A deep learning and large language hybrid workflow for omics interpretation

Profiling molecular panorama from massive omics data identifies regulatory networks in cells but requires mechanistic interpretation and experimental follow up. Here we combine deep learning and large language model reasoning to develop a hybrid workflow for omics interpretation, called LyMOI. LyMOI incorporates GPT-3.5 for biological knowledge reasoning and a large graph model with graph convolutional networks (GCNs). The large graph model integrates evolutionarily conserved protein interactions and uses hierarchical fine-tuning to predict context-specific molecular regulators from multi-omics data. GPT-3.5 then generates machine chain-of-thought (CoT) to mechanistically interpret their roles in biological systems. Focusing on autophagy, LyMOI mechanistically interprets 1.3 TB transcriptomic, proteomic and phosphoproteomic data and expands the knowledge of autophagy regulators. We also show that LyMOI highlights two human oncoproteins, CTSL and FAM98A, for enhancing autophagy upon treatment with disulfiram (DSF), an antitumour agent. Silencing these genes in vitro attenuates DSF-mediated autophagy and suppresses cancer cell proliferation. Strikingly, DSF treatment with Z-FY-CHO, a CTSL-specific inhibitor previously used for preventing SARS-CoV-2 infection, potently inhibits tumour growth in vivo.