Persistent spectral based ensemble learning (PerSpect-EL) for protein-protein binding affinity prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions.
• Machine learning models show promise for PPI analysis, but require effective molecular featurization.
• Current featurization methods present challenges for learning models in PPI prediction.

Purpose of the Study:

• To develop a novel representation and featurization method for PPIs using persistent spectral attributes.
• To introduce PerSpect-based ensemble learning (PerSpect-EL) models for predicting PPI binding affinity.
• To establish a new benchmark in machine learning for PPI analysis.

Main Methods:

• Generated a sequence of Hodge (or combinatorial) Laplacian matrices at various scales via a filtration process.
• Extracted PerSpect attributes, representing statistical and combinatorial properties of spectral information from these matrices, for PPI characterization.
• Employed 1D convolutional neural networks (CNNs) for each PerSpect attribute, stacking them in an ensemble learning framework (PerSpect-EL).

Main Results:

• Systematically evaluated the PerSpect-EL model on the SKEMPI and AB-Bind datasets.
• Achieved state-of-the-art performance in PPI binding affinity prediction.
• Demonstrated superior performance compared to all existing models evaluated.

Conclusions:

• The PerSpect-EL model offers a powerful new approach for PPI binding affinity prediction.
• Persistent spectral featurization significantly improves machine learning model performance for PPIs.
• This method represents a breakthrough in computational approaches to understanding protein interactions.