antibody drug industry and enable treatments for previously "undruggable" diseases. Source
Platform
Oct 20, 2025

Anthropic Launches Claude for Life Sciences

Anthropic unveiled Claude for Life Sciences, marking its first formal entry into life sciences research. Built on Claude Sonnet 4.5 and optimized for laboratory protocols, the platform integrates with Benchling, PubMed, 10x Genomics, and Synapse.org to streamline R&D processes from literature reviews to regulatory submissions. Anthropic aims for Claude to support "a meaningful percentage of all life science work globally," reducing tasks that previously took days to mere minutes. Source
Partnership Infrastructure Drug Discovery
Sep 10, 2025

Absci Accelerates AI-Driven Drug Discovery with Oracle and AMD

Absci announced a collaboration with Oracle Cloud Infrastructure and AMD to accelerate generative AI-driven drug discovery. Using OCI's AI infrastructure with AMD Instinct MI355X GPUs, Absci has reduced inter-GPU latency to 2.5 µs and achieved terabytes-per-second throughput for large-scale molecular dynamics simulations. The company claims its Integrated Drug Creation Platform cuts drug development timelines by 14 months and costs by 75%, with Phase 1/2a trials for hair regrowth therapy ABS-201 planned for December 2025. Source
Platform
Dec 5, 2025

GenSyntax: Product-Contextualized LLM for Prokaryotic Genomes

Researchers unveiled GenSyntax, a specialized LLM trained on 49,250 annotated prokaryotic genomes for whole-genome decoding. This product-contextualized model advances genomic sequence analysis and functional annotation capabilities for microbial research, demonstrating enhanced performance on prokaryotic genome interpretation tasks. Source
Drug Discovery Platform
Dec 4, 2025

Cambridge Uses GPT-4 as AI Scientist for Drug Combination Discovery

Cambridge University researchers used GPT-4 as an "AI scientist" to identify novel drug combinations for cancer treatment. The study demonstrated that LLMs can explore hypothesis spaces that human researchers might miss due to cognitive biases. Professor Ross King noted the AI predicted combinations "no one would have found apart from randomly trying things," showcasing LLM capabilities beyond pattern recognition. Source
Platform
Dec 4, 2025

Nature Biotechnology Highlights Agentic AI in Lab Automation

Nature Biotechnology's 2025 research review highlighted how agentic AI systems using LLMs are streamlining laboratory workflows, including CRISPR system selection and gene transfer pathway discovery. The journal emphasized that with resources like ToolUniverse, LLM-powered AI co-scientists will increasingly guide hands-on research, moving beyond advisory roles to active experimental design. Source
Platform
Dec 4, 2025

Science Immunology Editorial: Are AI Immunologists Ready?

Science Immunology published a perspective piece examining whether LLMs are ready for immunology research applications. The editorial addresses the "meteoric rise" of LLMs and evaluates their readiness for analyzing immune system data and assisting immunological discovery, while cautioning about current limitations in specialized domains requiring deep mechanistic understanding. Source
Platform Clinical Trials
Dec 3, 2025

Medical-Specific LLM Outperforms GPT-4o on Clinical Tasks

John Snow Labs' medical-specific LLM (under 10 billion parameters) outperformed GPT-4o on clinical tasks when evaluated by medical doctors using the CLEVER framework. The smaller, domain-trained model was preferred 45-92% more often on factuality, clinical relevance, and conciseness, demonstrating that specialized healthcare LLMs can surpass larger general-purpose models through focused training on medical corpora. Source
Regulatory Clinical Trials
Dec 3, 2025

Responsible AI Framework for LLM-Driven Clinical Decision Support

A collaborative framework for responsible AI in LLM-driven clinical decision support systems for precision oncology was published in npj Precision Oncology. The framework addresses ethical implementation, patient data handling, and regulatory alignment using real-world patient data, providing guidelines for healthcare institutions deploying LLM-based diagnostic and treatment recommendation systems. Source
Platform Drug Discovery
Dec 2, 2025

RFdiffusion3 Released for Protein Design

The Institute for Protein Design released RFdiffusion3, an AI foundation model capable of generating proteins that interact with any molecular type. The tool offers 10-fold faster performance over RFdiffusion2 and uses atom-level diffusion for unprecedented precision in designing enzymes, biosensors, and gene therapy tools. This release follows Nobel recognition for protein design advances. Source
Platform
Dec 2, 2025

Biomedical Knowledge Graph Construction Using GPT-4o

Researchers developed IP-RAR (Integrated and Progressive Retrieval-Augmented Reasoning) and constructed BioStrataKG, a stratified knowledge graph from large-scale biomedical articles using GPT-4o mini. The system demonstrates enhanced cross-document question answering and multihop reasoning for biomedical knowledge extraction, published in GigaScience. Source
Platform
Dec 1, 2025

European Commission Releases Largest Human Biology Dataset for AI

The European Commission's Joint Research Centre released the largest dataset of human biology for data-driven AI techniques. This resource, comprising hundreds of thousands of entries, aims to enhance biomedical research through cutting-edge AI and LLM applications, supporting Europe's AI health research infrastructure development. Source
Drug Discovery Platform
Jan 11, 2026

DrugCLIP: 10 Million Times Faster Virtual Drug Screening

Tsinghua and Peking University researchers unveiled DrugCLIP, AI-driven screening enabling virtual screening of 10,000 proteins and 500 million compounds in single day, generating 2 million small-molecule hits. System achieves 10 million-fold speed increase versus current methods by combining contrastive learning and dense retrieval, enabling cross-screening of 10+ trillion protein-molecule pairs. Published in Science (DOI: 10.1126/science.ads9530), platform openly accessible for global drug discovery. Converts protein pockets and molecules into mathematical vectors for rapid matching, validated by computational and laboratory testing. Source
Drug Discovery Infrastructure Partnership
Jan 9, 2026

Zealand Pharma Secures Access to Denmark's Gefion AI Supercomputer

Zealand Pharma entered agreement with DCAI to use Gefion, Denmark's flagship AI supercomputer, for accelerating drug discovery. Access provides unprecedented computational power for large-scale modeling, prediction, and optimization of drug candidates. Agreement supports Zealand's Metabolic Frontier 2030 strategy targeting five launches, 10+ clinical pipeline programs, and industry-leading cycle times from idea to clinic by 2030. Represents European pharmaceutical companies securing sovereign AI computing capabilities for competitive advantage and data sovereignty. Source
Regulatory Clinical Trials Platform
Jan 6, 2026

Utah Launches First State-Approved AI Prescription Renewal Pilot

Utah launched first state-approved artificial intelligence program for autonomous prescription renewals. Twelve-month pilot permits AI to evaluate patient history and clinical data to approve refills for 190 common chronic medications including diabetes and hypertension treatments. Program developed by Doctronic in partnership with Utah's Office of Artificial Intelligence Policy and Department of Commerce. Controlled substances, ADHD medications, and injectables excluded for safety reasons. Represents first regulatory approval for autonomous AI medication management without physician review for each refill. Source
Clinical Trials Platform Regulatory
Jan 2, 2026

Stanford/Harvard Release Major AI Clinical Safety Benchmark

Stanford and Harvard researchers released NOHARM benchmark evaluating 31 large language models on 100 real primary care cases across 10 medical specialties. Study found top-performing AI models make 12-15 severe errors per 100 cases, while worst-performing systems exceed 40 severe errors. Models from Google, OpenAI, Anthropic, Meta, and medical platforms (AMBOSS, Glass Health) were compared against board-certified internal medicine physicians. Two-thirds of American physicians currently use LLMs in clinical practice, with one in five consulting systems for second opinions. Benchmark publicly available at bench.arise-ai.org. Source
Funding Platform Drug Discovery
Dec 30, 2025

Insilico Medicine: First AI-Driven Biotech IPO on Hong Kong Exchange

Insilico Medicine (3696.HK) listed on Hong Kong Stock Exchange, becoming first AI-driven biotech to go public under Main Board Chapter 8.05 rules. IPO raised HKD 2.277 billion ($293M USD), marking largest biotech IPO in Hong Kong for 2025. Public offering oversubscribed 1,427x, locking subscription funds exceeding HKD 328 billion. Shares jumped 25-45% on trading debut. Capital allocation: 48% clinical R&D, 20% early-stage drug discovery, 15% new generative AI models, 12% automated lab expansion. Company holds 300+ peer-reviewed papers, 700+ patents/applications, featured in Nature Index 2024 top 100 global institutions and Nature Biotechnology December 2025 cover. Source
Platform
Dec 24, 2025

Cornell Study: LLMs Transform Scientific Publishing Landscape

Cornell University study published in Science analyzed 2+ million papers across arXiv, bioRxiv, and SSRN (January 2018-June 2024), demonstrating LLM adoption dramatically increased scientific output. Researchers using LLMs posted 33% more papers on arXiv, exceeding 50% increase on bioRxiv and SSRN. Largest benefit for non-native English speakers: Asian-affiliated institutions increased output 43-89% depending on platform. Critical finding: growing AI-written content makes it harder for decision-makers to distinguish meaningful work from low-value content. Study also identified literature search advantage: AI-powered tools surfaced newer papers and relevant books more effectively than traditional search methods. Source
Drug Discovery Platform
Dec 23, 2025

AI-Generated Antibodies Tolerated by Human Immune Cells

Drug Discovery World reported AI-generated antibodies successfully tolerated by human immune cells, representing critical milestone toward clinical applications. Development follows December 9 Nature article on AI-designed antibodies approaching clinical trials, demonstrating rapid progression from computational design to biological validation. Progress indicates accelerating pathway from AI protein design to therapeutic candidates. Source
Regulatory Platform
Dec 19, 2025

AI Transforms Medical Genetics: Comprehensive Review Published

Genes journal published comprehensive review examining AI's integration with medical genetics, highlighting transformative role in diagnostic precision, non-invasive molecular profiling, and predictive medicine. Review emphasizes AI as generative tool in therapeutic design accelerating drug discovery, protein engineering, and precision gene editing, while addressing ethical challenges including data privacy, algorithmic bias, and dual-use biosecurity risks. Source
Funding
Dec 17, 2025

Life Sciences Industry Prioritizes AI Investment for 2026

FTI Consulting survey of 300 US life sciences decision-makers reveals 59% plan increased investment in AI and LLM initiatives for 2026, with R&D receiving 51% focus. Survey indicates 70% optimistic outlook despite fundraising uncertainty, with AI viewed as critical competitive advantage. Nearly 6 in 10 companies increasing AI/LLM budgets signals industry-wide recognition of transformative potential. Source
Platform Clinical Trials
Dec 16, 2025

Mount Sinai V2P: AI Decodes Disease-Causing Genetic Variants

Mount Sinai researchers developed V2P (Variant-to-Phenotype), phenotype-specific AI model published in Nature Communications that connects genetic variants to disease types with improved accuracy. System provides "clearer window" into how genetic changes translate into disease, enabling better prioritization of genes and pathways for therapeutic development. Represents advancement over general variant prediction tools by categorizing disease associations. Source
Platform
Dec 15, 2025

John Snow Labs Medical LLMs Win InfoWorld Technology of the Year

John Snow Labs' Medical LLMs recognized by InfoWorld for industry-leading accuracy and clinical impact. Peer-reviewed research shows models outperform GPT-4.5 and Claude 3.7 by 61-200% in factuality, clinical relevance, and conciseness while operating at fraction of cost. Platform includes 2,500+ pre-trained medical models with HIPAA, NIST AI RMF, and EU AI Act compliance, demonstrating domain-specialized superiority over general-purpose LLMs. Source
Regulatory
Dec 11, 2025

EFPIA Report: AI Governance Challenges in Medicine Lifecycles

European Federation of Pharmaceutical Industries and Associations published report highlighting governance challenges and policy actions needed to implement AI regulation across medicines' lifecycles. Report addresses gap between AI innovation pace and regulatory frameworks, emphasizing need for coordinated approach to ensure safety and efficacy while enabling advancement in pharmaceutical development. Source
Regulatory
Dec 11, 2025

European Commission Publishes Digital Omnibus on AI Regulation

European Commission released Digital Omnibus proposal for AI regulation, analyzed by Member States with criticisms regarding planned delays for key EU AI Act duties. Development represents evolving regulatory landscape attempting to balance innovation with safety requirements in AI deployment across healthcare and life sciences sectors. Source
Platform Clinical Trials
Dec 9, 2025

Comprehensive Medical LLM Benchmarking Framework Published

CMC journal published comprehensive evaluation of leading Medical-LLMs including GPT-4Med, Med-PaLM, MEDITRON, PubMedGPT, and MedAlpaca across diverse medical datasets. Study introduces domain-specific categorization system aligning models with optimal applications in clinical decision-making, documentation, drug discovery, research, patient interaction, and public health. Addresses deployment challenges including trustworthiness and explainability requirements. Source
Drug Discovery Platform
Dec 9, 2025

AI-Designed Antibodies Approach Clinical Trials

Nature reports that AI-designed antibodies are on the cusp of clinical trials, just one year after first proof-of-concept. Multiple teams using proprietary and open-source tools created antibodies with therapeutic properties comparable to commercial drugs. Chang Liu (UC Irvine) called these advances "remarkably powerful" and capable of "democratizing antibody engineering." This represents acceleration from concept to clinical-ready molecules in under 12 months. Source
Drug Discovery Platform
Dec 9, 2025

AI-Generated Antimicrobials Target Bacterial Membrane Microdomains

Scientific Reports published research using generative neural networks to design antimicrobial compounds targeting specific bacterial membrane microdomains, particularly cardiolipin-rich domains. The study modeled bacterial membranes with distinct lipid distributions and assessed AI-generated candidates via free-energy calculations, demonstrating selective targeting potential for next-generation antimicrobials. Source
Clinical Trials Platform
Dec 8, 2025

NetraAI Platform Demonstrates Precision Clinical Trial Enrichment

Nature npj Digital Medicine introduced NetraAI, an explainable AI platform integrating dynamical-systems modeling, evolutionary feature selection, and LLM-generated insights. Applied to a Phase II ketamine trial for treatment-resistant depression (n=63), it identified high-effect-size patient subpopulations and substantially enhanced treatment effect detection compared to traditional ML approaches, offering a pathway for prospective trial enrichment. Source
Drug Discovery Platform Regulatory
Dec 5, 2025

Frontiers Review: Generative AI in Antimicrobial Resistance

Frontiers in Microbiology published comprehensive review examining LLMs and protein language models for anticipating antimicrobial resistance pathways, designing novel agents, and guiding interventions informed by evolutionary dynamics. The review emphasizes robustness, explainability, and equitable predictions while addressing biosafety concerns around dual-use risks of generative AI in designing resistant strains. Source
Platform Drug Discovery
Dec 4, 2025

AI-Powered Biofoundries Accelerate Protein Engineering

Current Opinion in Biotechnology published review on AI-driven biofoundries accelerating the design-build-test-learn cycle in synthetic biology. Language models, generative AI, and active learning drive protein engineering, with emerging foundational biological models enabling multiscale design from DNA to cells. Cloud biofoundries and multi-AI agents advance toward self-driving laboratories. Source
Clinical Trials Platform
Dec 1, 2025

GPT-4 Achieves 97.9% Accuracy in Clinical Trial Patient Screening

New England Journal of Medicine published results showing GPT-4-based RECTIFIER system achieved 97.9% accuracy screening heart failure patients for clinical trials versus 91.7% for human specialists. The retrieval-augmented generation system processed average 12 clinical notes per patient and could significantly accelerate enrollment timelines while reducing operational costs. Source
Clinical Trials Regulatory
Nov 30, 2025

AI-Literacy Training Enhances Physician-LLM Diagnostic Collaboration

Randomized controlled trial in Pakistan (60 physicians, January-May 2025) showed AI-literacy training enabled physicians using LLMs to achieve 71.4% diagnostic reasoning scores versus 42.6% with conventional resources (27.5 percentage point difference, P<0.001). Study demonstrated physician-AI complementarities, with trained physicians surpassing LLM-alone performance in 38% of cases, highlighting importance of structured training before clinical deployment. Source
Platform Clinical Trials
Nov 30, 2025

Nature Review: LLMs Transforming Biomedicine and Healthcare

A comprehensive review in Nature npj Precision Oncology examined the current state of LLMs across biomedicine and healthcare. The paper explores practical applications including genomics (Evo, gLM, Caduceus), transcriptomics (scGPT), protein structure prediction (ESM-2, ESM3), and clinical decision-making, while addressing ethical concerns and technical challenges for real-world implementation. Source
Clinical Trials Platform
Nov 23, 2025

Bilingual Clinical Drafting AI Agent Deployed with EHR Integration

Publication describing on-premises LLM integration with electronic health records for clinical documentation. The system addresses the rare real-world implementation of LLMs in clinical workflows, providing bilingual capabilities for medical documentation generation while maintaining data privacy through local deployment architecture. Source
Platform Drug Discovery
Nov 26, 2025

Autonomous AI Agent "kai" Conducts Single-Cell Biology Analyses

A bioRxiv preprint introduced "kai," an agentic AI system that uses LLMs to plan and execute single-cell omics analyses through iterative Jupyter notebook interactions. The system combines retrieval-augmented generation across 7,000+ APIs and 6,000 notebooks, demonstrating improved robustness against code errors compared to one-shot generation. kai can autonomously formulate and address research questions on single-cell datasets without human intervention, representing a shift from AI assistance to autonomous scientific discovery. Source
Platform Drug Discovery
Nov 25, 2025

MIT Releases Open-Source BoltzGen for Generative Drug Discovery

MIT researchers released BoltzGen, a fully open-source generative AI model that designs novel protein binders for drug discovery targets from scratch. Building on Boltz-2, it is the first model to unify protein structure prediction with de novo drug candidate generation while incorporating physical and chemical constraints. The model was validated across 26 therapeutic targets in 8 wet labs spanning academia and industry, including tests against traditionally "undruggable" disease targets. Source
Drug Discovery Platform
Nov 24, 2025

UNC Achieves 200-Fold Potency Improvement Using AI-Guided Drug Design

UNC Eshelman School of Pharmacy's Center for Integrative Chemical Biology and Drug Discovery reported that AI-guided generative methods discovered compounds targeting a critical tuberculosis protein in six months, achieving 200-fold enzyme potency improvement in just a few iterations. The lab also released DELi, the first open-source software rivaling commercial tools for analyzing DNA-encoded library data. Source
Clinical Trials Platform
Nov 22, 2025

Multimodal LLM Achieves 93% Accuracy in Clinical Trial Patient Matching

Researchers published validation results for a multimodal LLM pipeline for clinical trial patient matching in Nature Communications Medicine. The system uses visual LLM capabilities to interpret medical records including scans, tables, and handwritten notes without lossy conversions. Results: 93% criterion-level accuracy on the n2c2 benchmark, 87% accuracy across 485 real-world patients from 30 sites matched against 36 trials, and 80% reduction in screening time versus manual chart review. Source
Partnership Funding Drug Discovery
Nov 19, 2025

AI Proteins Secures $41.5M for Generative Miniprotein Therapeutics

Boston-based AI Proteins closed a $41.5 million Series A led by Mission BioCapital and Santé Ventures to advance de novo miniprotein therapeutics designed using generative AI. The company has generated molecules against 150+ targets with multiple programs showing in vivo proof-of-concept, following a $400M research collaboration with Bristol Myers Squibb announced in December 2024. Source
Regulatory
Nov 19, 2025

WHO Europe Issues First Regional AI Healthcare Assessment

The World Health Organization's European office published its first comprehensive assessment of AI adoption and regulation across 50 of 53 member countries. Key findings: 86% of countries cite legal uncertainty as their top barrier to healthcare AI adoption, 78% cite affordability concerns, and fewer than 10% have established liability standards for AI in health settings. WHO Regional Director Hans Kluge warned that without clear data privacy protections and AI literacy investments, the technology risks deepening health inequities. Source
Regulatory Clinical Trials
Nov 18, 2025

Nature Perspective Reveals FDA's Internal LLM "Elsa"

A perspective article in npj Digital Medicine examined AI modernization of clinical trials and disclosed that the FDA has deployed an internal LLM called "Elsa" (powered by Anthropic's Claude) to help staff accelerate clinical protocol reviews and shorten scientific evaluation times. The paper detailed how LLMs enable AI-driven eligibility optimization, reinforcement learning for real-time protocol adaptation, and digital twin modeling for clinical trials. Source

No news items match your search criteria.