Oversampling to overcome overfitting: exploring the relationship between data set composition, molecular descriptors,...

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Traditional biological assays are time-consuming, necessitating faster methods for compound screening.
• High-throughput screening (HTS) is crucial in biomedical research and drug discovery for rapid compound evaluation.
• Imbalanced datasets, common in cytotoxicity prediction, pose challenges for predictive modeling due to overfitting.

Purpose of the Study:

• To improve the classification of compound cytotoxicity using machine learning.
• To investigate the effectiveness of support vector machines (SVM) with various molecular descriptors and sampling techniques.
• To address overfitting issues in imbalanced datasets for cytotoxicity prediction against the Jurkat cell line.

Main Methods:

• Utilized support vector machines (SVM) as the primary machine learning method.
• Employed diverse molecular descriptors, including 4D-FPs, MOE (1D, 2D, 2.5D), noNP+MOE, and CATS2D.
• Applied different sampling techniques, particularly oversampling, to handle imbalanced cytotoxicity data.
• Compared SVM model performance against CATS2D-based random forest models.

Main Results:

• SVM models trained on oversampled datasets demonstrated improved predictive abilities.
• The models showed enhanced performance on both training and external test sets compared to previous literature.
• Oversampling effectively mitigated overfitting towards the majority (inactive) class in imbalanced datasets.

Conclusions:

• Machine learning, specifically SVM with oversampling, offers a robust approach to predict compound cytotoxicity.
• The chosen molecular descriptors and sampling strategies significantly impact predictive model performance.
• This method accelerates compound evaluation, supporting efficient drug discovery pipelines.