Sequence-based chromatin activity modeling and regulatory impact prediction of genetic variants in farmed animals using deep learning

Abstract

Noncoding genomic variations are crucial for the genetic regulation of traits; however, their functional impact in farmed animals remains underexplored due to limited genomic resources and the absence of tailored computational tools. Here, we present a deep learning-based framework that utilizes functional genomic data to generate genome-wide predictions of the regulatory impact of noncoding variants in cattle, chicken, pig, and Atlantic salmon. By leveraging chromatin profiles such as assay for transposase-accessible chromatin, DNase I hypersensitive site, and chromatin immunoprecipitation sequencing data, we train and optimize separate deep networks for each species, achieving robust sequence modeling accuracy specific to each. Motif analysis confirms that the models capture regulatory grammar, while in silico saturation mutagenesis experiments provide meaningful interpretations of the functional impact of putative causal variants. Furthermore, functional scores derived from these models predict expression quantitative trait loci causal variants and enhance genomic prediction performance. Our findings highlight the transformative potential of sequence-to-function models in prioritizing causal variants and improving genomic prediction for livestock and aquaculture animals.