Computational Metabolomics in Disease Biomarker Discovery

Abstract

Metabolomics -- the systematic identification and quantification of small molecules (< 1,500 Da) in biological samples - provides the most direct window into the biochemical state of cells, tissues, and organisms. The metabolome reflects the integrated output of genome, transcriptome, proteome, and environmental exposures, making it ideally suited for disease biomarker discovery, drug response monitoring, and personalised medicine. Yet metabolomics remains computationally challenging: LC-MS and NMR data contain tens of thousands of features per sample, of which only 30-50% can be confidently identified, and disease-relevant signals must be distinguished from the dominant variance driven by age, sex, diet, medication, and circadian rhythms. We present the Computational Metabolomics Biomarker Framework (CMBF), evaluating five computational approaches -- targeted pathway-based analysis, untargeted feature-based machine learning, network-based metabolite prioritisation, deep learning spectral interpretation, and multi-omics metabolomics integration -- across four biomarker discovery tasks (disease diagnosis, disease progression monitoring, treatment response prediction, and prognosis). Our Metabolomics Biomarker Score (MBS) measures biomarker discovery sensitivity, validation reproducibility, mechanistic interpretability, clinical translation potential, and analytical robustness. Deep learning spectral interpretation achieves the highest MBS (0.926) through direct analysis of raw mass spectra without preprocessing information loss, while multi-omics integration achieves the highest mechanistic interpretability (0.960) by linking metabolite signatures to upstream genetic and proteomic drivers.