Fig. 5

Imputation performance using SAVER, MAGIC and our SCA-MAGIC. a Recovery of marker genes on the Splatter dataset analyzed in Fig. 2b–e. For each method, we measure the average correlation between the marker genes after imputation and an indicator vector for the rare cells. MAGIC achieves significantly higher correlation using SCA as a base embedding (SCA-MAGIC). b Visualizing gene-gene relationships in cytotoxic T cell data after imputation using SAVER, MAGIC or our SCA-MAGIC. SCA-MAGIC better recovers the inverse relationships between CD8 and CD4 and between granzyme B and granzyme K, with gamma-delta T cells expressing neither CD8 nor CD4 and lower granzyme expression for T helper cells. c Scatter plot showing dropout recovery in the cytotoxic T cell dataset at various dropout rates. A fixed percentage of nonzero transcript measurements were set to zero, and the mean imputed values of these removed transcripts were assessed for each gene. While MAGIC and SCA-MAGIC perform similarly on most genes, SCA-MAGIC consistently performs better on a subset of them, measured by dropout rate. d A closer look at the genes where SCA-MAGIC significantly outperforms MAGIC in recovering dropouts at the 90% dropout rate. They include many key marker genes such as CD8A, CD8B, CD4, emphTIGIT, and TRDV2