Expanding the DNA Motif Lexicon of the Transcriptional Regulatory Code

Abstract

Cis-regulatory sequences in metazoan genomes contain elaborate combinations of transcription factor (TF) motifs. Stereospecific arrangements of simple motifs constitute composite elements (CEs) that enhance DNA-protein interaction specificity and enable combinatorial regulatory logic. Despite their importance, CEs remain underexplored. We integrate genome-scale computational discovery (CEseek) with customized massively parallel reporter assays (MPRAs) to uncover cell type-specific and shared CEs, corroborated by coincident and/or cooperative TF binding in vivo and in vitro. A deep learning model, GRACE, trained on MPRA data at single-nucleotide resolution, learns the CE lexicon and predicts activities of variants and alternative stereospecific CE configurations not included in the experimental design. Comparative analysis with models trained on orthogonal features such as chromatin accessibility, demonstrates convergence within the CE lexicon while resolving how individual motifs contribute to transcription versus chromatin accessibility. The expanded lexicon enables CE-aware predictions and interpretation of the impact of disease-associated non-coding variants in. the human genome.