Optimized model architectures for deep learning on genomic data

Abstract In computational biology, there is a lack of agreement on the optimal design of deep learning architectures, such as types and number of layers, often resulting in non-optimal design choices. We introduce GenomeNet-Architect , an architectural design framework that researchers can use to optimize deep learning models for genome sequence data. Relative to the best-performing baseline, GenomeNet-Architect reduces the read-level misclassification by 19%, with 32% faster inference and 83% fewer parameters.