Deep learning uncovers conserved regulatory logic and element dosage underlying stable gene expression in grasses

Gene expression programs are widely conserved across grasses despite rapid evolutionary turnover of cis-regulatory element (CRE) sequences, yet the mechanisms that maintain expression stability remain unclear. In this study, we generated chromatin accessibility (CA) maps and matched transcriptomes for five orthologous tissues across five Poaceae species, and trained deep learning models to predict CA from DNA sequence. Models trained in one species accurately predicted CA in rice, with cross-species correlations frequently exceeding 0.70, indicating that the regulatory logic mapping sequence to CA remains conserved despite extensive non-coding divergence. We developed Basenji-HMM, which combines deep-learning-derived mutational effect profiles and a hidden Markov model to delineate high-impact candidate CREs (cCREs) at nucleotide resolution. Integrating Basenji-HMM with motif scanning and footprinting identified complementary cCRE classes and revealed strong positional and tissue-dependent regulatory architectures. Although only 27%-36% of rice cCREs have detectable sequence orthologs in other grasses, cCRE abundance around orthologous genes is conserved across species and aligns with conserved CA and expression. These results support a "regulatory equivalence" model, wherein compensatory CRE turnover may maintain element dosage under conserved regulatory logic, thereby preserving stable gene expression across grass species.