Dbert2_LR: A deep learning-based model for predicting cis-regulatory elements in crops

Cis-regulatory elements (CREs) are key determinants of gene expression underlying complex agronomic traits. However, accurately identifying CREs in crops with large and highly repetitive genomes (such as upland cotton) remains extremely challenging. To address this, we developed Dbert2_LR, a novel hybrid deep-learning framework that integrates the pretrained genomic foundation model DNABERT-2 with parallel bidirectional RNN and LSTM networks. This architecture efficiently captures both deep contextual dependencies and local sequence patterns in DNA, enabling high-accuracy classification of promoters, enhancers, and non-regulatory sequences. Systematic evaluations on Arabidopsis thaliana and upland cotton show that Dbert2_LR outperforms multiple benchmark models, achieving macro-averaged F1 scores of 0.890 and 0.637, respectively. More importantly, in-silico saturation mutagenesis (ISM) analysis confirmed that the model's decisions strongly depend on known transcription factor binding motifs such as TATA-box, DOF, and E-box, revealing its biological interpretability and helping overcome the “black-box” nature of deep learning. To facilitate broader use, we also developed user-friendly A_cre and C_cre prediction systems. This study provides a powerful new tool for functional annotation of complex crop genomes and lays the foundation for CRE-based molecular breeding design. • Dbert2_LR integrates DNABERT-2 with hybrid recurrent networks to achieve state-of-the-art CRE prediction accuracy in crops. • The model demonstrates robust performance and generalization even in the highly repetitive and complex genome of upland cotton. • In-silico saturation mutagenesis reveals biological grammar and accurately identifies functional editing sites in the rice IPA1 promoter. • User-friendly web platforms (A_cre and C_cre) were developed to facilitate genome-wide functional annotation for breeders.