Sequence-based Protein-Protein Interaction Prediction Using Multi-kernel Deep Convolutional Neural Networks with Protein Language Model

Abstract

Predicting protein-protein interactions (PPIs) using only sequence information represents a fundamental problem in biology. In the past five years, a wide range of state-of-the-art deep learning models have been developed to address the computational prediction of PPIs based on sequences. Convolutional neural networks (CNNs) are widely adopted in these model architectures; however, the design of a deep and wide CNN architecture that comprehensively extracts interaction features from pairs of proteins is not well studied. Despite the development of several protein language models that distill the knowledge of evolutionary, structural, and functional information from gigantic protein sequence databases, no studies have integrated the amino acid embeddings of the protein language model for encoding protein sequences.In this study, we introduces a novel hybrid classifier, xCAPT5, which combines the deep multi-kernel convolutional accumulated pooling siamese neural network (CAPT5) and the XGBoost model (x) to enhance interaction prediction. The CAPT5 utilizes multi-deep convolutional channels with varying kernel sizes in the Siamese architecture, enabling the capture of small- and large-scale local features. By concatenating max and average pooling features in a depth-wise manner, CAPT5 effectively learns crucial features with low computational cost. This study is the first to extract information-rich amino acid embedding from a protein language model by a deep convolutional network, through training to obtain discriminant representations of protein sequence pairs that are fed into XGBoost for predicting PPIs. Experimental results demonstrate that xCAPT5 outperforms several stateof-the-art methods on binary PPI prediction, including generalized PPI on intra-species, cross-species, inter-species, and stringent similarity tasks. The implementation of our framework is available athttps://github.com/anhvt00/MCAPS