Image-based tracking of individual cardiomyocytes for scRNA-seq prevents artificial cell clustering
Since common microfluidic or droplet-based scRNA-seq methods do not allow scRNA-seq of whole cardiomyocytes, we employed the ICELL8 platform, which uses a large-bore nozzle dispenser to distribute single cells into 5184 nanowells for further processing. Due to the high physical density of cardiomyocytes, several modifications of the standard protocol were necessary to accomplish adequate loading of nanowells according to the Poisson distribution (Suppl. Fig. 1A, B). Critical steps included frequent dispensing with intermittent gentle mixing of the cell suspension to avoid sedimentation of cardiomyocytes and a reduction of the calculated cell concentration from 1 to 0.2 cell per dispense volume (= 50 nl). Usually, we achieved a loading of 450–750 intact cardiomyocytes per chip, which is below the theoretical value of ~ 1800 cells but still within an acceptable range (Fig. 1a). We assume that high fragility of adult cardiomyocytes and low concentrations of cells are main reasons for suboptimal loading.
Initial analysis of scRNA-seq data from 586 cardiomyocytes suggested the existence of two distinct classes of cardiomyocytes, which clustered separately in PCA and t-SNE plots (Fig. 1b, Suppl. Fig. 1C). However, we noted that clustering was mostly driven by strong differences in the absolute number of genes detected per cardiomyocyte (Fig. 1b). Since the ICELL8 approach includes an imaging step ensuring processing only of nanowells containing single cells, we correlated images of individual cardiomyocytes with the outcome of sequencing reactions and localization in clusters 1 and 2. All cells within cluster 1, characterized by low numbers of detected genes, had a small size, were not rod-shaped anymore, and showed undefined cellular silhouettes, which results from leakage of the Cell Tracker dye (Fig. 1c). We concluded that such cardiomyocytes had encountered cellular damage causing partial RNA degradation, which eventually leads to inefficient RNA-seq library preparation and decreased numbers of detected transcripts. Thus, we changed the experimental design and excluded all cells from subsequent analysis showing any signs of cellular damage.
In a second set of experiments, only scRNA-seq data from undamaged individual cardiomyocytes were analyzed as indicated by image-based analysis. Additional quality control criteria included the total number of detected genes (“features”), percentage of dropout, percentage of mitochondrial transcripts, percentage of non-unique alignments, and presence of cardiomyocyte markers (Suppl. Fig. 1D). In total, we analyzed 2767 micro-wells containing cardiomyocytes, of which 715 harbored a single intact rod-shaped cardiomyocyte (Fig. 1d). After cell lysis, reverse transcription, barcoding, cDNA amplification, library preparation, and sequencing, we obtained 0.6 M reads and the detection of 3.9 k genes per cell on average.
Bioinformatical evaluation using the “scater” R package  and PCA analysis revealed only a low degree of heterogeneity (8% and 1% variance for first and second principle components, respectively) among adult cardiomyocytes of healthy mice, which was not sufficient to drive formation of distinct clusters (Fig. 1e, Suppl. Fig. 1E). Remarkably, different cell sizes had no effects on the distribution of cardiomyocytes within the PCA cluster (Fig. 1e, Suppl. Fig. 1E). Taken together, our findings suggest that cardiomyocytes of different sizes are remarkably homogenous and do not form distinct subpopulations. Furthermore, our data indicate that rigorous quality control, which, in case of highly fragile cardiomyocytes, needs to comprise image-based assessment, is essential to avoid technical artifacts that might suggest non-existing heterogeneity.
Mono- and multi-nucleated cardiomyocytes express similar sets of genes
Multi-nucleated cardiomyocytes are assumed to be larger than mono-nucleated cardiomyocytes . To corroborate these reports and validate our own data, we plotted the size of cardiomyocytes in relation to the number of nuclei. As expected, we observed a clear correlation of size and nuclei numbers (Fig. 1f, Suppl. Fig. 1F). The lack of substantial transcriptional heterogeneity among differentially sized cardiomyocytes already indicated that the number of nuclei has only marginal effects on the transcriptome of individual cardiomyocytes. To explore potential differences or similarities between mono- and multi-nucleated cardiomyocytes in more detail, we separated cardiomyocytes into groups based on the number of nuclei taking advantage of metadata acquired during the procedure. We employed the MAST R package, which detects differentially expressed genes using the Hurdle model for calculation of statistical significances at the single-cell level . In addition to single-cell violin plots, we applied pseudobulk visualization to depict results analogous to more familiar, conventional bulk RNA-seq (Fig. 2a, b). The number of nuclei did not correlate with the number of mapped reads in each group, indicating that similar sequencing qualities and depths were reached thereby excluding a bias due to the RNA content of cardiomyocytes (Suppl. Fig. 1G, H).
Differential gene expression analysis of mono-, bi-, and multi-nucleated cardiomyocytes revealed only a relatively low number of significantly regulated genes (Fig. 2a, b) including Hes1 and Egr1, which have distinct roles in hypoxia responses [28, 33]. Both genes were slightly down-regulated in mono-nucleated cardiomyocytes. However, log-fold-changes of differentially expressed genes between the groups were minor, and the PCA and t-SNE analysis did not identify clustering of cardiomyocytes with different number of nuclei (Figs. 2c, d, 3a). No meaningful and statistically significant enrichments of Gene Ontology terms or pathway were identified, indicating that the number of nuclei has no profound effect on the composition of the cardiomyocyte transcriptome. Surprisingly, the total read count per cell, which corresponds to the initial mRNA content , did not differ between mono- and binucleated cardiomyocytes (Fig. 3b), suggesting that the presence of additional nuclei in cardiomyocytes does not lead to a proportional increase of transcripts.
Although rod-shaped cardiomyocytes did not show sufficient transcriptional heterogeneity to generate distinct clusters in PCA or t-SNE plots, we wanted to know whether individual groups of genes were differentially expressed, considering differences in read count numbers per cell higher than 70% of the mean expression across the whole data set as significant. We detected a significant expression of cell-cycle regulating genes such as cyclins and cdk’s in some cardiomyocytes (Fig. 3c), which is surprising since adult cardiomyocytes barely cycle . The list of detectable (more than five mapped sequencing reads) cell-cycle-related genes included Ccni, Ccnl2, Ccnk, Ccnd3, Ccnh, Ccny, Ccnd2, Ccnl1, Ccna2, Cdk4, and Ccng1 as stimulatory cell cycle; and Cdkn2d, Cdk2ap1, Inca1, and Cdkn1b as inhibitory cell cycle-related genes. Interestingly, expression of individual cell-cycle regulatory genes was randomly distributed within the population and no individual cardiomyocyte expressed a full set of cell-cycle genes. Moreover, no correlation to the number of nuclei was evident (Fig. 3d).
Cardiac hypertrophy induces heterogeneous transcriptional responses in cardiomyocytes
To investigate whether pathological conditions might induce heterogeneity in rod-shaped cardiomyocytes, we induced cardiac hypertrophy by applying transverse aortic constriction (TAC) . Cardiomyocytes isolated from hypertrophic hearts 8 weeks after TAC were clearly different from cardiomyocytes of healthy hearts (WT) (Suppl. Fig. 2A–D; Fig. 4a, b). Bioinformatical analysis using the MAST package demonstrated expression of cardiac marker genes and revealed that the total number of genes detected per cell was comparable between normal and hypertrophic cardiomyocytes, excluding major technical biases. 184 genes were differentially expressed (FDR < 5%) between WT (basal) and TAC conditions (Suppl. Table 1), which caused a clear separation in the PCA and t-SNE analysis (variances of 5% and 2% in the first and second PCA components, respectively). However, we noted an overlap in the PCA plot, containing cells from both TAC and WT conditions, suggesting that not all cardiomyocytes responded equally to hypertrophy.
Top differentially expressed genes between basal and TAC conditions included hypoxia response and muscle-related genes such as Nppa, Nppb, Hif1α, Egr1, Acta1, Hes1, and Ankrd1 (Fig. 4c, d). In addition, we found up-regulation of VEGFA and STAT pathway components such as Stat2. Likewise, TAC cardiomyocytes were enriched (p < 5%) for the GO terms “Hypoxia response via HIF-activation”, as well as for different inflammation- and signaling pathways-related terms (Suppl. Fig. 3A). Furthermore, we calculated cell-associated coefficients of transcriptional variation for both WT and TAC cardiomyocytes, and plotted them to the number of sequencing reads (Fig. 4e). The resulting scatter plot revealed dramatically increased transcriptional variation after TAC-induced cardiac hypertrophy. In addition, we generated single-cell interactome maps (see methods), which also revealed a dramatic increase of gene–gene co-expressions in TAC versus baseline conditions (236 pairs for WT and 716 pairs for TAC) (Fig. 5a; Suppl. Fig. 3B), indicating increased transcriptional activity of cardiomyocytes during hypertrophy.
Heterogeneity of cardiomyocytes in hypertrophic hearts is driven by hypoxic responses
PCA analysis of hypertrophic cardiomyocytes indicated the existence of two partially connected clusters correlating with Hif1α expression (Fig. 5b), which is a crucial transcriptional factor mediating hypoxic responses . Clustering was performed using the k-means algorithm (Fig. 5c) . Cardiomyocytes in cluster 1 (“Hif1αhigh”) were transcriptionally more active compared to cluster 2, resulting in a high number of differentially expressed genes between both clusters (Suppl. Table 2). Cluster 1 cardiomyocytes were enriched for “Angiogenesis”, probably as a consequence of Hif1α expression, while cluster 2 cardiomyocytes were enriched for “Striated Muscle Contraction” (Suppl. Fig. 3C). To analyze the impact of Hif1α expression on the transcriptional profile of hypertrophic cardiomyocytes, we set a threshold of minimal Hif1α expression greater than 70% of the mean expression across the data set. According to this definition, ~ 41% of cardiomyocytes expressed Hif1α and ~ 59% did not. Both groups exhibited similar expression levels of Tnni3, Tnnt2, Myh6, and Myh7 (Suppl. Table 3, Suppl. Fig. 4A) but more than 2000 genes were differential expressed with an FDR < 1% (Suppl. Fig. 4B; Suppl. Table 4). The majority of deregulated genes were found in the Hif1αhigh” group, consistent with higher transcriptional activity in these cardiomyocytes. Furthermore, cardiomyocytes in the Hif1αhigh” group showed higher expression of Egln2 (also called Phd1)  and Vegfa . The concomitant up-regulation of Egln2 and Vegfa in cluster 1 was clearly evident by pseudobulk analysis (Fig. 4c) and single-cell visualization (Fig. 4d, e). In addition, cardiomyocytes in the Hif1αhigh” group were enriched for Ldha, Pgk1, Pfkl, and Hk2 transcripts (Fig. 5f), which are known targets of Hif1α. No differences in average numbers of nuclei were found in Hif1α+ compared to Hif1α− cardiomyocytes (Suppl. Fig. 4C).
HIF1α stabilization in cardiomyocytes inversely correlates with distance to vessels in hypertrophic hearts
Since our data indicated that a substantial amount of the cardiomyocyte heterogeneity in hypertrophic hearts might be driven by hypoxic responses, we wondered whether during hypertrophic growth, some areas of the myocardium encounter low oxygen levels probably due to heterogeneous vessel growth. Cells undergoing hypoxic responses were detected by immunofluorescence staining for HIF1α and vascularization was assessed by staining for the endothelial cell marker CD31. The normal heart did not show HIF1α expression in nuclei under baseline conditions and was characterized by a well-organized vascular network with wide and long blood capillaries (Fig. 6a1). In contrast, hypertrophic hearts contained areas with patches of endothelial Hif1α+ nuclei located in substantially smaller capillaries lacking obvious interconnections (Fig. 6a2, a3). Co-staining with the nuclear cardiomyocyte marker PCM1  revealed that such areas also contained cardiomyocytes undergoing hypoxic responses as indicated by HIF1α localization in nuclei (Fig. 5b). Expression of HIF1α was inhomogeneous and showed a “patchy” pattern (Fig. 6c). To quantify the inverse correlation between HIF1α expressing cardiomyocytes and the presence of vessels in the proximity, we counted the number of Hif1α+ cells as well as the average area of blood vessels (CD31+ area) per view field. A correlation of R2 = 0.54 was calculated between both parameters confirming that HIF1α-expressing cells are preferentially located in areas with a comparatively low degree of capillarization (Fig. 6d).