AI-human discovery engines for mechanistic cancer biology

The National Institutes of Health Office of the Director awarded Weill Medical College of Cornell University $1.185 million on September 1, 2026, under the Trans-NIH Research Support program (CFDA 93.310) to develop an AI-human discovery engine for mechanistic cancer biology.

The project builds, deploys, and benchmarks an agentic, closed-loop system that learns from human multi-omic datasets, constructs virtual-tumor foundation models to simulate cancer biology, and autonomously designs and executes experiments in patient-derived organoids using high-throughput robotics. A multi-agent architecture of hypothesis, critique, planning, and analysis agents will generate, test, and refine causal mechanisms in continuous cycles, governed by a scientist-in-command principle where human researchers steer objectives, interrogate the AI's reasoning, and ensure scientific rigor and ethical oversight. The engine shifts discovery from a literature-first, animal-first, and hypothesis-first paradigm to one that is data-first, human-first, and AI-collaborative, explicitly designed to move beyond correlation to uncover causal mechanisms through training on interventional data from perturbation screens and adversarial criticism. The system will initially be deployed to investigate the mechanistic basis of early lung adenocarcinoma progression. Success will be measured by the engine's ability to generate replicable mechanistic discoveries and by creation of a blueprint—including open-source code, protocols, and governance practices—that allows other research groups to adopt this paradigm.

Performance occurs in New York, New York. The period of performance runs through August 31, 2031.

Abstract

The process of biomedical discovery has not kept pace with the data, models, and instruments available today. We are constrained by a 20TH-century scientific method while possessing 21ST- century tools. This project seeks to RE-engineer the scientific method itself for mechanistic cancer biology by building, deploying, and benchmarking the first AI-human discovery engine. This innovative system will shift discovery from a literature-first, animal-first, and hypothesis-first paradigm to one that is data-first, human-first, and AI-collaborative. The engine is designed as an agentic, closed-loop system that learns directly from human multi-omic datasets, builds virtual-tumor foundation models to simulate cancer biology, and autonomously designs and executes experiments in patient-derived organoids using high-throughput robotics. A multi-agent architecture of hypothesis, critique, planning, and analysis agents will generate, test, and refine causal mechanisms in a continuous cycle. This entire process is governed by a scientist-in- command principle, where human researchers steer objectives, interrogate the AI's reasoning, and ensure scientific rigor and ethical oversight. By training the system on interventional data from perturbation screens and designing IT for adversarial criticism and human auditability, the engine is explicitly built to move beyond correlation to uncover causal mechanisms. To demonstrate its power and refine its capabilities, the engine will initially be deployed to investigate the mechanistic basis of early lung adenocarcinoma progression, a critical area of clinical need. The project's success will be measured by the engine's ability to generate replicable mechanistic discoveries and by the creation of a blueprint-including open-source code, protocols, and governance practices-that will allow other research groups to adopt this new paradigm. This work aims not only to deliver key insights into cancer biology but also to fundamentally change how biomedical science is conducted, making AI a true collaborative partner in a human-led scientific enterprise.