Generative Adversarial Networks (GANs) for Robustness Testing of Transcriptomic Perturbation Signatures

Robustness testing of transcriptomic perturbation signatures is critical for ensuring reliable biomarker discovery and drug response prediction, yet traditional methods often fail to account for complex, high-dimensional noise and batch effects. Here, we propose a novel framework based on Generative Adversarial Networks (GANs) to systematically evaluate and enhance the robustness of transcriptomic signatures. The GAN generator learns to produce realistic perturbations of gene expression profiles, while the discriminator distinguishes authentic biological signals from noise-induced or artifactual variations. By adversarially challenging perturbation signatures with increasing levels of synthetic but biologically plausible noise, our approach quantifies signature stability, identifies vulnerable features, and generates augmented training data to improve model generalization. Applied to publicly available transcriptomic datasets, the GAN-based robustness test reveals previously undetected signature degradation under moderate technical noise and batch shifts, outperforming conventional perturbation methods. This framework provides a scalable, unsupervised strategy for stress-testing and refining transcriptomic signatures, with direct implications for precision medicine and toxicogenomics.