Circularly shifted filters enable data efficient sequence motif inference with neural networks

Motivation Nucleic acids and proteins often have localized sequence motifs that enable highly specific interactions. Due to the biological relevance of sequence motifs, numerous inference methods have been developed. Recently, convolutional neural networks (CNNs) achieved state of the art performance because they can approximate complex motif distributions. These methods were able to learn transcription factor binding sites from ChIP-seq data and to make accurate predictions. However, CNNs learn filters that are difficult to interpret, and networks trained on small data sets often do not generalize optimally to new sequences. Results Here we present circular filters, a novel convolutional architecture, that contains all circularly shifted variants of the same filter. We motivate circular filters by the observation that CNNs frequently learn filters that correspond to shifted and truncated variants of the true motif. Circular filters enable learning of non-truncated motifs and allow easy interpretation of the learned filters. We show that circular filters improve motif inference performance over a wide range of hyperparameters. Furthermore, we show that CNNs with circular filters perform better at inferring transcription factor binding motifs from ChIP-seq data than conventional CNNs. Contact markus.kollmann@hhu.de