PARP1pred: a web server for screening the bioactivity of inhibitors against DNA repair enzyme PARP-1

Abstract

Cancer is the leading cause of death worldwide, resulting in the mortality of more than 10 million people in 2020, according to Global Cancer Statistics 2020. A potential cancer therapy involves targeting the DNA repair process by inhibiting PARP-1. In this study, classification models were constructed using a non-redundant set of 2018 PARP-1 inhibitors. Briefly, compounds were described by 12 fingerprint types and built using the random forest algorithm concomitant with various sampling approaches. Results indicated that PubChem with an oversampling approach yielded the best performance, with a Matthews correlation coefficient > 0.7 while also affording interpretable molecular features. Moreover, feature importance, as determined from the Gini index, revealed that the aromatic/cyclic/heterocyclic moiety, nitrogen-containing fingerprints, and the ether/aldehyde/alcohol moiety were important for PARP-1 inhibition. Finally, our predictive model was deployed as a web application called PARP1pred and is publicly available at https://parp1pred.streamlitapp.com, allowing users to predict the biological activity of query compounds using their SMILES notation as the input. It is anticipated that the model described herein will aid in the discovery of effective PARP-1 inhibitors.

Keywords: DNA repair; PARP-1; QSAR; cheminformatics; machine learning; webserver.

Table 1. Twelve different sets of fingerprint descriptors derived from the PaDEL-Descriptor software

Table 2. Descriptions of SMARTS patterns and substructures from the top 20 Gini indices

Figure 1. Overall workflow of the development of the webserver for PARP-1 inhibitors

Figure 2. Illustration of the relationship between molecular weight (MW) and Ghose-Crippen-Viswanadhan octanol-water partition coefficient (LogP). Blue and orange represent active and inactive compounds. The size of the circle refers to the pIC₅₀ value, which is the negative logarithmic of the IC₅₀ concentration (nM).

Figure 3. Box plots of Lipinski's rule-of five descriptors comparing between active and inactive groups. The dashed line represents cut-off values indicating drug-like molecules: molecular weight (MW) < 500, Ghose-Crippen-Viswanadhan octanol-water partition coefficient (LogP) < 5, number of hydrogen bond donors (NumHDonors) < 5, number of hydrogen bond acceptors (NumHAcceptors) < 10. A circle represents the mean, and an asterisk indicates a significant difference between two groups (p < 0.05).

Figure 4. Heat maps of the MCC values of the training, CV, and test sets for each data sampling approach. (A) Balanced undersampling, (B) balanced oversampling, and (C) imbalanced non-class weight. Abbreviations: MCC, Matthews correlation coefficient; CV, cross-validation; gaussianNB, Gaussian Naive Bayes; LBMC, light gradient boosted machine; MLP, multi-layer perceptron; SVC, C-support vector; XGB, extreme gradient boosting

Figure 5. Heat maps of MCC_train−MCC_CV and MCC_train−MCC_test for each data sampling approach. (A) Balanced undersampling, (B) balanced oversampling, (C) imbalanced non-class weight. Abbreviations: MCC, Matthews correlation coefficient; CV, cross-validation; gaussianNB, Gaussian Naive Bayes; LBMC, light gradient boosted machine; MLP, multi-layer perceptron; SVC, C-support vector; XGB, extreme gradient boosting

Figure 6. Plot of PCA scores for applicability domain analysis. The score plot indicates the distribution of chemical space of the internal (green) and external (red) datasets, which were used to determine the applicability domain of the PARP-1 inhibitors dataset.

Figure 7. Feature importance plot as rationalized by Gini index obtained from random forest model using oversampling

Figure 8. Crystal structure of the catalytic domain of PARP-1 (PDB ID 1UK0) and the interaction network between PARP-1 and olaparib (PDB ID 7KK4). The alpha-helical subdomain (HD) is shown in light orange color while the ADP-ribosyl transferase subdomain (ART) is shown in wheat color. Hydrogen forming network (blue solid line), π-π (green dashed line), and hydrophobic (grey dashed line) interactions between key amino acids within the nicotinamide binding site and olaparib

Figure 9. Screenshot of the PARP1pred webserver before (A) and after (B) entering the SMILES input. Notice that after submission of the SMILES notation the corresponding molecular fingerprints are computed whereby the trained predictive model is applied to classify the query molecule as active or inactive. In this case, the query molecule is classified to be active.