Coping with unbalanced class data sets in oral absorption models.

Area of Science:

• Drug Discovery
• Computational Chemistry
• Machine Learning in Cheminformatics

Background:

• Class imbalance is prevalent in drug discovery datasets, particularly for oral absorption.
• Existing models are often biased towards highly absorbed compounds, limiting their utility for early-stage drug candidates which are frequently poorly absorbed.

Purpose of the Study:

• To investigate strategies for handling unbalanced class data in drug discovery.
• To improve the generalization of classification models for predicting drug absorption.

Main Methods:

• Applied undersampling of the majority (high absorption) class in training datasets.
• Utilized misclassification costs within classification and regression tree (C&RT) analysis.
• Validated methods on a dataset of 645 drug and drug-like compounds with known human intestinal absorption percentages.

Main Results:

• Undersampling achieved a balanced 50:50 training set, enhancing accuracy for poorly absorbed compounds.
• Misclassification costs improved class predictions by reducing false positives and false negatives.
• Standard accuracy metrics were found to be misleading for unbalanced datasets; alternative measures offer more realistic performance assessments.

Conclusions:

• Undersampling and misclassification costs are effective strategies for addressing class imbalance in drug absorption prediction.
• These methods yield more robust and industrially applicable classification models.
• The study highlights the importance of appropriate performance metrics for unbalanced datasets.