pdCSM-PPI: Using Graph-Based Signatures to Identify Protein-Protein Interaction Inhibitors

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Protein-protein interactions (PPIs) are crucial in biological processes and represent promising drug targets.
• Developing selective drugs for PPIs is challenging due to the large, lipophilic nature of interacting molecules.

Purpose of the Study:

• To introduce pdCSM-PPI, a novel machine learning approach for identifying small molecule inhibitors of PPIs.
• To evaluate the performance of pdCSM-PPI across various PPI targets and compare it with existing methods.

Main Methods:

• Utilized a graph-based representation of small molecules for machine learning model development.
• Applied the approach to 21 distinct PPI targets, developing both interaction-specific and a generic predictive model.
• Validated model performance using metrics such as MCC, F1 scores, Pearson's correlation, and AUC on blind test datasets.

Main Results:

• Interaction-specific models achieved high accuracy (MCC/F1 up to 1, Pearson's correlation up to 0.87), surpassing previous approaches.
• The generic model accurately predicted IC50 values (Pearson's correlation of 0.64) and distinguished active from inactive compounds (AUC of 0.77).
• Achieved high sensitivity (76%) and specificity (78%) in identifying active compounds.

Conclusions:

• pdCSM-PPI is an effective tool for guiding the efficient screening of novel PPI inhibitors.
• The freely available web server and API facilitate broader accessibility and application in drug discovery research.