Optimizing antimalarial discovery: A robust machine learning framework for chalcone bioactivity prediction using ensemble classifiers and oversampling strategies

Abstract

• ExtraTrees-KlekotaRothCount framework predicts chalcone potency with 100% accuracy. • Oversampling proves essential for stabilizing models on small imbalanced datasets. • Nitrogen heterocycles identified as primary structural drivers of antimalarial activity. • Systematic benchmark of 1,440 models optimizes antimalarial discovery pipelines. • Pharmacophoric fingerprints outperform binary descriptors in bioactivity classification. The rapid emergence of drug-resistant Plasmodium falciparum necessitates efficient computational tools to prioritize novel antimalarial scaffolds. This study presents a robust machine learning framework for predicting the bioactivity of chalcone derivatives, specifically addressing the challenge of dataset imbalance in quantitative structure-activity relationship (QSAR) modeling. 1,440 model combinations were systematically evaluated, comprising twelve classification algorithms, twelve molecular fingerprint types and three sampling strategies, on a curated dataset of 251 chalcones. The results demonstrate that the ExtraTreesClassifier, combined with KlekotaRothCount fingerprints and an oversampling strategy, achieves superior predictive performance (MCC = 1.0; Accuracy = 100% on independent test sets), significantly outperforming standard approaches. Feature importance analysis revealed that nitrogen-containing heterocycles and specific aromatic configurations are the primary structural drivers of antimalarial potency. Unlike previous benchmarks, these findings are consistent with the view that oversampling is beneficial for stabilizing model performance in small, imbalanced chemical datasets. These findings provide a validated, high-precision computational tool for the rational design of next-generation antimalarial agents.