Multimodal deep representation learning for protein interaction identification and protein family classification

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Protein-protein interactions (PPIs) are vital for biological processes and disease.
• Understanding PPIs is crucial for drug discovery and understanding disease mechanisms.
• Existing methods for PPI discovery are limited by the scarcity of experimentally validated interactions compared to protein sequence data.

Purpose of the Study:

• To develop a novel computational framework for predicting protein-protein interactions (PPIs).
• To integrate diverse data types, including protein sequence and PPI network topology, for enhanced prediction accuracy.
• To establish a new benchmark for PPI prediction using a multi-modal deep representation learning approach.

Main Methods:

• A multi-modal deep representation learning framework was developed.
• Incorporated protein physicochemical features and graph topological features from PPI networks.
• Utilized a stacked auto-encoder and a continuous bag-of-words (CBOW) model with metapaths.
• Employed supervised deep neural networks for PPI identification and protein family classification.

Main Results:

• Achieved high PPI prediction accuracy, ranging from 96.76% to 99.77% across eight species.
• Demonstrated superior performance compared to existing computational methods for PPI prediction.
• Successfully identified protein-protein interactions and classified protein families using the proposed framework.

Conclusions:

• This work presents the first multi-modal deep representation learning framework specifically designed for analyzing PPI networks.
• The proposed framework offers a powerful and accurate approach for predicting protein-protein interactions.
• The findings have significant implications for advancing biological research and drug development through improved PPI understanding.