protein2vec: Predicting Protein-Protein Interactions Based on LSTM.

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Predicting protein-protein interactions (PPIs) is crucial in molecular biology.
• Traditional methods using gene ontology (GO) semantic similarity rely on manual features and oversimplify protein-GO term relationships.
• Existing approaches often fail to capture the complex interplay between proteins and their annotated GO terms.

Purpose of the Study:

• To develop a novel method, protein2vec, for improved PPI prediction.
• To overcome limitations of traditional semantic similarity measures in capturing hidden information within GO and complex protein-GO relationships.
• To integrate GO term information and known PPIs effectively for enhanced prediction.

Main Methods:

• Applied network embedding on the GO network to generate feature vectors for GO terms.
• Utilized Long Short-Time Memory (LSTM) to encode protein-specific GO term feature vectors into protein vectors.
• Employed a feedforward neural network to predict PPIs using the generated protein vectors.

Main Results:

• The protein2vec method generates effective vector representations for proteins based on their annotated GO terms.
• The model successfully integrates GO term information and known PPI data.
• Experimental results demonstrate that protein2vec significantly outperforms commonly used traditional semantic similarity methods in PPI prediction.

Conclusions:

• protein2vec offers a more sophisticated approach to PPI prediction by leveraging network embedding and deep learning on GO terms.
• The method effectively addresses the limitations of traditional semantic similarity measures.
• protein2vec shows strong potential for advancing biological network analysis and understanding protein functions.