StrongGeneAgent: An Evidence-Grounded Conversational Agent for Gene-Centric Pan-Cancer Multi-Omics Analysis

Background: Single-gene analysis remains a common entry point for investigating cancer biology, yet integrating evidence for a target gene across molecular layers, tumor types, and clinical endpoints remains fragmented across specialized tools and workflows. Here, we present StrongGeneAgent, an evidence-grounded conversational framework for natural-language, gene-centric multi-omics analysis across cancer types. Methods: StrongGeneAgent organizes 64 standardized analytical workflows across 33 TCGA cancer types, spanning transcriptomic, epigenomic, genomic, proteomic, pathway, immune-microenvironment, and pharmacogenomic evidence. Open-ended queries are resolved into molecular entities, cancer contexts, and analytical intents; decomposed when necessary; and routed to compatible deterministic analyses. Quantitative findings are retrieved from structured analytical records rather than generated by the language model, while matched visualizations are produced from the same evidence source, preserving correspondence between questions, estimates, and figures. Results: StrongGeneAgent reproduced the underlying analytical findings in 198 of 200 benchmark questions and maintained high fidelity across repeated, composite, invalid-input, and multi-turn queries. Cross-layer interrogation of CCNB1 in hepatocellular carcinoma demonstrated the integration of expression, clinicopathological variation, survival, protein abundance, and pathway activity around a single molecular target. In a ten-participant usability study, all 50 predefined tasks were completed, with a mean System Usability Scale score of 86.25. Conclusions: StrongGeneAgent provides a scalable interface for transforming single-gene cancer questions into structured, traceable, and interpretable multi-omics evidence across tumor types.