Functional similarity of non-coding regions is revealed in phylogenetic average motif score representations

Here we frame the cis-regulatory code (that connects the regulatory functions of non-coding regions, such as promoters and UTRs, to their DNA sequences) as a representation building problem. Representation learning has emerged as a new approach to understand function of DNA and proteins, by projecting sequences into high-dimensional feature spaces, where the features are learned from data by a neural network. Inspired by these approaches, we seek to define a feature space where non-coding regions with similar regulatory functions are nearby each other. As a first attempt, we engineered features based on matches to biochemically characterized regulatory motifs in the DNA sequences of non-coding regions. Remarkably, we found that functionally similar promoters and 3 UTRs could be grouped together in a feature space defined by simple averages of the best match scores in (unaligned) orthologous non-coding regions, which we refer to as phylogenetic average motif scores. Perhaps most important, because this feature space is based on known motifs and not fit to any data, it is fully interpretable and not limited to any particular cell type or experimental context. We find that we can read off known regulatory relationships and evolutionary rewiring from visualizations of phylogenetic average motif score representations, and that predicted regulatory interactions based on neighbors in the feature space are borne out in transcription factor deletion experiments. Phylogenetic averages of match scores to known motifs is a baseline for representation learning applied to non-coding sequences, and may continue to improve as databases of motifs become more complete.