Many common diseases, including cancer, cardiovascular and neurological diseases, arise from multi-scale biology spanning genes, cells, tissues and organs. Despite advances in omics and imaging, analyses are often siloed by modality, obscuring cross-organ interactions. Foundation models: large artificial intelligence (AI) systems trained on heterogeneous data, have been useful in biology but are usually single-modality. We will use Deep Latent Variable Path Modelling (DLVPM) to integrate raw, multi-organ magnetic-resonance imaging (MRI) with variant-level genotypes (single-nucleotide variants, SNVs) and environmental covariates (alcohol, diet, activity, deprivation). In DLVPM, each data type is a node in a directed graph; deep networks learn latent factors directly from raw data and the cross-modal links between them. The trained representation is a reusable multimodal foundation model that we will deploy in two ways: (1) Outcome prediction at scale: predicting all major clinical outcomes, including cancers, cardiovascular events, stroke, dementia, diabetes, and mortality; quantifying gains for incident risk beyond established scores. (2) Mechanistic investigation: interrogating the learned graph to identify cross-organ pathways and indirect/mediated effects. For example, we will test whether factors learned from raw abdominal, cardiac and brain MRI reveal a liver-heart-brain comorbidity axis and how genetic burden and environmental exposures load on it; we will then relate these factors to incident outcomes. For example, we will ask whether liver steatosis/inflammation factors precede and mediate cardiac remodelling (reduced left-ventricular ejection fraction) and greater cerebral small-vessel disease (elevated white-matter hyperintensity load), increasing risk of incident heart failure, stroke and dementia; interpretability will draw on attention/saliency (cross-modal attention, integrated gradients/SHAP) alongside formal mediation.
