Personal portfolio / Featured project
I reanalyze the published single-cell and spatial transcriptomic data from Tung et al. to examine how vessel-associated cells and their RNA profiles differ across sampled stages after myocardial infarction in mice, and where activation-associated RNA appears in heart sections.
This matters because forming blood vessels is not sufficient for cardiac repair: vessels also need supporting cells and an appropriate tissue environment to remain stable. Tung and colleagues investigated these changes in the context of progressive vascular loss after myocardial infarction.
Study: Tung et al. (2023), Spatiotemporal signaling underlies progressive vascular rarefaction in myocardial infarction, Nature Communications.
My reanalysis asks what conclusions the deposited measurements support. In the single-cell analysis, I test how filtering and clustering affect those conclusions. The study’s sampling structure and incomplete animal-identification information limit what I can establish about consistency across mice.
I use single-cell RNA-seq to characterize vessel-supporting cell clusters and test how their inferred identities depend on filtering and clustering. With spatial transcriptomics, I examine whether an activation-associated RNA signature is stronger in tissue regions with molecular signs of injury.
Because the deposited datasets do not provide recoverable animal-level replication for these comparisons, the results describe patterns in the available cells and tissue sections but cannot estimate how consistently those patterns recur across mice.
The supporting work covers count-matrix QC, annotation, descriptive expression comparisons, GO enrichment and spatial analysis in a modular Nextflow workflow.
Scientific limitations are stated beside the findings.
The authors use a mouse heart-attack model to study how vessel-supporting cells respond to injury and where tissue changes occur. These two experiments sample hearts at different endpoints, rather than follow the same heart through recovery.
| Experimental arm | Reported design | Main analysis limitations |
|---|---|---|
| Injury-series single-cell RNA-seqDesign and evidence: injury series | CD146-selected cells, enriched for vessel-supporting mural cells. SS, D3, D7 and D42 |
The authors’ expression tests compare cells, but do not show how consistently a change occurs across mice. Differences between time points could also come from how the samples were processed. |
| Spatial transcriptomicsDesign and evidence: spatial transcriptomics | RNA measured at known positions in transverse heart sections. SS, D3, D7 and D14 |
Each spot measures a mix of cells. More spots from the same heart give a more detailed view, not more independent hearts. Differences across time can also reflect which heart was sampled and how its tissue was processed. Comparing damaged tissue with a distant (“remote”) area compares two regions of the same injured heart—not an injured mouse with an uninjured one. |
| Question | Analysis | Answer and evidence |
|---|---|---|
| 01Which observations pass quality checks, and what remains uncertain? | Input verification, RNA QC and doublet assessment Separate available measurements from the retained analysis cohort. |
|
| 02Which broad cell identities are supported? | Normalization, clustering, multi-marker annotation and sensitivity analysis Use marker evidence to interpret clusters, then check dependence on analysis choices. |
|
| 03How do captured populations and pericyte expression vary with injury time? | Per-library composition and descriptive expression contrasts Distinguish which cells are represented from how their RNA measurements differ. |
|
| 04Which biological processes are associated with the expression differences? | GO functional enrichment and contributing-gene expression Connect gene-list overlap to the actual genes and measurements behind it. |
|
| 05What does tissue location add? | Spatial transcriptomics and within-section molecular comparison Recover tissue position without treating mixed spots as identified individual cells. |
|
Broad pericytes form a larger share of the captured injury libraries than of SS. D7 increased genes emphasize extracellular-matrix organization, whereas D42 increased genes emphasize cytoplasmic translation. The D3 within-section association places activation-associated RNA along a molecular injury gradient. Together these are selected-cell and mixed-tissue associations, not repeated animal effects or a causal explanation for vessel loss.
These data describe selected cells and individual tissue sections. They do not resolve whether changes repeat across mice, whether more captured cells mean more cells in the heart, or whether a pericyte program causes vessel loss.
The sensitivity results show which interpretations depend on filtering and annotation. Mixed spatial measurements leave the cells producing the RNA signal unresolved.
The earlier operational reproduction tested a cycling pericyte-like candidate, not the new broad-pericyte population. Its recorded verdicts remain separate:
| Earlier claim | Historical operational verdict | Boundary |
|---|---|---|
| Pericyte-like identity | Reproduced | Held-out marker checks under the older candidate rule. |
| Cycle, matrix and vascular programs | Not assessable | D7 had 13 candidates, below the prespecified 20-cell minimum. |
| Temporal representation | Reproduced | Captured cycling-candidate fractions, not tissue abundance. |
| Infarct association | Partly reproduced | Molecular spatial proxy; cellular localization remained not assessable. |
Those candidate counts were SS 0, D3 56, D7 13 and D42 1. New broad-pericyte counts do not retroactively repair the old program comparison or change its verdict. Historical four-claim report · Historical definitions and amendments.
I propose matched, mouse-identified RNA and tissue measurements with sham controls and injury times balanced across processing batches. Per-cell matrix expression and independent tissue cell counts would distinguish within-pericyte changes from changing mixtures or capture; matched coverage/perfusion would test the functional association.
The paper already contains EdU, lineage tracing, histology and perfusion evidence. The proposed addition is matched animal-level replication and cell-resolved attribution, not a claim those assays are missing. Existing source assays, design, controls and contrasting outcomes. A causal test and its sample-size justification require separate prospective design; no new experiment was performed here.
I connect the experimental question to input checks, modular analysis, documented cell annotations and results others can inspect. I used AI assistance for implementation and review. Reanalysis of these data does not provide independent experimental validation.
Workflow: deposited counts → QC/doublets → full-gene normalization and clustering → reviewed annotation → descriptive contrasts and GO → bounded spatial analysis → linked figures, tables and report.
Source run: output/portfolio/pc001-full-2026-09-12-01. Complete-run task count: 116. The linked execution record distinguishes actual resume/fresh-run checks from synthetic tests; workflow completion alone does not certify the biology.
Question 01 · Analysis evidence and its limits
Finding 1.1 · Recorded evidence
Each point is a deposited barcode. Detected endogenous genes and mitochondrial RNA fraction describe different aspects of its RNA measurement. The frozen primary rule requires more than 200 genes, less than 15% mitochondrial RNA and positive endogenous counts.
The primary cohort retains 17,107 of 19,106 observations: SS 3,594/3,890; D3 8,379/9,775; D7 2,474/2,634; D42 2,660/2,807. These are filtering differences, not evidence that injury caused poor cell quality.
Passing QC does not prove that a barcode is a single cell or a pericyte. Predicted doublets remain in the primary cohort; filtered inputs do not supply a newly verified empty-droplet background.
Compare each alternative with the primary rule using the same deposited-library denominator. A taller bar means more retained observations, not better data.
The stricter and more permissive mitochondrial cutoffs, detected-gene cutoff, high-complexity removal and predicted-doublet removal expose which observations depend on filtering choices.
Similar retained totals can hide different cells. Downstream identity and expression sensitivities, not retention alone, assess the impact on biological interpretation.
Scrublet compares observed profiles with simulated two-cell mixtures. A model flag is a warning, not a confirmed doublet.
SS: 15 flags, 0 missing calls; D3: 9 flags, 0 missing calls; D7: 2 flags, 0 missing calls; D42: 10 flags, 0 missing calls.
Similar-cell mixtures can be missed, and assumed loading rates affect calls. The primary analysis retains flags; the removal sensitivity and missing-call status stay explicit. Sample-tag demultiplexing is not claimed.
The Methods describe dataset-tailored gene, RNA-count and mitochondrial thresholds.
I use separately frozen Scanpy processing and expose high-complexity and doublet sensitivities. These new QC diagnostics are not an exact reproduction of the authors’ selection.
Different filters can retain different cells. Neither recipe recovers missing animal identities or resolves possible time–processing confounding.
Finding 1.2 · Recorded evidence
| Library | Deposited | Retained |
|---|---|---|
| SSGSM7898091 | 3,890 | 3,594 |
| D3GSM7898092 | 9,775 | 8,379 |
| D7GSM7898093 | 2,634 | 2,474 |
| D42GSM7898094 | 2,807 | 2,660 |
Deposited barcodes are starting observations; retained barcodes pass my RNA-quality rules. Neither column counts mice.
Each matrix contains 32,287 gene-expression features, including two constructs; the new single-cell analysis preserves 32,285 endogenous genes. D3 alone has four sample-tag features. I used the RNA counts to calculate the QC and doublet diagnostics in finding 1.1.
Four tag features do not establish four mice or assign every cell to an animal. Missing identities prevent animal-level inference; they do not mean only one mouse was harvested. Empty-droplet background measurements remain unverified.
Methods: single-cell library construction; Methods: single-cell RNA-seq analysis
The Methods distinguish sample hashing, custom-reference alignment and dataset-specific filtering.
D3 tag features and retained construct features agree with the reported input preparation. The SS comparator belongs to this same CD146-selected series.
Input correspondence does not reproduce demultiplexing, establish biological replication or justify tuning filters to match the authors’ final cell count.
Finding 1.1. My executed QC recipe differs from the source processing; it is preprocessing evidence, not a biological reproduction verdict.
Finding 1.2. This is an input and inference-boundary audit; no matching publication figure is asserted.
Question 02 · Analysis evidence and its limits
Finding 2.1 · Recorded evidence
Pericytes support capillaries; VSMC means vascular smooth muscle cells. Schwann cells support nerves; fibroblasts are matrix-producing stromal cells. The embedding arranges similar expression profiles near each other. Its shapes and distances do not by themselves establish cell identity, lineage or separate biological states.
Reported primary broad-identity counts: VSMC: 11,684; pericyte: 4,735; Schwann: 432; fibroblast: 256. Confidence counts: supported: 17,107. The annotation protocol requires concordant markers and competing evidence, not cluster number or Mcam alone.
A supported broad cluster label is not a purity guarantee for every barcode. Selection and processing shape the observed population; no new activated-pericyte classifier is claimed.
Read detection and expression together. A marker concentrated in a few cells is different from a concordant program across most of a cluster; follow the plotted legend for dot size and color.
Kcnj8, Abcc9, Colec11 and Vtn support pericyte interpretation; Pdgfra with Lum/Dcn/Cd34 supports fibroblast interpretation, and Plp1/Cnp supports Schwann interpretation. Acta2/Tagln/Myh11 require additional mural context because contractile programs overlap.
Absent Cspg4 or positive matrix/contractile genes do not automatically exclude pericytes. Dropout, ambient RNA and mixed cells remain competing explanations in the linked decisions.
In panel b, colors identify the authors’ annotated populations and separate projections show sampling times. Fig. 2b: annotated injury series; Results: pericyte marker interpretation and Supplementary Fig. 3c–d
The authors use marker and other evidence to identify pericytes and an injury-associated subset. Supplementary Fig. 3c–d shows pericyte marker and shared program evidence.
The authors used Seurat with integration; I used Scanpy without integration. I interpret cell identities through marker expression rather than similarity between UMAP shapes. The broad annotations shown here do not establish the authors’ activated-pericyte state.
Finding 2.2 · Recorded evidence
Broad-identity agreement compares names assigned to common cells. Adjusted Rand index compares cluster partitions, so the two measures answer different questions. Newly retained or excluded cells are reported separately.
Across alternative analyses, 88.0% to 98.9% of cells present in both the primary and alternative analysis retain the same broad cell-type label. I compared 14 filtering and clustering configurations. The linked decisions preserve uncertain and mixed clusters as unassigned for composition.
Stable labels can still be wrong. A cluster containing two cell types differs from cells expressing both programs. Excluding three cell-division genes tests only a limited part of cell-cycle influence.
The source Methods describe variable-gene selection, integration, clustering and marker-informed grouping.
My prespecified alternatives test the dependence of clustering and annotation on QC, graph and limited mitotic-panel choices. Every variant is retained, including unfavorable or unresolved cases.
There is no matching publication plot for this 14-variant audit. It is an extension, not reproduced robustness or independent biological confirmation.
Finding 2.1. The source informs marker interpretation; its cell labels are not silently transferred to my cells.
Finding 2.2. This robustness audit is an explicitly new analysis. Compare common-cell memberships, not UMAP rotations.
Question 03 · Analysis evidence and its limits
Finding 3.1 · Recorded evidence
Each library has its own retained-cell denominator. A larger fraction means a larger share of this selected sample—not automatically more cells in the heart.
Broad pericytes among all retained cells: SS: 466/3,594 (13.0%); D3: 2,234/8,379 (26.7%); D7: 894/2,474 (36.1%); D42: 1,141/2,660 (42.9%). These are captured-library fractions, not tissue abundance. The table records every broad identity and percentage-point differences against SS. Uncertain or mixed annotations remain unassigned. Exact numerators and denominators.
Dissociation, CD146 selection, filtering and annotation affect representation. These are not activated-pericyte fractions, replicated animal estimates or evidence of absolute tissue expansion.
Panel b shows how source-annotated cells occupy expression space at each time. Fig. 2b: annotated injury series; Results: pericyte marker interpretation and Supplementary Fig. 3c–d
The authors identify an injury-associated pericyte subset in their integrated series.
A scatterplot’s density is not a test of tissue abundance. The authors’ activated-state fraction and my broad-pericyte fraction are different quantities and cannot be compared as interchangeable estimates.
Finding 3.2 · Recorded evidence
Each panel displays the six largest absolute descriptive ratios, not the Col1a1 example below. Bars are log2 ratios of mean normalized counts; group sizes and eligibility remain in the linked status table. Each comparison uses that variant’s supported broad pericytes.
At D7, Col1a1 has mean normalized counts 7.73 versus 1.702 at SS; detection is 59.6% versus 25.3%, with a descriptive log2 ratio of 2.119. Some expression differences depend on clustering. For Nrip2 at D42, the log2 ratio relative to SS changes from 1.6623 in the primary analysis to -1.2139 with an alternative clustering seed (1702). Example gene values.
The frozen log2 mean ratio uses a 0.1 pseudocount in normalized-count units. Small means can produce large ratios. A pooled difference can reflect changing state proportions, changes within states, or both; time–processing ambiguity also remains. Across the reported assessable variant/contrast slots, effect-direction agreement on common eligible genes ranges from 87.6% to 97.4%. Each variant uses its own reviewed pericytes; this is not common-cell refitting or an animal-level significance test. Effect-direction and magnitude sensitivities.
Panels e and f summarize temporal comparisons across all pericytes. The strip plots and heatmaps summarize genes, not independent mice. Fig. 2e–f: temporal pericyte expression; Results: matrix and vascular programs
The paper reports changes in matrix-associated and vascular-function gene expression after injury.
These panels are not the activated-subset-versus-other-pericyte comparison in panel c or Supplementary Fig. 3d. My descriptive thresholds, processing and identities differ from the authors’ Wilcoxon analysis.
Finding 3.1. My fractions describe independently annotated broad groups within retained CD146 libraries, not the source activated-subset frequency.
Finding 3.2. Related biological question, different processing and estimand details; no exact DE reproduction or animal-level confirmation is claimed. Recorded contrast status: D3_vs_SS: descriptive, 2,234 injury versus 466 SS cells, 7,224 eligible genes; D42_vs_SS: descriptive, 1,141 injury versus 466 SS cells, 4,414 eligible genes; D7_vs_SS: descriptive, 894 injury versus 466 SS cells, 5,427 eligible genes.
Question 04 · Analysis evidence and its limits
Finding 4.1 · Recorded evidence
Gene Ontology associates genes with biological processes. Enrichment asks whether a term is more common in the selected list than in the eligible mapped background. Increased and decreased lists are tested separately. Bars show −log10 of the adjusted overlap probability: longer bars mean smaller adjusted probabilities, not greater pathway activity.
Genes with lower expression at D3 relative to SS are enriched for muscle system process. Genes with higher expression at D7 relative to SS are enriched for extracellular matrix organization. Genes with higher expression at D42 relative to SS are enriched for cytoplasmic translation.
| Comparison / direction | Term |
|---|---|
| D3_vs_SS / increased | GO:0009895: negative regulation of catabolic process |
| D3_vs_SS / increased | GO:0019730: antimicrobial humoral response |
| D3_vs_SS / increased | GO:0010718: positive regulation of epithelial to mesenchymal transition |
| D3_vs_SS / increased | GO:0010717: regulation of epithelial to mesenchymal transition |
| D3_vs_SS / decreased | GO:0003012: muscle system process |
| D3_vs_SS / decreased | GO:0090066: regulation of anatomical structure size |
| D3_vs_SS / decreased | GO:0006936: muscle contraction |
| D3_vs_SS / decreased | GO:0006941: striated muscle contraction |
| D7_vs_SS / increased | GO:0030198: extracellular matrix organization |
| D7_vs_SS / increased | GO:0043062: extracellular structure organization |
| D7_vs_SS / increased | GO:0045229: external encapsulating structure organization |
| D7_vs_SS / increased | GO:0034097: response to cytokine |
| D7_vs_SS / decreased | GO:0000122: negative regulation of transcription by RNA polymerase II |
| D7_vs_SS / decreased | GO:0045892: negative regulation of DNA-templated transcription |
| D7_vs_SS / decreased | GO:1902679: negative regulation of RNA biosynthetic process |
| D7_vs_SS / decreased | GO:0009792: embryo development ending in birth or egg hatching |
| D42_vs_SS / increased | GO:0002181: cytoplasmic translation |
| D42_vs_SS / increased | GO:0140236: translation at presynapse |
| D42_vs_SS / increased | GO:0140241: translation at synapse |
| D42_vs_SS / increased | GO:0140242: translation at postsynapse |
| D42_vs_SS / decreased | GO:0009792: embryo development ending in birth or egg hatching |
| D42_vs_SS / decreased | GO:0043009: chordate embryonic development |
| D42_vs_SS / decreased | GO:0009790: embryo development |
| D42_vs_SS / decreased | GO:0000122: negative regulation of transcription by RNA polymerase II |
84 of 84 contrast/direction slots are assessable; 315 primary term tests meet the frozen adjusted overlap-probability threshold.
Status of each comparisonThe leading D7 matrix/structure terms reuse contributors; they are not independent confirmations. Synapse-named translation terms overlap ribosomal annotations, not evidence of neuronal identity or synapse formation. D3 epithelial-to-mesenchymal-transition annotations do not demonstrate lineage conversion. Genes enter the list when their descriptive expression ratio meets the fixed twofold threshold. Enrichment identifies associated annotations; it does not measure pathway activity. An unassessable comparison cannot be interpreted as an absence of enrichment.
The Methods distinguish DAVID GO Direct enrichment from averaging expression of genes associated with selected GO terms.
Fig. 2e–f displays GO-associated temporal expression. Fig. 5b contains enrichment plots for a different spatial gene-list question; it is not a direct counterpart of my single-cell enrichment plot.
My retained GO/MGI resources and explicit backgrounds differ from the authors’ 2023 DAVID resources. This is a versioned functional characterization, not exact recovery of the published enrichment table.
Finding 4.2 · Recorded evidence
Rows are contributing genes; columns are D3, D7 and D42, each compared with SS. Colors show the descriptive log2 mean ratio: red is higher and blue lower than SS. The term is an annotation summary, not a measurement of protein production. Use the term and comparison named in the figure to find the corresponding rows in the contributor table.
For the contributor figure’s selected term, cytoplasmic translation (GO:0002181; D42_vs_SS, increased), the selected-list overlap is 17/44 genes, versus 143/4,035 in the background. Adjusted overlap probability: 1.57e-10; enrichment ratio: 10.9.
Related terms share genes, and some annotations describe upstream roles rather than direct participation. A term can hide differing gene-level effects. The example view is not a substitute for all contributors and the complete selected/background lists.
The heatmap shows relative normalized expression; the strip plot summarizes GO-associated genes over time. Fig. 2e–f: temporal pericyte expression; Results: matrix and vascular programs; Methods: pathway enrichment and expression scoring
The authors connect matrix and vascular themes to particular genes with different temporal expression patterns.
Neither a scaled expression heatmap nor an enrichment statistic is a functional assay. My full gene and resource tables are needed to judge whether a similar term rests on similar evidence.
Finding 4.1. Keep term overlap, contributing-gene expression and measured biological function distinct.
Finding 4.2. A repeated process label is not independent support when the same genes drive related terms.
Question 05 · Analysis evidence and its limits
Finding 5.1 · Recorded evidence
Each spot combines RNA from nearby cells. The injury index rises with injury-associated RNA and falls with heart-muscle RNA, standardized within that section. Check each map’s color scale and section label.
The activation-associated score averages normalized log-expression of Loxl2, Postn, Timp1, Mmp14, Thy1, Ptn and Ecscr. The injury index uses separate injury-associated and heart-muscle genes. Their maps show where each RNA pattern is stronger within a section.
The index is not an independently annotated infarct boundary, and the scores do not identify which cells produced the RNA. Coordinates provide location, not histological segmentation. SS/D3/D7 are the shared modality times; D14 is not D42.
Panel a links source spatial clusters to complementary tissue images. Fig. 5a: tissue-positioned data and imaging; Results: why spatial position matters; Methods: spatial processing and label transfer
The authors distinguish infarct-associated from remote regional programs using integrated spot data and imaging context.
My analysis compares RNA scores within each section. It does not estimate cell-type composition at each spot or reproduce the authors’ regional clustering and communication analysis.
Finding 5.2 · Recorded evidence
Each dot is one retained spatial spot. The horizontal axis is the continuous molecular injury index; the vertical axis is the activation-associated RNA score. D3 is primary. The companion text/table supplies median- and quartile-group differences and rank correlation; those group contrasts are not plotted as bars here. Moran’s I describes spatial similarity between neighboring spots.
At day 3, the activation-associated RNA signature is stronger where the injury-related RNA index is higher. These measurements cannot identify the cells producing the signal.
SS: 591 retained spots, median difference -0.00762, quartile difference -0.0203, rank correlation -0.0137; D3: 1,068 retained spots, median difference 0.299, quartile difference 0.584, rank correlation 0.742; D7: 1,245 retained spots, median difference 0.596, quartile difference 0.985, rank correlation 0.606; D14: 947 retained spots, median difference 0.507, quartile difference 0.632, rank correlation 0.66.
Section summary tableThe association may reflect changes in the mix of cells contributing RNA. It cannot establish accumulation of activated pericytes or a causal mechanism. D7 provides an additional descriptive comparison; SS and D14 provide uninjured and later context.
Panel i maps predicted cell-associated scores. The separate microscopy experiment in Fig. 2d examines EdU-positive pericytes in the injury region. These are different measurements.
The paper combines cell-division imaging with reference-based spatial predictions to support its biological interpretation.
Prediction scores are not direct cell counts, and my fixed-panel proxy is neither that label-transfer model nor the microscopy experiment. Reusing the same study does not provide independent confirmation of cellular localization.
Finding 5.1. My fixed-panel spatial scores are a molecular proxy, not the paper’s histologically informed regional assignments.
Finding 5.2. I retained the earlier project’s gene sets and scoring rules. The observed RNA association leaves the location of activated pericytes unresolved.
