PerturbAgent: An Agentic AI system for Analysis and Prediction of Genetic Perturbations

Kanglu Pei, Sisi Qu, Philip Torr, Jonathan G. Hedley, Christian Schroeder de Witt

Track: Regular Track (Page limit: 6-8 pages)

Keywords: LLM-based Agents, Genetic perturbation, Interpretable AI, Automated Scientific Discovery, Agentic System Evaluation

TL;DR: PerturbAgent is a multi-agent system that automates analysis, prediction, and interpretation for genetic perturbation studies, evaluated by our new MAST++ framework for general agentic performance and combined with biological validity assessment.

Abstract: We introduce PerturbAgent, a large language model (LLM)-based multi-agent system for single-cell genetic perturbation studies. In biomedical research, understanding cellular responses to perturbations is essential for interpreting gene function and regulatory pathways in single-cell data. Existing methods focus only on either single-cell analysis pipelines or perturbation prediction models, and often lack this necessary biological interpretation. PerturbAgent addresses these limitations, targeting both analysis and prediction tasks while also generating comprehensive biological interpretations with results grounded in mechanisms, pathways, and existing knowledge. We further propose MAST++, a general framework that evaluates agentic performance across profile, reasoning, perception, interaction, and memory, and complement it with biological validity assessments. On public single-cell Perturb-seq and RNA-seq datasets, PerturbAgent reliably achieves high task completion and delivers citation-backed biological summaries, representing progress toward practical and interpretable agent workflows for scientific discovery.

Submission Number: 44