PLNet: Persistent Laplacian neural network for protein-protein binding free energy prediction

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Topology-based modeling advances molecular prediction, especially protein-ligand binding.
• Predicting protein-protein interactions (PPIs) binding free energy is challenging due to ineffective topological feature use and data limitations.

Purpose of the Study:

• To introduce a novel machine learning framework for predicting PPIs binding free energy.
• To address limitations in current methods for capturing topological information in PPI prediction.

Main Methods:

• Developed the Persistent-Laplacian Neural Network (PLNet) framework.
• Encoded protein chains using persistent Laplacian features and protein language model embeddings.
• Assembled a new benchmark dataset (P2P) with 6886 protein complexes.

Main Results:

• Achieved a Pearson correlation of 0.80 on the P2P dataset using leave-out-protein-out cross-validation.
• PLNet demonstrated superior performance compared to a gradient-boosting decision tree baseline.
• Highlighted the advantage of PLNet in capturing complex topology-aware descriptors for PPI prediction.

Conclusions:

• The PLNet framework offers a promising approach for predicting PPIs binding free energy.
• Integrating persistent Laplacian features enhances the prediction of complex molecular interactions.
• The P2P dataset serves as a valuable resource for future PPI research.