MaTPIP: A deep-learning architecture with eXplainable AI for sequence-driven, feature mixed protein-protein...

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence in Biology

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions and have broad applications in drug discovery and therapeutics.
• Predicting PPIs from protein sequences remains a significant challenge in computational biology.

Purpose of the Study:

• To introduce MaTPIP, a novel deep-learning framework for accurate sequence-based protein-protein interaction prediction.
• To enhance the generalization capability of PPI prediction models, particularly for cross-species applications.

Main Methods:

• MaTPIP integrates pre-trained Protein Language Model (PLM)-based features with curated protein sequence attributes.
• The framework incorporates both granular amino-acid level (2D) and whole-protein level (1D) features.
• A hybrid deep learning architecture combining Convolutional Neural Networks (CNNs) and Transformer components is employed.

Main Results:

• MaTPIP significantly outperformed existing methods on human and cross-species PPI datasets.
• Achieved state-of-the-art performance in novel PPI prediction scenarios, improving key metrics like Area Under ROC Curve and average precision.
• Established new benchmark scores in cross-species PPI prediction for multiple organisms, including Mouse, Fly, Worm, Yeast, and E.coli.

Conclusions:

• MaTPIP effectively combines manually curated features with PLM-extracted features for sequence-based PPI prediction.
• The framework demonstrates robust generalization capabilities, particularly for predicting cross-species protein-protein associations.