An Improved Conditional Wasserstein GAN With Gradient Penalty for Gene Expression Profiling Data Augmentation Based...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Healthcare

Background:

• Small sample sizes in gene expression profiling data can cause overfitting in medical diagnosis.
• Existing Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) methods lack control over sample generation and training stability.

Purpose of the Study:

• To propose an Improved CWGAN-GP to address limitations in data augmentation for gene expression profiling.
• To enhance the quality and stability of generated gene expression data for improved medical diagnosis.

Main Methods:

• Implemented a data segmentation strategy using sample influence scores to prioritize boundary and outlier samples.
• Introduced a depth feature constraint based on Pearson correlation coefficient to guide feature extraction and stabilize training.
• Utilized an encoder to extract deep features and applied constraints between noise and deep features.

Main Results:

• The Improved CWGAN-GP generates higher quality synthetic gene expression data.
• The proposed method demonstrates superior stability during the training process compared to existing approaches.
• Empirical evaluations on six public datasets validate the effectiveness of the enhanced model.

Conclusions:

• The Improved CWGAN-GP effectively overcomes limitations of standard CWGAN-GP for gene expression data augmentation.
• The data segmentation and depth feature constraint strategies lead to more explicit decision boundaries and stable training.
• This approach offers a promising solution for improving the reliability of machine learning models in medical diagnosis using gene expression data.