ABSTRACT: Brain circuits are assembled from a large variety of morphologically and functionally diverse cell types. It is not known how the intermingled cell types of individual brain regions differ in their expressed genomes. Here we describe an atlas of cell type transcriptomes of the adult retina. We found that each adult cell type expresses a specific set of genes, including a unique set of transcription factors, forming a “barcode” for cell identity. Cell type transcriptomes carry enough information to categorize cells into corresponding morphological classes and types. Surprisingly, several barcode genes are eye disease-associated genes that we demonstrate to be specifically expressed not only in photoreceptors but also in particular retinal circuit elements such as inhibitory neurons as well as in retinal microglia. Our data suggest that distinct cell types of individual brain regions are characterized by marked differences in their expressed genomes. To test whether amplifications were linear, we examined the relationship between gene expression values and the amounts of RNA in a cell group. We pooled together two retinas from the Chrnb4-GFP (cone) and six retinas from the ChAT-Cre × Rosa26-LSL-RFP (starburst amacrine cell) mouse lines. Five mixtures of GFP:RFP cells (“titration mixtures”) were then FACS-sorted with the following cone-starburst compositions: 0:200, 50:150, 100:100, 150:50 and 200:0. The total number of cells in each mixture was 200. Five mixtures from different mice were sorted 3 times to give biological triplicates. Next 20 genes with the highest specificity ratio for cones and another 20 for starburst cells were chosen and linear curves were fitted to the data pairs of expression value:cone number or expression value:starburst cell number. For both cone- and starburst-enriched genes, gene expression values increased linearly with an increase in the number of cones or starburst cells in the mixtures. The linear correlation coefficients were independent of the extent to which genes were expressed in the pure cone and starburst cell groups. This suggests that A practical use of the finding that amplification is linear is to separate the transcriptome of a cell group that contains cells from different types into its cell-type components. Assuming, for example, a cell group (Group A) containing a mixture of two cell types, Type 1 and Type 2, as well as a futher cell group (Group B) containing only Type 1 (Supplementary Figure 10). We wish to determine the transcriptome of Type 2 but we have no cell group (Group C) that contains Type 2 only. Or more generally, assume a cell group (Group A) containing a mixture of cell types, Type 1, Type 2…Type N, as well as a further cell group (Group B) containing only Type 1. We wish to determine the mixed transcriptome of Group C that contains Type 2….Type N, without Type 1. This separation of cell-type transcriptomes is useful for two reasons. First, ~80% of mouse retinal cells are rods30 and it is expected that transcripts from rods will be present in the solution after retina dissociation, leading to rod contamination of the transcriptome of each cell group. We wished to eliminate rod contamination. Second, we found that cones in some mouse lines are labeled together with a type of amacrine cell and we needed to determine the amacrine cell transcriptome alone. Due to linearity, the following procedure can be used to separate Type 1 from the rest of the cell types in a mixture (Group A). A set of cell type-specific genes for Type 1 is determined. This may be known a priori or from the dataset itself. The ratio of the expression value in Group A to the expression value in Group B is calculated for each Type 1-specific gene. The mean ratio is the estimate of the number of Type 1 cells in Group A. Next, the transcriptome (the expression values of all genes, which we treat as a vector) of Group B is multiplied by the mean ratio and the result is subtracted from the transcriptome of Group A. All negative numbers are then removed from this “unmixed” transcriptome by setting them to zero followed by normalization with the Group B (Type 1) transcriptome using quantile normalization. The quantile normalized “unmixed” transcriptome forms the prediction of the transcriptome of Group C. Note that if Group A contains only two cell types, Type 1 and Type 2, then the transcriptome of Group C is the transcriptome of Type 2 cells. We have demonstrated and quantitatively evaluated this procedure for the cone:starburst mixtures described above. In addition to the titration mixtures, we also made three “test mixtures” consisting the following cone:starburst cell numbers: 157:43, 75:125 and 183:17. Cones served as the Type 1 cells and starburst cells as Type 2. In our experiment, we had one pure cone group (200:0) and one pure starburst cell group (0:200) as well as six mixtures (50:150, 100:100, 150:50, 157:43, 75:125 and 183:17. First, we predicted the mixture composition using different numbers of cone-specific genes. Since we chose 20 cone-specific genes, the use of single genes presents 20 possible ways of predicting the number of cones in the mixture. Using two genes for the prediction provides 190 possible ways of predicting the number of cones in the mixture. The number of possible predictions when using k genes is: 20!/((20-k)! k!). When more than one gene is used for the prediction, the prediction is the mean of the individual gene-based predictions. Error was calculated as the mean of the absolute values of the deviation from the cone cell numbers in the mixtures. The mean and standard deviation of the prediction error decrease with the use of increasing number of genes. With 10 genes used for the prediction, for example, the prediction error will be approximately 12±3% (s.d.), regardless of which 10 genes are chosen, as long as they are cell type-specific. We pooled together two adult retinas from the Chrnb4-GFP (cone) and six adult retinas from the ChAT-Cre × Rosa26-LSL-RFP (starburst amacrine cell) mouse lines. Five mixtures of GFP:RFP cells (“titration mixtures”) were then FACS-sorted with the following cone-starburst compositions: ChAT.200: cone 0, ChAT.150: cone 50, ChAT.100: cone 100, ChAT.50 : cone 150, ChAT.0: cone 200. The total number of cells in each mixture was 200. Five mixtures from different mice were sorted 3 times to give biological triplicates (indicated by _1, _2, _3). X1, X2 and X3 stand for the following ratios: X1: ChAT.43 : cone 157, ChAT.125 : cone 75, ChAT.17 : cone 183. Arc represents as internal batch control to compare with other gene arrays within our dataset.
