PepLM-GNN: A graph neural network framework leveraging pre-trained language models for peptide-protein binding...

Area of Science:

• Biochemistry
• Computational Biology
• Bioinformatics

Background:

• Accurate prediction of peptide-protein interactions (PepPI) is crucial for peptide drug research and understanding biomolecular regulation.
• Existing computational methods struggle with the non-Euclidean nature of PepPI data and exhibit poor generalizability, especially in cold-start scenarios.

Purpose of the Study:

• To develop a novel computational framework, PepLM-GNN, for accurate and generalizable prediction of peptide-protein interactions.
• To address the limitations of existing methods in handling non-Euclidean data and cold-start scenarios in PepPI prediction.

Main Methods:

• Integrated a pre-trained language model (ProtT5) with a hybrid graph neural network (GNN).
• Constructed graphs using ProtT5-extracted semantic features for peptides and proteins as heterogeneous nodes.
• Employed Graph Convolutional Networks (GCN) for local feature aggregation and Graph Isomorphism Networks (GIN) for global interaction capture, effectively handling non-Euclidean data.

Main Results:

• PepLM-GNN demonstrated highly accurate and robust performance in PepPI prediction compared to existing advanced methods.
• The framework effectively addressed the cold-start problem by enhancing generalizability for novel peptides, proteins, and binding pairs.
• Successfully applied PepLM-GNN to virtual peptide drug screening, showcasing its potential for drug discovery.

Conclusions:

• PepLM-GNN offers a powerful and accurate approach for predicting peptide-protein interactions.
• The framework's ability to handle non-Euclidean data and improve generalizability makes it valuable for future peptide drug development and functional proteomics research.