Human-supervised Agentic AI for Hypothesis Generation and Experimental Assistance in Drug Repurposing

It introduces the RepurAgent hierarchical multi-agent architecture, in which a supervisor agent and a planning agent coordinate four specialized sub-agents (research, prediction, data, and report) under a human-in-the-loop design with episodic memory and retrieval-augmented generation, grounded in data, tools, and standard operating procedures specific for drug repurposing developed within the REMEDi4ALL consortium. Relative to prior computational drug repurposing work focused largely on rapid hypothesis generation, this work extends agentic AI across a broader lifecycle including drug candidate suggestion, experiment design, assay data analysis, and iterative candidate refinement. The system architecture and its grounding sources are explicitly described in the abstract, and the system is open source and publicly deployed, though this document provides only the abstract and revision notes without implementation details.

In the Acute Myeloid Leukemia scenario, a blinded expert evaluation indicated that RepurAgent produced substantially more novel and mechanistically credible candidates compared to a vanilla LLM baseline. It brings in external blinded expert evaluation as a validation approach rather than relying only on internal metrics. The basis is the blinded expert evaluation design; the abstract does not report the number of evaluators, the rating scale, or specific statistics.

In a retrospective COVID-19 antiviral screen, RepurAgent acted as an adaptive experimental collaborator, prioritizing compounds with AUC-ROC up to 0.99 without predefined thresholds and flagging confounders missed in manual review. It demonstrates agent participation in the assay data analysis stage, not only in hypothesis generation. The AUC-ROC of up to 0.99 is a quantitative figure given in the abstract, but the setting is a retrospective screen and the abstract does not state compound counts or validation cohorts.

For Multiple Sulfatase Deficiency, RepurAgent prioritized 81 high-confidence candidates from 5000 compounds, which were further corroborated by domain experts. It compares agent output against domain expert judgment in a rare-disease repurposing setting where data are relatively scarce. The candidate count and total compound number are explicit in the abstract; expert corroboration is qualitative, and the abstract reports no experimental validation results.

The work targets the early and middle stages of drug repurposing and applies to research settings that have the relevant data, tools, and standard operating procedures together with domain expert supervision; in the revision notes the authors explicitly narrow their claims, removing language implying end-to-end autonomous discovery or coverage of the "entire" repurposing lifecycle, and state that the value lies in orchestrating a multi-step, multi-tool workflow rather than in a completely new model or algorithm for screening.

This document contains only the abstract and revision notes, without figures, sample sizes, comparator settings, or experimental validation details, so the actual biological validity of the candidates in each scenario cannot be assessed; which endpoints are proxies and what they can and cannot establish is stated by the authors to be explicit in the full text but is not expanded in the abstract; additionally, whether the Acute Myeloid Leukemia and Multiple Sulfatase Deficiency candidates underwent wet-lab validation, and whether the retrospective COVID-19 screening results replicate prospectively, remain open questions worth watching.