A CNN-BiLSTM Model with Multi-Level Feature Fusion for Multifunctional Peptide Prediction

A novel framework for multifunctional peptide prediction is introduced in this study, which extracts features from peptide sequences through multiple perspectives. Firstly, peptide sequences are converted into low-dimensional dense vectors using embedding encoding. Subsequently, the embedded sequences are input into a convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) network independently, allowing for the extraction of both local and temporal features, respectively. The diverse features obtained from these two modules are then fused to comprehensively capture the rich information embedded in peptide sequences. Finally, peptide classification is performed using several fully connected layers. The superior performance of the model is enabled by the proposed feature fusion strategy, with accuracy and absolute true values of 0.621 and 0.556 being achieved, respectively, which surpasses existing methods. This novel feature fusion approach is provided with new insights for the identification and application of multifunctional peptides.