Polarity Collapse in the Hypertrophied Heart: A Cross-Species, Cell-State-Resolved Test

The wall of a healthy heart is not a stack of parallel fibers. Layer by layer, cardiomyocytes fan out at subtly different angles — a 3D "twist" of roughly 10–15° between adjacent layers — and the whole organ wrings like a towel when it contracts. In failing hearts, this architecture collapses into a rigid, uniform alignment with layer-to-layer angle spread of only 5–7°, a change that actually makes individual cells contract faster in a dish while destroying the organ's torsional pumping action (Masuyama et al., intravital heart imaging). Nobody knows which molecular program coordinates this tissue-scale reorganization. Developmental biologists have spent decades dissecting exactly this class of problem — how a tissue-wide "compass" (planar cell polarity, or PCP) orients thousands of cells relative to one another — but their machinery has never been connected to cardiac remodeling.

This proposal makes a specific claim: pathological alignment remodeling in the hypertrophied heart is a polarity collapse, analogous to what happens in developmental epithelia when the tissue-scale directional cue is lost and local cell–cell coupling defaults cells into uniform, parallel orientation (Matis et al.). Testing this claim faces two obstacles that the four research directions conveniently supply. First, whole-tissue ("bulk") molecular measurements average over many cell types and would hide which cells are losing their compass — the masking problem documented directly in diseased human hearts (Reichart et al., cardiomyopathy snRNA-seq atlas). Second, the polarity genes are conserved, but the non-coding regulatory DNA that controls when and where they turn over rapidly between human and mouse (Yue et al., Mouse ENCODE), so a mouse answer may not be a human answer — 42.3% of human cardiomyocyte genes downstream of one cardiac transcription factor were simply absent from the mouse model of the same mutation (Steimle et al., Holt-Oram snRNA-seq). The program below therefore pairs orientation-resolved imaging with single-nucleus sequencing in mice, maps the results onto existing human atlases and bulk cohorts with a divergence-aware deconvolution method, and tests the mechanism with a drug already known to bind the receptor sitting at the fork between canonical and polarity Wnt signaling.

Core Claim

The claim has three parts, each falsifiable on its own. First, the loss of myocyte twist under pressure overload is driven by failure of the non-canonical Wnt/PCP module: tissue-scale directional information (the analog of the Fat/Dachsous/Four-jointed gradient system that orients microtubule tracks and vesicle trafficking in fly wing epithelium Matis et al.) degrades, while local PCP coupling remains, producing locally coherent but globally uniform alignment — exactly the geometry Masuyama observed. Second, this failure is molecularly visible only in specific cardiomyocyte stress states that bulk transcriptomics averages away. Third, the regulatory wiring of this module diverges enough between mouse and human that mouse-only inference would misidentify the human-relevant drug targets. The unifying quantitative object is a polarity collapse index,

$$C = 1 - \sigma_{\angle}/\sigma_{\angle}^{\text{healthy}},$$

where $\sigma_{\angle}$ is the layer-to-layer angle standard deviation measured by intravital two-photon imaging; healthy hearts sit near $C = 0$ and dilated cardiomyopathy (DCM) hearts near $C \approx 0.5$–$0.6$, using the 10–15° versus 5–7° values reported by Masuyama et al..

The mechanistic motivation comes from work on how polarity is maintained under physical deformation, not just established. Banavar et al. showed in budding yeast, via coupled 3D stochastic simulations and continuum mechanics, that a polarity cap forms by symmetry breaking but cannot stay anchored at the growing tip without active mechanical feedback from the cell wall. Cardiomyocytes have no cell wall; their analog is the integrin/focal-adhesion/costamere complex, with titin ($TTN$) as the strain-bearing linker. Our model is that chronic pressure overload saturates this feedback, the tissue-scale directional cue degrades, and alignment homogenizes. This also gives a natural explanation for why uniform alignment looks "adaptive" in reductionist assays — Masuyama's micropatterned cells showed higher one-dimensional contraction velocity ($P < 0.01$) — while being lethal geometry at the organ level: local coupling, unguided, optimizes the local metric.

Research Program

The program has four workstreams. The first builds a paired orientation–transcriptome atlas in mice. We run a transverse aortic constriction (TAC) time course in $mT/mG$ mice — the same model used to show empagliflozin's benefit in pressure overload Yuan et al. — imaging at 2, 4, and 8 weeks with motion-stabilized two-photon microscopy, and adding tissue clearing plus light-sheet imaging of the same hearts to extend orientation measurement beyond the ~100 µm scattering limit into mid-myocardial and endocardial layers, which intravital imaging cannot reach. Immediately after imaging, we perform snRNA-seq on the imaged region, so every nucleus carries an approximate 3D position and local angle statistics from its neighborhood. We define a "polarome" gene set — mammalian Fat/Dachsous homologs, core Frizzled/Dishevelled/VANGL/Prickle components, and the costamere/focal-adhesion mechanotransducers — and ask which cell states lose polarome expression or, via immunostaining for VANGL2/Dishevelled membrane asymmetry, lose polarized localization while the proteins are still present. The developmental expectation is explicit: asymmetry of core PCP proteins at cell interfaces is the readout of a functioning compass Matis et al.; uniform membrane staining means the compass is off even if the cell looks structurally intact. Orientation itself is quantified with the nematic tensor framework developed for non-vectorial alignment axes in 3D tissues quantification of nematic cell polarity.

The second workstream tests the mechanism with a perturbation. Empagliflozin binds Frizzled-4 and blocks canonical β-catenin signaling, attenuating hypertrophy markers in TAC mice Yuan et al. — but Frizzled receptors are shared between the canonical arm and the non-canonical PCP arm, and that study measured only the canonical reporter. We repeat the TAC + empagliflozin design with the polarity readouts added: PCP protein asymmetry, $C$, and polarome expression in the stressed cardiomyocyte states from workstream one. Three outcomes are informative: restoration of asymmetry and twist (the drug re-opens the polarity arm), further suppression of both arms (therapeutic effect is purely canonical, contradicting our model), or asymmetric loss only in specific cell states (masking is real and state-specific).

The third workstream carries the result across species. We map the mouse polarome cell states onto the human atlas of ~880,000 nuclei from 61 failing and 18 control hearts, which resolved ten cell types and 71 states and showed genotype-specific cardiomyocyte depletion and profibrotic fibroblast ($vFB2$) expansion that bulk profiling completely obscured Reichart et al.. Conservation is scored with the CACIMAR approach, which computes reference-free conservation scores for cell types, regulatory networks, and intercellular interactions and showed in retinal regeneration that cell types can preserve ligand–receptor interactions while their internal marker genes and networks diverge completely CACIMAR — in that dataset, 11 of 15 zebrafish retinal cell types were strongly conserved in mouse while nearly 2,500 markers were species-specific. We then ask whether the putative polarity-collapse states exist in human DCM hearts, and whether their regulatory networks or only their markers diverge. To probe the enhancer layer directly, we fine-tune a Nvwa-style multitask sequence model — Nvwa predicts cell-type-specific expression from 13kb promoter sequences across 480 cell types in 8 species at AUROC 0.78 and AUPR 0.59, and found 94.9% of transcription factor motifs homologous while their higher-order syntax diverges Nvwa — on polarome and epistatic-locus enhancers, testing whether predicted activity in stressed cardiomyocyte states transfers between human and mouse sequences.

The fourth workstream resolves the masking question in existing human data and is where the methodological contribution lives. The epistasis study that discovered non-additive genetic control of left ventricular mass (loci near $CCDC141$, $TTN$, $IGF1R$, $TNKS$, invisible to additive GWAS) also showed that these genes' co-expression collapses in bulk transcriptomes of 313 failing versus 136 healthy hearts — but bulk cannot distinguish real network disassembly inside cardiomyocytes from an artifact of shifting cardiomyocyte-to-fibroblast ratios Epistasis regulates genetic control of cardiac hypertrophy. We deconvolve that cohort with a divergence-aware ensemble: proportions $\pi$ are estimated by solving

$$\hat{\pi} = \arg\min_{\pi \ge 0}\Big\| b - \sum_k w_k R_k \pi \Big\|^2, \qquad w_k \propto \text{CSCT}_k \cdot \text{CSRN}_k,$$

where $b$ is the bulk count vector, $R_k$ is a single-cell reference matrix from species $k$ (human atlas and mouse TAC atlas), and $w_k$ weights each reference by the CACIMAR conservation scores for the relevant cell types and networks rather than by reconstruction error alone — the extension that existing ensemble methods lack. SCDC already showed that weighting multiple references matters within a species, reaching Pearson 0.991 on controlled mixtures and recovering a beta-cell/HbA1c correlation ($p = 0.0019$) that single-reference tools missed SCDC, and Count Bridges showed that respecting the integer geometry of counts improves on continuous methods like CIBERSORTx and MuSiC, though no raw decimal error values are reported on Lacuna for Count Bridges Count Bridges. We use Count Bridges as the count-faithful estimator inside our ensemble. After deconvolution, we recompute co-expression within cardiomyocyte-inferred proportions: if connectivity is restored, the "collapse" was compositional; if it persists, it is a genuine within-cardiomyocyte network failure, and we test whether the collapsed genes overlap the polarome.

Proposed Experiments

The decisive time-course prediction is ordering: if polarity collapse is causal, loss of PCP protein asymmetry in stressed cardiomyocyte states should precede geometric homogenization (rising $C$), which should precede overt dysfunction. TAC hearts at 2 versus 4 versus 8 weeks give three time points; the model predicts asymmetry loss at the earliest point where $C$ is still near zero. A parallel clinical framing motivates sensitivity: adverse ventricular structural changes begin at blood pressures considered normal, with no clear threshold de Marvao et al., high-resolution MRI, though no numerical regression coefficients are reported on Lacuna for that study; we therefore test whether $C$ or PCP asymmetry separates mild-overload from sham hearts at time points where bulk fetal-gene markers ($Nppa$, $Nppb$, $Myh7$ — the canonical hypertrophy markers re-induced by pressure overload Morioka et al. and suppressed by empagliflozin Yuan et al.) do not.

For human validation despite tissue scarcity — the Holt-Oram study that revealed 42.3% human-specific TBX5 targets rested on a single patient Steimle et al. — we use human iPSC-derived cardiomyocytes. The epistasis paper established the assay: dual siRNA knockdown of $CCDC141$ and $IGF1R$ in diseased hiPSC-CMs produced an 8.5% non-additive reduction in median cell size at $P < 1 \times 10^{-4}$, measured only as cross-sectional area in microfluidic channels. We repeat that knockdown on both uniformly micropatterned substrates and 3D layered cultures engineered to reproduce physiological twist, adding orientation and PCP-asymmetry readouts. If the epistatic loci interact with the polarity machinery, the non-additive effect should appear in alignment phenotypes, not only cell size.

Baselines

The naive reference point is the standard bulk pathological-hypertrophy readout: ventricular expression of the fetal gene program ($Nppa$, $Nppb$, $Myh7$), as used in the TAC literature Morioka et al. and Yuan et al.. It is the field's default diagnostic of maladaptation and requires no new instrumentation. We will run it on the same TAC hearts and time points, with matched sample numbers, and count our orientation metrics as beating it only if $C$ or PCP asymmetry separates groups (per-mouse classification, pre-registered threshold) at a time point where the fetal-gene panel does not, or predicts later decompensation (chamber dilation) better in a held-out cohort.

The strongest competitor for the masking workstream is SCDC ensemble deconvolution SCDC, which reported Pearson 0.991 against ground-truth controlled mixtures and was uniquely able to replicate the beta-cell/HbA1c clinical correlation ($p = 0.0019$) among its peers; it is the method to beat on within-species deconvolution of the 313-heart cohort. We run SCDC unmodified on the same bulk data, same references, same compute and tuning budget, and additionally in a cross-species configuration (mouse reference, human target) it was never designed for. Our divergence-weighted ensemble beats it if it recovers known ground-truth composition shifts — the $vFB2$ expansion and cardiomyocyte depletion documented by Reichart et al. on hearts where paired snRNA-seq and bulk exist — with lower proportion error, and remains calibrated in the cross-species configuration where SCDC's error-only weighting degrades. The key ablation runs the identical estimator with single-species references (human-only and mouse-only) versus the conservation-weighted dual-species ensemble; the difference in proportion error and in the within-cardiomyocyte co-expression estimate isolates exactly the ingredient this proposal adds. For enhancer prediction, Nvwa's reported AUROC of 0.78 and AUPR of 0.59 Nvwa on cross-species expression-from-sequence is the baseline our polarome-focused fine-tuning must at least match on general cell types while improving on stressed-cardiomyocyte states.

Why not just…?

Why not just use mouse TAC models and standard readouts? Because the specific molecular layer we target is the one that turns over fastest between species: roughly half the human genome cannot be aligned to mouse, concentrated in regulatory regions Yue et al., human cardiac enhancer usage is largely absent at orthologous mouse cell types Li et al., brain cCRE atlas, and 42.3% of human TBX5 cardiomyocyte targets were never seen in the mouse knockout Steimle et al.. The mouse arm of this proposal is not a shortcut to human conclusions; it is one end of a measured divergence axis, and the CACIMAR scores tell us which conclusions are allowed to cross it.

Why not just run snRNA-seq on failing human hearts and skip the mouse and the imaging? Because snRNA-seq destroys exactly the phenotype under study: nuclei suspensions discard all 3D coordinates and orientation, which is why Reichart et al. could characterize 71 cell states but say nothing about myocyte twist or planar polarity. And primary human diseased tissue is so scarce that the only prior single-nucleus cardiac regulatory-divergence dataset rests on $n = 1$ Steimle et al.. The paired imaging-then-sequencing design in mice, plus deconvolution of the existing 313-heart bulk cohort Epistasis paper, is the affordable way to get human-relevant, orientation-linked answers.

Why not just use existing deconvolution (SCDC, CIBERSORTx, MuSiC) on the bulk cohort? For within-species, human-reference deconvolution, SCDC is a genuine threat and is treated as the baseline to beat, not an argument to dismiss — the proposal's masking claim would survive even if SCDC alone resolves the co-expression question, and we will report that honestly if it does. The gap we target is two-fold: single-reference and reconstruction-error-only methods have no mechanism for cross-species reference mismatch, which CACIMAR shows is severe (nearly 2,500 species-specific markers in an ostensibly conserved tissue), and their accuracy is bounded by reference completeness — if the atlas lacks the early mechanically stressed cardiomyocyte state our mouse time course is designed to capture, those reads get misassigned to canonical states. The single-versus-dual-species ablation settles empirically whether this matters here, rather than by argument.

Why not just measure cell size or ejection fraction instead of 3D orientation? Because the simplest metrics point the wrong way: uniformly aligned cardiomyocytes generate significantly higher one-dimensional contraction velocity and deformation distance ($P < 0.01$) in vitro, which reads as beneficial in a dish, while representing the loss of the physiological 10–15° twist that enables organ-level wringing Masuyama et al.. Any screen built on cross-sectional area or single-axis contractility would actively select for the pathological geometry.

Decisive Checks

Three outcomes would cleanly decide the core claim. If PCP asymmetry loss precedes geometric homogenization in the TAC time course, and if empagliflozin's Frizzled-4 binding restores polarity readouts rather than only suppressing β-catenin, the polarity-collapse mechanism is supported and FZD4/PCP modulators gain a human-relevant, orientation-based efficacy endpoint. If instead asymmetry and geometry move together or geometry fails first, the collapse is mechanically driven and the PCP framing dies — but the paired atlas, the divergence scores, and the deconvolution of the 313-heart cohort remain independently valuable, settling whether epistatic co-expression collapse is real network disassembly or a fibroblast-ratio artifact, and quantifying for the first time how much of the hypertrophic polarity program a mouse is actually permitted to model for humans.

References

  1. Homogeneous 2D and 3D alignment of cardiomyocyte in dilated cardiomyopathy revealed by intravital heart imaging
  2. Microtubules provide directional information for core PCP function
  3. Pathogenic variants damage cell composition and single cell transcription in cardiomyopathies
  4. A comparative encyclopedia of DNA elements in the mouse genome
  5. Single-nuclei transcriptomics reveals TBX5-dependent targets in a patient with Holt-Oram syndrome
  6. Coordinating cell polarization and morphogenesis through mechanical feedback
  7. Empagliflozin improves pressure-overload-induced cardiac hypertrophy by inhibiting the canonical Wnt/β-catenin signaling pathway
  8. Quantification of nematic cell polarity in three-dimensional tissues
  9. CACIMAR: cross-species analysis of cell identities, markers, regulations, and interactions using single-cell RNA sequencing data
  10. Inferring predictive genetic models and regulatory elements by deep learning of cross-species single-cell gene expression landscapes
  11. Epistasis regulates genetic control of cardiac hypertrophy
  12. SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
  13. Count Bridges enable Modeling and Deconvolving Transcriptomic Data
  14. Adverse changes in left ventricular structure begin at normotensive systolic blood pressures: a high resolution MRI study
  15. Changes in Contractile and Non-Contractile Proteins, Intracellular Ca2+ and Ultrastructures During The Development of Right Ventricular Hypertrophy and Failure in Rats.
  16. A comparative atlas of single-cell chromatin accessibility in the human brain