Hybrid deep learning framework for enhanced prediction of linear B-cell epitopes using forward-based feature filtering

Linear B-cell epitopes represent specific amino acid sequences on antigens that elicit immune responses and serve as critical targets in vaccine development. Despite their importance, accurate computational prediction remains a challenge due to complex sequence dependencies and pronounced class imbalance in available datasets. This study introduces a novel hybrid deep learning framework that integrates a Convolutional Neural Network (CNN) with a Bidirectional Long Short-Term Memory (BiLSTM) network, supported by a forward filter-based feature selection strategy. The framework utilizes physicochemical descriptors derived from protein sequences and applies ADASYN oversampling to mitigate class imbalance. The proposed model was evaluated on Dengue and Zika virus datasets using three feature subsets (50, 100, and 150 features). The subset of 100 features delivered the best results. On the Zika dataset, the CNN-BiLSTM model achieved 70.64% accuracy, 77.51% sensitivity, 63.67% specificity, and 76.05% AUC. On the Dengue dataset, it obtained 69.96% accuracy, 73.96% sensitivity, 65.65% specificity, and 79.34% AUC. These results surpass the performance of individual CNN and BiLSTM models, as well as existing state-of-the-art approaches such as DeepLBCEPred and LBCE-BERT, demonstrating enhanced discriminative capability. An ablation study further validates the complementary contributions of CNN and BiLSTM components, while feature-wise ROC and Precision–Recall analyses underscore the relevance of selected descriptors. Overall, this work presents a robust and interpretable deep learning pipeline with direct implications for epitope-based vaccine design. Future work will focus on improving cross-virus generalization and incorporating biologically informed interpretability mechanisms.