Deep Learning for Accurate Diagnosis of Viral Infections through scRNA-seq Analysis: A Comprehensive Benchmark Study

Abstract

This presentation examines a comprehensive benchmark study of deep learning approaches for diagnosing viral infections through single-cell RNA sequencing (scRNA-seq). Using a dataset of more than 200,000 immune cells across COVID-19, influenza, and dengue, the study tested models including PCA, SAVER, scVI, scGPT, and contrastiveVI. The pipeline incorporated preprocessing, batch control, and dimensionality reduction for robust model evaluation. Performance was assessed using clustering metrics (ARI, NMI), neighborhood consistency (k-NN accuracy), and structural cohesion (S-score). Results indicate contrastiveVI as the most accurate model for infection classification, with scGPT providing strong batch harmonization capabilities. UMAP projections were used to validate clustering visually. These findings underscore the role of batch control and careful model selection in building reliable AI pipelines for infection-specific scRNA-seq diagnostics, offering methodological insights for future biomedical applications.