Article Figures & Data
Figures
- Figure S1. Overview of the experimental setup.
(A) Barcoding scheme across N rounds of multicolor imaging. Each detected fluorescent spot is registered as a “1” in its corresponding bit position. Then, RNA species are identified by matching the barcodes to a preselected gene codebook. (B) Experimental workflow. First, the gene panel and codebook are designed. Then, probes are hybridized onto the sample, which has been mounted and fixed onto a coverslip that is in turn placed in a flow cell connected to an automated fluidic system. The sample is imaged in five color channels at multiple fields of view and z-positions using an inverted epi-fluorescence microscope. Images are analyzed to identify RNA molecules and their spatial coordinates as well as cell boundaries. These data result in single-cell count matrices that can then be analyzed with standard bioinformatics tools to achieve goals such as cell type identification.
- Figure S2. Investigation of RNA tissue quality on multiplexed error-robust fluorescence in situ hybridization measurements.
Average transcript density is shown as a function of the tissue RIN score for each tissue sample. There was a clear correlation between the RIN score and transcript density up until a RIN score of about 5, after which the transcript density plateaued. Thus, we opted to exclude datasets before the plateau and did not consider the kidney and pancreas multiplexed error-robust fluorescence in situ hybridization datasets with the RIN score below four in this study.
- Figure 1. Sample multiplexed error-robust fluorescence in situ hybridization data acquisition workflow.
(A) Low-resolution image of the DAPI channel for a mouse kidney tissue sample. Red box indicates zoomed-in region displayed in B-E. (B, C) DAPI (B) and (C) cell boundary antibody stain channels for the zoomed-in region. (D) Multiplexed error-robust fluorescence in situ hybridization signal for a single barcode bit channel in the same zoomed-in region. (E) Positions of decoded mRNA transcripts (colorful dots) and segmented cell boundaries (black) in the same zoomed-in region after running data through image analysis pipeline, with each color representing a unique gene species. Panels (B, C, D) show raw images with brightness and contrast levels selected for ease of visualization.
- Figure 2. Comparison of multiplexed error-robust fluorescence in situ hybridization (MERFISH) data with bulk RNA-seq from Tabula Muris Senis.
(A, B) Bulk RNA counts per gene in the mouse (A) liver and (B) kidney between MERFISH sample replicates. (C, D) Comparison of bulk RNA counts per gene in the mouse (C) liver and (D) kidney between MERFISH and RNA-seq. (E, F) Comparison of bulk MERFISH RNA counts per gene with pseudo-bulk RNA counts from Visium in the mouse (E) liver and (F) kidney. In (C, D, E, F), counts per gene were averaged across replicates for each technology.
- Figure 3. Quality control of single-cell multiplexed error-robust fluorescence in situ hybridization data.
(A) Sample cropped image of mouse liver tissue with segmented cell boundaries (white, magenta), decoded RNA transcripts (gray), DAPI nuclear stain (blue), and cell boundary antibody stain (green). White boundaries indicate cells that have passed the filtering stage, whereas magenta boundaries indicate cells that have been thrown out. (B, C, D) Histograms of (B) cell areas, (C) RNA transcript count per cell, and (D) average DAPI per cell for the image in (A). Magenta indicates minimum or maximum cutoff values for filtering criteria. (E) Sample cropped image of mouse kidney tissue. (F, G, H) Histograms of (F) cell areas, (G) RNA transcript count per cell, and (H) average DAPI per cell for the image in (E). White arrows in (A) and (E) indicate examples of poor cell segmentation.
- Figure 4. Comparison of single-cell multiplexed error-robust fluorescence in situ hybridization (MERFISH) data with Tabula Muris Senis.
(A, B) Histograms of total RNA transcript count per cell in the mouse (A) liver and (B) kidney for the gene panel, for MERFISH (blue) and scRNA-seq (orange). (C, D) Histograms of dropout rates per cell in the mouse (C) liver and (D) kidney for the gene panel. (E, F) Comparison between MERFISH and scRNA-seq of per-gene fraction of total cells of the whole population of each dataset that have nonzero counts for the mouse (E) liver and (F) kidney. Here, each dot represents a single gene. (G, H) Mean transcript count of pancreas marker genes in the mouse (G) liver and (H) kidney, for MERFISH (blue) and scRNA-seq (orange). A liver marker gene (Hmgcs2) and kidney marker gene (Kcnj1) are shown for comparison. Standard error of the mean across cells was negligible and too small to visualize.
- Figure S3. Investigation of molecular crowding in multiplexed error-robust fluorescence in situ hybridization measurements.
(A, B) Distribution of median counts per gene across cells with nonzero counts in the mouse liver (A) and kidney (B). Text annotations highlight the top two most abundant genes in each tissue. (C, D) Distribution of counts per cell for top two most abundant genes in the mouse liver (C) and (D) kidney, in cells that registered over 100 total counts (top) or in cells that registered under 100 total counts (bottom).
- Figure 5. Statistical analysis of single-cell multiplexed error-robust fluorescence in situ hybridization (MERFISH) and Tabula Muris Senis RNA count distributions.
(A, B) Comparison of mean RNA counts in MERFISH versus scRNA-seq in mouse (A) liver and (B) kidney across cells with nonzero counts. Each dot represents a gene, and only genes possessing at least 50 cells with nonzero counts are shown for each tissue. The dotted black line shows the x = y line, indicating one-to-one concordance between MERFISH and scRNA-seq. (C, D) Mean–variance relationship for RNA counts in MERFISH (blue) and scRNA-seq (orange) for the mouse (A) liver and (B) kidney, across cells with nonzero counts. Each dot represents a gene, and only genes possessing at least 50 cells with nonzero counts are shown for each tissue. The dotted line indicates the x = y line, which represents the Poisson scenario where the mean equals the variance. (C, inset) Distribution of MERFISH RNA counts across cells with nonzero counts for Hmgcs2, a liver hepatocyte marker gene. The red line indicates the best fit to a negative binomial distribution. (D, inset) Distribution of MERFISH RNA counts across cells with nonzero counts for Kcnj1, a kidney loop-of-Henle cell marker gene. Red line indicates the best fit to a negative binomial distribution. Error bars in (A) and (B) represent the standard error of the mean.
- Figure 6. Single-cell and spatial analysis of the multiplexed error-robust fluorescence in situ hybridization (MERFISH) liver sample.
(A) UMAP plots of MERFISH data colored by manually annotated clusters or by normalized, log-transformed, and scaled expression of example marker genes. (B) Spatial plot of MERFISH dataset alone, colored by manually annotated cell types in (A). Inset highlights spatial co-localization of periportal endothelial cells and hepatocytes (gray and brown, respectively) and pericentral endothelial cells and hepatocytes (olive and pink, respectively). (C) DAPI stain of the liver sample. White box indicates the same inset region as in panel (C). (D) Spatial plot of MERFISH dataset using scANVI predicted cell type labels. Legend is the same as in panel (B). (E) Cell type composition for scRNA-seq and MERFISH datasets. Each point in (A, B, D) represents a single cell. (F) Confusion matrix of MERFISH cell type annotations between the manual method and scANVI predictions. The rows for other endothelial cells and hepatocytes (“o-EC” and “o-hep”) are blank because none of the manual annotations were in these groups. In addition, the row for bile duct epithelial cells is noisy because none of the scRNA-seq reference cells possessed this annotation. Cell type abbreviations are as follows: “IC,” “immune cell”; “o-EC,” “other endothelial cell”; “KC,” “Kupffer cell”; “HSC,” “hepatic stellate cell”; “o-hep,” “other hepatocyte”; “PP-hep,” “periportal hepatocyte”; “PC-hep,” “pericentral hepatocyte”; “PP-EC,” “periportal endothelial cell”; “PC-EC,” “pericentral endothelial cell”; “BD-EC,” “bile duct epithelial cell.”
- Figure S4. Co-occurrence probability analysis of periportal and pericentral hepatocytes and endothelial cells.
(A, B, C, D) Co-occurrence probability ratio of (A) pericentral endothelial cells, (B) pericentral hepatocytes, (C) periportal endothelial cells, and (D) periportal hepatocytes with other periportal or pericentral endothelial cells or hepatocytes. Cell type abbreviations are as follows: “PP-hep,” “periportal hepatocyte”; “PC-hep,” “pericentral hepatocyte”; “PP-EC,” “periportal endothelial cell”; “PC-EC,” “pericentral endothelial cell.” (A, B) For small length scales (<500 μm), pericentral endothelial cells and hepatocytes tend to co-localize with each other with R > 1 (A, B). (C, D) This also holds true for periportal endothelial cells and hepatocytes (C, D). In contrast, periportal and pericentral cells do not co-localize, with R < 1. At larger length scales (>500 μm), the co-occurrence probabilities trend to R = 1, indicating no strong co-localization effects.
- Figure S5. Integration of multiplexed error-robust fluorescence in situ hybridization (MERFISH) mouse liver and kidney data with Tabula Muris Senis scRNA-seq using scVI and scANVI.
(A) UMAP plot of MERFISH liver data (gray) with Tabula Muris Senis (teal) using scVI. Surprisingly, the two technologies are not particularly well mixed, suggesting that the underlying statistics of the two modalities are systematically different enough that constructing a good joint embedding may be difficult. (B) UMAP plot of liver integration in (A) with cell type annotations predicted using scANVI. (C) UMAP plot of MERFISH kidney data (gray) with Tabula Muris Senis (teal) using scVI. Just like the liver data, the kidney data fails to exhibit good mixing of the two modalities. (D) UMAP plot of kidney integration in (C) with cell type annotations predicted using scANVI. Cell type abbreviations in (B) are as follows: “IC,” “immune cell”; “o-EC,” “other endothelial cell”; “KC,” “Kupffer cell”; “HSC,” “hepatic stellate cell”; “o-hep,” “other hepatocyte”; “PP-hep,” “periportal hepatocyte”; “PC-hep,” “pericentral hepatocyte”; “PP-EC,” “periportal endothelial cell”; “PC-EC,” “pericentral endothelial cell.” Cell type abbreviations in (D) are as follows: “EC-PT,” “epithelial cell of proximal tubule”; “IC,” “immune cell”; “per,” “pericyte”; “KLH-EC,” “kidney loop-of-Henle epithelial cell”; “KCD-EC,” “kidney collecting-duct epithelial cell”; “KDCT-EC,” “kidney distal-convoluted-tubule epithelial cell”; “EC,” “endothelial cell,” “pod,” “podocyte.”
- Figure S6. Integration of multiplexed error-robust fluorescence in situ hybridization (MERFISH) mouse liver and kidney data with Tabula Muris Senis scRNA-seq using Harmony and Symphony.
(A) Spatial plot of MERFISH mouse liver dataset, colored by cell type annotations predicted by Symphony. (B) Spatial plot of MERFISH mouse kidney dataset, colored by cell type annotations predicted by Symphony. Qualitatively, the results are similar to the scANVI results (Figs 6D and 7D), reproducing the structure of the liver (Fig 6D) and the artifactual population of podocytes in the kidney medulla (Fig 7D). (C) Confusion matrix of MERFISH cell type annotations between the manual method and Symphony predictions for the mouse liver. The rows for other endothelial cells and hepatocytes (“o-EC” and “o-hep”) are blank because none of the manual annotations were in these groups. In addition, the row for bile duct epithelial cells (“BD-EC”) is noisy because none of the scRNA-seq reference cells possessed this annotation. (D) Confusion matrix of MERFISH cell type annotations between the manual method and Symphony predictions for mouse kidney. Like with scANVI (Figs 6F and 7F), Symphony generally is consistent with manual annotation, as can be seen by the higher values along the main diagonal. However, Symphony produces slightly more off-diagonal values than scANVI and also misclassifies podocytes and pericytes (Fig S6D) just like scANVI.
- Figure 7. Single-cell and spatial analysis of multiplexed error-robust fluorescence in situ hybridization (MERFISH) kidney sample.
(A) UMAP plots of MERFISH data colored by manually annotated clusters or by normalized, log-transformed, and scaled expression of example marker genes. (B) Spatial plot of MERFISH dataset alone, colored by manually annotated cell types in (A). Inset shows an example of a podocyte cluster in the kidney cortex region (olive, black arrow). (C) DAPI stain of kidney sample. White box indicates the same inset region as in panel (C). (D) Spatial plot of MERFISH dataset using scANVI predicted cell type labels. Black arrow indicates falsely predicted podocytes distributed in a ring-like structure around the medulla. Legend is the same as in panel (B). (E) Cell type composition for scRNA-seq and MERFISH datasets. Each point in (A, B, D) represents a single cell. (F) Confusion matrix of MERFISH cell type annotations between the manual method and scANVI predictions. Cell type abbreviations are as follows: “EC-PT,” “epithelial cell of proximal tubule”; “IC,” “immune cell”; “per,” “pericyte”; “KLH-EC,” “kidney loop-of-Henle epithelial cell”; “KCD-EC,” “kidney collecting-duct epithelial cell”; “KDCT-EC,” “kidney distal-convoluted-tubule epithelial cell”; “EC,” “endothelial cell”; “pod,” “podocyte.”
- Figure S7. Investigation of systematic differences in RNA expression profiles between multiplexed error-robust fluorescence in situ hybridization (MERFISH) and scRNA-seq.
(A) Distribution of mean cosine similarity values using the full 307-gene panel between manually annotated MERFISH kidney podocytes and cells in the scRNA-seq kidney reference labeled as podocyte (blue) or endothelial cell (orange). (B) Distribution of mean cosine similarity values using the full 307-gene panel between manually annotated MERFISH liver periportal hepatocytes and cells in the scRNA-seq liver reference labeled as periportal hepatocyte (blue) or pericentral hepatocyte (orange). (C) Same as (A) but only using seven podocyte marker genes (Wt1, Actn4, Synpo, Dag1, Foxc1, Podxl, Mme). (D) Same as (B) but only using four periportal hepatocyte marker genes (Cyp2f2, Pck1, Hal, and Cdh1). Mean cosine similarity values in (A, B, C, D) were obtained by calculating the cosine similarity between each individual cell in a pair (e.g., MERFISH podocyte versus scRNA-seq podocyte) and then averaging across the scRNA-seq cells, yielding a mean cosine similarity value for each MERFISH cell and a distribution of mean cosine similarity values across MERFISH cells.
In this Issue
Subjects
Related Articles
Cited By...
- Accurate single-molecule spot detection for image-based spatial transcriptomics with weakly supervised deep learning
- Increased spatial coupling of integrin and collagen IV in the immunoresistant clear cell renal cell carcinoma tumor microenvironment
- Feasibility of Inferring Spatial Transcriptomics from Single-Cell Histological Patterns for Studying Colon Cancer Tumor Heterogeneity
- A high-resolution transcriptomic and spatial atlas of cell types in the whole mouse brain