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

Abstract Poly ADP-ribose polymerase 1 (PARP1) is an attractive therapeutic target for cancer treatment. Machine-learning scoring functions constitute a promising approach to discovering novel PARP1 inhibitors. Cutting-edge PARP1-specific machine-learning scoring functions were investigated using semi-synthetic training data from docking activity-labelled molecules: known PARP1 inhibitors, hard-to-discriminate decoys property-matched to them with generative graph neural networks and confirmed inactives. We further made test sets harder by including only molecules dissimilar to those in the training set. Comprehensive analysis of these datasets using five supervised learning algorithms, and protein-ligand fingerprints extracted from docking poses and ligand only features revealed two highly predictive scoring functions. The PARP1-specific support vector machine-based regressor, when employing PLEC fingerprints, achieved a high Normalized Enrichment Factor at the top 1% on the hardest test set (NEF1% = 0.588, median of 10 repetitions), and was more predictive than any other investigated scoring function, especially the classical scoring function employed as baseline. Scientific Contribution We present the first PARP1-specific machine-learning scoring functions for structure-based virtual screening. A particularly rigorous evaluation, including test sets with novel molecules and a much higher proportion of challenging property-matched decoys, reveals the most predictive scoring function for this important therapeutic target. Typically, narrow machine learning analyses would have likely missed this promising PARP1-specific scoring function, which is now released with this paper so that others can use it for prospective virtual screening. Key Points A new scoring tool based on machine-learning was developed to predict PARP1 inhibitors for potential cancer treatment. The majority of PARP1-specific machine-learning models performed better than generic and classical scoring functions. Augmenting the training set with ligand-only Morgan fingerprint features generally resulted in better performing models, but not for the best models where no further improvement was observed. Employing protein-ligand-extracted fingerprints as molecular descriptors led to the best-performing and most-efficient model for predicting PARP1 inhibitors. Deep learning performed poorly on this target in comparison with the simpler ML models.