A foundation model for mapping the phenomic and genetic landscape of cerebral small vessel disease biomarkers

Cerebral small vessel disease (CSVD) is a leading cause of age-related cognitive decline and neurological disorders, yet its precise characterization in large populations has been constrained by reliance on subjective neuroimaging ratings. To address this, we developed CSVDtransformer, a foundation model that simultaneously quantifies six key CSVD biomarkers from structural brain MRI. In 3,718 subjects, the model achieved excellent accuracy (mean AUC = 0.904) in measuring periventricular and deep white matter hyperintensities, Fazekas scores, enlarged perivascular spaces, lacunar infarcts, and cerebral microbleeds. Validation across two independent, external datasets (N=568) confirmed its robust generalizability. As a clinical decision-support tool, it augmented neurologist assessment relative accuracy by 20%. Application to 59,772 UK Biobank participants revealed distinct associations of these quantified biomarkers with incident stroke, dementia, and psychiatric disorders. Large-scale multi-omics analysis identified 1,365 significant plasma protein correlates and 14 novel genetic loci for these CSVD biomarkers. These associations implicate pathways of endothelial dysfunction, inflammation, and lipid metabolism. Mendelian randomization analyses provided evidence for causal relationships between specific vascular-metabolic proteins and CSVD biomarkers, such as positive effect of EFEMP1 and negative effect of EPO on CSVD. Furthermore, drug-target enrichment analysis highlighted the potential for targeting TFPI and EPO to address vascular dysfunction associated with CSVD. Our study establishes CSVDtransformer as a scalable foundation model that deciphers the complex systemic biology of cerebral microvascular health.