An integrated cell atlas of 2.4 million cardiac cells across 209 individuals in health and disease

Figures

• 

  Fig. 1. Schematic overview of the composition of HeartMap. The numbers of nuclei and subclusters assigned to each cell type based on a global map are provided. Figure created in BioRender; Datar, Y. https://biorender.com/j2v5l8w (2026).

• 

  Fig. 2. Batch correction on the HeartMap atlas using scANVI. a−c, UMAP representation of all nuclei without batch correction colored by individual (a), cell type (b) and study (c). d−f, UMAP representation of all nuclei with batch correction from scANVI colored by individual (d), cell type (e) and study (f). CM, cardiomyocyte.

• 

  Fig. 3. Cell type composition by disease. Box plots show the distribution of cell type proportions across all ventricular (LV, RV, AX, IVS) samples, with the center of the box being the median and the lower and upper bounds being the 25th and 75th percentiles, respectively. Outliers are shown as points in the box plots. Only ventricular samples are included as other anatomy samples are only represented by non-failing controls. Each data point represents an independent biological sample, defined as a unique combination of study, individual and anatomy (Methods). Sample sizes: DCM (n = 114 samples); HCM (n = 15); ARVC (n = 29); ICM (n = 15); AMI (n = 6); CAD-HF (n = 1), Control (n = 151). Lines between groups denote a statistically credible change in proportions at FDR < 0.05, as calculated using scCODA with the neuronal population set as the reference group (Methods). The two colors in a given line depict the two groups compared in the given pairwise test. CM, cardiomyocyte; FDR, false discovery rate. Source data

• 

  Fig. 4. Sample-level embeddings of HeartMap. a−d, UMAP representations of pseudobulk embeddings across all cell types per sample with annotations based on study of origin (a), anatomy (b), sex assignment (c) and disease (d). Each point in the UMAP represents one sample in the study. In d, samples with acute myocardial infarctions are labeled. e,f, UMAP representations restricted to the most prevalent cell types, including ventricular cardiomyocytes, fibroblasts and endothelial cells, which are shown column-wise with annotations for disease (e) and anatomy (f), respectively. Source data

• 

  Fig. 5. Heatmap of the differentially expressed genes in each pairwise phenotypic comparison. For each disease versus control or disease versus disease comparison, the numbers of differentially expressed genes in each cell type are shown. DE was tested using limma-voom with the model Expr ~ 0 + disease + study + anatomy + sex, with a duplicateCorrelation() effect for individual donor, where ~ indicates ‘is modeled as a function of’. Differentially expressed genes are defined as having a two-sided Benjamini−Hochberg FDR-adjusted P < 0.05 based on both CellBender and CellRanger counts, a low chance of being from background contamination (background heuristic <0.5) and a concordant direction of effect in both CellBender and CellRanger. Statistical testing was performed only for pairwise comparisons with at least three samples per group and at least 25 nuclei per sample for the given cell type. CM, cardiomyocyte; FDR, false discovery rate. Source data

• 

  Fig. 6. Venn diagrams demonstrating the overlap of strongly differentially expressed genes in DCM versus non-failing control samples across multiple reprocessed studies. For each cell type, strongly differentially expressed genes (|logFC| > 1.5, two-sided Benjamini−Hochberg FDR-adjusted P < 0.05 and background heuristic <0.5 in both comparator groups) were identified in each study separately using limma-voom DE testing (Methods). For Chaffin et al.10 (n = 11 DCM, n = 16 controls) and Koenig et al.9 (n = 11 DCM, n = 23 controls), the model was ~ 1 + disease + sex. For Reichart et al.12 (n = 44 DCM, n = 6 controls), the model was ~ 1 + disease + sex + anatomy with duplicateCorrelation() for individual when applicable (Methods). Genes with directionally consistent effects between studies are shown in the overlap; exact P values for specific gene comparisons are provided in Supplementary Table 3. In cases where a gene had a strong but directionally inconsistent effect between studies, it was included in the tabulation for each study-specific count but excluded from the intersection. CM, cardiomyocyte; FDR, false discovery rate. Source data

• 

  Fig. 7. Fibroblast investigation based on subclustering and compositional analysis. a, UMAP plot of 590,066 fibroblast nuclei colored by subcluster. The COL22A1-enriched subcluster is colored in red (cluster 3), and the TNC-enriched subcluster is colored in dark orange (cluster 7). b, UMAP with overlay of log-normalized expression for canonical fibroblast genes (DCN and NEGR1), general activated fibroblast genes (POSTN and FAP) and subcluster-specific genes (COL22A1 and TNC). c, Dot plot depicting the percent of nuclei in each subcluster expressing the genes in b (size) and the average log-normalized expression, normalized by gene across subcluster (shade). d, Box plots show the distribution of fibroblast subcluster proportions across samples, where the center line represents the median; box bounds represent the 25th and 75th percentiles; box bounds extend to the minimum and maximum values within 1.5× the interquartile range; and points beyond the whiskers represent outliers. Each data point represents an independent biological sample defined as a unique combination of individual study anatomy (Methods). Sample sizes: n = 202 non-failing controls, n = 114 DCM, n = 15 HCM, n = 29 ARVC, n…

• 

  Extended Data Fig. 1. Batch correction evaluation metrics used to assess the overall scores of the integration methods based on the ‘individual’ and ‘study’ covariate keys. a, b, Scoring was conducted using ‘individual’ (a) and ‘study’ (b) as the covariate keys. Columns keys are as follows: Method = Batch-correction method used, HVG by batch? = using the top 2,000 highly variable genes across each batch, instead of across all nuclei, for batch-correction, # PCs = number of principal components, # LVs = number of latent variables. A normalized scale bar is provided to show how the scores compare to each other (see Methods). Source data

• 

  Extended Data Fig. 2. Global aggregation of 2.4 million nuclei integrated across individual and study using scANVI. a, UMAP representations of the atlas revealing common and rare cell types. Nuclei are colored by Leiden cluster at resolution 2.0 (left most), cell type (collapsing Leiden clusters), anatomical regions, and disease phenotype (right most). b, Dot plot depicting the expression of the top marker genes for each global cell type by highest nuclei-level estimate logFC that are protein-coding and have an AUC-score > 0.6. The size of each point represents the percent of nuclei in the given cluster expressing the gene at non-zero levels and the shading represents the average log-normalized expression of each gene in each cluster, normalized across clusters using min-max normalization. Source data

• 

  Extended Data Fig. 3. Cell type composition by study. Box and whisker plots show the distribution of cell type proportions, with the center of the box being the median and lower and upper bounds being the 25th and 75th percentiles, respectively. Outliers are shown as points in the box plots. Sample is considered as a unique combination of ‘study’ + ‘individual’ + ‘anatomy’ (see Methods). Source data

• 

  Extended Data Fig. 4. Pairwise Euclidean distances from sample-level pseudo-bulk to controls. For each sample, Euclidean distance to each control sample of the same sex and anatomy was calculated based on the top 30 principal components (PCs) derived from the top 500 highly variable genes (HVG) (a), the variance stabilization transformed expression (see Methods) of top 500 highly variable genes (b), and the variance stabilization transformed expression of all 14,951 genes with appreciable expression (see Methods) (c). The distribution of these distances to controls is shown specifically for each of the 6 acute myocardial infarction samples. For reference, samples of other diseases are pooled together to see the distribution of their distances to controls. The gray horizontal line represents the median pairwise Euclidean distance for control samples compared to other control samples of the same sex and anatomy. DCM, dilated cardiomyopathy; HCM, hypertrophic cardiomyopathy; ARVC, arrhythmogenic right ventricular cardiomyopathy; ICM, ischemic cardiomyopathy; AMI, acute myocardial infarction. Source data

• 

  Extended Data Fig. 5. UMAP by cell type based on covariates. Plots shows the UMAP representations colored by study of origin, anatomical groups, sex-assignment, and disease status (columns from left to right). Each point on the plot represents one sample, defined as unique combination of study, individual, and anatomy. CM, cardiomyocyte; EC, endothelial cell; VSMC, vascular smooth muscle cell. Source data