Single-cell imaging exposes a hidden layer of cardiomyocyte variability
Human induced pluripotent stem cell-derived cardiomyocytes have become a workhorse for disease modeling, yet their readouts are often collapsed into a single well average. CardioShift was built to challenge that habit. The framework is designed to analyze high-content single-cell imaging data while preserving the nested structure of the experiment, so individual cells are not mistaken for independent biological replicates when they are actually clustered within the same well.
That statistical distinction is not cosmetic. In cardiomyocyte phenotyping, a well average can flatten rare but biologically meaningful subpopulations, especially when contractility and calcium handling vary widely from cell to cell. The preprint makes the case that this variation is not background noise to be smoothed away; it is often the dominant signal.
A framework built to avoid pseudoreplication
CardioShift is presented as an open-source bioimage informatics framework for mapping single-cell distributions in hiPSC-derived cardiomyocytes. Its central goal is to handle pseudoreplication bias, the error that occurs when cells are analyzed as if each one were an independent experimental unit. In reality, a single well can contain many cells shaped by the same differentiation batch, culture conditions, and imaging context, so variance must be interpreted at the correct level.
The analysis was applied to two large datasets: 32,818 Ca2+ transients and 10,719 contractility-lipid profiles. Those numbers matter because they are large enough to reveal structure within the population rather than just a mean response. Across most functional parameters, cell-to-cell variability emerged as the largest source of variance. In practical terms, the distribution of responses carried more information than the average response itself.
A one-number summary can hide a phenotypic split right in front of the microscope.
Calcium handling and contraction do not move in lockstep
The most important message from the dataset is that traditional well averages can mask condition-dependent phenotypes. Calcium transients and contractility measures are related, but they do not always shift together in a uniform way across every cell. A population average can make a culture appear stable even when distinct subgroups are moving in different directions.
CardioShift addresses that problem by resolving the single-cell distribution rather than forcing every cell into a composite value. This approach is especially relevant in cardiomyocyte biology, where maturation state, metabolic shift, excitation-contraction coupling, and membrane transport can vary substantially across cells differentiated under the same nominal conditions. Heterogeneity is not merely a nuisance in that setting; it may be a signature of incomplete maturation, divergent cell states, or differential responses to experimental perturbation.
The framework’s value lies in making those differences visible without inflating the apparent sample size.
Lipid accumulation marked a distinct cellular state
Unsupervised clustering added a second layer of insight. Beyond simple averaging, the analysis identified condition-dependent shifts in functional subpopulations, including a distinct cellular state that accumulated lipids while maintaining baseline contraction kinetics. This subgroup was missed by both well-level analysis and single-cell averaging.
That finding is important because it separates metabolic phenotype from mechanical output. A cardiomyocyte can appear contractile-normal while still drifting into a lipid-loaded state, and a mean readout would likely miss that divergence. In the context of cardiac cell modeling, this raises immediate questions about how metabolic remodeling is tracked and whether apparently normal beating masks an underlying change in cell state.
The biology is plausible. Lipid handling and contractile function share the same cellular machinery but do not always fail in parallel. A cell can preserve its calcium-driven contraction pattern while undergoing metabolic reprogramming that changes substrate use, storage, or stress susceptibility. CardioShift makes that separation easier to detect.
Why the analytic layer matters for disease modeling
hiPSC-derived cardiomyocytes are widely used for disease modeling, drug response studies, and screening workflows. Their usefulness depends on whether the assay captures true biology or just a diluted average. When one cell group shifts functionally while another remains stable, a pooled readout can produce a reassuring mean that conceals meaningful pathology.
CardioShift offers a way to move beyond that ceiling. By explicitly accounting for nested variance, it can help investigators distinguish three different situations: a genuinely uniform response, a mixed population with hidden subgroups, and a misleading mean that obscures a phenotype present only in part of the culture. In a screening context, that difference could change whether a compound looks active, inactive, or selectively active in a subpopulation.
The preprint positions the framework as more than a software package. It is an argument for treating single-cell imaging data as a structured biological hierarchy rather than a pile of interchangeable datapoints.
Open-source access and the limits of the current evidence
The authors provide CardioShift as an open-source resource, with code available through GitHub. That lowers the barrier for adoption in other cardiomyocyte imaging pipelines and makes the analytical approach easier to test against independent datasets. The preprint also makes clear that the work is a preprint and has not undergone peer review, so its methods and conclusions still need external validation.
Even with that caveat, the message is concrete: cell-to-cell variation is large, actionable, and easy to underestimate when data are compressed into averages. For hiPSC-CM research, that means the unit of analysis must match the biology under study, not the convenience of the summary statistic.
The remaining question is not whether heterogeneity exists. It is how many important phenotypes have already been flattened by the average before anyone noticed them.