Depthwise Convolutions using Physicochemical Features of DNA for Transcription Factor Binding Site Classification: Physicochemical Features for DNA-Protein Classification with Depthwise Convolutions

Abstract

Classifying DNA sequences based on a nucleotide representation has enjoyed considerable success with the advancement of Deep Learning, as the proper usage and combination of different layers and architecture choices led to an increase in performance. The most common approaches rely on convolutional, recurrent and attention layer types. Moreover, the inclusion of further information in addition to the nucleotide sequence provides increases in performance, even though the methods of combining the input feature representations with distinct model structures could pose a challenge. To examine this topic, we applied depthwise separable convolutional layers to a physicochemical DNA sequence representation and train models to detect the binding sites of DNA binding proteins. While convolutional kernels learn the local feature patterns of motifs, the behaviour of the depthwise separable convolution better exploits the feature, shape and physicochemical information that could be stored in the input representation. Our models with depthwise separable convolution achieve increases in accuracy compared to the convolutional and nucleotide-based approaches on several datasets.