The research, led by Mari Tolonen, Ziwei Xu, and Jakub Sedzinski of the University of Copenhagen together with Bianca Dumitrascu’s group at Columbia University, tackled a problem that has long frustrated developmental biologists: how to quantify the behavior of thousands of individual cells inside a densely packed, actively remodeling tissue. Morphology is often treated as a downstream consequence of fate decisions, a passive readout of the genetic program. The team set out to test a bolder idea, namely that the dynamic geometry of a cell might serve as a complementary, and potentially predictive, indicator of its state, in the same way that transcriptomic profiles define molecular identity.

Their experimental system was the Xenopus mucociliary epithelium, a bilayered embryonic skin that closely resembles the cellular composition of the mammalian airway. The researchers dissected animal cap explants from embryos at early blastula stages, when the tissue was still multipotent, and mounted them on fibronectin-coated dishes where they flattened into stratified epithelia amenable to long-term optical access. Embryos had been injected with messenger RNAs encoding membrane-localized mNeonGreen and nuclear H2B-RFP, allowing both cell boundaries and nuclei to be visualized. Volumetric confocal time-lapse imaging then captured the full course of differentiation, from NF stage 9 through stage 30, with z-stacks acquired every three to five minutes over 16 to 22 hours across three independent experiments.

Turning those raw movies into quantitative data required a substantial computational pipeline. Nuclei were segmented in three dimensions using a custom-trained StarDist model, while membrane signals were first denoised with a CARE network and then segmented slice by slice with Cellpose, with individual 2D contours stitched into 3D volumes based on maximal pairwise overlap. Nuclear objects were tracked over time using TrackMate, and the team integrated a convolutional neural network called Oneat to detect cell division events, which were used to impose branching points on trajectories. This combination substantially improved mitosis detection compared with TrackMate’s native algorithm, raising branching correctness from 0.122 to 0.328. On average, trajectories spanned at least 9.5 hours in a third of cells, providing sufficient temporal coverage for the analyses that followed.

From the segmentations and tracks, the researchers computed a rich set of time-resolved features for each cell: three-dimensional shape descriptors derived from point clouds resampled from surface meshes, including surface area, eccentricity, and orientation; conventional two-dimensional descriptors such as area, Feret diameter, and solidity taken from each membrane object’s central slice; nuclear shape metrics; spatial position; and instantaneous movement. The result was a high-dimensional morphodynamic feature space, conceptually analogous to a transcriptomic state vector, in which each cell’s identity at each moment was represented by dozens of quantitative measurements.

When the team explored this feature space with unsupervised methods, the picture was sobering. Principal component analysis of more than 4,200 cell trajectories revealed no distinct clusters corresponding to cell types or developmental stages. Cells were distributed continuously, with only slight shifts over time, and standard clustering metrics such as the Adjusted Rand Index and Normalized Mutual Information were uniformly low. A simple K-nearest-neighbor classifier reached a mean accuracy of just 65.8 percent against a stratified baseline of 54.0 percent, with balanced accuracy of only 28.1 percent. Unlike cultured cells, where shape heterogeneity is high and morphology alone can resolve distinct states, cells in this densely packed epithelium share constrained shapes and largely collective movement, making them hard to distinguish by any single feature.

The breakthrough came from supervised learning combined with lineage backtracking. After live imaging ended, the samples were fixed and immunostained for markers of multiciliated cells, goblet cells, small secretory cells, and basal cells, while ionocytes were identified by their characteristic morphology and lack of marker expression. By aligning the immunostained images with the final live-imaging frame, the team assigned terminal fates to roughly 75 percent of all tracked trajectories, producing a ground-truth dataset of 2,576 curated lineages. They then trained an XGBoost gradient-boosted classifier and a multinomial logistic regression model, treating each cell at each time point as an independent observation labeled by its track’s endpoint fate, with absolute experimental time included as an additional input.

The supervised models achieved moderate but consistent performance across 20 independent train-test splits. XGBoost reached a mean accuracy of 80.6 percent and a balanced accuracy of 61.4 percent, while logistic regression achieved 70.3 percent accuracy with a comparable balanced accuracy of 62.9 percent. Performance was strongest for the abundant epithelial lineages: basal cells scored an F1 of 0.90, goblet cells 0.72, and multiciliated cells 0.56 under XGBoost, whereas the rare ionocytes and small secretory cells remained difficult to classify despite class balancing with SMOTE-based oversampling and undersampling. Notably, classifier confidence and accuracy rose steadily during development, with mean confidence increasing from 0.80 at around five hours to 0.89 by 22 hours, and class-balanced accuracy climbing from 50.4 to 75.1 percent. Intriguingly, predictions for multiciliated cells peaked at roughly 15 hours, with 80.6 percent accuracy, and then declined, suggesting these cells reach a morphodynamic plateau before terminal maturation, after which their features increasingly overlap with those of goblet cells.

Feature importance analysis revealed which aspects of cellular behavior carried the predictive signal. Normalized Z position, reflecting a cell’s height within the apico-basal axis of the epithelium, was the single most informative feature, consistent with the invariant layer preferences of different lineages: goblet cells remain in the superficial layer, basal stem cells stay confined to the base, and intercalating lineages ascend. The nucleus-to-membrane centroid offset was the most predictive morphological variable in the XGBoost model, likely capturing the transition from a flat, migratory configuration to a vertically oriented shape during radial intercalation. Absolute experimental time was also a strong predictor, particularly for small secretory cells, which emerge late in development. Perhaps most strikingly, movement features contributed almost nothing to fate prediction, apparently because migration within the tissue is largely collective rather than fate-specific. Nuclear shape features proved more informative than membrane shape, which the authors attribute to the higher segmentation fidelity of nuclei.

The broader significance of this work lies in what it suggests about the nature of cell state itself. Fate, the study indicates, is not defined by discrete morphological states that cells snap into; instead, morphodynamic signatures of identity emerge gradually along continuous, overlapping trajectories, and only become legible through time-integrated, supervised analysis. Because morphodynamic tracking provides dense temporal sampling with positional context, something neither single-cell RNA sequencing nor spatial transcriptomics can fully deliver, it could complement molecular fate mapping by filling temporal and spatial gaps. The authors note that their pipeline, built from open tools including StarDist, Cellpose, TrackMate, and Oneat, can be extended to other developing tissues and combined with transcriptomics or targeted perturbations to uncover causal links between gene expression, signaling, and phenotype. Future models that incorporate neighboring cells and local tissue geometry may sharpen predictions further. For now, the message is clear: the shape a cell takes, where it sits, and when it does so are not mere consequences of its destiny but readable, quantitative clues to what that destiny will be.

Subject of Research: Machine learning prediction of cell fate from single-cell morphodynamic features during Xenopus mucociliary epithelial differentiation

Article Title: Single-cell morphodynamics predict cell fate decisions during mucociliary epithelial differentiation

Article References: Tolonen, M., Xu, Z., Beker, O., Kapoor, V., Dumitrascu, B., & Sedzinski, J. (2026). Single-cell morphodynamics predict cell fate decisions during mucociliary epithelial differentiation. Molecular Systems Biology, 22(7), 1040-1069. https://doi.org/10.1038/s44320-026-00212-x

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00212-x

Keywords: cell fate, cell differentiation, Xenopus laevis, mucociliary epithelium, live imaging, morphodynamics, machine learning, XGBoost, cell tracking, segmentation, radial intercalation, developmental biology

The logic behind ICE models is deceptively simple. When two species respond similarly to a wide range of chemicals, the toxicity of a new compound in one species can be estimated from its measured toxicity in the other. The challenge lies in demonstrating that the correlation is strong enough, and based on enough matched data points, to be trustworthy. For this model, the authors gathered acute toxicity data for both species from publicly available databases, pairing 96-hour LC50 values, the concentrations lethal to half of the exposed animals over four days, for the same chemicals in frog tadpoles and in juvenile rainbow trout.

Raw database values cannot simply be pooled. The researchers curated the dataset carefully, converting units where necessary so that all values were directly comparable, and calculating geometric mean values whenever multiple LC50 values existed for the same chemical. They also checked the paired data for potential outliers that might distort the regression. After this cleaning process, the final analysis rested on 138 LC50 values covering 69 organic chemicals, a respectable sample size for a model intended to support regulatory hazard assessment.

The statistical outcome was clear. Spearman’s rank correlation coefficient between frog and trout toxicity values was 0.86, with a p-value below 0.0001, indicating a highly significant relationship across the chemicals tested. The regression itself yielded an R-squared of 0.74, also with p below 0.0001, and the model was judged statistically robust according to the evaluation criteria built into the US EPA’s Web-ICE software, the standard tool for this class of interspecies estimation models. In practical terms, knowing how toxic a chemical is to rainbow trout explains roughly three quarters of the variation in its toxicity to African clawed frog tadpoles.

Why build a frog model from fish data at all? The answer lies in how the two species are actually used in testing. Rainbow trout are among the fish species most commonly employed for acute toxicity testing worldwide, so trout LC50 values exist for a very large number of chemicals, including many pesticides and industrial organics. The African clawed frog, by contrast, is primarily used in endocrine assays, tests that examine how chemicals interfere with hormones rather than how quickly they kill. Acute lethality data for the frog are therefore comparatively rare, even though the species is a standard laboratory organism with a long history in aquatic toxicology.

That scarcity matters in specific regulatory contexts. Acute toxicity information for Xenopus laevis may be needed to set test concentrations for endocrine assays, ensuring that animals in those experiments are exposed to meaningful but not uniformly lethal doses. Where required, frog LC50 values also serve as inputs for aquatic risk assessments for amphibians. Without a model like this one, obtaining such a value could mean commissioning an additional vertebrate study, with all the ethical, financial and temporal costs that entails. The authors note that by applying this robust ICE model, additional vertebrate studies can be avoided altogether, a point that resonates strongly with ongoing efforts across toxicology to reduce animal testing.

The new model fits into a broader and rapidly maturing framework. The US EPA maintains the Web-ICE platform, which hosts a family of interspecies correlation estimation models linking standard test species to one another, and recent work has focused on uncertainty analysis and updated user guidance for these models, including their behavior with low-toxicity compounds. Other recent evaluations have examined how ICE models can increase the taxonomic diversity of chemical assessments while reducing reliance on animal testing under statutes such as the Toxic Substances Control Act. Adding a fish-to-amphibian model extends this toolkit to a class of vertebrates, amphibians, that sits between fish and reptiles in many assessment schemes but has historically been underrepresented in acute toxicity databases.

Amphibians occupy an awkward position in chemical regulation. Their aquatic larvae share exposure routes with fish, yet their biology differs in ways that could plausibly alter sensitivity, from permeable skin and developmental transitions to distinct metabolic pathways. Earlier comparative studies have examined the relative acute and chronic sensitivity of fish and amphibians, and case studies have asked whether fish assays and standardized frog embryo assays are protective enough for amphibian larvae. The strong correlation reported here between trout and frog acute toxicity suggests that, at least for the organic chemicals covered by the 69-compound dataset, rainbow trout data capture much of the interspecies variation in lethal sensitivity, lending empirical weight to the practice of using fish as a surrogate for amphibian acute hazard.

The underlying data themselves tell a story about the history of ecotoxicology. The matched values compiled for the model draw on decades of published acute toxicity testing, spanning legacy compounds and modern pesticides alike. Amphibian toxicity testing has its own long lineage, from early studies of pesticides on tadpoles of various species to the development of standardized protocols such as the Frog Embryo Teratogenesis Assay-Xenopus, known as FETAX, and the suite of OECD test guidelines covering the amphibian metamorphosis assay, the larval amphibian growth and development assay, and the Xenopus eleutheroembryonic thyroid assay. Each of these endpoints probes a different aspect of chemical hazard, but none generates the simple 96-hour LC50 as routinely as the fish acute toxicity test does, which is precisely the gap the ICE model is designed to bridge.

For the model to be useful in practice, its limitations need to be respected. An ICE model is only as reliable as the chemicals it was trained on; predictions for compounds with modes of action or physicochemical properties far outside the 69-chemical training set carry greater uncertainty, and interspecies correlations generally perform better for chemicals within the range of toxicity values used to fit the regression. The authors describe the model as a non-testing method to generate an amphibian LC50 value when a rainbow trout value is available, for use in hazard and risk assessment. Read that way, the equation does not replace amphibian testing where it is genuinely required, but it offers a defensible, quantitative alternative when the question is whether a frog study is needed at all, or what concentrations a planned assay should cover.

The implications extend beyond the laboratory. Chemical registration programs worldwide routinely require acute fish toxicity data, meaning trout LC50 values already exist for thousands of substances in regulatory dossiers. A validated trout-to-frog model effectively unlocks that existing data trove for amphibian hazard characterization, at zero additional cost to laboratory animals. As environmental agencies continue to confront the question of how well fish-based assessments protect other aquatic vertebrates, models like this one provide the quantitative evidence needed to make those judgments transparent. For a species whose embryos and tadpoles have taught generations of biologists about development, the African clawed frog may now also help redefine how chemical safety is assessed, one regression equation at a time, with fewer animals in the tanks.

Subject of Research: An interspecies correlation estimation model predicting acute chemical toxicity to the African clawed frog using rainbow trout toxicity data

Article Title: An interspecies correlation estimation (ICE) model to predict acute toxicity of organic chemicals to the African clawed frog (Xenopus laevis) using rainbow trout (Oncorhynchus mykiss) data

Article References: Weltje, L., & Savaliya, D. (2026). An interspecies correlation estimation (ICE) model to predict acute toxicity of organic chemicals to the African clawed frog (Xenopus laevis) using rainbow trout (Oncorhynchus mykiss) data. Ecotoxicology, 35(8), Article 170. https://doi.org/10.1007/s10646-026-03165-x

Image Credits: AI Generated

DOI: 10.1007/s10646-026-03165-x

Keywords: Xenopus laevis, rainbow trout, ICE model, Web-ICE, acute toxicity, LC50, ecotoxicology, amphibians, animal testing reduction, organic chemicals, risk assessment, regulatory toxicology