Skip to content

Latest commit

 

History

History
38 lines (29 loc) · 2.87 KB

02-general.md

File metadata and controls

38 lines (29 loc) · 2.87 KB

General

The cell-type annotation framework implemented in this project consists on supervised classification using two approaches:

  • Azimuth: This method implemented in Seurat V4 classifies by cells by mapping the query dataset onto a supervised PCA space constructed from modal-weighted neighbors calculated by combining RNA and protein data.

  • Hierarchical scPred: This approach is based on two methods:

    • scPred**:** a supervised classification method that learns a low dimensional representation of a reference dataset. Unlike azimuth (which relies on a lazy classifier), scPred trains binary classifiers for each cell type using a one vs. all approach. The query dataset is projected onto the training low dimensional space and labels are assigned according to probability scores. For more details see https://genomebiology.biomedcentral.com/articles/10.1186/s13059-019-1862-5

    • Hierarchical progressive learning (scHPL): this method decomposes the classification problem into smaller sub-tasks according to a hierarchy representing cell-type relationships. Cells are classified progressively using a top-bottom approach instead of using a flat classifier. For more details see https://www.nature.com/articles/s41467-021-23196-8

The cell-type classification is guided using a CITE-seq reference containing:

  • 161,764 human PBMCs
  • 20,953 genes
  • 224 protein markers (CITE-seq)
  • 31 cell types
Lymphoid populations Myeloid populations
CD4 Naive, CD4 TCM, CD4 TEM, CD4 CTL, Treg, CD4 Proliferating CD14 Mono, CD16 Mono
CD8 Naive, CD8 TCM, CD8 TEM, CD8 Proliferating cDC1, cDC2, pDC, ASDC
gdT, MAIT, ILC, dnT HSPC
NK, NK_CD56bright, NK Proliferating Platelet, Eryth
B naive, B intermediate, B memory, Plasmablast Doublet

"This PBMC reference dataset was generated as part of the Hao and Hao et al, Cell 2021 paper. It is comprised of data from eight volunteers enrolled in an HIV vaccine trial from which three time point samples were taken at day 0, 3, and 7 following vaccination. All 24 samples were processed with a CITE-seq panel of 228 TotalSeq A antibodies to generate single-cell RNA and ADT data. The data were then integrated using metholody described in the pre-print linked above to generate a weighted nearest neighbor (WNN) representation of the RNA and protein data jointly. This WNN representation is used in the Azimuth app to assign celltypes, embed in the reference UMAP, and impute protein levels for the query dataset". Hao Y, et al. Cell. (2021)