CBS: a hybrid deep-learning framework and online platform for plant promoter activity prediction and candidate screening

Promoter sequences play central roles in regulating gene expression, and accurate prediction of promoter activity is important for crop improvement, synthetic regulatory element design, and gene-circuit optimization. However, promoter activity is shaped by local cis-regulatory motifs, contextual dependencies, and sequence-wide combinatorial effects within standardized promoter windows. Moreover, domain shifts across species and assay systems (leaf and protoplast) make robust regression challenging. Here we present CBS, a hybrid deep-learning framework that performs end-to-end regression on 170-bp promoter sequences while integrating GC content, species identity, and strand orientation as auxiliary inputs. On the leaf test set, CBS achieves R² = 0.7638, RMSE = 0.8398, and Pearson r = 0.8790; on the protoplast test set, it achieves R² = 0.7367, RMSE = 0.6208, and Pearson r = 0.8673, outperforming the baseline models evaluated under the same split and training protocol. In an additional comparison with the pretrained sequence model DNABERT2, CBS also remained superior on both datasets. DNABERT2 achieved R² = 0.7563 on Leaf and 0.6795 on Protoplast, but did not show lower overall prediction error than CBS. Ablation studies show that BiLSTM provides the dominant sequential backbone, whereas the SSM module and attention mechanism deliver consistent incremental gains across both datasets. Finally, we deploy CBS as an online platform, PlantCRE, offering single-sequence and batch prediction with interactive visualization to facilitate promoter activity estimation, candidate comparison, and downstream planning. By supporting candidate comparison and batch-level screening, PlantCRE provides a practical online workflow for data-driven regulatory element selection in crop improvement. The present study demonstrates robust performance within the benchmarked species and assay settings; zero-shot transfer to unseen species remains future work.