xCAPT5: protein-protein interaction prediction using deep and wide multi-kernel pooling convolutional neural networks...

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Predicting protein-protein interactions (PPIs) from sequence data is a significant challenge.
• Existing computational methods have not fully exploited the rich information within protein language model embeddings.
• A need exists for advanced neural networks to extract multifaceted representations from protein sequences.

Purpose of the Study:

• To introduce xCAPT5, a hybrid classifier designed to predict PPIs.
• To leverage T5-XL-UniRef50 for generating comprehensive amino acid embeddings.
• To develop an efficient neural network for extracting complex interaction features.

Main Methods:

• Utilized the T5-XL-UniRef50 protein large language model for sequence embeddings.
• Employed a multi-kernel deep convolutional siamese neural network to capture interaction features.
• Integrated the XGBoost algorithm for enhanced classification performance.
• Applied depth-wise concatenation of max and average pooling features for efficient learning.

Main Results:

• xCAPT5 effectively extracts crucial features with low computational cost.
• The model demonstrates superior performance in binary PPI prediction.
• Achieved excellent results in cross-validation across multiple benchmark datasets.
• Showcased robust generalization capabilities in various species and similarity contexts.

Conclusions:

• This study pioneers the extraction of informative embeddings from large protein language models using deep convolutional networks.
• xCAPT5 outperforms current state-of-the-art methods for binary PPI prediction.
• The proposed method offers a powerful new tool for computational biology and drug discovery.