Adversarially Robust Transformer Models for Analyzing COVID-19 Single-Cell RNA-seq Data

Single-cell RNA sequencing is a technology that measures gene activity in thousands of individual cells from each patient. These datasets are extremely detailed but also very large, noisy, and difficult to analyze. Differences in data collection across studies can further complicate analysis, making it challenging to build models that work reliably for different patient groups. In this project, we developed an artificial intelligence (AI) method to help interpret single-cell data more accurately. The method uses a type of deep learning model called a Transformer, combined with adversarial training techniques that make the model more robust to noise and technical variation. Robustness is important because single-cell data often contain many zeros, measurement errors, and batch effects, which can cause traditional models to perform poorly. We applied this method to two publicly available COVID-19 single-cell datasets. The model was able to predict which patients had COVID-19 with high accuracy, even when the data came from different sources. The approach also provides a way to examine which groups of cells contribute most to the predictions, offering an interpretable view of the patterns the model learns from the data. This work demonstrates how modern AI tools can improve the analysis of complex single-cell datasets. Although the project focuses on COVID-19 data, the methodology can be applied to other studies that use single-cell sequencing and face similar challenges with data quality and variability.