SyntheVAEiser: augmenting traditional machine learning methods with VAE-based gene expression sample generation for improved cancer subtype predictions

Abstract

The accuracy of machine learning methods is often limited by the amount of training data that is available. We proposed to improve machine learning training regimes by augmenting datasets with synthetically generated samples. We present a method for synthesizing gene expression samples and test the system's capabilities for improving the accuracy of categorical prediction of cancer subtypes. We developed SyntheVAEiser, a variational autoencoder based tool that was trained and tested on over 8000 cancer samples. We have shown that this technique can be used to augment machine learning tasks and increase performance of recognition of underrepresented cohorts.

Keywords: Cancer subtyping; Data augmentation; Feature engineering; Gene expression; Generative modeling; Molecular subtyping; Sample synthesis; Synthetic data; Transcriptomics; Variational autoencoder.