adverSCarial: assessing the vulnerability of single-cell RNA-sequencing classifiers to adversarial attacks

Several machine learning (ML) algorithms dedicated to the detection of healthy and diseased cell types from single-cell RNA sequencing (scRNA-seq) data have been proposed for biomedical purposes. This raises concerns about their vulnerability to adversarial attacks, exploiting threats causing malicious alterations of the classifiers' output with defective and well-crafted input.

With adverSCarial, adversarial attacks of single-cell transcriptomic data can easily be simulated in a range of ways, from expanded but undetectable modifications to aggressive and targeted ones, enabling vulnerability assessment of scRNA-seq classifiers to variations of gene expression, whether technical, biological, or intentional. We exemplify the usefulness and performance with a panel of attack modes proposed in adverSCarial by assessing the robustness of five scRNA-seq classifiers, each belonging to a distinct class of ML algorithm, and explore the potential unlocked by exposing their inner workings and sensitivities on four different datasets. These analyses can guide the development of more reliable models, with improved interpretability, usable in biomedical research and future clinical applications.

adverSCarial is a freely available R package accessible from Bioconductor: https://bioconductor.org/packages/adverSCarial/ or https://doi.org/10.18129/B9.bioc.adverSCarial. A development version is available at https://github.com/GhislainFievet/adverSCarial.