Pathogenic variants damage cell composition and single cell transcription in cardiomyopathies

1Dilated cardiomyopathy (DCM), a prevalent disorder occurring in 1:250 individuals, is characterized by left ventricular (LV) dilatation, cardiomyocyte loss with fibrotic replacement, and impaired contractility. Arrhythmogenic cardiomyopathy (ACM) similarly incites ventricular dysfunction, often with more prominent right ventricular (RV) involvement, high arrhythmogenic burden, and fibrofatty accumulations. Both disorders can arise from genetic causes. DCM genes encode proteins involved in contractility (titin; TTN, troponin T; TNNT2, troponin C; TNNC1, tropomyosin; TPM1, and filamin C; FLNC), that regulate cardiac splicing (RNA-binding motif protein; RBM20) or calcium sequestration (phospholamban; PLN), and maintain cytoskeletal (desmin; DES) or nuclear (lamin A/C; LMNA) integrity. ACM genes often encode desmosome proteins including plakophilin-2 (PKP2) and desmoplakin (DSP). The cardiomyocyte-specific expression and damaging effects of pathogenic variants (PVs) in many DCM and ACM genes propel the development of arrhythmias and heart failure (HF), a highly morbid condition affecting 23 million individuals worldwide.

2We hypothesized that PVs in different genes would evoke distinct single-cell molecular phenotypes. To address this, we studied the molecular signals underlying HF pathogenesis using snRNAseq of human hearts with advanced DCM and ACM compared to non-failing donor (control) hearts. We revealed differences in the cellular landscape and transcriptional changes between several DCM and ACM genotypes. By leveraging machine learning approaches we illuminated genotype-specific molecular responses, as validated by reconstructing the underlying PVs using snRNAseq data.

3We studied LV and RV tissues (Fig. 1A), obtained prior to any mechanical support in 61 patients and 18 controls (Tables S1, S2) including 12 controls previously reported. Thirty-eight independent PVs were identified in three DCM genes (LMNA; n=12, RMB20; n=8, TTN; n=12) and in the ACM gene PKP2 (n=6), while no PVs were detected in eight DCM patients (PVneg) (Fig. 1B). Analyses were performed for these five genotypes individually (n=46), for aggregated DCM genotypes (LMNA, RBM20, TTN, PVneg), ACM (PKP2), and controls. Additionally, we generated data from 15 PVs across PLN, BAG3, DES, FLNC, FKTN, TNNC1, TNNT2, TMP1, and DSP (Table S1, S2). Due to few recurrent PVs, these genes were excluded in downstream analyses, except where indicated.

4Males predominated among patients (60%) and controls (72%) (Fig. S1A). The mean age of patients was 48.4±4.3 years exclusive of RBM20 (mean age 32.9±14.6 years). Clinical manifestations indicated similar LV dysfunction in LMNA, TTN, and PVneg patients, but greater LV dilatation and reduced systolic contraction in RBM20, and preserved LV with reduced RV function in PKP2. LMNA and TTN patients received more pacemaker/resynchronization therapies than those with other genotypes (Table S1).

5Nuclei were isolated as described from full thickness LV free wall (FW), apex (AP), septum (S) and RV free wall (RV) (Fig. 1C). We compared ~500,000 high-quality nuclei from patients’ ventricular tissues and ~380,000 nuclei from controls (Fig. S1B-E). After pre-processing and quality control filtering, data were integrated across samples using Harmony prior to constructing manifolds using Uniform Manifold Approximation and Projections (UMAPs). Clustering identified ten major cardiac cell types encompassing ventricular cardiomyocytes (CM), fibroblasts (FB), adipocytes (AD), pericytes and smooth muscle cells (mural, MC), endothelial (EC), myeloid and lymphoid (immune), neuronal (NC) and mast cells (Figs. 1C) with 71 distinct cell states. States of the same cell type shared a transcriptional profile but also expressed distinct genes, which implied biological differences.

6Cell type abundances in sample replicates were highly correlated (Pearson coefficient 0.74 - 0.99) (Figs. S2A, B). Cell composition, states, and transcript counts across the FW, AP, and S showed high similarities, and therefore are reported grouped together as LV (Figs. S2C, D).

7Using center log ratio- (CLR) transformed abundance of cell types, we considered the effects of sex on LV and RV cell compositions in DCM (10 female, 29 male) and control (7 female, 11 male) hearts (Fig. S2E). Myeloid cells showed a modest sex-specific difference (FDR=0.016). Only LMNA tissues (from 7 males, 5 females) were sufficient for genotype-specific, sex-associated analyses. Female LV myeloid cells were increased (FDR 0.05), paralleling the findings in DCM vs controls. In addition, RV endothelial cells showed a modest female-specific increase between LMNA patients and controls (FDR 0.048).

8The proportions of LV nuclei across the genotypes demonstrated depletion of CMs except in LMNA, and increased ECs and immune cells except in PKP2 (Figs. 1D, S3). In RVs (Fig. S4), CMs were depleted in the DCM subgroup except for TTN, while ECs were increased in LMNA, TTN, and RBM20, and immune cells were not changed. FBs were not increased in LVs and RVs (Figs. 1D), despite histopathological findings of fibrosis, which implied the acquisition of a secretory rather than a proliferative phenotype.

9Individual-level abundances for cell types (LV, Table S3; RV, Table S4) and cell states (Table S5) are provided for controls and patients. Different cell type abundances in disease compared to control LVs remained significant using a linear model adjusting for age. Pairwise cell type ratios in disease LVs compared to controls confirmed loss of CMs and showed accompanying increased ECs, lymphoid and myeloid cells, and altered ratios (FB and mural compared to CM) with quantitative, genotype-specific differences (Figs. 1E). RBM20 and PVneg, respectively showed greatest shifts in EC:CM (8- and 10.3-fold), myeloid:CM (9.8- and 14.6-fold), and lymphoid:CM (12.4- and 15.2-fold). By contrast, pairwise ratios of all cell types compared to CM were modest or unchanged in LMNA.

10Disease and control ventricles exhibited the previously described (vCM1.0-5;) and four new cardiomyocyte states: vCM1.1, 1.2, 1.3, 3.1 (Fig. 2A-B, Table S6). Across all vCM cell states (Fig. S5), differentially expressed genes (DEGs) in disease connoted increased stress (for example, NPPB) and apoptosis (Tables S7-12, Fig. S6A). While these findings implied late-stage transcriptional convergence, 20-40% of DEGs were genotype-specific (Fig. S5). Only PVneg reduced MYH6 expression, increasing MYH7:MYH6 ratios (Fig. S6B). Conversely, only vCMs with PVs downregulated SMYD1 (Figs. 2C, S6C-E, Table S13), a cardioprotective muscle-specific histone methyltransferase that regulates sarcomere assembly and mitochondrial energetics, and ADRB1 (β1-adrenergic receptor; Fig. S6C) indicative of adrenergic activation, which is therapeutically targeted by β-blockade medicines. Genotype-selective responses included upregulation of FNIP2 (Fig. 2D), inhibiting AMP-activated protein kinase activity and oxidative metabolism in LMNA and PKP2, and CPEB4, an RNA-binding protein that activates glycolysis and stimulates fibrosis in LMNA, RBM20, and PKP2 (Fig. S6F).

11Proportional differences in vCM states (Figs. S6G, S7, S8) varied among genotypes and controls and were more prominent in RVs than LVs, perhaps reflecting hemodynamic differences. PVs generally increased vCM1.1 and 1.2, with accompanying decreased vCM3.1 in LMNA and TTN (Figs. 2B, S7A, S8A-D). State-enriched DEGs for vCM1.1 included MYO18B, required for sarcomere formation, XPR1, regulating phosphate homeostasis, and IGF1R and ROR1, involved in cell survival. vCM1.2 had heightened expression of electrophysiologic genes including PCDH7 that enables intercellular contacts, but when overexpressed, impedes synaptic currents, CAMK2D modulating excitation-contraction coupling, and ANK2, harboring arrhythmogenic PVs. vCM1.3 was modestly increased in RBM20, associated with the largest numbers of DEGs in all genotypes, and enriched for BMP receptors (BMPR1A, BMPR1B; Figs. S7B, D) and GPC5 that regulates BMP signaling. vCM3.1, reduced in LMNA and TTN, expressed early-adaptive transcriptional regulators of stress responses (ATF3, ATF4) and cardioprotection (NR4A3; Fig. S7E). Only PVneg depleted vCM2, enriched for SH3RF2, an anti-apoptosis regulator of the JNK pathway. As DCM and ACM genes are highly and often selectively expressed in CMs, these disease-associated vCM states and DEGs defined intrinsic responses to PVs as well as cell-cell responses to CM death, depressed contractile performance, and arrhythmogenicity.

12We identified four previously characterized and two new ventricular FB states (vFB1.1, vFB1.2; Fig. 2E, Table S14). Across all vFB states, upregulated DEGs included genes involved in extracellular matrix (ECM) remodeling (Figs. S9-11, Table S15-21). LMNA, TTN, and PKP2 increased fibrogenic signaling receptor, EGFR and PKP2 also increased AGTR1 (Fig. S10A) that enables EGFR-transactivation. The expression of profibrotic TGFβ2 was universally increased (Fig. S10B). DCM hearts were enriched for PCOLCE2, promoting insoluble collagen formation, and LMNA, TTN, and PVneg downregulated metalloproteinase inhibitors TIMP1 and 3 (Fig. S10C). Notably, collagen genes showed genotype-specific expression. COL4A1 and COL4A2 were upregulated in LMNA, TTN and PKP2 while COL4A5 and COL24A1 were enriched in PVneg. Thus, although ECM upregulation and collagen deposition (Figs. 2F, S10D-F) was a shared feature, genotype and chamber influenced composition and organization.

13FB state abundance was also altered across genotypes (Figs. S11A, S12A-D). vFB1.1 and vFB1.2 expressed canonical vFB1.0 genes but vFB1.1 also up-regulated APOD, APOE and APOO, typifying lipogenic fibroblasts and CST3 (Fig. S13A), involved in matrix remodeling. vFB1.2-enriched genes are related to actin filament assembly (DAAM1) and chondrocyte differentiation (GPC6) (Fig. S13B). vFB2 expressed prominent profibrotic genes including TGFβ-targets (increased in DCM LVs) and fibrogenic IL11 (highest in RBM20 LVs) (Figs. 2H, S14A-C). vFB2 was increased in TTN and PVneg and modestly in other DCM LVs (Figs. 2G-H, S11). vFB3, diminished in hearts with PVs, expressed proinflammatory cytokine genes (CCL2) and genes related to OSM signaling (IL6ST, OSMR, and oncostatin M-target genes) (Figs. 2I, S14D). The resultant increased ratio of vFB2:vFB3 and altered DEGs suggested dysregulation of fibroblast-to-macrophage interactions that would promote deleterious ECM remodeling, particularly in TTN LV and RV (Figs. S12C, D).

14Three previously described pericyte (PC1-3) and three smooth muscle cell states (SMC1.1, SMC1.2, SMC2, Fig. 3A, Table S22) were identified. DEGs across all states (Fig. S15, Tables S23-28) showed upregulation of the sodium channel SCN3A, with unknown vascular functions, and the noncoding antisense ADAMTS9-AS2, with concordant downregulation of ADAMTS9, a metalloprotease involved in ECM remodeling (Figs. 3B, S16A). Overall DEGs implied that disease evoked signals to synthesize specific ECM and integrin components.

15DEGs in PC states included downregulation of two central signaling receptors, NOTCH3 and PDGFRB, in PKP2 and TTN PC1 (Fig. 3B). NOTCH3 is required for SMC maturation and deficiency causes pericyte dysfunction and arteriovenous malformations. Notch signaling regulates PDGFRB, necessary for angiogenesis and PC recruitment.

16DEGs in SMCs subdivided the previously identified canonical SMC1 into two states. SMC1.1, had higher expression of ACTA2, MYH11, and TAGLN. SMC1.2, strongly expressed ITGA8, required for maintenance of SMC contractile phenotype, and ATP10A, suggesting vascular stiffness and increasing diastolic dysfunction (Fig. 3B) Methylation of the ATP10A locus in SMC decreases with age and atherosclerosis. SMC1.2 was enriched in LMNA, TTN and PVneg RVs (Figs. S17, S18). SMC2 expressed high levels of genes involved in collagen and elastic fiber formation (ELN, LAMA2) and MYH10, demarcating dedifferentiated SMCs with secretory properties (Figs. S16B, S17C). LMNA and PKP2 SMC2 upregulated SLIT3, a stimulator of fibroblast activity, and ECM synthesis and collagen formation genes (Figs. 3B-C). Collectively, disease remodeling of MC indicated modulation of PDGF and NOTCH signaling receptors and synthesis of selective ECM and integrin components.

17Characterization of ECs identified seven previously described cell states (EC1, 2, 5-8 and Mesothelial) (Figs. 3D, S19-21, Table S29). Shared DEGs occurred across all and between genotypes (Table S30-35) with clear RV and LV differences (Fig. S19). Dysregulation of genes encoding factors involved in EC fate, blood vessel development and angiogenesis (NOTCH4, FLT1, FGFR1, RGCC) in disease LVs indicated that vascular remodeling was a common HF feature.

18EC composition in DCM RVs and LVs increased EC5 in comparison to controls (Fig. S21A), while genotype-specific cell state ratios in LVs generally decreased EC1 relative to EC2, EC5, and EC6 (Figs. 3F, S21C). Assessment of selective gene enrichment scores for proliferating cells, endothelial-to-mesenchymal transformation (EMT), and cell death showed no difference in disease compared to controls (Fig. S22; Table S36).

19EC7 had the most DEGs and highest proportion of unique DEGs across all genotypes. Although initially defined as atrial-enriched EC7 expressed endocardial-enriched genes (SMOC1, NPR3, and POSTN (Figs. S20B-C), which prompted a revised annotation to endocardial cells. Interestingly, upregulated genes in EC7 from DCM LVs and PKP2 RVs (Fig. 3G) encoded secreted proteins involved in myocardial stress-adaptation (NRG1), CM force production (EDN1), and endocardial expansion during development or after cardiac injury (BMP6). Notably, BMP6 was selectively upregulated in the endocardium of DCM LVs and ACM RVs (Fig. 3E). Conversely, INHBA, a TGFB superfamily member involved in atrioventricular canal development, and cell adhesion molecule OPCML were downregulated. In PKP2, an unconventional myosin promoting cell adhesion (MYO10), and POSTN, were upregulated in both ventricles. These data highlighted the involvement of the endocardium in chamber-specific remodeling of cardiomyopathies.

20We classified 14 subclusters of myeloid cells comprising macrophages (MP), monocytes (MO), conventional dendritic (cDC1 and cDC2) and proliferating myeloids (Figs. 4A, S23-27, Tables S37-45). Distinct gene sets were unique to each genotype and were particularly pronounced in PVneg and PKP2 in LVs (Fig. S24).

21Although disease increased myeloid cells, the proportions of proliferating myeloids (Figs. 4B-C, S26-27) were consistently lower compared to controls, implying monocyte-infiltration. Tissue resident MPs LYVE1highMHCIIlow and LYVE1lowMHCIIhigh (Fig. S25C) were the most abundant myeloid cells (Figs. S27A-B). MP LYVE1lowMHCIIhigh were increased in the RV of LMNA with similar trends in other genotypes (Fig. S27B). Proportions of MPOSM were modestly decreased with downregulation of OSM in TTN ventricles (Fig. S28A), which would attenuate the MPOSM-vFB3 signaling axis, and mirrored the decreased OSM pathway activation score observed in vFB3 (Fig. S14B).

22Prominent antigen presenting activities, were evident in MOVCAN, MOCD16, cDC1 and cDC2, also occurred in MPFOLR2, akin to tumor-associated MP. Across antigen-presenting MP, RBM20 LVs showed the highest presentation of antigens based on MHCII genes (Figs. S28B-C) and more abundant cDC2 compared to the other genotypes (Fig. S27). PKP2 LVs upregulated MPISG (Fig. S27) with interferon-stimulated genes (ISG), perhaps contributing to inflammatory PKP2 phenotypes.

23We classified 15 lymphoid cell states, including T- and natural killer (NK)-cell subsets, innate lymphoid cells (ILC), B-cells, plasma cells plus proliferating lymphoids (Figs. 4D, S29-32, Table S46). Our experimental design did not enrich for immune cells, and few (<40) genotype specific DEGs were identified (Fig. S29). Proliferating lymphoids were rare and abundance was unchanged in disease (Figs. S30, 32A-D, Table S47-52).

24The cardiac complexity of CD4+ and CD8+T-cells included naive (CD4Tnaive), activated (CD4Tact), regulatory (CD4Treg) and cytotoxic (CD8Tcytox), transitional (CD8Ttrans), terminal effector (CD8Tte) and effector memory (CD8Tem) cells (Fig. S31B). We detected increased CD4Tact only within DCM samples (Figs. S32A-D). However, DEGs indicated lymphocyte activation (cytokines IFNG, CCL3, CCL4 and signaling molecules CBLB, FYN, TXNIP), and maturation (cell surface receptors CD69, CXCR4), particularly in PKP2 NK−, CD4+T, and CD8+T-cells (Fig. 4E).

25CD4+helper T-cells are critical drivers in cardiomyopathy and myocarditis pathogenesis. In LMNA CD4Tact we identified upregulation of TBX21, important for Th1 polarization. Conversely, Th2-polarization transcription factor GATA3 was downregulated in PVneg (Fig. 4F).

26Analysis of cardiac NCs was limited by rarity of this cell type (Figs. 5A, S33, Table S53-57). Across all NC states, DEGs indicated increased NFATC2 in LVs with PVs and PKP2 RVs, and genotype-selective enrichment of LRRK2, an activator of a neurotoxic cascade. Other upregulated DEGs function in proteoglycan synthesis for neuronal myelination and axon regeneration (XYLT1 and HS3ST4), and a complement inhibitor (SUSD4) that impacts neural function and morphology. Genotype-specific DEGs were highest in PKP2.

27We identified previously described and three new NC states (Figs. 5B-C, S34, S35) that were genotype and chamber-selective: NC1.1 in LMNA and TTN LVs; NC1.2 in TTN and RBM20 LVs and in LMNA, TTN, and PKP2 RVs. NC1.1 was distinguished by the highest upregulation of NFATC2. NC1.2 upregulated genes associated with electrocardiogram intervals (SLC35F1 and AJAP1; Fig. 5D), IGFBP5, involved in neuronal apoptosis and autophagy, and the ion channel and heart rate regulator KCNK1. NC1.3 was enriched for the neuromodulator receptor GALR1, and phosphodiesterases PDE10A and PDE3B that participate in neuroprotection and signaling. Dysregulated expression of genes involved in stress-responses and electrophysiology may account for characteristic life-threatening arrhythmias in DCM and PVs in PKP2.

28Similar to NCs, limited adipocytes were captured. We identified three adipocyte states previously found in controls: canonical AD1.0, expressing lipid metabolism genes; stromal AD2, expressing ECM genes; and immune AD3, expressing OSMR and cytokine responsive genes (Figs. 5E-F, S36, Table S60-64). While AD3 was not detected in disease, a fourth identified state, AD1.1 was almost exclusive to disease hearts (Figs. 5G, S37). Compositional analysis identified increased proportions of AD1.1 in PKP2 LVs and RVs concurrent with decreased LV proportions of AD1.0. LMNA and RBM20 RVs, unlike LVs, also increased proportions of AD1.1 (Figs. S38A-D).

29DEGs between AD1.1 and AD1.0 revealed changes in fatty acid metabolism pathways (Fig. 5H). AD1.1 showed downregulation of DGAT2, encoding a triglyceride forming enzyme, and G0S2 and MGLL, encoding lipolysis regulators. Conversely ABHD5, a positive lipolysis regulator, PDK4, a kinase promoting the shift from glucose to fatty acid metabolism, and CIDEA, a regulator of adipose tissue energy expenditure, were upregulated. PKP2 RVs which typically display fibrofatty replacement also showed an enrichment of Gene Ontology biological processes for apoptosis and cell death (Fig. 5I). These data implied genotype-specific state transitions or replacement of canonical adipocytes in DCM and ACM.

30Genome-wide association studies (GWAS) have identified common genetic variants associated with DCM. We selected candidate genes from 15 previously identified DCM loci (Table S67) and examined expression across cell types and genotypes (Fig. S39). Overall, GWAS genes were more often DEGs in LV and RV than expected by chance in snRNAseq data (LV: OR=7.0, P=0.0007; RV: OR=6.1, P=0.0009, one-sided Fisher’s exact test). Multiple genes showed cell type specific expressions, with the majority highly enriched in CM (ALPK3, BAG3, FHOD3, FLNC, HSPB7, MLIP, SLC6A6, SMARCB1, SVIL and TTN). Among these we observed both genotype- and chamber-specific expression differences. Heat shock protein HSBP7 was reduced in TTN LVs and in LMNA and PKP2 RVs. SLC6A6, a taurine and amino acid transporter with cardioprotective effects, was increased in both LV and RV across all genotypes except for PKP2 and PVneg. CDKN1A, a cell cycle regulator and modulator of apoptosis was increased in LMNA CMs from LVs but not RVs.

31GWAS genes that were more broadly expressed across cell types included MTSS1, encoding a putative actin-cytoskeletal interactor. MTSS1 showed highest and unchanged expression in myeloids and was widely increased in mural cells as well as in FBs in LMNA, TTN, and PKP2 (Fig. S39), suggesting influences beyond direct effects on CM function. SLC39A8, a lowly expressed cardiac solute carrier, was unchanged in CMs but increased in LMNA and TTN LV ECs. We suggest that cell-specific expression changes of GWAS genes may improve interpretation of their biologic effects.

32By examining the expression of genes encoding for receptors and ligands, we inferred intercellular signaling and communication. We initially quantified the probability of cell-cell interactions and compared signaling between cell states, and then aggregated information to produce cell-specific and across-all-cell-types data, for each genotype relative to controls. This sequential approach accounted for differential abundances of cell states.

33We detected aberrant intercellular signaling across disease (Figs. 6A, S40), including upregulation of BMP, FN1, Collagen, EGF, IGF and TGF pathways that promote fibrosis. Signaling dependent on VEGF, NOTCH and ANGPT was also increased in disease, implying vascular remodeling. Strikingly, genotype-selective increases in intercellular signaling pathways were also identified in LVs (Figs. 6B, S40A, B): EDN in LMNA, the proinflammatory IL6 in TTN, BAFF/LIGHT (denoting TNF signaling) in RBM20, pro-inflammatory CCL and TNF in PKP2, and the immune modulator BTLA in PVneg. Some of these intercellular signaling pathways were similarly dysregulated in RV, but chamber-specific changes were also observed (Figs. S40C, D, S41).

34We also identified genotype-specific differences in the cells sending and receiving signals in disease pathways (Figs. 6C). For example, IGF signaling (Fig. S42), crucial in cell growth and CM hypertrophy, showed increased FB autocrine (highest in PKP2) and paracrine FB to CM signaling, paralleling findings in experimental models. In addition, IGF signaling from myeloid cells to CMs only occurred in TTN, PKP2 and PVneg, an interaction that might promote muscle repair.

35The source of BMP signaling changed in a genotype-specific manner (Figs. 6C, S43). BMP signaling from MC to CM was increased in TTN LVs and RVs but was downregulated in PVneg and LMNA LVs and other RVs (Figs. S41, S43). Notably, BMP signaling in ECs originated solely from EC7, likely depending on BMP6 upregulation (Figs. 3E, G). EDN signaling from EC7 to CM and MC was highly genotype-selective, occurring in LMNA LV and PKP2 RV (Fig. S44).

36NRG signaling (comprising NRG1-3 and ERBB receptors) showed multifaceted changes. Disease LVs markedly attenuated autocrine NRG signaling in CMs, while EC and CM signaling was upregulated (highest in LMNA) in all genotypes except RBM20 (Fig. 6C, S45). Additionally, NRG3-ERBB4 interactions identified in controls shifted in disease to NRG1-ERBB4 and NRG1-ERBB3 in a genotype-specific manner, consistent with changes in NRG1/3 expression in EC7 (Figs. 3G, S46). This predicted NRG/ERBB shift may provide compensatory responses to adverse remodeling in cardiomyopathies.

37We applied machine learning approaches to snRNAseq data to further advance the recognition of cell and genotype-specific transcriptional patterns. Cell-specific neighborhood graphs showed more connectivities among single-nuclei transcriptomes from PVneg hearts and hearts with PVs in the same gene, compared to PVs from different genes (Fig. S47). Subsequently, we generated a graph attention network (GAT) for multinomial classification of genotypes trained on four major informative LV cell types: CMs, FBs, ECs, and myeloids (Fig. S47A-B). The GAT predicted LMNA, TTN, RBM20, PKP2, and PVneg genotypes with high accuracy. Among LV samples, the genotype prediction accuracy differed by cell types: CMs: 0.93, FBs: 0.92, myeloids: 0.85, ECs: 0.79 (Figs. 6D, S48, Tables S69-70; corresponding RV data Fig. S47C). Aggregation of genotype predictions obtained from these four LV cell types strengthened the correct prediction of genotypes, resulting in a high confidence model (Fig. 6D). Of note, three (H10, H22 and H33) of the four lower prediction probabilities occurred in samples with both a primary and secondary PVs, assigned as such by prior genotype and clinical phenotype review (Table S1). Moreover, as this machine learning model independently confirmed a genotype- and cell type-specific transcriptional signature, we concluded that these snRNAseq datasets accurately described the molecular responses to PVs and unexplained causes of DCM and ACM.

38Our analyses of snRNAseq of LV and RV samples illuminated the cell types and states, molecular signals, and predicted intercellular communications that characterized DCM and ACM. In comparison to control hearts, we identified differences at multiple levels, including changes in the proportions of cell types and states, additional cell states, and differential gene expression, substantially expanding earlier insights achieved by bulk tissue analyses. Across all genotypes, disease hearts demonstrated some common dissimilarities from control hearts, often with graded differences between LVs and RVs. Despite studying hearts from patients with advanced disease who received diverse therapies, congruent transcriptional signatures emerged for different PVs within the same gene and varied between genotypes.

39The differences between genotype groups and controls reflected differences in mean expression and not differences in variance. This was true for each genotype group including PVneg. Transcriptional signatures were complex, diversified both by the proportions of canonical and stressed cell states and by differentially expressed genes within the same states. While interrogation of these datasets provides ongoing opportunities for discovery, our findings provided substantial evidence that genotype influenced pathological remodeling of the heart. These results upend a prevalent dogma that HF results from a common final pathway and can guide the future development of therapies with selective targets to enhance personalized medicine.

40Despite anatomical and histopathological differences between DCM and ACM, we identified shared changes in the cellular composition of ventricular tissues, albeit with DCM LV features largely mirrored in ACM RVs. Cardiomyopathies were depleted in CMs while EC and immune cell populations were increased. FBs did not expand, but increased states that augmented ECM gene expression and collagen deposition. Based on these changes and cytokine profiles (TGFβ activation, increased IGF signaling, decreased CCL2 expression; Figs. 2H-I, 6C), we predicted cell-cell interactions and key molecules that are appropriate for mechanistic studies to causally link differential expression with adverse cardiac remodeling.

41Disease CMs exhibited loss of the canonical state vCM1.0, and the emergence of genotype-enriched cell states with DEGs. Many of these responses are associated with stress-induced contractile, metabolic, and electrophysiologic properties that are prominent clinical manifestations of some genotypes. For example, attenuated expression of SMYD1, a critical organizer of sarcomere structure, epigenetic and metabolic remodeling, occurred only in CMs with PVs, and was unaltered in PVneg samples. With its global impact on myocardial function, dysregulation of SMYD1 might contribute to earlier presentation and poorer outcomes of cardiomyopathy patients with, versus without, PVs. LMNA CMs had the greatest expansion of vCM1.2, enriched for genes encoding Ca2+ regulators and molecules with electrophysiologic functions, while RMB20 and LMNA genotypes had the highest expression of MYL4, a sarcomere protein associated with atrial fibrillation. These data suggested molecular mechanisms whereby particular genotypes convey increased risks for arrhythmias and sudden cardiac death in patients.

42Shifts in FB states explained the paradoxically increased fibrosis in cardiomyopathies, without expansion of overall FB populations. DCM LVs and ACM RVs showed increased proportions of vFB2, enriched for genes that modulate ECM composition, turnover, stiffness and fibrotic scarring and reciprocally decreased proportions of vFB3 that express transcripts suppressing fibrosis. Notably, fibrogenic genes within vFB2 were highly expressed in RBM20, a genotype within our cohort with the poorest ventricular function and youngest age for HF diagnosis and cardiac transplantation. Strategies to manipulate proteins encoded by these genes may attenuate the prominent myocardial fibrosis that characterizes cardiomyopathies.

43Unexpectedly, mural cells (PCs, SMCs) showed no increase compared to ECs across cardiomyopathies. DEG analysis suggested that these cells promoted vascular remodeling and dysfunction. PCs diminished PDGFRB expression and showed aberrant NOTCH signaling in vascular beds, while genotype selective DEGs in SMCs upregulated contractile genes, augmenting fiber formation. Together these molecular signals may contribute to microvascular dysfunction that occurs in cardiomyopathy patients and adversely influences ventricular performance.

44Among EC, EC7 demarcated the endocardium and had the most DEGs in disease compared to control. Little is known about this myocardial layer in HF pathogenesis. The endocardium forms via dynamic regulation of NOTCH, neuregulin and BMP signals. Pediatric heart diseases can exhibit pathological expansion of the endocardium (denoted endocardial fibroelastosis) which diminishes cardiac performance and is associated with progression to HF. Our data suggested similar dysregulation of these molecular pathways in adult-onset cardiomyopathies with increased BMP6 and NRG1 in EC7 and decreased INHBA. These signals may cause pathological changes in endocardium in DCM and ACM hearts and contribute to myocardial dysfunction.

45Diseased hearts increased myeloid cells. Expansion of immune cell populations might arise from recruitment of circulating cells or proliferation of resident cells. In our samples proliferating myeloid cells decreased, as previously observed in other DCM studies. Surprisingly, myeloid recruitment via the CCR2/CCL2 axis, primarily mediated via vFB3, was decreased in all samples, with particularly dysregulated fibroblast-to-macrophage interactions in TTN. Although distinguishing between proliferation versus recruitment late in disease was problematic, previous studies indicate peak cytokine expression precedes the emergence of HF. Analyses of earlier time points during disease progression will be important to discern these key signals.

46To confirm and expand the conclusion that genotype-specific signals meaningfully contributed to disease pathogenesis, we employed machine learning strategies. Using GAT to classify patients’ genotypes from each cell type, we found that CMs, FBs, ECs, and myeloid cells provided the highest and most discriminatory information. Harnessing LV and RV data from these cell types, we independently predicted the established genotype of each patient with high accuracy. Moreover, among the four samples with the lowest genotype predictive probability, three samples carried two PVs, indicating that the model detected subtle transcriptional differences with additional influences.

47While expanding machine learning models to much larger datasets will undoubtedly improve accuracy, these early analyses supported the conclusion that PVs in different genes evoked cell type and state-specific responses that altered inter-cellular communications and promoted distinct disease pathways. We recognize that pathways may converge, but even in advanced disease, our data indicated that genotypes promoted specific transcriptional signals that likely contributed to distinct as well as common manifestations of genetic cardiomyopathies.

48Future studies are needed to comprehensively define the molecular pathophysiology of cardiomyopathies and HF, including assessments of age, sex, and ancestral influences, other DCM and ACM genotypes, additional cardiac regions, and longitudinal analyses to identify initiating and secondary processes. We also expect that the deployment of strategies that upsample conduction system and other rare cell types, and incorporation of techniques to characterize the epigenome, proteome, and spatial relationships between cell types, states, and gene expression will also be highly informative. To promote these initiatives, we freely provide all datasets and an interactive platform (https://cellxgene.cziscience.com/collections/e75342a8-0f3b-4ec5-8ee1-245a23e0f7cb) with cell type and state annotations. We expect that these resources will advance mechanistic studies to improve treatments of cardiomyopathies and enable HF prevention strategies.

49Detailed information on human subject studies, experimental methods, data access, codes, algorithms, and computational programs used in this manuscript is provided in the Supplementary Materials.

50Human studies were performed using protocols that were reviewed and approved by the ethics boards of participating institutions. DCM and ACM ventricular samples were collected from genotyped patients undergoing mechanical support (n=15) or heart transplantation (n=31) and from deceased donors with non-failing hearts as previously described. Nuclei from full-thickness LV and RV regions were isolated and processed for snRNAseq as described. Data were mapped to the human genome (GRCh38), processed to remove doublets and identify nuclei that met high quality standards, and harmonized to remove batch effects. Manifolds were constructed using Uniform Manifold Approximation and Projections (UMAPs) for all and individual cell types.

51Differential abundance analyses of cell types and states were performed using centered log ratio transformation including a linear model. Differential gene expression between disease and control tissues were deduced using a pseudobulk approach and EdgeR. Comparative analyses assessed cell type and cell state abundances and differential gene expression between disease samples and controls and between genotypes. Selected genes with differential expression were validated using single molecule fluorescent in situ hybridization or quantitative immunohistochemistry.

52We investigated cell-cell communication using CellChat. The expression of genes previously identified through genome wide-association studies of DCM was assessed in diseased and control tissues. We interrogated transcriptional datasets with machine learning tools to generate cell type features that distinguished PVs within the same gene from PVs within different genes and used these data to generate a graph attention network (GAT). The accuracy for GAT analyses of randomly selected patient data to assign the correct, clinically assigned genotype was assessed.