COVER: conformational oversampling as data augmentation for molecules.

Area of Science:

• Computational chemistry
• Machine learning
• Toxicology

Background:

• Neural networks struggle with small, imbalanced datasets, leading to overfitting and poor minority class prediction.
• Predictive toxicology requires models balancing sensitivity and specificity for accurate risk assessment.

Purpose of the Study:

• Introduce conformational oversampling to balance and augment datasets for toxicity prediction.
• Improve neural network performance on imbalanced toxicological data.

Main Methods:

• Generate multiple molecular conformations to enhance existing data.
• Utilize these conformations for dataset balancing and oversampling without artificial data generation.

Main Results:

• Conformational oversampling effectively balances imbalanced datasets.
• Facilitates improved neural network training for predictive toxicology.
• Achieved state-of-the-art performance on the Tox21 dataset.

Conclusions:

• Conformational oversampling is a viable strategy for improving predictive toxicology models.
• This method enhances dataset quality for machine learning applications in chemical safety.