A CONVERSATIONAL MULTI-AGENT AI FRAMEWORK FOR INTEGRATED MULTI-OMICS ANALYSIS AND BIOMEDICAL DISCOVERY

Pankaj Rajdeo, Shunya Asanuma, Michal Kouril, Peixin Lu, Jichao Chen, Aman Chadha, Surya Prasath, Bruce J Aronow, Nathan Salomonis

Abstract: Single-cell and spatial transcriptomics offer unprecedented opportunities to decipher the mechanisms of disease and tissue development, however, this process requires teams of experts, iterative trial-and-error and reasoning across modalities. Here we present a conversational agentic framework for integrated multi-omics reasoning and biomedical discovery. It is deployed as a hierarchical multi-agent architecture in which a stateful supervisor decomposes natural-language research questions into parallel, tool-grounded tasks executed by specialized agents for single-cell analysis, spatial transcriptomics, literature synthesis, drug repurposing, and functional enrichment. A virtual filesystem enables efficient inter-agent communication while preventing context degradation, and the framework maintains long-term memory for personalized analytical context across iterative sessions. For therapeutic hypothesis generation, the framework includes Direction-Aware Repurposing and Targeting (DART), a method that distinguishes perturbations that reverse disease transcriptional programs from those that reinforce or disrupt them, enabling cell-type–resolved therapeutic prioritization and safety profiling. Applied to 39 datasets, spanning fetal development, tissue homeostasis and 15 pathological conditions, the system generates publication-ready results with complete parameter provenance for reproducibility. We demonstrate autonomous discovery that identifies known IPF disease drivers, cell-communication networks, and validated FDA-approved drugs. Notably, DART predicts Saracatinib (SRC inhibitor) as a top therapeutic candidate, now in Phase 1b/2a trials while identifying cell-type-specific safety risks and enables efficacy and safety profiling at single-cell resolution. This architecture is tissue-agnostic, offering a blueprint for agentic AI systems that integrate foundation models with executable, verifiable scientific discovery. We instantiate this framework as LungChat, deployed for lung biology under the NHLBI LungMAP Consortium (https://chat.lungmap.net).

Submission Number: 94