A historical journey of metabolite-protein interaction discovery: from data harmonization to AI-driven prediction.

Metabolites are life-sustaining small molecules produced by living organisms. They interact with proteins involved in metabolism, signalling, and gene regulation, called metabolite-protein interactions (MPIs). This review traces the history of MPI research, from curated resources and early cheminformatics to harmonized identifiers, proteome-scale structural models, and artificial intelligence-driven prediction, while highlighting persistent challenges that continue to limit mechanistic interpretation of metabolomics. Early small-molecule-protein interaction prediction tools (e.g. Molpat and Catalyst) and resources (e.g. ChEMBL and BindingDB) were typically biased towards drug-like molecules. As drug-centred research continued, a revolution in large-scale metabolomics enabled high-throughput profiling of metabolite levels across physiological and disease states. However, these advances also introduced major data integration challenges such as data fragmentation, unresolved metabolite identities, and limited physiological context. Subsequent metabolite-centric resources (e.g. HMDB) and high-throughput screens applied to MPI detection (e.g. thermal proteome profiling) have partially addressed this bias. Proteome-scale structure prediction (e.g. AlphaFold) has further incentivized research into the effects of metabolites on protein structure and function. Nevertheless, the complexity of the biological response also depends on, e.g. exposure, access, and target expression. Looking ahead, MPI research is likely to be shaped by structure-aware deep learning and the integration of MPIs with comprehensive single-cell multi-omics data and host-microbe modelling. These advances may turn metabolomic signals into causal, testable hypotheses, enabling robust systems-level MPI maps for identifying intervention points and designing new treatments. We propose a historically structured roadmap centred on standards-driven data integration and calibrated, structure-aware modelling to support mechanistic, systems-level MPI maps.