DeepISO: deep learning-powered prediction of protein-protein interaction rewiring generated by alternative splicing

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in biology

Background:

• Protein isoforms, arising from the same gene, can alter protein interaction networks.
• Evaluating these isoform-specific effects computationally across the proteome is a significant challenge.
• Existing protein-protein interaction (PPI) prediction tools often do not account for isoform diversity.

Purpose of the Study:

• To develop a computational framework, DeepISO, for predicting isoform-specific protein-protein interactions (PPIs).
• To enable a more accurate, proteome-wide evaluation of how protein isoforms impact interaction networks.
• To establish a new benchmark for isoform-specific PPI prediction.

Main Methods:

• DeepISO employs a deep learning architecture integrating two graph convolutional neural networks (GCNs).
• It combines GCN outputs using a random forest model, further refined by a logistic regression model.
• The framework uniquely leverages AlphaFold-predicted protein structures and ESM2 protein language model embeddings.

Main Results:

• DeepISO demonstrates superior performance in predicting isoform-specific PPIs compared to state-of-the-art methods.
• The integration of structural and language model data significantly enhances prediction accuracy.
• The framework provides a robust tool for analyzing isoform-driven rewiring of protein interaction networks.

Conclusions:

• DeepISO represents a significant advancement in predicting isoform-specific protein interactions.
• This tool facilitates a deeper understanding of the functional consequences of protein isoforms.
• The approach sets a new standard for computational analysis of isoform effects in biological systems.