Evaluating the Contribution of Genome 3D Folding to Variation in Human Height Using Machine Learning

Genome-wide association studies (GWAS) have identified thousands of variants associated with complex traits, yet the majority lie in noncoding regions, making it difficult to determine their functional impact. Alterations to the three-dimensional (3D) spatial interactions among gene regulatory elements are increasingly recognized as a mechanism by which genetic variants influence gene expression. However, experimentally evaluating whether variants disrupt 3D-genome structure is not feasible at GWAS scale. To address this, we developed a computational framework that integrates GWAS summary statistics with predictions from the Akita sequence-based deep learning model of 3D chromatin contacts. We applied the framework to 9,917 genomic regions associated with human height, assessing both individual variants and haplotypes for their predicted impact on 3D genome architecture. Only a small fraction of height-associated haplotypes had substantial predicted disruption of 3D folding (17 regions, 0.17%, exceeded a disruption score of 0.1). Considering all common variants in a haplotype together generally produced greater perturbations than individual variants, but several highly divergent regions were driven by single variants. We highlight a variant that disrupts the binding motif at a confirmed CTCF binding site and is predicted to modify 3D genome contacts with the LCOR promoter, suggesting that 3D-genome-mediated disruption of gene regulation underlies the association with height. This work presents a scalable and interpretable strategy for integrating 3D genome modeling with GWAS, enabling investigation of this important regulatory mechanism in the connection of non-coding genetic variation to complex traits.