Benchmarking Adversarial Robustness of Deep Learning Models for Perturbation Response Prediction (e.g., scGen, CellOT)

The ability to predict single-cell perturbation responses—how individual cells transcriptionally react to drugs, genetic edits, or environmental changes—is a critical frontier in computational biology. Deep learning models such as scGen and CellOT have emerged as promising frameworks for this task, leveraging latent space mappings and optimal transport to model counterfactual cellular states. However, the practical utility of these models in high-stakes applications (e.g., drug discovery or toxicology) depends not only on their predictive accuracy but critically on their adversarial robustness: their stability and reliability when confronted with small, worst-case input perturbations, noisy measurements, or out-of-distribution samples. Despite the proliferation of perturbation response models, no standardized benchmark exists to evaluate their vulnerability to adversarial attacks. This work presents the first systematic benchmarking framework for assessing the adversarial robustness of deep perturbation response predictors. Using scGen and CellOT as representative architectures, we evaluate model performance under various threat models, including gradient-based attacks on gene expression inputs and latent space interventions. We measure robustness via task-relevant metrics: preservation of biological state identities, fidelity of counterfactual generation, and collapse of out-of-distribution robustness. Our results reveal that both models exhibit significant vulnerability to imperceptible input perturbations, leading to biologically implausible or entirely inverted perturbation responses. Furthermore, we find that standard robustness techniques (adversarial training, input denoising) trade off between nominal accuracy and robust generalization. This benchmark establishes critical safety-aware evaluation protocols for perturbation response models and provides actionable insights for developing more reliable and trustworthy deep learning methods in single-cell genomics.