PCPredG: Protein Complex Prediction Using Graphlet Features.

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Protein complexes are fundamental to cellular functions and organismal responses.
• Predicting these complexes is vital but challenging, with limited existing methodologies.
• Understanding protein-protein interactions (PPIs) is key to deciphering biological pathways.

Purpose of the Study:

• To develop a novel computational method, PCPredG, for predicting 3-node protein complexes.
• To leverage 5-node graphlet features for enhanced prediction accuracy.
• To compare the performance of traditional machine learning classifiers with state-of-the-art deep learning models.

Main Methods:

• Utilized the CORUM protein complex repository for data curation.
• Employed MCODE and MCL clustering algorithms for sample preparation.
• Trained Random Forest (RF) and Support Vector Machine (SVM) classifiers.
• Implemented Graph Convolutional Networks (GCN) with polarized message-passing and Graph Attention Networks (GAT).
• Evaluated models using 10-fold cross-validation with varying positive-negative sample ratios (1:1 to 1:10).

Main Results:

• The PCPredG method demonstrated effective 3-node protein complex prediction.
• Random Forest (RF) classifier achieved the best performance in both balanced and imbalanced datasets.
• Deep learning models (GCN, GAT, and their ensemble) were also implemented and evaluated.
• A 10-fold quality consensus was applied to assess model robustness on hold-out data.

Conclusions:

• PCPredG offers a promising approach for predicting 3-node protein complexes from PPI networks.
• Random Forest remains a highly effective classifier for this prediction task, even with imbalanced data.
• The study contributes to advancing computational methods for understanding protein complex formation and function.