Causally-anchored multi-omic deep learning recovers exercise-responsive and ageing-causal genes from human physical activity

Abstract Physical activity is among the most robust epidemiological correlates of reduced mortality and multi-morbidity, yet the molecular mechanisms through which exercise exerts these effects in humans remain incompletely resolved. This study combines causally-anchored multi-omic Mendelian Randomisation (MR) with graph-based deep learning for gene prioritisation in human exercise-ageing biology, using accelerometer-derived vigorous physical activity (VPA) in the UK Biobank as exposure. Combining multi-omic MR across five molecular layers, it asks whether causal inference and deep learning can recover exercise-responsive and potentially ageing-causal genes that were independently identified in prior studies. Enrichment for experimentally exercise-responsive genes was undetectable in the raw MR signal (p = 0.97) yet was recovered by the graph model (p = 0.007, reproducible across all initialisations); and the convergence between VPA MR-anchored and ageing-causal genes (significant on its own at 1.6-fold; p = 0.023) was likewise recovered by the graph model where p-value and effect-size ranking could not. The model further reproduced established acute exercise-responsive immune and lipid-metabolic programmes, supporting its recovery of genuine signal. Extending the prioritised genes to formal causal testing, systematic cis-MR with colocalisation across the eight convergent genes and four ageing outcomes identified cathepsin F ( CTSF ) as causally associated with exceptional longevity, with concordant positive estimates in the protein and LD-clumped expression arms and colocalisation support at the protein level. The contribution is therefore twofold: a model-free, like-for-like convergence between exercise-anchored and ageing-causal genes; and a graph-based method that recovers this convergence, together with exercise-responsive biology, beyond the reach of per-gene MR ranking. Abstract Figure Graphical abstract This study introduces a causally-anchored deep learning approach to integrating multi-omic human data. The analytical pipeline begins with an exome-wide association study (ExWAS) and Mendelian randomisation (MR) with vigorous activity as exposure and omics markers as outcomes. MR serves as a causal anchor to prioritise genes from features derived across five molecular layers: CpG methylation, bulk and single-cell gene expression, protein levels, and glycan traits. These causally-anchored priors are unified within a supervised graph attention network (GAT) that integrates a protein-protein interaction network to learn a systems-level representation of the exercise-ageing axis. The model’s prioritised genes are then tested for statistical overlap with two independently derived reference sets: exercise-responsive genes (from an independent multi-omic human exercise study) and potentially ageing-causal genes (CpGs causal for ageing by epigenome-wide MR, mapped to genes). The analysis demonstrates that the graph model recovers enrichment for both reference sets that is not detectable in the raw MR signal. Created with Biorender.co