Completing sparse and disconnected protein-protein network by deep learning.

Area of Science:

• Computational systems biology
• Bioinformatics
• Network biology

Background:

• Protein-protein interaction (PPI) prediction is crucial for understanding cellular processes.
• Existing network-level PPI prediction methods often assume connected training networks, limiting their applicability to sparse, disconnected real-world data.
• Predicting PPIs from incomplete and fragmented networks remains a significant challenge.

Purpose of the Study:

• To develop a novel computational method for predicting protein-protein interactions (PPIs) using deep learning.
• To address the limitations of existing methods by effectively handling sparse and disconnected protein-protein interaction networks.
• To improve the accuracy and scope of PPI prediction by integrating heterogeneous data sources.

Main Methods:

• Developed a deep neural network with an autoencoder-like architecture to simulate interactome evolution.
• Utilized a regularized Laplacian kernel applied to a transition matrix derived from an evolved PPI network.
• Input layer of the neural network received zero input, mimicking a lack of prior knowledge.

Main Results:

• Achieved significant improvements in PPI prediction accuracy, with AUC increases of up to 8.4% for yeast and 14.9% for human data compared to baselines.
• Demonstrated the method's ability to leverage complementary information from disconnected networks and integrate multiple heterogeneous data sources.
• Integration of six heterogeneous feature kernels for yeast data further improved prediction performance by up to 2%.

Conclusions:

• The proposed evolution deep neural network combined with a regularized Laplacian kernel effectively completes sparse and disconnected PPI networks.
• This method facilitates the integration of heterogeneous data sources for enhanced PPI prediction.
• The approach offers a powerful tool for advancing systems biology and understanding complex cellular interactions.