Neural networks with circular filters enable data efficient inference of sequence motifs

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) have achieved state of the art performance. These methods were able to learn transcription factor binding sites from ChIP-seq data, resulting in accurate predictions on test data. However, CNNs typically distribute learned motifs across multiple filters, making them difficult to interpret. Furthermore, networks trained on small datasets often do not generalize well to new sequences.

Here we present circular filters, a novel convolutional architecture, that convolves sequences with circularly permutated 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 full-length motifs and allow easy interpretation of the learned filters. We show that circular filters improve motif inference performance over a wide range of hyperparameters as well as sequence length. Furthermore, we show that CNNs with circular filters in most cases outperform conventional CNNs at inferring DNA binding sites from ChIP-seq data.

Code is available at https://github.com/christopherblum.

Supplementary data are available at Bioinformatics online.