Hybrid Approach for DNA-Binding Protein Prediction Using Contextual Sequence Modelling

Accurate identification of DNA-binding proteins (DBPs) is essential for understanding gene regulation and protein function. This study proposes a deep learning framework that integrates multimodal feature representations—sequence-based, physicochemical, and protein language model embeddings (ProtBERT, ESM-1b)—to achieve precise DBP prediction. The model integrates amino acid composition, PSSM, and transformer-based embeddings using an attention-based fusion mechanism. A hybrid CNN-BiLSTM-BiGRU backbone combined with a Transformer + SHARP module enabled advanced feature learning and interpretability. Capsule networks were also tested for encoding and reconstruction, though their classification ability was limited. The framework achieved 100% precision, recall, F1-score, and MCC on an independent test set, with training accuracy of 99.02% in the first epoch and validation accuracy stabilizing at 100% within five epochs. DBPs were found to have longer sequences, higher isoelectric points, and greater hydrophilicity than non-DBPs. Training loss dropped from 1.5 to 0.2, and capsule networks reduced reconstruction loss from 0.5 to below 0.1 but produced a modest AUC of 0.52. The integration of multimodal features with deep learning provided robust and interpretable DBP predictions. Attention-based fusion captured complementary information, and the hybrid CNN-BiLSTM-BiGRU-Transformer architecture improved generalization. While capsule networks showed strong feature encoding, their limited classification performance highlighted challenges in DBP separation. This framework offers a state-of-the-art and interpretable approach for DBP prediction, achieving perfect classification metrics and supporting applications in protein annotation, drug discovery, and gene regulation.