IMCP-net: An imbalance-aware multi-modal contrastive framework for metabolic pathway prediction on NAFLD

Non-alcoholic fatty liver disease (NAFLD) entails profound metabolic reprogramming across carbohydrate, amino acid, and lipid pathways. Accurate delineation of these metabolic alterations is hindered by severe class imbalance and heterogeneous molecular representations. To address this challenge, an imbalance-aware multimodal contrastive framework, IMCP-Net, is proposed to integrate SMILES-derived chemical descriptors with molecular graph-based structural features via cross-modal attention. A dual contrastive module-combining class-frequency weighting with hard-sample margin loss-enhances pathway discriminability under imbalanced conditions. Pre-trained on the KEGG compound database and refined through Bayesian optimization, IMCP-Net achieved an accuracy of 94.65%, precision of 76.14% and macro-F1 score of 70.92%, outperforming state-of-the-art baselines. Transfer learning on a clinical NAFLD cohort revealed that moderate data augmentation coupled with imbalance loss functions yield superior generalization. Collectively, IMCP-Net establishes a robust paradigm for imbalance-aware metabolic pathway prediction, offering mechanistic insights into metabolic dysregulation underlying NAFLD.