Unadjusted transcriptional changes in ischemic border zone of cmAKO mice

To examine the role of cardiomyocyte α1A-ARs after heart injury, we created a cardiomyocyte-specific α1A-AR knockout mouse line (Myh6-CrexAdra1a^fl/fl or cmAKO). We induced a myocardial infarction (MI) in this model by permanent LCA ligation at 8–12 weeks of age. In the same mice as described previously [9] we collected tissue from the border zone of the infarcted left ventricle, or a matched location in the sham surgery group, three days post-infarction and measured bulk gene expression by RNAseq (Fig 1A). To identify outliers and unexpected sources of variation that might bias our results, we performed principal component analysis. Samples were well stratified by treatment across PC1, which accounted for 98.4% of the variation in gene expression (Fig 1B). We proceeded with differential expression analysis using DESeq2 [19], modeling gene expression as the sum of the effects of genotype, treatment, and their interaction. LCA ligation produced the greatest number of differentially expressed genes (DEGs), followed by the interaction of cmAKO and LCA ligation, whereas cmAKO alone yielded the fewest DEGs (Fig 1C). These findings recapitulate our prior studies, where we found that α1A-knockout magnified the adverse effects of LCA ligation but elicited minimal phenotypic differences from WT mice in the absence of cardiac insult [9].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Transcriptomic profiling of cardiomyocyte-specific α1A-AR knockout mice following myocardial infarction (MI).

(A) Schematic of the experimental design for inducing MI in cmAKO or WT mice and subsequent bulk RNAseq analysis from tissue collected at the infarct border zone and matched sham locations. Figure created in Biorender. (B) Principal Component Analysis (PCA) of RNA counts showing stratification of samples by treatment along PC1, accounting for 90% of the variance. (C) Differential expression analysis indicating MI as the condition with the highest number of differentially expressed genes (DEGs), followed by the cmAKO and MI interaction; upregulated genes are shown in blue and downregulated genes in red. (D) Volcano plot of DEGs specific to the cmAKO response to LCA ligation, highlighting genes like Timp4 (fibroblasts) and Aplnr (cardiomyocytes) with p < 0.05 and fold change > 1.5. (E) Gene Ontology (GO) enrichment analysis showing biological processes significantly influenced by MI, cmAKO, and their interaction, such as extracellular matrix organization and cardiac muscle cell development. Terms with identical p-values have been consolidated.

https://doi.org/10.1371/journal.pgen.1011807.g001

When examining the genes that are differentially expressed in the cmAKO × myocardial infarction interaction term (the unique response of cmAKO to LCA ligation), we observe a large number of genes that are enriched in specific cell types, such as Timp4 (fibroblasts) [20], Zbtb16 (cardiomyocytes) [21], and Aplnr (cardiomyocytes) [22] (Fig 1D and S1 Table). Furthermore, gene set enrichment analysis (GSEA) of all interaction-associated DEGs using clusterProfiler [23] revealed strong enrichments of cardiomyocyte- and immune-specific pathways, with the first and third most significant pathways being “striated muscle cell differentiation” (Padj = 3.7e-09) and “adaptive immune response” (Padj = 1.5e-08). Likewise, genes associated with LCA ligation were enriched for cardiac fibrosis and CM energy metabolism processes, including “extracellular matrix organization” (Padj = 1.03e-22) and “cellular respiration” (Padj = 7.92e-24) (Fig 1E and S3 Table). Taken together, these suggest that cell-type compositional changes may contribute to bulk expression changes in the heart. Considering the abundance of expression changes related to cell-type-specific pathways and mechanisms, we next sought to estimate sample-specific changes in cellular composition.

Cell type-specific markers in snRNAseq

To guide the interpretation of our cmAKO model, we developed a cell-type-specific gene expression reference panel. To this end, we performed single nucleus RNAseq from the left ventricles of WT mice (n = 6). Because cell type deconvolution methods have been shown to be sensitive to the specific marker sets and reference dataset used [24], and given the limited application of snRNAseq to deconvolution algorithms, we elected to conduct a rigorous data cleaning and marker selection strategy for our snRNASeq. This began with in silico removal of ambient RNA and low-quality nuclei before proceeding with dimensional reduction, and clustering, retaining 20,968 nuclei assigned to 11 clusters (Figs 2A and S1). Next, we sought to connect our nuclei clusters to known cell types by comparing cluster-specific expression markers to known cell-type markers. Cluster markers were identified as genes whose within-cluster expression was significantly higher than out-of-cluster expression in one-versus-all testing using findMarkers from scran [25]. The top 15 most significant marker genes per cluster were supplied to ToppGene to look for enrichment of existing cell type markers. After excluding clusters with less than 500 nuclei or those which presented unclear cell-type associations, the final dataset contained 20,061 nuclei assigned to five cell-type clusters. In order of most-to-least abundant, we retained clusters of endothelial cells, cardiomyocytes, fibroblasts, macrophages, and a joint cluster of pericytes and smooth muscle cells (Fig 2B). Further, we found that our marker list recapitulates standard proteomic and transcriptomic cardiac cell-type markers, including Myh6 [26] and Tnnt2 [27] for cardiomyocytes, Egfl7 [28] and Fabp4 [29] for endothelial cells, and Dcn and Col1a1 [30] for fibroblasts (Fig 2C and S2 Table)

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Development and characterization of a cell-type-specific gene expression reference panel from single-nucleus RNA sequencing (snRNAseq).

(A) Generation of the reference panel involved identifying five major cell types in snRNAseq data from pooled nuclei of left ventricles of untreated C57BL/6J mice. After preprocessing, 15 marker genes per cell type were selected based on significance using the findMarkers function from scran. Figure created in Biorender. (B) Annotation and clustering within the snRNAseq reference identified five clusters, each labeled with broad cell type names based on known markers. Post-processing included dimensional reduction and exclusion of clusters with fewer than 400 nuclei or unclear cardiac cell type annotations, resulting in 20,061 nuclei across the five clusters. (C) Specificity of cell cluster gene expression markers demonstrated by the top three markers for each cell type, selected by log-fold-change in one-versus-all testing between clusters. Marker visualization includes point size proportional to the nuclei expression and color coding by average expression within each cluster.

https://doi.org/10.1371/journal.pgen.1011807.g002

Cell-type composition of mouse left ventricular tissue

Cell-type deconvolution is a method to infer how much RNA comes from each cell type present in bulk RNA sequencing data produced from heterogeneous tissue. As such, it is important to note that these approaches do not estimate the number of each type of cell nor their total volumes. For our pipeline, we chose to utilize MuSiC, one of several existing deconvolution methods, which iteratively tests for the combination of cell type proportions whose summed expression profiles best explains the overall tissue expression profile. (Fig 3A). To inform development of our analysis pipeline and to test its performance, we first applied it to several bulk RNAseq datasets derived from purified fractions of major cardiac cell types (Fig 3B). We found that our MuSiC-based method was able to accurately predict the makeup of every cardiomyocyte, endothelial cell, and fibroblast sample (Fig 3C). Confident in our ability to identify major cell types of the heart, we then applied our pipeline to the bulk RNAseq data from our experimental groups. In the untreated WT mice, we estimated that the major cell type by transcript abundance was cardiomyocytes (79%), followed by fibroblasts (8.4%) and macrophages (5.8%). For both genotypes, cardiomyocyte abundance decreased in the treatment group, while fibroblasts and macrophages rose in relative abundance. For each of these changes, these compositional shifts were more pronounced in cmAKOs when compared to the WT group (Fig 3D).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Inference of cell-type proportions from bulk RNAseq and dynamics during myocardial infarction in α1A-AR-KO mice.

(A) Overview of the reference-based deconvolution approach which utilizes gene expression markers to estimate the proportions of cell types contributing to bulk gene expression. Figure created in Biorender. (B) Generation of cell-type enriched bulk RNAseq data involved isolating cardiomyocytes, fibroblasts, and endothelial cells from left ventricles of untreated WT mice using Langendorff perfusion, gravity sedimentation, and CD31-bead binding. Figure created in Biorender. (C) Validation of the deconvolution pipeline using bulk RNAseq from pure cell type fractions as ground-truth samples, depicted in a ternary plot where each dot represents a replicate, color-coded by intended enriched cell type, indicating high accuracy of the composition estimates. (D) Estimation of cardiac cell type composition changes post-myocardial infarction (MI) and in α1A-AR knockout (cmAKO) mice, showing increased proportions of macrophages and fibroblasts and decreased proportions of cardiomyocytes, with more pronounced changes in cmAKO mice. (E) Results from a Dirichlet regression modeling cell type proportion as influenced by cmAKO, MI, and their interaction, revealing significant effects on several cardiac cell types. CMs = Cardiomyocytes; ECs = Endothelial Cells; FBs = Fibroblasts; WT = Wild-Type; MI = Myocardial Infarction.

https://doi.org/10.1371/journal.pgen.1011807.g003

Traditional models and association tests, such as the Pearson correlation [31], are prone to identifying spurious correlations when applied to contexts where measures are interrelated, as with cellular composition data [32]. To understand this, imagine a scenario where a treatment causes cardiac fibroblasts to proliferate while having minimal effect on adjacent cell types. If researchers were to quantify the number of each cell type in a treated heart, they would find that the absolute number of non-fibroblasts remains constant. However, since the total number of cells in the heart increased, the relative proportion of non-fibroblasts would seem to decrease compared to an untreated heart. As such, standard statistical tests applied to these compositional measures would report that the treatment is associated with both an increase in fibroblasts and a decrease in non-fibroblasts, despite there being no absolute increase in the number of non-fibroblasts.

To address this and correctly quantify the effect of each treatment and genotype on the compositional changes, we used a Dirichlet regression model, which can accommodate the sum-to-one constraint and covariance structure of compositional data [33,34]. Our model considered the effects of each genotype and treatment additively, as well as the interaction of the two terms. We found that MI and the cmAKO × MI interaction are the only terms to have significant effects on major cell type proportions. Specifically, MI and cmAKO × MI were both significantly associated with the abundance of cardiomyocytes (p = 0.0051 for MI; p = 0.0024 for cmAKO × MI), endothelial cells (p = 0.00053 for MI; p = 0.0089 for cmAKO × MI), and pericytes/SMC (p = 0.036 for MI; p = 0.01 for cmAKO × MI). Additionally, only cmAKO × MI was found to be significantly associated with the proportion of macrophages (p = 0.026) (Fig 3E).

The more pronounced cellular compositional changes in cmAKO mice compared to WT following MI suggest that α1A-ARs play a critical role in maintaining cardiac cellular homeostasis during stress. The significant loss of cardiomyocytes and increase in fibroblasts in α1A-KO + LCA mice, which was attenuated in WT + LCA animals, indicates that α1A-AR signaling may protect cardiomyocytes from stress-induced death and limit the fibrotic response. These findings provide a novel cellular mechanism that may underlie the cardioprotective effects of α1A-ARs previously observed in various models of cardiac injury.

Cell-type-adjusted differential gene expression

Variation in either cell state or abundance [35] can be responsible for differences in the overall transcriptional profiles of cellularly heterogeneous tissue (Fig 4A). Given the differences in cellular abundance in our samples, we became interested in discerning the contribution of these compositional differences to the transcriptional changes we observed in response to LCA ligation and CM-α1A KO. However, directly including cell type proportions as covariates in statistical models can lead to inflated associations since proportions are constrained to sum to one and thus inherently correlated [36,37]. To address this, we evaluated different transformations of the proportions, including principal components analysis and centered log-ratio (CLR), which converts proportions to an unconstrained space while preserving their relationships [38].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Impact of cell type composition on differential gene expression during myocardial infarction in α1A-AR-KO mice.

(A) Differences in either cell type abundances or novel states can lead to gene expression differences in bulk tissue. Figure created in Biorender. (B) Schematics of one simulation of compositional changes in differential expression; bulk RNAseq samples were generated with varying cardiomyocyte proportions (30-70%). All samples were compared against the 50% CM group with and without cell-types in the model. (C) DESeq2 simulations show increases in differentially expressed genes with compositional differences; including cardiomyocyte proportions in the analysis ablates this effect. Each dot is one simulation of 500 genes. (D) Upset plot showing DEGs related to each variable after adjusting for cell-type composition. Cardiomyocyte proportion is the second strongest associated variable. (E) Volcano plot showing differential expression of genes associated with cmAKO × MI interaction after adjusting for cell type proportions. (F) Changes in GO term associations and their significance before and after cell type adjustment, depicted through a scatter plot of Q-scores. Many myocardial infarction-associated GO terms are attributed to changes in cell type abundance. (G) and (H), Listing of top GO terms post-adjustment showing significant associations with myocardial infarction and cmAKOxMI (G), and with changes in cardiomyocytes and fibroblasts (H).

https://doi.org/10.1371/journal.pgen.1011807.g004

To model the inclusion of compositional estimates in gene expression analysis, we simulated bulk RNAseq sample groups composed of a range of underlying cellular proportions (Fig 4B). In silico mixtures were generated by sampling expression of individual genes from each cell-type cluster according to pre-specified cell-type ratios. We began with 50% cardiomyocytes and equal proportions of each remaining cell type from our snRNAseq panel. We then introduced stepwise changes in cardiomyocyte proportion from a range of 30–70% cardiomyocytes. Then, for non-control groups, we doubled the expression of 10% of random genes to serve as true positives. We then tested for differential gene expression between the 50% cardiomyocyte group and each other group. In these comparisons, we found a considerable effect of compositional changes on differential expression. Strikingly, a 10% reduction in the major cell type, a change smaller than that seen in cardiomyocyte abundance during MI in our data, caused 20% of transcripts to be identified as differentially expressed genes (DEGs) (Fig 4C).

When we repeated the analysis including different representations of the cell type proportions in the model (raw percentages, principal components analysis (PCA), or centered log-ratio (CLR) transformations), we found that including any of these compositional covariates dramatically improved the accuracy of DEG detection compared to the unadjusted model (S2 Fig), with PCA offering the strongest overall performance. However, given that CLR transformation still provided substantial improvements over the unadjusted model and is an established, is theoretically grounded method for compositional data analysis, and offers more direct biological interpretation related to specific cell types compared to the abstract components derived from PCA, we elected to utilize CLR adjusted models for analyzing our experimental data [38]. Nonetheless, incorporating compositional information via any of these methods clearly shifted the attribution of expression changes away from the sample group comparisons (which were blind to composition) and towards the compositional covariates themselves. These results demonstrate how comparisons of bulk gene expression can be skewed when sample groups have confounding variation in cell composition and how including compositional terms in DEG analysis enables more accurate attribution of expression changes.

Applying our cell abundance correction method significantly impacted the interpretation of gene expression changes, revealing that cardiomyocyte abundance was associated with the expression of hundreds of genes (Fig 4D). While LCA ligation remained the most influential variable overall, accounting for cell-type proportions clarified the source of many transcriptional signals (Fig 4D; S3 Table).

In the comparison between WT MI and WT Sham hearts, correction attributed a substantial portion of the initial differential expression signature to changes in cellular composition. Hundreds of genes exhibited reduced significance, with many losing statistical significance altogether after adjustment (S3 Table). For instance, the apparent strong upregulation of the epicardial marker Upk3b [39] (Padj = 4e-42 to 8e-22) and several complement cascade genes like Cfb and C4b (approx. 20 orders of magnitude reduction) in the uncorrected analysis likely reflected shifts in cell abundance and inflammatory cell infiltration, respectively, rather than direct transcriptional regulation within resident cell types due to infarction. This reattribution was mirrored in GSEA findings, where the association between MI and terms related to cardiomyocyte energy metabolism (e.g., “oxidative phosphorylation,” Padj = 7.92e-24 to 3.29e-19; “generation of precursor metabolites and energy,” Padj = 7.26e-19 to 3.27e-13) weakened considerably, suggesting these changes were largely driven by the loss of cardiomyocyte populations (Fig 4F, G; S4 Table).

Conversely, the correction unveiled a smaller, yet biologically insightful, set of genes whose association with MI strengthened (S1, S3 Tables). Notably, Lgals3, previously implicated in heart failure and the post-MI remodeling response [40–43], became highly significant only after correction (Padj = 0.16 to 1e-7), highlighting it as a potential candidate for future investigation. These findings aligned with GSEA results showing strengthened associations for terms related to protein synthesis (e.g., “ribonucleoprotein complex biogenesis,” Padj = 3.45e-8 to 3.04e-14; “cytoplasmic translation,” Padj = 3.08e-7 to 4.25e-12), consistent with recent novel evidence of altered translational activity in cardiomyocytes post-MI [44] (Fig 4H)

Within the cmAKO × MI interaction contrast, fewer genes were impacted by the correction compared to the main MI effect, but the adjustment still refined the interpretation (S3 Table). Key genes initially strongly associated with the interaction, such as the cell cycle regulator Zbtb16 (Padj = 9e-15 to 2e-6) and the inflammatory modulator Timp4 (Padj = 4e-18 to 1e-10), showed reduced significance. Given the high expression specificity of these genes in cardiomyocytes [21] and cardiac fibroblasts [20] respectively, this suggests their apparent genotype-specific response was largely influenced by differences in cell composition shifts (e.g., cardiomyocyte loss or fibroblast/immune infiltration) between WT and cmAKO mice post-MI. Accordingly, GSEA terms related to energy metabolism and immune activation (e.g., “response to bacterium,” “generation of precursor metabolites and energy”) weakened in their association with the cmAKO × MI interaction after correction (S4 Table).

Notably, the composition-adjusted analysis revealed strengthened associations between the cmAKO × MI interaction and processes pertaining to cellular stress, structure, and metabolism (Fig 4G, H). GSEA highlighted pathways related to cell adhesion (“actin filament organization,” Padj = 1.77e-02 to 4.29e-04; “cell-substrate adhesion,” Padj = 0.155 to 1.38e-03) and cell cycle regulation/stress response (“G1/S transition of mitotic cell cycle,” Padj = 0.293 to 2.9e-02; “DNA damage response via p53 mediator,” Padj = 0.435 to 4.11e-02), several of which became significant only after correction. Individual genes driving these enrichments included Cdkn1a, a known regulator of cardiomyocyte cell cycle [45,46] implicated in heart failure [47], which increased in significance (Padj = 6.04e-03 to 4.7e-06). Tent5b, involved in apoptosis and glycolysis [48] relevant to the cmAKO metabolic phenotype, also became significant (Padj = 0.38 to 3.71e-04). Similarly, Fhl1, linked to cardiomyopathy and muscle metabolism [49,50], gained significance (Padj = 0.477 to 5.27e-04), potentially connecting α1A-AR’s anti-glycolytic function [51] to Fhl1 activity. These newly prominent genes and processes suggest that α1A-ARs may exert cardioprotective effects by influencing cytoskeletal integrity, cell cycle control, and stress response mechanisms—functions obscured in the uncorrected analysis by confounding compositional changes (S3 Table).

Collectively, these results highlight the critical importance of accounting for cell-type composition in bulk RNA-seq analysis (Fig 4). By distinguishing genuine transcriptional responses within cell types from changes arising due to altered cellular abundances, our method refines the interpretation of the myocardial infarction response, particularly in the context of the cmAKO model. Genes whose significance increased after correction represent potentially crucial intrinsic regulatory mechanisms underlying α1-adrenergic cardio protection, while those losing significance reflect secondary population shifts. Thus, our composition-aware approach provides a more biologically accurate understanding of myocardial transcriptional responses to injury.

Spatial expression of cell type and cell state markers

To validate the ability of our model to predict regulation of transcripts within distinct cell types, we used RNAscope paired with immunohistochemistry. We designed probes for Zbtb16 and Pik3r1, two highly regulated transcripts in the bulk RNAseq data that experienced large reductions in significance in their association with the cmAKO interaction with MI upon adjusting for shifts in cellular composition (Fig 5A). We then localized those probes within cardiomyocytes using immunofluorescent staining for sarcomeric alpha-actinin as an additional control for our methods.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Fluorescent in-situ hybridization experimentally validates composition-aware differential gene expression results.

(A) Volcano plot showing expression changes associated with the unique response of cmAKO mice to myocardial infarction before (yellow) and after (blue) including terms for cardiomyocyte and fibroblast abundance in the DESeq2 model. Two highly significant genes, Zbtb16 and Pik3r1, are representative of a pattern of reduced significance in the updated model. (B) RNAscope of Pik3r1 (cyan) and Zbtb16 (purple) overlayed with IHC showing DAPI (blue) and Actinin (red) in mouse left ventricular tissue. Representative images of the border zone, infract region, and a remote region from sham surgery control wild-type mice, wild-type mice after LCA ligation and cmAKO mice after LCA ligation. (C) Adjusting for cardiomyocyte abundance minimizes differences in spatial expression of Pik3r1 and Zbtb16 in the infarct region, indicated by the post-adjustment expression increases in cmAKO MI and reductions in WT MI. Five representative regions were evaluated in each zone from (B) and the number of expression spots were tallied for each gene. Actinin was used as a proxy for cardiomyocyte abundance and spot counts were normalized to actinin abundance (blue) or not (yellow). To control inter-slide technical variation, spot counts were normalized to slide-matched measures in a remote region (shown on the y-axis).

https://doi.org/10.1371/journal.pgen.1011807.g005

Our predictive model indicated that Zbtb16 (a zinc finger transcription factor), and Pik3r1 (PI 3-kinase regulatory subunit 1) were regulated in a genotype-specific manner within cardiomyocytes after MI. Both transcripts are known to be expressed in cardiomyocytes [21,52], but no previous studies have identified them as downstream targets of α1-ARs. We found that both Zbtb16 and Pik3r1 were expressed in cardiomyocytes under all conditions. Control hearts (Fig 5B) displayed uniform transcript expression across tissue regions. Consistent with our predictions, transcript abundance of cardiomyocyte Zbtb16 was low in ligated WT mouse hearts compared to the robust upregulation in hearts from ligated cmAKO mice. The pattern of Pik3r1 expression was identical, although the magnitude of differential expression was qualitatively lower, also consistent with our predictions.

To quantify these effects, we evaluated five representative subregions within the border zone, infarct region, and remote region (or matched locations) of each slide. For each subregion, we tallied the number of expression spots of each gene, as well as the area occupied by actinin, which was used as a proxy for total cardiomyocyte cross-sectional area. To account for differences in cardiomyocyte abundance (S3 Fig), we then normalized expression spot counts to the area occupied by actinin in each subregion. Due to the technical variation present between each image, we further normalized these values to those from the remote region of each sample. We then observed that accounting for cardiomyocyte abundance reduced the differences in spatial expression both Zbtb16 and Pik3r1 between cmAKO and WT mice during myocardial infarction (Fig 5C). This finding is consistent with the predictions made in the composition-aware differential expression analysis, which suggested that Zbtb16 and Pik3r1 expression in bulk tissue was affected by cell type proportion shifts. See S6 Table for full results of quantification and parameters needed to reproduce them.

Ethics statement

All experiments were approved by the IACUC review boards of their respective institutions and mice housed in AAALAC-accredited vivariums. The specific relevant review boards are University of North Carolina at Chapel Hill Institutional Animal Care and Use Committee (IACUC), the UCLA Chancellor’s Animal Research Committee, and the Medical College of Wisconsin Institutional Animal Care and Use Committee.

Mouse husbandry and model generation

cmAKO mice: C57BL/6J and ROSAmT/mG (stock # 007676) mice were purchased from Jackson lab. The cmAKO mouse line was generated by breeding Myh6-Cre (originally from the lab of E. Dale Abel at the University of Iowa and provided by Leslie Leinwand at the University of Colorado) to Adra1a flox/flox mice with loxP sites flanking the first coding exon (constructed in the Paul C. Simpson lab at the University of California, San Francisco/San Francisco VA Medical Center) [77,78]. All mice were backcrossed regularly and maintained on a C57BL/6 genetic background. Twelve- to 16-week-old males were used to generate the myocardial infarction and sham-operated mouse models. Floxed mice (αMHC-Creneg/α1Afl/fl) were used as wild-type (WT) controls for cmAKO mice.

snRNAseq experiments: C57BL/6J mice (JAX stock 000664) were purchased from Jackson Laboratory, Bar Harbor, Maine, and bred to produce pups.

Pure Cell fractions: For experiments involving pure cell-type fractions, C57BL/6J mice were obtained from Jackson Laboratories at 7 weeks of age and allowed to acclimatize for at least two weeks.

Myocardial infarction model

Mice were subjected to permanent LCA ligation as previously described [9]. In brief, to induce myocardial infarction (MI), a left thoracotomy was performed at the fourth-fifth intercostal space, followed by permanent ligation of the left anterior descending (LAD) coronary using a 7/0 non-absorbable ethylene suture. Occlusion was verified by anemia and akinesis of the apex and anterior-lateral wall, after which the thorax was closed in layers. After extubation, mice were kept warm until fully recovered. For LCA ligation, sham surgery, and terminal cardiectomy, mice were anesthetized by inhalation of isoflurane (2%). For postoperative analgesia, 5 mg meloxicam/kg body weight was applied every 24 hours for the first 72 hours post-surgery.

Tissue collection and preparation

After sacrifice, the heart was quickly excised and sectioned perpendicular to the long axis of the left ventricle. The infarct border zone was visualized under a dissecting microscope and dissected out then immediately snap frozen in liquid nitrogen. Tissue was taken from an anatomically analogous location in sham-operated hearts using an identical process. RNA was extracted and shipped to Novogene (Sacramento, CA) for bulk RNAseq as described below.

To fix heart tissue for immunohistochemistry, mice were heparinized, and the heart was perfused with 10 mL of PBS followed by 20 mL of 4% paraformaldehyde (PFA)–PBS through a 23-gauge butterfly needle, then excised and placed in 4% PFA-PBS for 24 hours before transfer to 70% ethanol. Hearts were then embedded in paraffin and 5 µm sections collected.

Fluorescent In-Situ Hybridization (FISH)

Formalin-fixed paraffin sections were used for localizing the cellular expression of various mRNA transcripts by in situ hybridization with the RNAscope Assay [79] as described by the manufacturer (Advanced Cell Diagnostics, Inc., Newark, CA). Briefly, slides were baked at 60°C for 1 h, deparaffinized with xylene and absolute ethanol, and pretreated with Target Retrieval Reagent, H₂O₂, and Protease Plus according to manufacturer specified conditions and times. This was followed by hybridization with RNAscope Probes for 2 h at 40°C in the HybEZ oven for detection of mRNA, or a negative control probe (S5 Table). RNAscope Multiplex Fluorescent Reagent Kit v2 (Advanced Cell Diagnostics, Inc., Newark, CA, Cat No. 323120) was employed for signal amplification and detection. This kit uses fluorescent probes for the development of a reaction product visible at the Cy5 and Cy7 channels. FITC was not used due to green channel autofluorescence common in the heart.

Immunofluorescence staining and image acquisition

Immediately following FISH, slides were blocked with 10% normal goat serum for 1 h and incubated with primary antibody overnight at 4°C (S5 Table) Next, slides were incubated with secondary antibody for 2 hr at RT. Slides were then rinsed in PBS and mounting medium with DAPI was applied and coverslips were attached and sealed with clear nail polish. FISH and IF staining was performed in the University of North Carolina Histology Research Core.

Stained sections were imaged with an Olympus VS200 slide scanner equipped with a motorized stage, an Olympus DP74 digital camera, and OlyVIA software (Olympus America Inc., Center Valley, PA, RRID: SCR_016167). These full scans were opened using Qupath (RRID: SCR_018257), and single field images were generated [80]. (Raw data at: https://doi.org/10.17632/bskhbpf9pp.1)

Image analysis and quantification

Quantification of RNAscope fluorescent in-situ hybridization (FISH) spots and immunohistochemistry (IHC) staining area was performed on single-field images generated from full slide scans using QuPath (v0.5.1) [80]. For each slide, five representative, non-overlapping 300x300 pixel regions of interest (ROIs) were manually selected within the infarct border zone, infarct region, and remote region (or anatomically matched locations in sham controls).

Within each ROI, automated spot detection was used in QuPath to count the number of FISH spots for Zbtb16 (Cy5 channel) and Pik3r1 (Cy7 channel). The total area (in pixels) positive for sarcomeric alpha-actinin IHC staining (FITC channel) was measured using intensity thresholding to serve as a proxy for cardiomyocyte area within the ROI.

To account for variations in cardiomyocyte density, the spot count for each gene (Zbtb16, Pik3r1) in each ROI was normalized by dividing it by the corresponding actinin-positive area within that same ROI, yielding a spot density value. To control inter-slide technical variability, the spot density value for each border zone and infarct region ROI was further normalized by dividing it by the average spot density calculated from the five remote region ROIs on the same slide. The final normalized values are presented in Fig 5C, and the raw quantifications and parameters needed to reproduce this analysis for each ROI are available in S6 Table.

Cardiac cell isolation for pure cell type fractions

As previously described, adult B6 mice were treated with heparin (100 USP units) for 20 minutes to inhibit blood coagulation followed by anesthesia with sodium pentobarbital (100 µL of 50 mg/ml dilution, intraperitoneal) [81]. Upon loss of rear foot reflex, hearts were removed and immediately submerged in ice-cold PBS to arrest the heart.

Cardiac cell isolation for snRNAseq

On two occasions, hearts were excised from 3 littermates collected at P21. If 2 females and 1 male were used for the first collection, then the reverse was done on the second collection, such that the final sequencing represents 6 hearts, 3 males and 3 females.

Hearts were extracted from euthanized mice by cutting the aorta just below the arch arteries, along with the other major vessels. Isolated hearts were washed in ice cold Kruftbruhe (KB) solution and secured by their aortas to 18 or 20-gauge cannulas then tied off with a 3–0 silk suture. Atria were removed with Vannas micro spring scissors. Cannulated ventricles were then hung from a Langendorff apparatus and perfused with calcium-free Tyrode’s buffer, followed by 25–50 mL of 1 mg/mL collagenase type II (Thermo Fisher, 17101015) dissolved in calcium-free Tyrode’s buffer. Both solutions were warmed to 37°C. Following perfusion, ventricular tissue was diced with dissection scissors, triturated in ice cold Kruftbrühe (KB) solution using a wide bore 1 mL pipette, and allowed to settle for 10 minutes on ice.

The supernatant was removed, and the loose pellet was resuspended in 5 mL of Lysis buffer prepared as previously described [82], with only one adjustment – 50 µl of 10% Triton-X-100 was added (final concentration 0.1%). Cells were incubated in Lysis buffer + Triton for 5 minutes on ice, after which they were homogenized with a Tissue Tearor electric tissue homogenizer (Model # 985370) at the second lowest setting for 20–30 seconds and left to sit again for another 5 minutes on ice. They were then transferred through a 15 mL glass Dounce homogenizer and further homogenized with 20 strokes of the A pestle and 20 strokes of the B pestle. Homogenized cell suspensions were sequentially filtered through a 70 µM, 40 µm, and 20 µm cell strainer to remove debris and undigested materials. Samples were then spun at 1000G for 5 minutes and resuspended in 1 mL of 2% BSA dissolved in D-PBS with RNaseOut (Invitrogen, 200U/mL). A small aliquot was set aside to serve as an unstained control for fluorescent activated cell sorting (FACS). The remainder of the suspension was stained with DAPI at 10 µg/ml for 5 minutes on ice. Samples were spun at 1000G for 5 minutes and resuspended in fresh 2% BSA-RNaseOut solution.

Following staining, nuclei were sorted on a BD FACSMelody at 4°C. Following standard protocols, forward and side scatters were used to remove doublets. Unstained controls were used to set the V450 gate. 432,000 nuclei were collected into a 2 mL centrifuge tube preloaded with 500 µL of 2% BSA-RNaseOut solution. Sorted nuclei were spun down at 1000G for 5 minutes, supernatant was removed, and samples were resuspended in 100 µL of 2% BSA-RNaseOut solution (Invitrogen) before proceeding to 10x library preparation.

Bulk and single nucleus RNAseq library preparation

snRNAseq: Nuclei were quantified with a Luna Fl cell counter (Logos Biosystems) and the volume was adjusted to obtain the ideal concentration of nuclei recommended by 10x Genomics (1000 nuclei/µL). Individual nuclei were paired with Chromium v3.1 gel beads and cDNA synthesis, barcoding, and dual index library preparation was performed using Chromium Next GEM V3.1 chemistry according to the manufacturer’s recommendation (10x Genomics). 10,000 nuclei were targeted for each sample with 13 cycles for cDNA amplification and 13 cycles for sample index PCR. The fragment size of cDNA and libraries was assessed using Agilent’s 5200 Fragment Analyzer System to verify product quality prior to sequencing with RIN > 7.0 used as a cutoff for sequencing.

Bulk RNAseq of purified fractions: Sequencing libraries were prepared using 3 μg total RNA from isolated RNA using the Stranded mRNA-Seq Kit (KAPA Biosystems).

Bulk RNAseq of whole tissue: mRNA-Seq libraries were constructed using 4 μg total RNA with the Stranded mRNA-Seq Kit (KAPA Biosystems) at Novogene (Sacramento, CA).

RNA sequencing

snRNAseq: 2 libraries (3 mice per library) were sequenced at the Roy J. Carver Biotechnology Center at the University of Illinois, Urbana Champaign on a NovaSeq 6000 using one S4 lane with 2X150nt reads. Samples were demultiplexed and mapped to the mm10 genome using Cell Ranger v6.1.1 (10X Genomics).

Bulk RNAseq of purified fractions: Bulk RNA fractions were sequenced 2x50 (paired end) on a Novaseq 6000 S2 chip at the Technology Center for Genomics and Bioinformatics (TCGB) at UCLA.

Bulk RNAseq of whole tissue: Samples were multiplexed with Illumina TruSeq adapters and run on a single 75-cycle paired end sequencing run with an Illumina NextSeq-500.

All bulk RNAseq samples were demultiplexed, then decoy-aware pseudo-aligned to the GRCm39 transcriptome with Salmon v1.10.2 [83] and read quality was checked with FastQC v0.12.1.[84] and summarized with MultiQC [85]. Then, samples were joined into a counts matrix with the txmeta [86] and tximport [87] packages in R version 4.3.1. Bulk RNAseq and snRNAseq data are publicly available as NCBI BioProjects under accessions PRJNA1122769 (bulk RNAseq) and PRJNA880279 (snRNAseq). The full Snakemake [88] analysis pipeline is publicly available on GitHub (https://github.com/guralbrian/bulk_decon).

Single nucleus RNAseq data analysis

Raw counts, barcode, and feature matrices were joined into a single Seurat object [89] for each replicate. Ambient droplet RNA and doublets were identified and removed in-silico with DropletUtils::emptyDrops [90] and scDblFinder::computeDoubletDensity [91], respectively. Sample replicates were merged into a single Seurat object, then filtered by feature count, transcript count, mitochondrial transcript percentage, and doublet score. Raw counts were normalized, scaled, applied to principal component analysis, and integrated by harmony::RunHarmony [92]. For clustering, the number of principal components included was determined by finding the first PC which exhibits cumulative percent variability greater than 90% or that which explains less than 5% of the total variability. Samples were then clustered by the standard Seurat methods, using the harmony reduction. See S2 Table for detailed transcriptional marker list.

We converted Ensembl IDs to gene symbols to match the bulk and snRNA-seq formats; among duplicate gene symbols, we kept the one with the highest average gene expression.

Before performing deconvolution, the 35,334 unique transcripts from the bulk RNAseq data and 19,883 from the snRNAseq data were subset to the 15,376 genes present in both datasets. Then, scran::findMarkers [25] was used to find markers for each Seurat cluster. Markers were ordered by adjusted p-value and the top 15 markers per cluster were retained. After, marker genes were manually queried in ToppGene [93] for associations with known cell types. High confidence cell type associations were annotated to the relevant cluster and small clusters or those with low confidence annotated were excluded from further analysis. After, five cell-type clusters were retained: endothelial cells, cardiomyocytes, fibroblasts, a joint vascular smooth muscle cell and pericyte cluster, a joint monocyte and macrophage cluster, and smooth muscle cells.

Deconvolution analysis

Deconvolution of all bulk RNAseq datasets was performed with MuSiC [94], an R package for deconvolution of bulk RNA sequencing data which uses single cell or single nucleus RNA sequencing data as a reference. Only genes present in the 15 markers per cell type identified with scran (S2 Table) were included in deconvolution analysis. Raw expression counts were applied to MuSiC::music_prop and proportion estimates from the weighted deconvolution were used in all further analysis.

Dirichlet model

To model the relationship between treatment or genotype with estimated cell type proportions, we used a Dirichlet regression model. The DirichReg function from the DirichletReg [34] package was run with a genotype-treatment interaction term with common parameterization.

Differential expression

Differential expression analysis was performed with DESeq2(19). Two iterations were performed: without covariates, and with covariates including cardiomyocyte and fibroblast proportions. The DESeq2 workflow included creating DESeqDataSet objects, filtering genes with more than 10 reads in at least four samples. Both iterations modeled gene expression as the product of the additive effects of genotype, treatment, and their interaction, with the second analysis iteration also including additive effects of each cell type proportion. Multiple testing correction was performed using a false discovery rate (FDR) threshold of 0.05 and effects of low-coverage transcripts were minimized by the lfcShrink function [95]. See S1 Table for full DESeq2 outputs.

DESeq2 results were validated with limma-voom [96,97] Raw counts were filtered to retain genes with at least 1 transcript per million (TPM) in at least 4 samples, normalized using calcNormFactors, and transformed with voom to account for mean-variance relationships. We created two linear models: one with genotype × treatment interaction terms alone, and another that additionally included CLR-transformed cell type proportions as covariates. Empirical Bayes moderation was applied using eBayes, and the same contrasts examined in the DESeq2 analysis were tested.

DESeq2 and limma-voom exhibit high concordance when DESeq2 is not run with lfcShrink, with logFC of all transcripts in the genotype-by-treatment contrast exhibiting a Spearman correlation coefficient of 0.77 before cell types were included in the model and 0.971 after cell types were added to the model (S4 Fig). Shrinkage reduces spurious correlations of low-abundance genes by accordingly adjusting their estimated log-fold changes [95] and mildly reduces the correlation coefficients between the two approaches to 0.701 and 0.892 for pre- and post-cell type inclusion.

Simulated differential expression

Bulk RNAseq samples were simulated with predefined cell type contributions by leveraging our snRNAseq dataset. First, for each cell type cluster (cardiomyocytes, fibroblasts, endothelial cells, macrophages, and pericytes/smooth muscle cells), we generated a transcriptomic profile by summing the expression counts for all nuclei in the cluster and normalizing by the total counts. This produced a probability distribution for each gene representing the likelihood of being selected from that cell type.To mimic typical bulk RNAseq data, we sampled transcripts from these distributions to yield a total of 25 million reads per sample [98] (i.e., a sample with 50% cardiomyocytes would sample the expression profile of the cardiomyocyte cluster 12.5 million times). To systematically investigate the impact of cellular composition on differential expression, we simulated 81 distinct conditions in which the proportion of cardiomyocytes varied from 30% to 70% in stepwise changes of 0.5% CM proportion to yield 81 groups ((70–30)/0.5 + 1 = 81). For each replicate sample within these conditions, proportions for fibroblasts and macrophages were adjusted in anti-correlation with the cardiomyocyte change (increasing as cardiomyocytes decreased, based on estimated factors derived from experimental observations), while the remaining proportion was distributed amongst other minor cell types based on reference baseline estimates. Random noise was added to these target proportions before normalizing the final composition vector for each sample to sum to one. Additionally, we doubled the expression of a set of 10% genes to serve as composition-independent expression positive controls (S2 Fig). For samples not belonging to the baseline (50% cardiomyocyte) group, the expression of these designated true positive genes was altered by applying fold changes drawn from a log-normal distribution (mean log2FC ≈ 0.58 [equivalent to 1.5x linear fold change], SD log2FC = 0.5). To test the effect of including composition, DESeq2 was run with two models:

For each model, sample groups were contrasted against the 50% cardiomyocytes group and the number of significantly differentially expressed genes was recorded. The analysis was repeated with a total of three different representations of cell type proportions in the model: untransformed proportions, centered-log-ratios of proportions (a transformation common in compositional statistics which divides each compositional part by the geometric mean of all parts to reduce co-correlation), and principal components from principal component analysis of cell type proportions. All simulated differential expression analysis were conducted in batches of 1000 genes. Model performance in accurately identifying the simulated differentially expressed genes was evaluated using the F1 score, calculated as the harmonic mean of precision (the fraction of predicted positive genes that were true positives) and recall (the fraction of true positive genes that were correctly identified), with higher scores indicating better classification accuracy.

GSEA analysis

Expression changes of individual genes were summarized into biological pathways by gene ontology (GO) enrichment using clusterProfiler v4.10.0 [23]. Gene names were converted to ENSEMBL IDs using the org.Mm.e.g.,db reference package v3.18.0 [99]. For each model and variable, all genes were ordered by their –log10(adjusted P-value) multiplied by the sign (i.e., positive or negative) of the log-fold change. These were provided to clusterProfiler::gseGO using biological pathway ontology and Benjamini-Hochberg p-value adjustment. See S4 Table for entire GSEA term dataset.

Figure generation

Schematics and diagrams (Figs 1A, 2A, 3A, 3B, 4A and 4B) were produced with Biorender. All other plots were generated using the following R packages: dot and bar plots were produced with ggplot2 [100], upset plots with ComplexUpset [101,102], UMAP plot with ggplot2 and tidyseurat [103], ternary plot with ggtern [104], and tables with gt [105] and gtExtras [106].