Gene expression measured in individual cells from the human pancreas, published by Baron et al. (2016) under GEO accession GSE84133. The archive holds count tables for human and mouse samples; this page follows the 8,569 cells from four human donors.
It traces those tables to the 8,569 × 2,000 design matrix used by the Workflow page From gene counts to cell groups: one row per cell, one column per selected gene. How the cells were captured and sequenced is taken from the study’s description; the preparation steps are recomputed here from the count tables, which already contain processed gene counts rather than sequencing reads.
Baron M, Veres A and colleagues
- DocumentedPancreatic cells → RNA sequencing
GEO GSE84133 documents inDrop capture and Illumina HiSeq 2500 sequencing.
- DocumentedReads → released gene counts
GEO supplies the processed count tables; read alignment and counting were done by the study.
- HeldDonor count tables
6 compressed tables; the 4 human tables are used.
- ReproducedNormalise → saved panel → scale
8,569 human cells, 2,000 variable genes, gene-wise scaling and clipping.
How to read the labels: Documented — described by the study. Held — files this page has. Reproduced — recomputed here from those files. Not recorded — the files do not show how the step was done.
What the instrument measures
The study coupled inDrop microfluidic cell capture with paired-end sequencing on an Illumina HiSeq 2500. Cellular RNA was barcoded before pooled sequencing, allowing expression to be assigned back to individual cells.
Schematic of the measurement steps described by the study.
Published acquisition protocol
Documented Baron et al. (2016), STAR Methods; GEO GSE84133. Protocol quantities below are from this source.
| Stage | Protocol |
|---|---|
| Sample preparation | Human islets from NDRI recovered in CMRLS for 24 or 48 hours. TrypLE Express dissociation used 37 °C for 18–20 minutes. |
| Cell capture | Before inDrop, the suspension was adjusted to 75,000 cells/mL with 16% Optiprep. Reported encapsulation throughput was 12,000 cells/hour. |
| Library preparation | Reverse transcription and linear amplification preceded library enrichment with 9–12 PCR cycles. |
| Sequencing | Illumina HiSeq 2500, paired-end reads: read 1 carries cell-identifying sequences; read 2 supplies the transcript sequence for mapping. |
| Reads to expression | Reads were screened for expected barcode/adaptor sequences, trimmed with Trimmomatic 0.32, assigned to cells, and mapped with Bowtie 1.1.1 for quantification. |
These steps are described by the study. The archive starts from the count tables they produced.
One acquisition unit: one captured cell
Each measurement unit is one captured cell. For each cell, the published tables give a barcode, a cell-type annotation and its gene counts. A donor table contains many cells.
What the fields mean
| Field | Unit | Meaning |
|---|---|---|
| Source row name | Identifier | Cell identity including donor and library information. |
| Barcode | Sequence label | Captured-cell barcode; use the source row name for a unique identifier. |
| Cell annotation | Category | Cell type assigned by the study; a label, not a measured expression value. |
| Gene-name columns | UMI count | Captured transcript counts assigned to each gene in a cell. |
In the files: 8,569 human cells × 20,125 gene counts, across 4 donor tables.
| File | Bytes | Rows | Selection |
|---|---|---|---|
| GSM2230757_human1_umifm_counts.csv.gz | 5,670,823 | 1,937 | Human · retained |
| GSM2230758_human2_umifm_counts.csv.gz | 4,960,352 | 1,724 | Human · retained |
| GSM2230759_human3_umifm_counts.csv.gz | 9,767,500 | 3,605 | Human · retained |
| GSM2230760_human4_umifm_counts.csv.gz | 4,253,067 | 1,303 | Human · retained |
| GSM2230761_mouse1_umifm_counts.csv.gz | 1,755,400 | 822 | Mouse · excluded |
| GSM2230762_mouse2_umifm_counts.csv.gz | 2,699,967 | 1,064 | Mouse · excluded |
A file, opened
First human donor table: first rows and first fields only. The gene columns continue beyond this excerpt.
barcode assigned_cluster A1BG A1CF A2M A2ML1 A4GALT A4GNT human1_lib1.final_cell_0001 GATGACGGAC-GGTGGGAT acinar 0 4 0 0 0 0 human1_lib1.final_cell_0002 GAGCGTTGCT-ACCTTCTT acinar 0 0 0 0 0 0 human1_lib1.final_cell_0003 CTTACGGG-CCATTACT acinar 0 0 0 0 0 0 human1_lib1.final_cell_0004 GATGTACACG-TTAAACTG acinar 0 0 0 0 0 0 human1_lib1.final_cell_0005 GAGATTGCGA-GTCGTCGT acinar 0 0 0 0 0 0 human1_lib1.final_cell_0006 AATCCCACG-ATTCGACG acinar 0 1 0 0 0 0 human1_lib1.final_cell_0007 AAAATCGTT-CTCCGCAT beta 0 0 0 0 0 0 human1_lib1.final_cell_0008 GACTCTTGAC-AAACTGGT acinar 0 1 0 0 0 0
What the values look like
Total transcript counts per human cell, in donor-table order. Different sequencing depths motivate per-cell normalisation. Donor colours stay fixed across this page.
All values in this summary
Cells per donor
| Group | Cells |
|---|---|
| human1 | 1937 |
| human2 | 1724 |
| human3 | 3605 |
| human4 | 1303 |
- Zero is an observed count: 90.62% of human entries are zero; a zero can mean the gene is switched off or that its molecules were not captured.
- Barcode alone is not a unique cell key: 282 barcode occurrences repeat across the merged tables. Even donor-plus-barcode labels have 91 repeats. The original source row names have 0 repeats and are used as cell IDs.
- Only human cells are used: the 2 mouse tables are excluded before the matrix is built.
- No acquisition timeline: cell row order and donor IDs are preserved; the count tables do not give acquisition times.
These steps follow the project’s notebook and stop at the matrix that goes into dimensionality reduction. Embeddings, clustering and their interpretation are on the Workflow page.
| Step | Evidence | Operation | Result |
|---|---|---|---|
| 0 | Documented | Sequencing reads → published gene-count tables | GEO processed product |
| 1 | Reproduced | Select human donor tables; exclude mouse tables | 8,569 cells × 20,125 genes |
| 2 | Reproduced | For each cell: log(1 + count / max(total count, 1) × 10,000) | 8,569 cells; 0 zero-depth cells |
| 3 | Reproduced | Score genes with mean > 0.0125 using variance/mean standardised within 20 mean-expression bins; compare the top-ranked panel with the saved Workflow panel | 12,050 eligible; 1,998 selected genes agree |
| 4 | Held | Keep the Workflow’s own selection of 2,000 genes, so this page describes exactly the genes it uses | 2,000 selected genes |
| 5 | Reproduced | For each selected gene, subtract mean across cells, divide by population SD + 1e−8; clip to [−10, 10] | 8,569 × 2,000; 11,035 entries clipped |
| 6 | Reproduced | Keep donor, barcode and cell type beside the matrix as labels, not as features | 8,569 row labels; 14 annotated types |
What remains
Counts by group
| Group | Input rows | Stored rows |
|---|---|---|
| human1 | 1937 | 1937 |
| human2 | 1724 | 1724 |
| human3 | 3605 | 3605 |
| human4 | 1303 | 1303 |
Gene selection
| Step | Genes remaining |
|---|---|
| Held human genes | 20,125 |
| Mean expression > 0.0125 | 12,050 |
| Selected variable genes | 2,000 |
Follow one cell
1 · counts as published
2 · depth normalisation and log transform
3 · the stored row
Example values at every shown step
The examples span low, middle and high sequencing depth and all four human donors. The charts show the first selected genes; the Design matrix section shows every selected gene. Mouse examples stop at the species step.
Checks against the supplied Workflow
| Check | Result |
|---|---|
| Cell and gene counts | Matches |
| Selected-gene mask equals workflow | Matches |
| Cell-type counts equal workflow | Matches |
| Original source cell IDs unique | Matches |
Re-ranking the genes here selects 1,998 of the Workflow’s 2,000. Only in the Workflow’s selection: IQSEC1, SPIRE2; only in the re-ranking: DISP2, HIVEP2. This page keeps the Workflow’s selection and recomputes the normalisation and scaling.
One row is one human cell; the 2,000 columns are selected genes in original archive gene order. Rows preserve human donor order, then the original order within each donor. The diverging scale shows signed, standardised expression. The strip beside the heatmap identifies donors.
Prepared values span -1.54802 to 10.00000. Heatmap display only: colour saturates at ±3; 416,203 values exceed that display range. The heatmap uses colour steps of 0.02362, and screen pixels compress rows (and, for wide matrices, columns). The selected-row chart and table show the values without that colour rounding.
Open this section to load the embedded matrix.
Hover a row or use the row selector.
Every feature of the selected row
| Feature columns | Unit | Meaning |
|---|---|---|
| 0–1999 | Standard deviations | Depth-normalised, log-transformed gene expression, gene-wise standardised and clipped |
Labels kept beside the matrix
| Label | Meaning |
|---|---|
| Source cell ID | Original row name in the published table, unique across all human cells |
| Donor and barcode | Retained annotations; their combination can repeat |
| Published cell type | Reference annotation, never an input feature |
| Total count | Normalisation denominator, retained for inspection |
- 90.62% of raw human counts are zero; a zero does not by itself show that a gene is switched off.
- Re-ranking the genes here differs from the Workflow’s selection by 2 genes; the Workflow’s selection is kept.
- No batch correction is applied. Donor variation can remain after per-cell depth normalisation.
- 11,035 scaled values are clipped at ±10; those magnitudes cannot be recovered from the prepared matrix alone.
- Published cell annotations are evaluation labels, not error-free biological truth.
Baron et al. (2016). A Single-Cell Transcriptomic Map of the Human and Mouse Pancreas Reveals Inter- and Intra-cell Population Structure. Cell Systems. GEO GSE84133; PMID 27667365.
The Workflow page From gene counts to cell groups uses these profiles for dimensionality reduction and clustering; its Insight page presents the results.