PeptideSGCL: Structure-Enhanced Graph-Transformer Encoding and Dual-Level Contrastive Learning for Peptide Property...

Area of Science:

• Biochemistry
• Computational Biology
• Materials Science

Background:

• Peptides are crucial in biology and medicine, but their hemolytic (Hemo) and nonfouling (NF) properties impact safety and application.
• Accurate prediction of these properties is vital for designing effective peptides.
• Existing multimodal methods struggle with long-range structural dependencies and intra-modal feature discrimination.

Purpose of the Study:

• To develop an advanced multimodal dual-contrastive learning framework for peptide property prediction.
• To improve the modeling of both sequence and structural information for peptides.
• To enhance the accuracy of predicting hemolytic and nonfouling properties.

Main Methods:

• Utilized ProtBERT as the sequence encoder.
• Developed a hierarchical GNN-Transformer for structural encoding, capturing local and long-range dependencies.
• Implemented a parallel graph spatial channel attention module to refine structural features.
• Employed an inter-intra hybrid supervised contrastive learning strategy for joint optimization.

Main Results:

• The proposed framework significantly outperformed baseline models in predicting both hemolysis and nonfouling properties.
• The enhanced structural encoder effectively captured complex structural patterns.
• The dual-contrastive learning strategy improved the quality of joint sequence-structure representations.

Conclusions:

• The developed multimodal dual-contrastive learning framework offers a superior approach for peptide property prediction.
• This method provides an effective tool for the rational design of functional peptides with desired safety profiles.
• The study highlights the potential of advanced representation learning for peptide-based biomedical applications.