Identification of Multi-functional Therapeutic Peptides Based on Prototypical Supervised Contrastive Learning.

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• High-throughput sequencing generates vast peptide data, requiring efficient computational methods for identifying multi-functional therapeutic peptides (MFTP).
• Existing methods struggle with class imbalance and effective sequence representation learning for MFTP prediction.

Purpose of the Study:

• To propose PSCFA, a novel computational framework for accurate MFTP prediction.
• To address challenges in sequence representation and class imbalance in therapeutic peptide identification.

Main Methods:

• Utilized a two-stage training scheme: prototypical supervised contrastive learning for feature extraction and feature augmentation for classifier refinement.
• Employed a prototype-based variational autoencoder to transfer knowledge from common labels to infrequent labels, enhancing feature representation.
• Focused feature augmentation on infrequent (tail) labels to improve classifier performance.

Main Results:

• PSCFA demonstrated significantly superior performance compared to existing methods in MFTP prediction.
• The method effectively enhanced feature space distribution uniformity and improved learning for underrepresented peptide classes.
• Achieved a significant advancement in the computational identification of therapeutic peptides.

Conclusions:

• PSCFA offers a robust and effective computational solution for identifying multi-functional therapeutic peptides.
• The proposed method advances the field of therapeutic peptide discovery by overcoming key challenges in machine learning approaches.
• This work paves the way for more accurate and efficient identification of novel therapeutic peptides from sequence data.