Self-supervised representations reveal the genetic architecture of human cortical folding

Abstract

Cortical folding emerges during fetal development, is under genetic control, and remains stable throughout life, offering a lasting window into early neurodevelopment. Conventional morphometric descriptors, however, only partially capture the shape variability of cortical folds. Here, we compare 56 regional self-supervised deep learning representations of cortical folds with classical sulcal morphometry via multivariate genome-wide association studies (GWAS) in 35,940 UK Biobank participants. The learned representations identified 567 independent genome-wide significant loci, versus 162 for classical morphometry, 87% of which were also detected by our approach. More than half of these associations replicated in the independent Adolescent Brain Cognitive Development (ABCD) cohort. Gene, gene-set, BrainSpan and single-cell expression enrichment converge on a shared prenatal window of neurogenesis and morphogenesis; spatial gene-association maps recapitulate known regional expression gradients, including for NR2F1. Together, these results establish self-supervised representations of cortical folding as a powerful phenotypic framework for the genetic study of neurodevelopment.