AlleleHMM: a data-driven method to identify allele-specific differences in distributed functional genomic marks

How DNA sequence variation influences gene expression remains poorly understood. Diploid organisms have two homologous copies of their DNA sequence in the same nucleus, providing a rich source of information about how genetic variation affects a wealth of biochemical processes. However, few computational methods have been developed to discover allele-specific differences in functional genomic data. Existing methods either treat each SNP independently, limiting statistical power, or combine SNPs across gene annotations, preventing the discovery of allele specific differences in unexpected genomic regions. Here we introduce AlleleHMM, a new computational method to identify blocks of neighboring SNPs that share similar allele-specific differences in mark abundance. AlleleHMM uses a hidden Markov model to divide the genome among three hidden states based on allele frequencies in genomic data: a symmetric state (state S) which shows no difference between alleles, and regions with a higher signal on the maternal (state M) or paternal (state P) allele. AlleleHMM substantially outperformed naive methods using both simulated and real genomic data, particularly when input data had realistic levels of overdispersion. Using PRO-seq data, AlleleHMM identified thousands of allele specific blocks of transcription in both coding and non-coding genomic regions. AlleleHMM is a powerful tool for discovering allele-specific regions in functional genomic datasets.