AI Agents Collaborate Like a Drug Development Organization to Propose Lung Cancer Treatment Strategies

A system has been developed in which AI agents collaborate like a drug development organization to propose lung cancer treatment strategies and analyze why discontinued clinical trials failed.

The National Biotechnology Policy Research Center said in a report published on the 22nd that U.S. researchers had developed a “Virtual Biotech” in which multiple AI agents collaborate. The study was published on the 17th (local time) in the international journal Science.

Virtual Biotech is organized around a virtual chief scientific officer (CSO) and includes specialized agents in four areas—target discovery, safety assessment, treatment selection and clinical development—as well as research support and scientific review agents.

The virtual CSO breaks users’ questions into specific tasks, assigns them to specialized agents and synthesizes the results. If the evidence is insufficient, it directs further analysis; users can review the analysis process and the code, data and reports generated.

Using more than 100 analytical tools, the researchers analyzed information on about 79,000 targets, 40,000 diseases, more than 100 million single-cell profiles, drugs and clinical trials.

An analysis of about 56,000 clinical trials found that drugs targeting genes selectively expressed in specific cell types were 48% more likely to reach Phase 4 than drugs targeting genes expressed broadly. The incidence of adverse events was an average of 32% lower.

The researchers noted, however, that these figures came from an observational analysis of existing clinical trials and do not mean that adopting Virtual Biotech increased the success rate of drug development by 48%.

In a lung cancer case study, the system identified increased B7-H3 expression in cancer-associated fibroblasts and a poor prognosis among patients with high expression, and proposed antibody-drug conjugates (ADCs) as a promising treatment approach.

In a case study of ulcerative colitis, the system analyzed why a discontinued Phase 2 trial of vixarelimab failed. It proposed that blocking OSMR alone may not sufficiently suppress inflammatory signals and suggested using a composite biomarker strategy that reflects the relevant signaling pathways.

Four human experts independently repeated eight analyses across the two disease case studies. Their conclusions generally agreed with the AI analyses, and no coding errors had a substantial effect on the conclusions.

The researchers said they plan to expand the scope of specialized agents to include molecular design and virtual screening, toxicity prediction and clinical decision-making. The current system, however, is limited to computational analysis, evidence integration and prioritization; it does not include conducting experiments or demonstrating treatment efficacy.