Predicting Self-Interacting Proteins Using a Recurrent Neural Network and Protein Evolutionary Information.

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Self-interacting proteins (SIPs) are vital for cellular functions.
• Experimental methods for identifying SIPs are often time-consuming and expensive.
• Developing efficient computational methods for SIP identification is a significant challenge.

Purpose of the Study:

• To introduce a novel computational method, RNN-SIFT, for predicting SIPs.
• To leverage protein evolutionary information for improved SIP prediction accuracy.
• To provide a freely accessible web server for utilizing the developed method.

Main Methods:

• Utilized Scale Invariant Feature Transform (SIFT) to extract key features from Position-Specific Iterated BLAST (PSI-BLAST) position-specific scoring matrices.
• Employed a Recurrent Neural Network (RNN) classifier for predicting self-interacting proteins based on extracted evolutionary features.
• Developed the RNN-SIFT-SIPs web server for public access to the prediction tool and datasets.

Main Results:

• Achieved high prediction accuracies of 94.34% for yeast and 97.12% for human datasets.
• Demonstrated superior performance of RNN-SIFT compared to Back Propagation Neural Network (BPNN), Support Vector Machine (SVM), and other existing methods.
• Validated the effectiveness of combining SIFT feature extraction with RNN classification for SIP prediction.

Conclusions:

• RNN-SIFT is a powerful and accurate computational tool for predicting self-interacting proteins.
• The method offers a significant improvement over existing computational approaches for SIP identification.
• The developed RNN-SIFT-SIPs web server facilitates further research in proteomics and bioinformatics.