Zero-Shot Metabolite Prediction from Gene Expression via Physics-Informed Graph Neural Networks

Abstract

Predicting metabolite concentrations from gene expression is instrumental for linking regulatory programs to metabolic phenotypes. Prior approaches rely on static enzyme-metabolite mappings and often omit data-driven learning or biochemical constraints, limiting their ability to generalize to new metabolites. We present GAZE (Graph Attention for Zero-shot metabolite Estimation), a physics-informed graph neural network that integrates enzyme expression, Enzyme Commission functional embeddings, and ChemBERTa metabolite descriptors within a unified metabolic graph (5,414 nodes, 16,307 edges). A Metabolite-Conditioned Reader uses each metabolites SMILES embedding to query learned pathway representations, enabling zero-shot prediction with no metabolite-specific parameters. We evaluate GAZE in three scenarios: (i) standard cross-validation on the Cancer Atlas of Metabolic Profiles (18,044 genes, 180 metabolites, 867 cell lines), achieving R2 = 0.816; (ii) leave-one-metabolite-out (LOMO) zero-shot evaluation across 50 held-out metabolites, where the physics-informed variant halves the median R2 deficit relative to a standard GNN baseline (-0.34 vs. -0.72), with 30% of unseen metabolites achieving positive R2; and (iii) external validation on an independent clear cell renal cell carcinoma tissue cohort (220 samples), where GAZE achieves median Spearman{rho} = 0.330 across 214 metabolites without fine-tuning. GAZE outperforms scCellFie, MEBOCOST, and UnitedMet across all evaluation settings.