Multi-omic deep learning identifies exercise-responsive ageing pathways in humans

Genome-wide association studies of physical activity traits have mapped numerous loci, yet the molecular mechanisms through which exercise influences human biology remain poorly defined. Mechanistic progress has been limited by heritability-dominated signals, siloed single-omic analyses, and the lack of integrative models that connect genetic associations to causal, system-level pathways. We introduce the first deep learning, multi-omic framework for exercise genomics, unifying causal inference, molecular topology, protein structure, and functional context within a supervised Graph Neural Network (GNN) with experimental validation. Using linkage disequilibrium-aware Mendelian randomisation with accelerometer-derived vigorous physical activity as the exposure, we integrated four omic layers--plasma proteomics, blood CpG methylation, blood single-cell transcriptomics, and plasma glycomics. The GNN prioritised a coherent, multi-omic network of exercise-responsive genes spanning glycosylation and immunity, apoptosis/stress signalling, growth and transcription factors, proteostasis/autophagy, metabolism, oxidative stress/redox, mitochondrial/oxidative phosphorylation, chromatin/epigenetic, translation/ribosome, RNA splicing/processing, DNA damage and repair, cell cycle, and cytoskeleton/ECM functions. Pathway annotation and alignment with ageing biology using differentially methylated regions and principal component features from five biological ageing clocks (Horvath, Hannum, DunedinPACE, PhenoAge, Proteomic Clock), used as contextual markers of ageing-related regulation, mapped to: 1) autophagy (Fas signalling); 2) stress and DNA damage response pathways (Hypoxia-Inducible Factor, apoptosis, p53, p38, PI3K, Ras, oxidative stress, DNA replication, purine/pyrimidine metabolism and biosynthesis, pentose phosphate, Rho GTPase, G protein, ubiquitin proteasome, Epidermal Growth Factor Receptor); 3) immune and inflammatory pathways (cytokine and interleukin signalling, T and B cell activation, Toll-like receptor signalling); 4) physiological adaptation (VEGF, Wnt, GnRH, Fibroblast Growth Factor, Platelet-Derived Growth Factor, Thyrotropin-Releasing Hormone, glutamate receptor, plasminogen, endothelin, heme biosynthesis, cholecystokinin, Corticotropin-Releasing Hormone signalling); 5) neuroendocrine and neurotransmitter pathways (acetylcholine, dopamine, oxytocin, opioid, {beta}-adrenergic signalling); 6) neurodegeneration pathways (Alzheimers, Parkinsons, Huntingtons disease); 7) protein metabolism (leucine, isoleucine, valine biosynthesis); and exercise-responsive epigenetic regulatory pathways (S-adenosylmethionine, thiamine, vitamin D biosynthesis, bZIP transcription, and circadian rhythm). Finally, we partially validated GNN predictions in humans, demonstrating acute exercise-induced shifts in plasma glycomic markers consistent with GNN-predicted glycomic remodelling. This work establishes a deep learning, multi-omic map of exercise-responsive pathways in humans and identifies actionable regulators that couple habitual vigorous physical activity to stress-resilient immunometabolic regulation and healthy ageing trajectories.