Comprehensive machine learning boosts structure-based virtual screening for PARP1 inhibitors.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Poly ADP-ribose polymerase 1 (PARP1) is a key therapeutic target in cancer treatment.
• Machine learning (ML) scoring functions offer a promising avenue for identifying novel PARP1 inhibitors.

Purpose of the Study:

• To investigate and develop cutting-edge, PARP1-specific ML scoring functions for enhanced drug discovery.
• To rigorously evaluate the predictive performance of these functions using challenging datasets.

Main Methods:

• Utilized semi-synthetic training data including known PARP1 inhibitors, property-matched decoys generated by graph neural networks, and confirmed inactives.
• Constructed difficult test sets with molecules dissimilar to the training data.
• Applied five supervised learning algorithms and analyzed protein-ligand fingerprints (PLEC) and ligand-only features from docking poses.

Main Results:

• Identified a highly predictive PARP1-specific support vector machine (SVM)-based regressor using PLEC fingerprints.
• This SVM model achieved a high Normalized Enrichment Factor at the top 1% (NEF1% = 0.588) on the most challenging test set.
• The developed scoring function demonstrated superior predictive power compared to other investigated functions and a classical baseline scoring function.

Conclusions:

• A novel PARP1-specific SVM-based ML scoring function utilizing PLEC fingerprints shows exceptional performance in virtual screening.
• This method significantly advances the discovery of potent PARP1 inhibitors for cancer therapy.
• The study highlights the effectiveness of carefully curated datasets and advanced ML techniques in drug discovery.