Predicting protein-protein interactions using high-quality non-interacting pairs.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Protein-protein interactions (PPIs) are fundamental to cellular processes.
• Current machine learning methods for PPI prediction are limited by low-quality negative sample generation.
• This leads to unsatisfactory prediction accuracy in existing models.

Purpose of the Study:

• To develop novel approaches for generating high-quality non-interacting protein pairs (negative samples).
• To enhance the accuracy and effectiveness of machine learning-based PPI prediction.
• To provide robust negative datasets for training more reliable PPI prediction models.

Main Methods:

• Proposed NIP-SS (Non-Interacting Pairs based on Sequence Similarity) and NIP-RW (Non-Interacting Pairs based on Random Walk) methods.
• NIP-SS selects dissimilar protein pairs while controlling degree distribution.
• NIP-RW utilizes random walks on PPI networks to identify non-interacting pairs.
• Auto covariance (AC) descriptor used for amino acid sequence encoding.
• Deep neural networks (DNNs) employed for PPI prediction.

Main Results:

• NIP-SS and NIP-RW successfully generated higher quality negative samples compared to existing strategies.
• The proposed methods led to more accurate PPI predictions.
• Experimental validation confirmed the effectiveness of the novel negative sample generation techniques.

Conclusions:

• Negative datasets generated by NIP-SS and NIP-RW reduce prediction bias and improve generalization ability.
• These methods can be integrated as plugins to enhance existing PPI prediction tools.
• Available codes and datasets facilitate further research and application.