Generating bulk RNA-Seq gene expression data based on generative deep learning models and utilizing it for data...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• High-throughput transcriptome sequencing is crucial for biomedical research.
• Limited sample sizes in bulk RNA-Seq data can reduce analysis reliability.
• Generative deep learning models offer potential solutions for data scarcity.

Purpose of the Study:

• To develop and evaluate generative deep learning models for bulk RNA-Seq gene expression data.
• To address challenges in sample acquisition and enhance data analysis reliability.
• To improve downstream task performance through data augmentation.

Main Methods:

• Utilized bulk RNA-Seq gene expression data.
• Constructed Generative Adversarial Networks (GANs) and Diffusion Models (DMs) with Min-Max and Z-Score preprocessing.
• Trained models on the largest dataset to date, evaluating with Maximum Mean Discrepancy (MMD).
• Employed SHAP (Shapley Additive exPlanations) for model interpretability.

Main Results:

• The Min-Max-GAN model generated data with high similarity to real data, outperforming other models.
• Achieved low MMD values (0.030 training, 0.033 independent dataset) on a large-scale dataset.
• Data augmentation using generated data significantly improved classification model performance.
• SHAP explanations enhanced the credibility of the generative model.

Conclusions:

• A GAN-based approach effectively generates bulk RNA-Seq gene expression data.
• This method enhances the performance and reliability of downstream transcriptome analysis tasks.
• The study provides a valuable tool for overcoming sample size limitations in transcriptomic research.