A multi-task convolutional deep learning method for HLA allelic imputation and its application to trans-ethnic MHC fine-mapping of type 1 diabetes

Conventional HLA imputation methods drop their performance for infrequent alleles, which reduces reliability of trans-ethnic MHC fine-mapping due to inter-ethnic heterogeneity in allele frequency spectra. We developed DEEP*HLA, a deep learning method for imputing HLA genotypes. Through validation using the Japanese and European HLA reference panels (n = 1,118 and 5,112), DEEP*HLA achieved the highest accuracies in both datasets (0.987 and 0.976) especially for low-frequency and rare alleles. DEEP*HLA was less dependent of distance-dependent linkage disequilibrium decay of the target alleles and might capture the complicated region-wide information. We applied DEEP*HLA to type 1 diabetes GWAS data of BioBank Japan (n = 62,387) and UK Biobank (n = 356,855), and successfully disentangled independently associated class I and II HLA variants with shared risk between diverse populations (the top signal at HLA-DR{beta}1 amino acid position 71; P = 6.2 x 10-119). Our study illustrates a value of deep learning in genotype imputation and trans-ethnic MHC fine-mapping.