Deep learning-based multi-functional therapeutic peptides prediction with a multi-label focal dice loss function.

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• The postgenomic era yields numerous peptide sequences, necessitating efficient methods for function identification.
• Predicting the functions of therapeutic peptides, especially multi-functional ones (MFTP), using sequence-based computational tools remains a significant challenge.

Purpose of the Study:

• To develop a novel computational method for predicting 21 categories of therapeutic peptides.
• To address the challenge of accurately predicting multi-functional therapeutic peptides (MFTP) from sequence data.

Main Methods:

• A deep learning model architecture (ETFC) comprising embedding, text convolutional neural network, feed-forward network, and classification blocks.
• Implementation of an imbalanced learning strategy using a novel multi-label focal dice loss function to handle dataset imbalance.
• Application of teacher-student-based knowledge distillation to analyze attention weights from a self-attention mechanism.

Main Results:

• The ETFC method demonstrated significantly superior performance in MFTP prediction compared to existing approaches.
• The novel multi-label focal dice loss effectively addressed the inherent imbalance in the multi-label dataset.
• The study successfully quantified the contributions of different activities toward MFTP prediction.

Conclusions:

• The developed ETFC method offers a robust and accurate solution for predicting therapeutic peptide functions.
• The findings advance the field of computational drug discovery by providing a powerful tool for identifying potential therapeutic peptides.
• The source code and dataset are publicly available, facilitating further research and application.