Genomic language models decode bidirectional promoter activity and reveal conserved regulatory patterns across plant species

Bidirectional promoters (BDPs), intergenic regions that coordinate transcription of oppositely oriented genes, pervade eukaryotic genomes, yet sequence signatures remain poorly defined. Existing approaches to BDP identification rely on distance thresholds and cannot infer activity from DNA. Here, we use Evo2, a genomic language model, to predict BDP activity and uncover sequence-encoded regulatory principles across four plant species (Arabidopsis, maize, rice, and wheat). We stratify BDPs by co-expression strength and compare Evo2-derived embeddings against k-mer profiles, transcription factor binding site matrices, and convolutional neural networks trained on sequence. Evo2 embeddings outperform all baselines, demonstrating that genomic language models capture regulatory information underlying BDP activity. Models trained on one or multiple species transfer to held-out species, and in planta validation in Spartina alterniflora confirms predicted high-activity BDPs. Sparse autoencoder analysis links Evo2 embeddings to nucleosome positioning and species-specific transcription factor landscapes, providing insight into sequence determinants underlying BDP activity . This work establishes how genomic language models can extract bidirectional promoter logic from sequence and identify conserved regulatory rules across plants. Bidirectional promoters (BDPs) are intergenic regions that initiate transcription of two oppositely oriented genes. Here, the authors report a framework that employs Evo2 genomic language model embeddings to predict BDP activity directly from raw DNA sequence.