BioHybridNet: An Interpretable Hybrid Deep Learning Model for Genomic Prediction Based on Biological Pathways

Abstract

The feasibility of deep learning in genomic prediction is hindered by the lack of biological
insight and knowledge of the models, and the traditional linear models are incapable of
finding all possible genetic interactions. To this end, we designed BioHybridNet, a hybrid
system consisting of biological pathways and GWAS-directed attention incorporated into an
interpretable model. We have a dynamic gating model that mixes linear and non-linear
prediction and is able to perform federated learning in a privacy-preserving manner. On the
UK Biobank and wheat genomic data, BioHybridNet had a mean R 2 increase of +14.5
percent over linear models and +8.2 percent over deep learning models, and recovered 92
percent of known disease loci a 34 percent increase in interpretability and offered novel
information such as quantifying the epistatic nature of Schizophrenia. This research has
been able to balance the accuracy and interpretability of genomic prediction, and future
studies aim at the integration of multi-omics and clinical translation.