Prompt-CBP: A Novel Prompt Learning-Based Model for Predicting Cross-Species Promoters

Promoter sequences across species exhibit both specificity and conservation. The specificity of promoter sequences typically leads to lower cross-species prediction performance compared to within-species prediction. However, conserved promoter motifs such as the TATA box, Inr, BRE, and DPE are crucial for predicting promoters across species. This study introduces a novel prompt learning-based model, Prompt-CBP, for cross-species promoter prediction. First, a prompt generator was employed to create targeted prompts based on conserved promoter features, which were combined with the promoter sequence to form a sequence pair. Second, DNABERT-2 was used to encode the sequence pair. Third, CNN was applied to extract local sequence features, while BiLSTM captured global features. Finally, a fully connected layer was utilized for binary classification at the output stage. During prediction, prompts guide the model to focus on and capture conserved features shared across promoter sequences, minimizing the influence of species-specific features. The experiment utilized promoter data sets from two prokaryotes and three eukaryotes for ″leave-one-species-out cross-validation″. The results demonstrated that, compared to methods without prompt learning, our approach improved the average ACC by 0.3213 and the average AUC by 0.4137. Additionally, with smaller sample sizes, its prediction performance was comparable to models trained on extensive promoter data sets.