a, Prefiltering of doublet-like cells. UMAP embedding shows all original E9.5 cells (n = 5,646), coloured by doublet score. Doublet scores indicate the extent to which a given single-cell transcriptome resembles a linear combination of two randomly selected cells (Methods and ref. 39). b, Histogram of doublet scores. Scores >0.24 were filtered from subsequent analyses. c, t-SNE embedding of E9.5 cells (n = 4,367) post-doublet filtering. Individual cells are coloured according to annotated Louvain cluster identities. d, Top 20 positively enriched transcripts for each Louvain cluster relative to all other clusters, as detected by a two-sided Wilcoxon rank-sum test. Reported transcripts are ranked by FDR-corrected P values (Benjamini–Hochberg). Exact sample sizes are given in Supplementary Table 1.