When a heart is overloaded for years, for example by high blood pressure or a leaking valve, its walls thicken and it eventually fails. Much of what we believe about the molecular side of this process comes from bulk RNA sequencing, which grinds up a piece of heart muscle and measures the average gene activity of millions of cells at once. Two widely cited bulk findings illustrate the problem. In failing human hearts, genes for mitochondrial energy production and the protective factor CIP are turned down. Separately, the co-expression network linking the hypertrophy genes CCDC141, IGF1R, TTN and TNKS falls apart. Both findings can be read as "the muscle cells themselves have rewired."
Both are equally consistent with a duller story. Failing hearts lose cardiomyocytes, the muscle cells, and gain fibroblasts and immune cells. Surviving cardiomyocytes also grow, so each one contributes more RNA. Any of these shifts changes the tissue average without a single cardiomyocyte changing its program. This note proposes a deconvolution method, meaning a way to computationally unmix a bulk sample into per-cell-type contributions. It also proposes a cardiac-specific test bed for checking such methods. The goal is to decide, on existing human cohorts, how much of each finding survives once the cell mix and cell size are accounted for.
Core Claim
Current deconvolution tools have two blind spots in heart tissue. First, they assume a fixed amount of RNA per cell of each type. Second, they assume the reference atlas contains every relevant cell state. Hypertrophy violates both. Cardiomyocytes enlarge, and failing hearts contain cardiomyocyte states that healthy-heart references lack. Our claim is that under these violations, standard tools produce spurious "cardiomyocyte-specific" changes that are really composition effects. We further claim that a model which (a) treats cardiomyocyte RNA content as an explicit per-sample unknown and (b) reserves a residual cardiomyocyte state for unreferenced disease programs can separate three components:
- changes in cell composition,
- changes in cardiomyocyte size,
- genuine within-cardiomyocyte reprogramming.
Cell number and RNA per cell cannot be fully separated from bulk data alone, because only their product is observed. We therefore do not pretend to estimate cell size. Instead, we report each biological conclusion as a curve over a plausible range of cardiomyocyte content ratios. A finding counts as cell-autonomous only if it holds across that whole range.
Grounding in Prior Work
Motivating studies: bulk findings with an open question. Epistasis regulates genetic control of cardiac hypertrophy used UK Biobank heart MRI to find non-additive interactions among CCDC141, IGF1R, TTN and TNKS. It then showed that their co-expression breaks down across 313 bulk-profiled hearts. It confirmed the interaction only by measuring the size of cultured stem-cell-derived cardiomyocytes. Cardiac ISL1-Interacting Protein… reported CIP loss and its link to oxidative phosphorylation across 194 bulk human hearts, plus a mouse pressure-overload dataset (GSE194149). Neither study could say which cells carry the change. We reanalyze exactly these cohorts rather than proposing new wet-lab biology.
Resources supplying references and ground truth. Pathogenic variants damage cell composition and single cell transcription in cardiomyopathies profiled about 880,000 nuclei from 61 failing and 18 control hearts. It documents the very composition shifts, such as cardiomyocyte loss and expansion of the matrix-secreting vFB2 fibroblast state, that could confound bulk data. Cells of the adult human heart supplies the healthy reference, including five ventricular cardiomyocyte states. We use the disease atlas both as a reference and, critically, as the source of simulated bulk samples with known answers.
Studies constraining the design. Performance of tumour microenvironment deconvolution methods in breast cancer… showed, over 38,000 simulated mixtures, that when one cell type dominates, methods misassign related cells to it. BayesPrism and DWLS had the lowest combined errors. Cardiomyocytes dominate heart RNA, so we expect the same trap. The spatial benchmark in A comprehensive benchmarking… warns of platform shifts between nucleus-based and whole-tissue assays. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level (POPPY) aligns bulk and single-cell embeddings but fails when disease states are missing from the reference. That failure mode is what our residual state is for.
Method we reuse. Count Bridges enable Modeling and Deconvolving Transcriptomic Data is a generative model that works directly on integer read counts. It samples latent single-cell counts that must sum exactly to the observed bulk. We reuse it as the within-cell-type generator.
The contrast with Count Bridges and POPPY is specific. We keep the integer bridge as a component, but add a per-sample cardiomyocyte content multiplier and an unreferenced cardiomyocyte state. We also change the question from "estimate proportions" to "certify that a within-type change is not a composition artifact."
Research Program
The model. Write the bulk count vector of sample $j$ as
$$y_j \approx \sum_{k} \pi_{jk}\, s_{jk}\, x_{jk},$$
where:
- $\pi_{jk}$ is the fraction of cells of type $k$,
- $s_{jk}$ is the RNA content per cell of type $k$ relative to healthy,
- $x_{jk}$ is the sample's own expression profile for type $k$.
Non-cardiomyocyte content factors are fixed near one. The cardiomyocyte factor $s_{j,\mathrm{CM}}$ is swept over a grid. At each grid value we fit $\pi$ and $x$ by the Count Bridges EM procedure with projection guidance.
The cardiomyocyte profile $x_{j,\mathrm{CM}}$ is a mixture of two parts. One part is drawn from atlas states, both healthy and failing. The other is a low-rank residual state learned across the cohort. Its magnitude is penalized, so it only absorbs signal that no referenced state explains.
The downstream test. On the deconvolved cardiomyocyte profiles we recompute two quantities: CIP and oxidative phosphorylation scores, and the four-gene co-expression network. Each conclusion is reported as a function of $s_{\mathrm{CM}}$.
The test bed. We build semi-synthetic bulk samples from the disease and healthy atlases, where the truth is known. In the "composition-only" scenario, cardiomyocyte programs are frozen at control donors while fibroblast and immune fractions and cardiomyocyte content are shifted to failing-heart levels. In the "autonomous" scenario, failing-donor cardiomyocyte profiles are inserted at control composition. Crossing the two scenarios lets us measure false and true discovery of within-cardiomyocyte change for every method.
We also include a harder variant. We withhold failing-heart cardiomyocyte states from the reference, to reproduce the reference omission that breaks POPPY.
Baselines
Naive reference. Differential expression and co-expression computed directly on bulk, as in the original cohort analyses. It defines how many findings need explaining.
Covariate adjustment. Proportions estimated by MuSiC or CIBERSORTx are entered as covariates in the bulk regressions. We expect this to be the most-used practical fix.
Strongest existing methods. BayesPrism and DWLS, chosen because the breast-cancer benchmark found them lowest in combined false positives and negatives under high purity. Both produce per-sample cell-type-specific expression, which is the exact output our question needs. The breast study gives no single numeric score, and we found no reported result on Lacuna for BayesPrism or DWLS on cardiac tissue.
Off-the-shelf Count Bridges. Run without the content multiplier or residual state. That paper reports it outperforms CIBERSORTx and MuSiC on bulk proportion recovery, but no reported figure appears in our file.
Ablations. We run our model with only the content multiplier, only the residual state, and neither. This isolates which ingredient matters.
Fairness. All methods receive the same references, the same gene set, and the same held-out donors. Tuning uses an equal hyperparameter-search budget on simulated samples built from donors disjoint from the evaluation donors.
What counts as beating the baselines. At matched power on "autonomous" samples, our model must cut the false discovery rate for within-cardiomyocyte change on "composition-only" samples by a margin clearly outside donor-resampling intervals.
Why not just…?
Why not just regress out estimated cell proportions? Covariate adjustment corrects for cell number, but it cannot see RNA-per-cell changes. It also inherits proportion errors from the purity trap, and those errors are correlated with disease status. The composition-only simulations test this directly. If covariate adjustment already controls false discoveries there, we will say so and recommend it over our model.
Why not just analyze cardiomyocytes in the snRNA-seq atlas and skip bulk entirely? This is the most serious threat, and we use the atlas for exactly this cross-check. But it has three limits:
- The atlas covers only end-stage explanted hearts, so it misses the compensatory phase present in the 194-heart cohort and the mouse pressure-overload data.
- Nucleus sequencing misses cytoplasmic and mitochondrially encoded transcripts, which are central to the energy-production signal.
- Nuclei do not report cell size.
Where the atlas and our deconvolved bulk agree on end-stage hearts, we gain confidence to extend the conclusions to earlier stages.
Why not just use BayesPrism or Count Bridges as-is? They may suffice. The ablation arm answers this. If the content multiplier and residual state add nothing on the omitted-reference test bed, the contribution shrinks to the benchmark itself plus the reanalysis, which remains useful.
Decisive Checks
The first check is the crossed semi-synthetic benchmark described above, which calibrates every method before any real data are touched.
The second is agreement on real hearts. On the failing hearts in the disease atlas, we compare within-cardiomyocyte co-expression of CCDC141, IGF1R, TTN and TNKS from our deconvolved bulk-like pseudo-samples against direct nucleus-level estimates.
The third is the real reanalysis. On the 313-heart and 194-heart cohorts we report, for each finding, the range of cardiomyocyte content ratios over which it survives. Three outcomes are possible:
- The co-expression breakdown survives only at implausibly low content ratios. It is then mostly composition.
- CIP loss survives across the whole range. It is then a cell-autonomous cardiomyocyte change.
- The residual state absorbs the oxidative phosphorylation loss. That would point to an unreferenced cardiomyocyte state as the carrier of the tipping-point signal, and it would be the most interesting outcome for future single-cell sampling of pre-failure hearts.
References
- Epistasis regulates genetic control of cardiac hypertrophy
- Cardiac ISL1-Interacting Protein, a Cardioprotective Factor, Inhibits the Transition From Cardiac Hypertrophy to Heart Failure
- Pathogenic variants damage cell composition and single cell transcription in cardiomyopathies
- Cells of the adult human heart
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data