GAN-based data augmentation for transcriptomics: survey and comparative assessment

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Genomics

Background:

• High-throughput sequencing generates vast transcriptomics data, but data scarcity limits deep learning for phenotype prediction.
• Data augmentation artificially expands training sets, but transcriptomic transformations remain undefined.
• Generative Adversarial Networks (GANs) offer a potential solution for generating synthetic transcriptomic samples.

Purpose of the Study:

• To analyze Generative Adversarial Network (GAN)-based data augmentation strategies for transcriptomics.
• To evaluate the impact of GANs on cancer phenotype classification performance.
• To assess the quality and utility of GAN-generated transcriptomic data.

Main Methods:

• Employed Generative Adversarial Networks (GANs) to generate augmented transcriptomic data.
• Trained binary and multiclass classifiers on original and augmented RNA-sequencing datasets.
• Evaluated classification performance using accuracy metrics and analyzed GAN-generated data quality.

Main Results:

• Augmentation strategies significantly improved classification performance, increasing accuracy from 94% to 98% (binary) and 70% to 94% (multiclass).
• Training classifiers with 1000 augmented samples showed substantial gains compared to using only 50 original samples.
• Richer GAN architectures and more extensive training yielded better augmentation performance and data quality.

Conclusions:

• GAN-based data augmentation is a powerful strategy to overcome data scarcity in transcriptomics.
• This approach substantially enhances deep learning model performance for cancer phenotype classification.
• Multiple performance indicators are necessary for a comprehensive assessment of generated data quality.