MP-MIENN: predicting multifunctional peptides by neural network with multiangle initialization embedding

Functional peptide has been utilized widely in the treatment of disease due to its high absorption rate, low toxicity, and biological activity, making it a possible substitute for traditional antibiotic medications in the biomedical field. A number of machine learning methods have been developed recently for the prediction of functional peptides. However, few studies take into account multifunctional peptide identification, and the majority mainly depends on statistical features. Therefore, in the imbalanced multi-label functional peptide datasets, we propose a novel predictor, MP-MIENN. Firstly, we use physicochemical and evolutionary information to describe the peptide sequence's initiation features from different perspectives. Secondly, to extract more discriminative features of peptide sequences of varying lengths, the features are combined and then fed into a deep neural network. Ultimately, a novel loss function is developed to substitute for the traditional cross entropy loss function in order to handle the class imbalance issues. The results demonstrate that our approach improves accuracy over existing approaches by 3.89% on publicly available peptide datasets, while significantly improving the model's efficacious capacity to capture sequence information.