Integration of interpretable multi-features and multi-loss functions for multi-functional therapeutic peptide...

Multi-functional therapeutic peptides (MFTP) play a crucial role in drug development, exhibiting properties such as anti-cancer, anti-inflammatory effects, and more. Artificial intelligence-based predictors have been developed to identify MFTP and achieve satisfactory performance. However, these predictors often heavily rely on latent sequence features, overlooking physicochemical patterns and facing challenges with dataset imbalances. In this study, we propose MFTP-MFML, a model that combines interpretable multiple features and loss functions. Firstly, embedding features with positional information are input to the bi-directional long short-term memory (BiLSTM) network, generating latent representation information while preserving the original sequence information. Secondly, physicochemical attributes are utilized to supplement the amino acid composition and physicochemical properties across different functions of therapeutic peptides, and latent representation information are concatenated with these physicochemical attributes to enhance classification. Thirdly, addressing class imbalances and capturing label correlations, integration loss is employed, incorporating focal loss, binary cross entropy loss, and dice loss. Fourth, to enhance the diversity of functions of therapeutic peptides, MFTP-Mixed-90, a benchmark dataset comprising 27 functions, is constructed. Finally, to evaluate the performance of the model, we compare it with other methods on PrMFTP dataset and MFTP-Mixed-90 dataset. Experimental results demonstrate that MFTP-MFML outperforms existing methods, effectively utilizing integrated features and loss functions. Our code and the datasets are available at https://github.com/wongsing/MFTP-MFML.