Improved Chemical Structure-Activity Modeling Through Data Augmentation.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Data augmentation improves predictive model generalization and robustness.
• Its effect on quantitative structure-activity relationship (QSAR) models requires further investigation.

Purpose of the Study:

• To evaluate the impact of data augmentation on QSAR model predictive power.
• To identify optimal data augmentation strategies for QSAR modeling.

Main Methods:

• Collected 8 diverse QSAR datasets (pIC50 values).
• Augmented training data by replicating and perturbing with Gaussian noise (pIC50 values, descriptors, or both).
• Evaluated effects across Random Forest, Gradient Boosting Machine, and Support Vector Machine radial algorithms with Morgan fingerprints and physicochemical descriptors.

Main Results:

• Data augmentation consistently increased predictive power by 10-15%.
• Perturbing compound descriptors or both descriptors and pIC50 values yielded the greatest RMSE reduction.
• Maximum performance gains were observed when training data was replicated once.

Conclusions:

• Data augmentation is effective for enhancing QSAR model performance.
• Replicating training data once with perturbed descriptors offers a good balance between performance and computational cost.
• Optimizing noise levels and managing increased training data are key considerations.