Extended Data Figure 7: Library complexity and fraction of X chromosome reads highlights clusters of collisions between cells from different tissues.

Density plots of the estimated library complexity (using the same equation implemented in Picard; left) and the representation of X-chromosome reads (right) in individual clusters. While most of the clusters defined by t-SNE are readily biologically interpretable, a small number of clusters (containing relatively few cells) were not easily characterized and are marked by an increase in both estimated library complexity and an unusual distribution of X chromosome to autosomal reads. These clusters are likely to be clusters of collisions; that is, cases in which two or more distinct cells share the same barcode as a consequence of the combinatorial indexing protocol. The black line is the global distribution for all cells in that time point. The grey lines show the results of randomly sampling an equal number of cells to the cluster in question. The coloured line marks the distribution for the cluster being interrogated. a, c, e, Most clusters show relatively similar distributions of library complexity (left) and a characteristic, bimodal distribution among cells in the ratio of X-chromosome to autosomal reads (reflecting our use of a pool of male (XY) and female (XX) embryos, right). b, d, Putative collision clusters show a clear increase in the average library complexity (left) and a unimodal rather than bimodal distribution of X-chromosome to autosomal reads (right). f, These features are not universally diagnostic (for example, cluster 10 at 2–4 h seems to show a strong, bona fide sex bias), but the combination of features is strongly predictive of clusters containing few cells and conflicting biological annotations based on gene or enhancer overlaps.