RESUMO
INTRODUCTION: Computational methods are crucial for efficient and cost-effective drug toxicity prediction. Unfortunately, the data used for prediction is often imbalanced, resulting in biased models that favor the majority class. This paper proposes an approach to apply a hybrid class balancing technique and evaluate its performance on computational models for toxicity prediction in Tox21 datasets. METHODS: The process begins by converting chemical compound data structures (SMILES strings) from various bioassay datasets into molecular descriptors that can be processed by algorithms. Subsequently, Undersampling and Oversampling techniques are applied in two different schemes on the training data. In the first scheme (Individual), only one balancing technique (Oversampling or Undersampling) is used. In the second scheme (Hybrid), the training data is divided according to a ratio (e.g., 90-10), applying a different balancing technique to each proportion. We considered eight resampling techniques (four Oversampling and four Undersampling), six molecular descriptors (based on MACCS, ECFP, and Mordred), and five classification models (KNN, MLP, RF, XGB and SVM) over 10 bioassay datasets to determine the configurations that yield the best performance. RESULTS: We defined three testing scenarios: without balancing techniques (baseline), Individual, and Hybrid. We found that using the ENN technique in the MACCS-MLP combination resulted in a 10.01% improvement in performance. The increase for ECFP6-2048 was 16.47% after incorporating a combination of the SMOTE (10%) and RUS (90%) techniques. Meanwhile, using the same combination of techniques, MORDRED-XGB showed the most significant increase in performance, achieving a 22.62% improvement. CONCLUSION: Integrating any of the class balancing schemes resulted in a minimum of 10.01% improvement in prediction performance compared to the best baseline configuration. In this study, Undersampling techniques were more appropriate due to the significant overlap among samples. By eliminating specific samples from the predominant class that are close to the minority class, this overlap is greatly reduced.