We downloaded raw read or UMI matrices for all datasets, from their respective sources. The one exception was the 3pV1 dataset from the PBMC analysis. These data were originally quantified with the hg19 reference, while the other two PBMC datasets were quantified with GRCh38. Thus, we downloaded the fastq files from the 10X website (Supplementary Table 8 ). We quantified gene expression counts using Cell Ranger11 ,41 v2.1.0 with GRCh38. From the raw count matrices, we used a standard data normalization procedure, laid out below, for all analyses, unless otherwise specified. Except for the L2 normalization and within-batch variable gene detection, this procedure follows the standard guidelines of the Seurat single cell analysis platform.
We filtered cells with fewer than 500 genes or more than 20% mitochondrial reads. In the pancreas datasets, we filtered cells with the same thresholds used in Butler et al7 : 1750 genes for CelSeq, 2500 genes for CelSeq2, no filter for Fluidigm C1, 2500 genes for SmartSeq2, and 500 genes for inDrop. We then library normalized each cell to 10,000 reads, by multiplicative scaling, then log scaled the normalized data. We then identified the top 1000 variable genes, ranked by coefficient of variation, within in each dataset. We pooled these genes to form the variable gene set of the analysis. Using only the variable genes, we mean centered and variance 1 scaled the genes across the cells. Note that this was done in the aggregate matrix, with all cells, rather than within each dataset separately. With these values, we performed truncated SVD keeping the top 30 eigenvectors. Finally, we multiplied the cell embeddings by the eigenvalues to avoid giving eigenvectors equal variance.
We filtered cells with fewer than 500 genes or more than 20% mitochondrial reads. In the pancreas datasets, we filtered cells with the same thresholds used in Butler et al7 : 1750 genes for CelSeq, 2500 genes for CelSeq2, no filter for Fluidigm C1, 2500 genes for SmartSeq2, and 500 genes for inDrop. We then library normalized each cell to 10,000 reads, by multiplicative scaling, then log scaled the normalized data. We then identified the top 1000 variable genes, ranked by coefficient of variation, within in each dataset. We pooled these genes to form the variable gene set of the analysis. Using only the variable genes, we mean centered and variance 1 scaled the genes across the cells. Note that this was done in the aggregate matrix, with all cells, rather than within each dataset separately. With these values, we performed truncated SVD keeping the top 30 eigenvectors. Finally, we multiplied the cell embeddings by the eigenvalues to avoid giving eigenvectors equal variance.
