Assessing the druggability of protein-protein interactions by a supervised machine-learning method

Area of Science:

• Biochemistry
• Pharmacology
• Computational Biology

Background:

• Protein-protein interactions (PPIs) are crucial in human diseases but challenging drug targets.
• Existing methods lack efficiency in selecting druggable PPIs from vast datasets.
• A novel approach is needed to holistically assess PPI druggability.

Purpose of the Study:

• To develop a machine learning methodology for efficient selection of druggable PPIs.
• To assess the druggability of PPIs by analyzing structural, chemical, and functional attributes.
• To utilize a supervised machine learning method, specifically support vector machine (SVM).

Main Methods:

• Selected 69 attributes covering structural, drug/chemical, and functional information of PPIs.
• Utilized a support vector machine (SVM) model trained on 30 known druggable PPIs (positive instances).
• Applied the SVM model to 1,295 human PPIs with solved tertiary structures.

Main Results:

• The best SVM model achieved 81% accuracy in discriminating druggable PPIs (82% sensitivity, 79% specificity) via cross-validation.
• Key attributes for discrimination included the number of interacting proteins and pathways.
• Identified promising druggable PPI candidates, such as SMAD4/SKI.

Conclusions:

• The developed SVM-based method effectively predicts druggable PPIs.
• This approach can accelerate the discovery of novel drug targets for human diseases.
• The method's predictive power is expected to increase with growing PPI data.