DeepAnnotation: A novel interpretable deep learning-based genomic selection model that integrates comprehensive functional annotations.

Abstract

Background: Genomic selection, which leverages genomic information to predict the breeding value of individuals, has dramatically accelerated the improvement of economically important traits. The growing availability of multiomics data in agricultural species offers an unprecedented opportunity to enrich this process with prior biological knowledge. However, fully harnessing these rich data sources for accurate phenotype prediction in genomic selection remains in its early stages.

Results: In this study, we present DeepAnnotation, a novel interpretable genomic selection model designed for phenotype prediction by integrating comprehensive multiomics functional annotations using deep learning. To capture the complex information flow from genotype to phenotype, DeepAnnotation aligns multiomics biological annotations with sequential network layers in a deep learning architecture, mirroring the natural regulatory cascade from genotype to intermediate molecular phenotypes-such as cis-regulatory elements, genes, and gene modules-and ultimately to phenotypes of economic traits. Comparing against 7 classical models (rrBLUP, LightGBM, KAML, BLUP, BayesR, MBLUP, and BayesRC), DeepAnnotation demonstrated significantly superior prediction accuracy (Pearson correlation coefficient increased by 6.4% to 120.0%) and computational efficiency for 3 pork production traits (lean meat percentage, loin muscle depth, and back fat thickness) using a dataset of 1,700 training Duroc boars and 240 independent validation individuals, each genotyped for 11,633,164 single-nucleotide polymorphisms (SNPs), particularly in identifying top-performing individuals. Furthermore, the interpretability embedded within our framework enables the identification of potential causal SNPs and the exploration of their mediated molecular mechanisms underlying trait variation.

Conclusions: DeepAnnotation is an open-source, interpretable deep learning approach for phenotype prediction, leveraging comprehensive multiomics functional annotations. Freely accessible via GitHub and Docker, it provides a valuable tool for researchers and practitioners in genomic selection.