Stanford's AI Co-Scientist Clears Peer Review — With 10,000 Labs Already Running It
Biomni, Stanford's general-purpose biomedical AI agent, lands in Science with a live user base and a benchmark score nearly double general LLMs. Plus: a universal coordinate system for cell biology in Nature, investors split on whether AI-bio valuations still track clinical reality, and Anthropic's own drug-discovery bet comes into sharper focus.
Stanford's Biomni became the first general-purpose biomedical AI agent to clear peer review, appearing in Science on July 9 with more than 10,000 labs already using it in production — a distribution-first strategy that beat commercial competitors to market. The technical story isn't a new model; it's the action space: researchers mined roughly 2,500 bioRxiv papers across 25 subdomains to assemble 150 specialized tools, 105 software packages, and 59 databases into one environment an agent can actually operate. In one documented case, the system turned 450+ continuous glucose monitoring files into cleaned data, visualizations, and cross-variable patterns in about 40 minutes — work the team estimated at 60+ human hours.
Lead — Biomni clears peer review, 10,000 labs deep
PapersStanford's Biomni — a general-purpose biomedical AI agent that accepts plain-language research questions and executes full workflows (literature review, hypothesis formation, dataset selection, code, result interpretation, experimental design) — appeared in Science on July 9, with 10,000+ labs already running it. The innovation is the action space, not the model: researchers analyzed ~2,500 bioRxiv papers across 25 biomedical subdomains to extract 150 specialized tools, 105 software packages, and 59 databases, unified into one environment. On a 443-question benchmark, Biomni scored ~57% versus ~30% for general LLMs and ~43% for coding assistants. In a documented case, it turned 450+ continuous glucose monitoring files into cleaned data, visualizations, and food-intake/body-temperature correlations in ~40 minutes — an estimated 60+ hours of human work. The open-source code and Biomni-R0 weights are public, and free access plus tool integration got it into labs before any commercial competitor shipped a comparable product — distribution as the moat, not model quality.
Nature publishes a universal coordinate system for cell biology
PapersNature published Universal Cell Embedding (UCE) on July 8 — a foundation model trained on 36 million cells spanning 1,000+ cell types, dozens of tissues, hundreds of experiments, and eight species, built on a new Integrated Mega-scale Atlas. UCE embeds novel cells zero-shot, with no per-dataset labeling, retraining, or fine-tuning required. Cell annotation has been an expensive, inconsistent bottleneck across labs; a published, zero-shot, cross-species baseline now gives the field a standard to measure every perturbation-prediction platform against.
Investors split on whether AI-bio valuations have detached from clinical reality
OpinionGEN surveyed investor views on AI-bio valuations, with analyst Andrii Buvailo framing the fork: either major investors hold genuine conviction and are buying ownership ahead of clinical results, or, as he puts it, 'the AI-discovery valuation cycle has fully decoupled from clinical proof.' Isomorphic Labs raised $2.1B in May with no clinical data to point to. The investors interviewed converge on an uncomfortable point — the model isn't the defensible asset. Playground's Jory Bell emphasizes data differentiation; Obvious Ventures' Rohan Ganesh argues a model 'that is accurate but does not move the pace or probability of clinical success is...meaningless'; Dimension's Simon Barnett questions whether frontier labs eventually subsume platform technology outright; and Converge Bio's Dov Gertz estimates seven more years before a computationally designed molecule reaches a patient.
Anthropic ships a drug discovery workbench — and starts developing its own drugs
Product launchAnthropic's Claude Science launch (June 30) got a deeper profile this week — the Boston Globe covered the company's own preclinical programs targeting neglected and rare diseases on July 13. The workbench spans 60+ functions across genomics, single-cell analysis, proteomics, structural biology, and cheminformatics, with native protein/molecule rendering and every result traced to its underlying code. Context that sharpens the picture: Anthropic acquired drug-discovery startup Coefficient Bio for roughly $400M in April, its board includes Novartis CEO Vas Narasimhan, and BMS deployed Claude to 30,000+ employees in May. Per Endpoints and CNBC, Anthropic's own stated rationale for developing drugs is that it needs to understand the industry from the inside to build tools that actually work for it.
Insilico shares jump on Takeda deal
Industry dataInsilico's Hong Kong-listed shares rose 13.5% on news of the Takeda partnership — a reaction driven by the $60M upfront payment, not any clinical milestone.
Fore Bio raises $67.4M without an AI pitch
FundingFore Bio raised $67.4M for BRAF-altered precision oncology from Novartis Venture Fund, SR One, OrbiMed, Medicxi, and Wellington — notably absent from the pitch was any AI framing; the round leaned on traditional target and biomarker clarity instead.
Lilly taps Tamarind Bio for TuneLab 2.0 inference
InfrastructureLilly selected Tamarind Bio to host inference infrastructure for TuneLab 2.0, its federated platform giving biotech partners access to models trained on Lilly's proprietary data.
Cell and gene therapy funding stays flat at ~$2B
Industry dataCell and gene therapy funding has held at roughly $2B annually since 2022, untouched by the AI capital wave — the constraint is manufacturing and delivery, not computation, and poor commercial performance plus regulatory uncertainty keep investors cautious regardless of AI framing.
The Onion Desk
AI drug discovery sector celebrates a perfect safety record across zero approved drugs. $7B in partnership commitments sit against an entirely theoretical therapeutic portfolio; one platform executive called the record 'unblemished and, frankly, unbeatable.' Industry leaders report target-to-candidate timelines compressed 'from four years to eighteen months,' and candidate-to-approval improvements characterized as progressing 'from an average of eight years to infinity.' Full satirical dispatch on Substack.