ctGAN: combined transformation of gene expression and survival data with generative adversarial network

Jaeyoon Kim, Junhee SeokBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces ctGAN, a novel generative adversarial network that transforms both gene expression and survival data. By enabling style transfer between different cancer types, ctGAN effectively augments limited datasets, leading to significantly improved survival analysis predictions, particularly for challenging cancer types like COAD.

What Problem Does This Paper Address?

Analyzing gene expression data for survival prediction in diseases like cancer is crucial but faces a significant hurdle: the "curse of dimensionality." This refers to the challenge of having a vast number of genes (features) compared to a relatively small number of patient samples. This imbalance often leads to overfitting in deep learning models, limiting their accuracy and reliability. While existing style-transfer generative models can create synthetic gene expression data, they typically ignore survival outcomes, making them less useful for direct clinical applications. This paper aims to bridge this gap by developing a model that can generate both gene expression and survival data, thereby improving the robustness and applicability of survival analysis in oncology.

Methods

The core of this work is the ctGAN (combined transformation Generative Adversarial Network). It's built upon the principles of cycle-consistent GANs (CycleGANs), which are known for their ability to perform unpaired image-to-image translation. The key innovation here is adapting this framework to handle tabular data like gene expression and survival time.

Here's a breakdown of the approach:

  • Cycle Consistency: ctGAN uses two generators and two discriminators to learn transformations between two "domains" (e.g., breast cancer data and colon cancer data). The "cycle consistency" principle ensures that if you translate data from domain A to domain B and then back to A, you should recover the original data. This is crucial for generating realistic and faithful data.
  • Combined Transformation: Unlike previous methods that focused solely on gene expression, ctGAN explicitly models and transforms both gene expression and survival time data simultaneously.
  • Generator and Discriminator Architecture: The authors modified standard GAN architectures to better suit gene expression data. They incorporated residual blocks (similar to those in ResNet) with skip connections into the generator to improve feature propagation. The discriminator also uses a multi-layer structure with fully connected layers and leaky ReLU activations.
  • Loss Functions: The model employs a combination of GAN loss (using mean squared error for stability), cyclic loss, and identity loss to guide the training process.
  • Training Strategy: To address the inherent instability in GAN training, they employed a strategy where the generator is trained for more epochs than the discriminator.

Datasets

The primary dataset used in this study is The Cancer Genome Atlas (TCGA). The authors selected 12 specific cancer types from TCGA, including breast invasive carcinoma (BRCA), which had the largest sample size among the chosen types. They also used the SCAN-B dataset for an external validation of their model's effectiveness on a different breast cancer dataset.

Key Insights and Ideas

  • Beyond Gene Expression: The most significant insight is that augmenting both gene expression and survival data, rather than just gene expression, leads to more impactful improvements in survival analysis. This highlights the interconnectedness of these data types for predictive modeling.
  • Style Transfer for Data Augmentation: The paper demonstrates that "style transfer" between different cancer types is a powerful technique for data augmentation, effectively overcoming the limitations of small sample sizes in high-dimensional genomic data.
  • Plausibility Matters: The authors emphasize that the generated data isn't just random noise; it's highly plausible. This is supported by visualization (t-SNE) and clustering evaluations, showing that generated and reconstructed data closely resemble real data distributions.
  • Cancer-Specific Improvements: While the model shows broad applicability, the improvements are particularly pronounced for certain cancer types (e.g., COAD), suggesting that the effectiveness of style transfer can be cancer-type dependent.

Key Results

  • Improved C-index: ctGAN demonstrated improved C-index values in nine out of 11 cancer types compared to using only real data.
  • Significant Performance Gains: For COAD, the median C-index increased by approximately 15.70% when using ctGAN-generated data.
  • Enhanced Log-Rank P-value: For COAD, integrating generated data significantly improved the log-rank p-value from 0.797 to 0.041, indicating a more statistically significant distinction between high-risk and low-risk groups.
  • Outperforming Existing Methods: ctGAN outperformed previous style-transfer models (trVAE and stVAE) in seven out of 11 cancer types, with statistically significant improvements in several cases.
  • Clustering Performance: ctGAN achieved the highest or second-best performance in 89.62% of clustering evaluation metrics (KNN purity, NMI, ARI, Silhouette, Dunn index), demonstrating the high quality and plausibility of the generated data.

Limitations and Future Work

  • Computational Cost: Training GANs, especially cycle-consistent ones, can be computationally intensive. While the authors mention modifications to improve stability, GAN training remains a complex process.
  • Generalizability to Other Data Modalities: The current work focuses on gene expression and survival data. Future work could explore extending this framework to other omics data (e.g., proteomics, epigenomics) or clinical variables.
  • Clinical Validation: While the results are promising for survival analysis, further validation in real-world clinical settings and prospective studies would be necessary to fully realize its potential for personalized treatment selection.
  • Interpretability: Like many deep learning models, understanding why ctGAN makes specific predictions can be challenging. Future research could focus on improving the interpretability of the generated data and the model's decision-making process.

Summary updated

BibTeX
@article{kim2024ctgan,
  author = {Jaeyoon Kim and Junhee Seok},
  title = {{ctGAN: combined transformation of gene expression and survival data with generative adversarial network}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae325},
}
Figure