pdCSM-PPI

pdCSM-PPI predicts small-molecule modulators of protein-protein interactions (PPIs) using graph-based molecular representations and machine learning to prioritize PPI inhibitors for drug discovery.

Key Features:

• Graph-Based Representation: Uses graph-based representations of small molecules to capture structural properties relevant to PPI binding.
• Machine Learning Models: Employs machine learning to distinguish active from inactive compounds targeting PPIs.
• Interaction-Specific Models: Developed separate models for 21 distinct PPI targets to provide target-specific predictions.
• Generic IC50 Predictive Model: Constructs a generic model for predicting IC50 values derived from insights across individual interaction-specific models.
• Performance Metrics: Achieved Matthews Correlation Coefficient (MCC) and F1 scores up to 1, Pearson correlations up to 0.87 for some models, and a generic-model Pearson correlation of 0.64 on a low-redundancy blind test set.
• Screening Performance: Demonstrated ability to distinguish active versus inactive compounds with AUC = 0.77, sensitivity = 76%, and specificity = 78%.

Scientific Applications:

• Interaction-Specific Inhibitor Identification: Prioritizes potential inhibitors for individual PPI targets using target-specific predictive models.
• IC50 Prediction: Predicts IC50 values across PPI-relevant compounds via a generic predictive model.
• Virtual Screening and Prioritization: Facilitates virtual screening by distinguishing active and inactive small molecules against PPI targets using reported AUC, sensitivity, and specificity metrics.

Methodology:

Applies graph-based molecular representations and machine learning to develop interaction-specific models for 21 PPI targets and a generic IC50 model, evaluated using MCC, F1, Pearson correlation, AUC, sensitivity, and specificity on reported test sets including a low-redundancy blind test set.