Abstract
Deciphering how DNA sequence specifies cell-type-specific regulatory activity is a central challenge in gene regulation. We present Bag-of-Motifs (BOM), a computational framework that represents distal cis-regulatory elements as unordered counts of transcription factor (TF) motifs. This minimalist representation, combined with gradient-boosted trees, enables the accurate prediction of cell-type-specific enhancers across mouse, human, zebrafish, and Arabidopsis datasets. Despite its simplicity, BOM outperforms more complex deep-learning models while using fewer parameters. We validate BOM's predictions experimentally by constructing synthetic enhancers from the most predictive motifs, demonstrating that these motif sets drive cell-type-specific expression. By providing direct interpretability and broad applicability, BOM reveals a highly predictive sequence code at distal regulatory regions and offers a scalable framework for dissecting cis-regulatory grammar across diverse species and conditions.