TRIAGE: An AI Scientist for Adversarial Target Falsification

Jiawei Xing

Additional Submission Instructions: For the camera-ready version, please include the author names and affiliations, funding disclosures, and acknowledgements.

Track: Track 3: AI Scientist Proposal Competition

Keywords: drug target discovery, adversarial agent, causal perturbation

TL;DR: An agentic target validation system that prioritizes the falsification of weak target hypotheses through historical evidence and in silico causal perturbation analyses.

Abstract: Despite massive investment in drug discovery, more than 90\% of drug candidates fail in clinical trials, often because the underlying target hypothesis proves ineffective or unsafe in humans. We present TRIAGE (Target Review via Iterative Adversarial Generation and Evaluation), a multi-agent framework that shifts target discovery from hypothesis generation to hypothesis falsification. TRIAGE combines a generative agent for target nomination with two adversarial agents for evidence retrieval and $\textit{in silico}$ perturbation analyses. By systematically challenging proposed targets, TRIAGE aims to filter out false positives before they advance into the costly downstream stages of drug development. We evaluate TRIAGE using benchmarks derived from historical clinical trial outcomes and public biomedical databases.

Submission Number: 15