PIPA: An Agent for Protein Interaction Identification and Perturbation Analysis

ICLR 2026 Conference Submission10885 Authors

Keywords: Scientific discovery agent, Cellular organelle, Protein-protein interaction, Large language model

Abstract: Protein–protein interactions (PPIs) play a fundamental role in the functioning of proteins and the formation of cellular pathways. Given their implication in numerous disease processes, PPIs represent important targets for therapeutic intervention. To enhance the efficacy and efficiency of PPI identification, pathway analysis, and disease target screening, we present PIPA, a tool-augmented intelligent agent that automates the entire PPI-centric discovery pipeline through a structured Plan-Select-Execute-Reflect loop, to assist biological researchers in discovering novel PPIs and predicting the effects caused by pathological mutations. PIPA integrates multiple tasks, including automated database retrieval, protein interaction identification, pathological mutation annotation, and interface perturbation prediction. To test its scientific discovery capabilities, PIPA was used to investigate the interaction between the proteins of the endoplasmic reticulum (ER) and Dynamin 2 (DNM2). The identified hypo-PPI with dynamin mutation, a potential target for therapeutic intervention of centronuclear myopathy, was experimentally validated in the wet lab. Furthermore, PIPA autonomously explored numerous PPI and mutation combinations to elucidate disease-related alterations in protein–protein interactions at the proteome scale. The scientific data used in this study has been developed into a benchmark named ER-MITO. As a novel framework for quantitatively evaluating agent capabilities on biologically meaningful tasks, ER-MITO benchmark moves beyond simplistic QA metrics to assess an agent's skill in (i) predicting novel inter-organelle PPIs and (ii) accurately retrieving and associating pathogenic mutations with proteins, further facilitating the evaluation of other AI systems and agents in terms of their capabilities for scientific discovery.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 10885