Mal-Predict: Machine learning-guided rapid virtual screening of compounds against selected targets of Plasmodium falciparum validated using molecular dynamics simulation

There is paucity of work on integration of simple predictive models with diversified screening dataset, coupled with robust validation, for antimalaria activity-prediction. To bridge these gaps, we trained predictive models (Random Forest (RF) classifier, XGBoost classifier) using public data from Chembl and PubChem. The RF (AUC = 0.912) was used to classify 1.9 million compounds from the drug-bank, natural-products and Enamine-Real databases into actives or inactives. The predicted actives were validated using docking and molecular dynamics against high-priority Plasmodium falciparum targets. Free energy calculations (MMGBSA; kcal/mol) revealed compounds EN52 (-48.05 ± 3.91) and NP83 (-52.67 ± 4.43) that are energetically more favored than the co-crystalized ligands, WLK (-35.05 ± 3.47) and MMV (-43.08 ± 3.14) for PfPKG and F/GGPPS proteins respectively. We deployed Mal-Predict, a tool to classify compounds as actives (or inactive), and further predict vina-scores for select P. falciparum protein targets. These findings would support prioritizing candidates for further investigations in early-stage drug-discovery for researchers.