Integrating Multi-View Residue Graph and Protein Language Model for Cell-Penetrating Peptide Prediction via...

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Drug Delivery Systems

Background:

• Cell-penetrating peptides (CPPs) are crucial for intracellular delivery but their discovery is hindered by high costs and slow experimental processes.
• Current computational methods for CPP prediction often lack biophysical interpretability and rely on simplistic sequence features.

Purpose of the Study:

• To develop an accurate, interpretable, and scalable computational tool for pre-screening and discovering novel cell-penetrating peptides.
• To integrate biophysical principles with advanced deep learning techniques for improved CPP prediction.

Main Methods:

• Developed DeepCPP, a dual-branch framework combining a biophysically informed residue graph with ESM-2 protein language model embeddings.
• Employed graph neural networks with global-local aggregation and cross-attention for feature fusion, enhanced by a Kolmogorov-Arnold Network (KAN) head.
• Curated a new benchmark dataset from CPPsite3 with cluster-controlled splits to ensure reliable generalization.

Main Results:

• DeepCPP significantly outperformed existing state-of-the-art methods in both CPP and peptide-function prediction tasks.
• Interpretability analyses revealed key patterns such as charge clustering, termini characteristics, and residue connectivity related to amphipathicity.
• The model demonstrated strong predictive accuracy and generalization capabilities on the curated benchmark.

Conclusions:

• DeepCPP offers a powerful and interpretable computational approach for accelerating the discovery of cell-penetrating peptides.
• The findings provide actionable insights into the structural determinants of CPP activity, guiding rational peptide design.
• This framework represents a significant advancement in computational drug delivery and peptide engineering.