Hybrid Class Balancing Approach for Chemical Compound Toxicity Prediction.

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Imbalanced datasets in drug toxicity prediction lead to biased computational models.
• Addressing data imbalance is crucial for accurate and cost-effective toxicity assessments.

Purpose of the Study:

• To evaluate a hybrid class balancing technique for improving computational toxicity prediction models.
• To assess the performance of different balancing schemes on Tox21 datasets.

Main Methods:

• Chemical structures (SMILES) converted to molecular descriptors (MACCS, ECFP, Mordred).
• Applied Individual (Oversampling/Undersampling) and Hybrid (ratio-based balancing) schemes.
• Evaluated eight resampling techniques, six descriptors, and five classification models across 10 bioassays.

Main Results:

• All class balancing schemes improved prediction performance by at least 10.01% compared to baseline.
• ENN technique with MACCS-MLP improved performance by 10.01%.
• SMOTE (10%) and RUS (90%) combination with ECFP6-2048 and MORDRED-XGB achieved 16.47% and 22.62% improvements, respectively.

Conclusions:

• Hybrid class balancing effectively enhances computational toxicity prediction accuracy.
• Undersampling techniques are particularly effective due to significant sample overlap.
• The proposed methods offer a robust approach to mitigate bias in toxicity prediction models.