Figure 3
From: Northstar enables automatic classification of known and novel cell types from tumor samples
Performance on large datasets and comparison with scVI and scmap on murine droplet kidney data. (A) Runtimes and percentage of correctly classified cells with increasing numbers of cells in the new dataset. (B) Runtime and (C) percentage of correctly assigned cells for northstar, scVI, and scmap with an incomplete atlas containing only a fraction of the 18 cell types. Northstar is much faster since it does not need to train a deep neural network. Northstar is also more accurate because of its atlas-aware clustering step. All algorithms used the same input data (1,766 cells as test, spread evenly across cell types). Northstar’s accuracy surpasses 90% as the atlas becomes more complete. (D) Typical confusion matrix of a northstar run with an incomplete atlas [star in (B) and (C)]. Most cells are classified correctly (green dots) or assigned to similar cell types (yellow dots), while a small number of cells are misclassified into a distinct cell type (red dots). Subsamples of the kidney droplet data from Tabula Muris Senis11 with 20 cells per type were used as atlas.
