2026-09-18
A biotech-shaped agent org annotated 55,984 trials: cell-type-specific targets were 48% more likely to reach market and had 32% fewer adverse events, then argued for a B7-H3 ADC.
A drug still takes on the order of a decade and billions of dollars to reach patients, and about 90% of candidates that enter Phase I never get approved. Failures cluster on weak efficacy and surprise toxicity. The evidence that should catch those failures already exists: genetics, single-cell and spatial assays, structure, clinical trials. What fails is the organisation. The data live in different disciplines and databases, decisions mix with departmental incentives, and it is hard to audit why a call was made.
James Zou's group at Stanford turns that into a multi-agent system: not a replacement for scientists, but a biotech-shaped org that folds cross-scale evidence into one inspectable reasoning chain.
The standing roster is small: a virtual CSO that decomposes questions, delegates, and synthesises, four scientific divisions (target identification, target safety, modality selection, clinical), plus staff and a scientific reviewer. The CSO does not run analyses itself. Each scientist agent has its own system prompt and a slice of more than 100 custom MCP tools, covering about 78,726 targets, 39,530 diseases, over 100 million single-cell profiles, Tahoe-100M perturbations, ClinicalTrials.gov, and PubMed.
The scale-out is for trial annotation. Open Targets lacked endpoints, and ClinicalTrials.gov often stores them as messy free text. The system dispatched 37,075 clinical-trialist agents in parallel on Phase II/III records. Each NCT followed a three-tier cascade: ClinicalTrials.gov, then PubMed, then press releases and regulatory notes, with the source stored on every field. Phase I success was defined algorithmically as whether a later Phase II existed for the same drugs and indication.
The B7-H3 case study turned off web search and fetch. The underlying model cutoff was January 2025, to limit pretraining leakage.
On a 100-trial manual audit, agent-human agreement was 89.7% (78/87) on primary endpoints, 83.9% on secondary endpoints, and 92.4% on adverse-event rates. Against 7,278 overlapping trials in Therapeutic Data Commons, primary-endpoint agreement was 85.3%. Disagreements clustered on multi-arm studies, p-values hugging 0.05, and single-arm designs that do not split cleanly into win/lose.
Across 55,984 trials, drugs against cell-type-specific genes were 40% more likely to move from Phase I to Phase II (OR 1.27), 48% more likely to reach Phase IV (market), and had 32% lower adverse-event rates across several organ classes. Expression bimodality tracked endpoint success and fewer terminations. The associations survived 1,000 permutations, mixed-effects models for indication and modality, and adjustment for genetic evidence; in the 74.6% of trials with no genetic support, the single-cell effects were as large or larger. On Tahoe-100M, an apoptosis signature associated with primary-endpoint success at OR 1.15.
The second task asked whether B7-H3 (CD276) is a lung-cancer target. Germline genetics were essentially empty; the CSO treated that as expected for a checkpoint. Single-cell data showed the strongest upregulation in fibroblasts (SCLC log2FC 1.94, LUAD 1.46). Spatially, immune cells were depleted around B7-H3-high spots. In TCGA LUAD, high expression tracked worse overall survival (HR 1.62) and disease-specific survival (HR 1.82). The pharmacology agent found no druggable pocket, so the CSO preferred an ADC over a small molecule or PROTAC. The run cost $46 in Anthropic credits. In August 2025, the B7-H3 ADC ifinatamab deruxtecan received FDA Breakthrough Therapy Designation, with a 48.2% objective response rate in Phase II SCLC.
The third case revisited the terminated ulcerative colitis Phase II MOONGLOW (vixarelimab against OSMRβ, stopped for futility at 79 patients). The system flagged the lack of OSMR-high enrollment and found OSMR-pathway amplification in fibroblast states from biologic non-responders. That run cost $54.
For target ranking, this moves single-cell features from descriptive colour to observational predictors of endpoints and safety, orthogonal to genetics. For agent builders, the thing that actually scaled is 37,075 cloned annotators, not 37,000 distinct experts; the standing org is still about eleven agents plus tools. The B7-H3 case is useful because a clinical readout that landed after the cutoff pointed the same way as the system's fibroblast, immune-exclusion, and ADC story. That is still hypothesis generation. It did not invent Merck's drug.
Agents misread and miscalculate, and errors propagate through the virtual org. A reviewer agent reduces that risk; humans still have to stay in the loop. The trial analysis is observational. Mixed-effects and permutations help, but indication, endpoint definition, and sponsor effects can remain. Coverage is ClinicalTrials.gov plus mappable molecular targets. The case studies are hypotheses that need wet lab and later trials. Pretraining leakage cannot be ruled out; a January 2025 cutoff and a literature check lower it, they do not zero it. This reading uses the bioRxiv full text of the same paper plus the Science abstract, whose headline numbers match.