Multinomial Convolutions for Joint Modeling of Sequence Motifs and Enhancer Activities

Abstract

Massively parallel reporter assays (MPRAs) have enabled the study of transcriptional regulatory mechanisms at an unprecedented scale and with high quantitative resolution. However, this realm lacks models that can discover sequence-specific signals de novo from the data and integrate them in a mechanistic way. We present MuSeAM (Multinomial CNNs for Sequence Activity Modeling), a convolutional neural network that overcomes this gap. MuSeAM utilizes multinomial convolutions that directly model sequence-specific motifs of protein-DNA binding. We demonstrate that MuSeAM fits MPRA data with high accuracy and generalizes over other tasks such as predicting chromatin accessibility and prioritizing potentially functional variants.