Fig. 4

SCA outperforms many other methods, both general and problem-specific, at rare and subtly defined cell type discovery on a large PBMC dataset with ground-truth cell labels. a UMAP plots computed from PCA, ICA, and SCA reductions of T cell scRNA-seq data for patient 1 from the Hao et al. [6] dataset. Cell type labels were determined by the original authors via parallel screening of 228 antibodies using CITE-seq. b Adjusted mutual information (AMI) of true cell labels with clusters output by each of the 11 methods tested in each patient (FiRE does not output a clustering but a rareness score and thus is not amenable to AMI analysis). For PCA, ICA, SCA, and scVI, we perform Leiden clustering with resolution 1.0 after reduction. SCA-based clusterings consistently have higher AMI with the true labels. c F1 scores for recovery of all T cell subtypes across all 8 patients of the dataset from Hao et al., from PCA, ICA, and SCA followed by Leiden clustering with resolution 1.0, and from nine other methods. For each clustering and cell type, the set of clusters best identifying that cell type was selected, and the F1 score reported