Undersampling techniques for large datasets.

Area of Science:

• Medicinal Chemistry
• Chemoinformatics
• Machine Learning

Background:

• DNA-Encoded Libraries (DELs) enable rapid synthesis and screening of billions of small molecules.
• Machine learning (ML) models benefit from DEL binding data for drug discovery.
• Class imbalance, with far more inactive than active compounds, poses a significant challenge for ML model training in DELs.

Purpose of the Study:

• To investigate and benchmark various undersampling strategies for the majority (inactive) class in DEL datasets.
• To assess the impact of these strategies on the performance of ML models trained on imbalanced DEL data.

Main Methods:

• Exploration of different undersampling techniques for the majority class.
• Benchmarking undersampling strategies against random selection.
• Prototyping and evaluation on two distinct DEL datasets.
• Testing with three different machine learning models.

Main Results:

• The 'max_sim' undersampling strategy demonstrated superior performance across evaluated metrics.
• Comparative analysis showed significant improvements over random selection in handling class imbalance.
• The developed pipeline was successfully implemented within the DELight package.

Conclusions:

• Undersampling strategies, particularly 'max_sim', are effective in mitigating class imbalance in DEL datasets.
• Improved ML model training using balanced DEL data can enhance hit identification and drug discovery efforts.
• The DELight package provides a practical tool for applying these strategies in DEL-based drug discovery.