Curated, multi-omic, ML-driven single-cell atlas for characterizing the human immune system across disease states

Abstract

Abstract The immune system plays a key role in various diseases such as cancer, autoimmune disorders and infections. It has been recently shown that cell surface protein markers on peripheral blood mononuclear cells (PBMCs) can be predictive of response to therapies1, but measurements of such markers is technically limited in the number of cells and markers that can be assessed. Single-cell RNA technology allows the measurement of the entire transcriptome in tens of thousands of cells, and therefore has the potential to survey the immune system in more depth and reveal rare cell states. However, no large single-cell datasets for immune cell states exist. Moreover, attempts to combine smaller distinct datasets are often hindered by batch effects stemming from differences in sample handling, single-cell technologies and computational analyses. We created a large curated multi-omic single-cell human PBMC atlas with clinical annotations from dozens of patients with several conditions across hundreds of thousands of cells. Our standardized end-to-end protocols and quality control processes provide a platform that allows production of large datasets with minimal batch effects. By combining single-cell RNA-seq (scRNA-seq) with surface marker identification by CITE-seq, we generate curated immune gene signatures and train a classifier to robustly identify many cell types and states across patients and different diseases, including rare cell populations. This curated, multi-omic clinically-annotated atlas is particularly suited for use by machine learning algorithms, and we are confident that as this data accumulates, it will be instrumental for inferring disease states and predict responses to various therapies.