SyntheVAEiser: augmenting traditional machine learning methods with VAE-based gene expression sample generation for...

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Machine learning accuracy is often constrained by limited training data availability.
• Effective data augmentation strategies are crucial for enhancing model performance in complex biological datasets.

Purpose of the Study:

• To introduce a novel method for synthesizing gene expression samples to augment existing datasets.
• To evaluate the capability of synthetic data generation in improving the accuracy of cancer subtype prediction using machine learning.

Main Methods:

• Development of SyntheVAEiser, a variational autoencoder-based tool for gene expression data synthesis.
• Training and testing the SyntheVAEiser tool on a dataset comprising over 8000 cancer samples.
• Assessing the impact of augmented datasets on the performance of categorical prediction tasks for cancer subtypes.

Main Results:

• Demonstrated that synthetic gene expression sample generation can effectively augment machine learning training datasets.
• Showcased a significant increase in the performance of cancer subtype recognition, particularly for underrepresented cohorts.
• Validated the utility of SyntheVAEiser in improving the accuracy of machine learning models in a real-world biomedical context.

Conclusions:

• Data augmentation using synthetically generated gene expression samples is a viable strategy to overcome data limitations in machine learning.
• SyntheVAEiser offers a powerful tool for enhancing the predictive accuracy of cancer subtype classification models.
• The proposed method shows promise for improving the identification and analysis of rare or underrepresented cancer subtypes.