Every cell watched: lineage analysis as governance

What the work claims

This is a methods grant with a hypothesis attached, led by Anna-Katerina Hadjantonakis at the Sloan Kettering Institute for Cancer Research (FY2026 award $439,849; project period 2024-03-01 to 2028-01-31). Its premise is that fluorescence 3D time-lapse imaging already produces data good enough to image, track, and measure every cell in a tissue or whole organism over hours to days, but that the analysis tools lag the microscopes. The project funds deep learning to close that gap, with three aims.¹

The first two aims are tool-building: a cell-tracking method that can deliver "substantial cell lineages" of hundreds to thousands of cells imaged over hours to days across "a wide range of model organisms and organoid cultures," and an automated landmark-based image-registration method that uses statistical templates and neural networks to integrate data across imaging modalities. The third aim is biology: the team states that it already discovered a novel Planar Cell Polarity scheme in C. elegans through cell tracking and deep learning of cell movement patterns, and proposes to dissect how that conserved polarity pathway coordinates with other pathways to organize diverse cell movements during development.¹

So the work sits in two epistemic categories at once, and a careful reading should keep them apart: a claimed result (a polarity discovery made from movement data) and two proposed capabilities (generalizable tracking and registration). Only the first is presented as achieved; the record contains no sample sizes, no validation benchmarks, and no preprints for the tools.

How it works

Cell tracking in long time-lapse movies is deceptively hard. A developing tissue divides constantly; cells move, deform, crowd, and briefly vanish behind neighbors. Following identity across thousands of divisions is the "essential first step" the abstract names, because lineage trees and measurements of dynamic cell behaviors both depend on getting identity right. Deep learning is proposed here as the route to accuracy at the scale of whole tissues, where hand-annotation cannot keep up.¹

The second aim attacks a subtler problem: heterochrony. Developmental processes run on different clocks in different individuals, so two specimens of "the same" stage can show combinatorially different configurations of anatomical landmarks, with complex systematic co-variance between them. The proposed solution is to learn statistical templates, effectively average reference shapes and their allowed variation, and use neural networks to register each new dataset against them. This is what makes cross-modality integration possible: the abstract explicitly frames registration as the bridge between light-microscopy movies and modalities such as electron microscopy, expansion microscopy, and spatial transcriptomics, which provide molecular information that movies lack.¹

The third aim shows what the tools are for. Planar Cell Polarity is the conserved pathway that orients cells within a sheet, the mechanism behind aligned hairs, oriented cilia, and coordinated convergent extension during embryogenesis. The team's claim is that movement patterns extracted by tracking, fed through deep learning, revealed a compound polarity scheme in the worm that was not visible to conventional assays, and that dissecting its regulators will explain how one core pathway drives diverse motile behaviors in different developmental contexts.¹

Where a skeptic should push

The single most load-bearing assumption is that tracking errors stay small enough over long movies that the downstream biology survives them. Lineage reconstruction is cumulative: a single identity swap early in a movie corrupts every descendant branch that follows from it, and measurements of "dynamic cell behaviors" inherit those errors silently. The abstract promises tracking across a "wide range of model organisms and organoid cultures," which is precisely the property that would make this a standard tool, and precisely what is hardest to deliver, because each tissue type brings its own imaging artifacts. No accuracy figures exist in the public record to test the promise against.

Second, the heterochrony templates are not neutral bookkeeping. A statistical template of development is built from some population of specimens, and it encodes that population's average trajectory and variance as the definition of normal timing. Build the template from one species' embryos and apply it to another context, and "developmental delay" becomes a property of the mismatch, not of the specimen. The abstract does not say what populations the templates will be trained on, and that omission matters more than it looks.

Third, the demonstrated element, the C. elegans polarity discovery, is claimed in one sentence in a grant abstract. C. elegans is a transparent, invariant-lineage organism about as friendly to cell tracking as biology gets; the leap from there to dense mammalian tissues or three-dimensional organoid cultures is the entire difficulty, and it is unproven. Treat the worm result as a promising existence proof, not as evidence the general tool works.

The developmental clock is a governance artifact

For platforms that compute on living neural tissue, this project's most consequential output is not the tracking; it is the template. Once statistical templates define what "the same developmental stage" means, every downstream judgment that depends on stage inherits that definition: maturation benchmarks for organoids, quality-control gates for commercial cultures, cross-lab comparability claims, and eventually any welfare threshold that keys on developmental maturity rather than chronological age. There is no reference trajectory for human neural organoid development, so the first credible template does not describe a standard; it creates one, with the toolmaker's training data as its silent constitution. Whoever curates the template owns the developmental clock, and the developmental clock will sit underneath questions the template's authors were never asked.

The opportunity is easy to understate. Total lineage records, every cell followed over days, are exactly the telemetry substrate a serious welfare-monitoring regime for living neural tissue would need. Today the moral-status debate is partly an instrumentation problem: we argue about capacities we cannot observe continuously. A tool that renders a developing neural culture fully observable converts speculation about trajectory, integration, and response into data. The same completeness also raises the inverted question, one the field has not had to answer while observation was sparse: if you can record everything a living developing system does, what may you record, how long may you keep the record, and who owns it? A complete observation record of a candidate sentient substrate is not obviously ethical by default; surveillance-completeness is a choice that arrives dressed as a capability.

The vendor angle follows directly. An analysis layer this general does not stay in the originating lab: it gets bundled into imaging platforms and organoid-platform software stacks as the default "developmental analysis" button, and its errors propagate as quietly as its templates do. If the field's maturation benchmarks are computed with a tracker that drifts on dense neural tissue, benchmark failures will be read as biology when they are artifacts. Procurement teams evaluating organoid platforms should be asking not only what the software does but whose developmental trajectory it encodes, on what data, with what disclosed error rates. That question has never appeared on a spec sheet, and this grant is a reason it should.

The bottom line

Established: cell tracking plus deep learning can surface developmental phenomena invisible to conventional assays, with the team's claimed C. elegans polarity discovery as the evidence, pending peer-reviewed publication. Proposed, not demonstrated: generalizable tracking across many organisms and organoid cultures, and statistical-template registration as a cross-modality standard. What would confirm the claim: published validation of the tracker on dense mammalian and organoid datasets with lineage accuracy reported against hand-curated ground truth, and templates released with their training populations documented. What would break it: accuracy that collapses outside transparent invariant-lineage organisms, or templates that smuggle one population's timing in as the universal norm. Watch the templates more than the tracker: the tool follows cells, but the template decides what development is supposed to look like, and that is the part someone will govern whether or not anyone chooses to.