We expected about 2% of all cells to be doublets. Preliminary exploratory analysis (including by generating simulated doublets) showed that most doublets would either form separate clusters, or would tend to end up at the fringes of other clusters (in graph embeddings, and in t-SNE). To eliminate many doublets, we (1) removed clusters classified with ambiguous labels; (2) removed cells classified with a different label from the majority of cells in its clusters; (3) removed outliers when clustering, typically on the fringes of clusters in t-SNE space.
