ABSTRACT
The bendability of genomic DNA impacts chromatin packaging and protein-DNA binding. However, we do not have a comprehensive understanding of the motifs influencing DNA bendability. Recent high-throughput technologies such as Loop-Seq offer an opportunity to address this gap but the lack of accurate and interpretable machine learning models still remains. Here we introduce DeepBend, a convolutional neural network model with convolutions designed to directly capture the motifs underlying DNA bendability and their periodic occurrences or relative arrangements that modulate bendability. DeepBend consistently performs on par with alternative models while giving an extra edge through mechanistic interpretations. Besides confirming the known motifs of DNA bendability, DeepBend also revealed several novel motifs and showed how the spatial patterns of motif occurrences influence bendability. DeepBend's genome-wide prediction of bendability further showed how bendability is linked to chromatin conformation and revealed the motifs controlling the bendability of topologically associated domains and their boundaries.
ABSTRACT
A common goal in the convolutional neural network (CNN) modeling of genomic data is to discover specific sequence motifs. Post hoc analysis methods aid in this task but are dependent on parameters whose optimal values are unclear and applying the discovered motifs to new genomic data is not straightforward. As an alternative, we propose to learn convolutions as multinomial distributions, thus streamlining interpretable motif discovery with CNN model fitting. We developed MuSeAM (Multinomial CNNs for Sequence Activity Modeling) by implementing multinomial convolutions in a CNN model. Through benchmarking, we demonstrate the efficacy of MuSeAM in accurately modeling genomic data while fitting multinomial convolutions that recapitulate known transcription factor motifs.
