AI-augmented discovery of autophagy regulators through mechanistic interpretation of multiomics data

Large-scale multiomics profiling has delineated dynamic molecular landscapes during autophagy, yet translating these complex datasets into mechanistic regulatory insights remains a major challenge. In our recent work, we developed LyMOI, a hybrid artificial intelligence workflow that combines graph-based deep learning and a large language model (LLM) for mechanistic interpretation of autophagy-related omics. The graph model integrates 1.3 TB of autophagy-associated multiomics datasets and prioritizes molecules of interest (MOIs) across 34 autophagy-specific conditions, and then LLM-based chain-of-thought (CoT) reasoning generates mechanistic hypotheses to interpret their potential roles in biological contexts. Using LyMOI, we identified essential regulators, including GIN4, ELM1, RVS167 and STE50, involved in yeast autophagy induced by nutrient deprivation. Furthermore, LyMOI revealed that two cancer-associated proteins, CTSL and FAM98A, are required for maintaining autophagy activity upon disulfiram (DSF) treatment. Silencing either CTSL or FAM98A attenuated DSF-induced autophagy and inhibited cancer cell proliferation. Notably, combination treatment with DSF and Z-FY-CHO, a CTSL-specific inhibitor previously developed against SARS-CoV-2 infection, potently suppressed tumor growth. Collectively, our work presents an LLM-powered platform with biologist-like reasoning for uncovering autophagy regulatory mechanisms.Abbreviations: ARGs: autophagy-regulating genes; ATGs: autophagy-related genes; BECN1: Beclin 1; CoT: chain-of-thought; CQ: chloroquine; CTSL: cathepsin L; DSF: disulfiram; GCNs: graph convolutional networks; GI: graph inference; IHC: immunohistochemistry; LC3-II: lipidated LC3; LGM: large graph model; LLM: large language model; MOIs: molecules of interest; NATs: normal adjacent tissues; SQSTM1/p62: sequestosome 1; 3-MA: 3-methyladenine.