Enhancing supervised analysis of imbalanced untargeted metabolomics datasets using a CWGAN-GP framework for data...

Area of Science:

• Computational Biology
• Bioinformatics
• Metabolomics

Background:

• Untargeted metabolomics is vital for biological system discrimination and biomarker discovery.
• Data analysis challenges include large sample size requirements and class imbalance, common in clinical studies.
• Class imbalance can reduce model generalizability, increase overfitting, and bias results.

Purpose of the Study:

• To investigate data augmentation using Conditional Wasserstein Generative Adversarial Networks with Gradient Penalty (CWGAN-GPs) for metabolomics data.
• To evaluate the quality of generated data and the performance of predictive models trained on augmented datasets.
• To assess the impact of data augmentation on class imbalance in metabolomics.

Main Methods:

• Utilized CWGAN-GPs for data augmentation of metabolomics datasets.
• Applied multiple quality assessment criteria for generated data.
• Evaluated supervised predictive model performance on datasets with and without augmented data.
• Tested models on benchmark datasets with varying class separation levels.

Main Results:

• CWGAN-GP models generated realistic metabolomics data, closely matching real samples and avoiding mode collapse.
• Augmenting minority classes with generated samples improved predictive model performance in imbalanced datasets with well-separated classes.
• Model performance improvements were modest for datasets with high class overlap.
• Results were consistent across synthetic datasets with varying class separation.

Conclusions:

• Data augmentation using CWGAN-GPs is a viable strategy to address class imbalance in metabolomics.
• The effectiveness of data augmentation is dependent on the degree of class separation within the dataset.
• Augmentation is not universally beneficial and its impact varies based on data characteristics.