TabPFN-SHAP metabolic networks reveal chronic liver disease progression and identify a hepatocellular carcinoma screening panel

Abstract The conventional biomarker discovery approach in metabolomics treats metabolites as isolated features, failing to capture their interdependencies. We applied a network-based method using Tabular Prior-Data Fitted Networks (TabPFN) integrated with SHapley Additive exPlanations (SHAP) to derive stage-specific metabolic networks capturing statistical metabolite co-contribution patterns and identify key metabolites. We analyzed 1312 participants across five stages (control, hepatitis, fibrosis, cirrhosis, and hepatocellular carcinoma (HCC)), quantitatively profiling 160 metabolites, and externally validated the results in 721 samples. We compared three approaches: (1) TabPFN-SHAP, (2) conventional statistics with machine learning, and (3) TabPFN-SHAP integrated with literature-validated biomarkers. TabPFN-SHAP–based approaches outperformed the conventional pipeline, revealing distinct stage-specific metabolic signatures with overlapping metabolites between consecutive networks, marking consistently prioritized features. A five-metabolite HCC-screening panel (sebacic acid, alpha-linolenic acid, phenylalanyl-tryptophan, arachidonic acid, and glycocholic acid) was identified. At the screening-oriented cutoff, the panel achieved an AUROC of 0.979 (95% CI: 0.966–0.991) for Control vs HCC and 0.955 (95% CI: 0.936–0.977) for Fibrosis/Cirrhosis vs HCC, with sensitivity of 0.937, specificity ≥0.855, accuracy ≥0.897, and negative predictive value ≥ 0.981 across both clinical scenarios, outperforming all comparative approaches. TabPFN-SHAP provides an interpretable computational framework for metabolite-panel discovery, with demonstrated utility for HCC screening.