Unlocking Immunological Insights: An Interpretable Deep Learning Approach to scRNAseq Analysis and Investigating Neonate Homeostatic Proliferation through Immune Profiling
- Davalos, Oscar Alejandro
- Advisor(s): Hoyer, Katrina K
Abstract
Currently, many deep learning approaches for single-cell RNA sequencing (scRNAseq) analysis suffer from a lack of interpretability. We present scANNA (single-cell ANalysis using Neural-Attention), a deep-learning cell annotation model that provides interpretability through neural attention. Through neural attention, scANNA is able to provide clear explanations regarding the reasoning behind its predictions. The generated attention weights enable us to comprehend the model's decision-making process. Our model's high accuracy is attributed to its neural attention architecture, which outperforms the capabilities of current state-of-the-art models. We conducted tests using viral infection and cancer datasets to assess scANNA's ability to handle diverse immunological datasets and ensure its reliability and effectiveness. ScANNA simplifies scRNAseq analysis for researchers, eliminating the need for specialized training and task-specific models. It saves time and enhances the research process. Neonatal immune homeostatic proliferation can lead to potential risks of immune dysregulation and related disorders in later stages of life. We analyzed CD8 T cells in neonatal mice to investigate gene expression patterns, subpopulations, and differentiation trajectories. We identified a subpopulation of CD8 T cells called Virtual Memory cells (TVM) that may play an important role in neonatal immune homeostasis and disease progression. Understanding the mechanisms of CD8 T cells during immune homeostatic proliferation will deepen our understanding of autoimmune disorders and the regulation of homeostasis.