OncoReasoner: An Interpretable Regulatory Network Inference Framework for HPV E6/E7-Induced Transcriptomic Perturbations Leveraging Large Language Models

Human papillomavirus (HPV) E6/E7 oncoproteins perturb host gene regulatory networks, driving oncogenesis. Existing computational methods often struggle to provide interpretable, chain-of-thought mechanistic explanations for observed transcriptomic changes. To address this, we introduce OncoReasoner, a novel framework that integrates biological expression analysis with the advanced reasoning capabilities of large language models (LLMs) and graph neural networks (GNNs). OncoReasoner comprises an Expression Encoder for rich gene embeddings, a Bio-LLM Reasoning Module for context-aware mechanistic explanations, and a Graph Refinement Module leveraging GNNs and prior knowledge for network consistency. Evaluated on diverse datasets, including GEO and TCGA, our framework significantly outperforms traditional statistical methods, GNNs, and other LLM baselines across differential gene expression classification, regulatory network edge prediction, and particularly, functional pathway reasoning. OncoReasoner notably achieves high accuracy in pathway identification and receives excellent expert ratings for its mechanistic explanations, demonstrating its superior ability to provide deep, accurate, and highly interpretable biological insights. An ablation study confirms the critical contribution of each module, and human evaluation further validates the qualitative excellence of its mechanistic explanations, marking a substantial advancement in explainable AI for cancer research.