Experimental Evolution / MulticellularitySnowflake Evolution Lab

A yeast cluster became 20,000 times larger. When did it become an individual?

Current research status: stable product v1.0.1 preserves the frozen v0.1 morphology study and frozen v1.0 genome-duplication result. The patch clarifies the explanatory animation without changing either analysis. Both studies reanalyse source-author data; neither is an independent wet-lab replication or a claim to have rediscovered multicellularity.

The question behind one of evolution's largest transitions

Every animal, plant, and mushroom is a society of cells whose evolutionary interests became organized into a larger individual. That transition happened independently many times, but its earliest stages are mostly inaccessible: the relevant ancestors are gone, and fossils rarely preserve the mechanics of fragile cell clusters.

Experimental evolution offers a narrower view. Snowflake yeast are Saccharomyces cerevisiae cells whose daughters remain attached after division, producing branching clonal clusters. In the Multicellularity Long-Term Evolution Experiment (MuLTEE), researchers repeatedly selected clusters that settled through liquid more quickly. Under anaerobic conditions, all five replicate populations eventually produced millimetre-scale bodies containing hundreds of thousands of cells.

Bozdag and colleagues reported that, after 600 daily selection rounds—about 3,000 cellular generations—the anaerobic clusters were approximately 20,000 times larger and 10,000 times tougher than their ancestor. The cells became increasingly elongated. That geometry first reduced packing strain and later allowed branches to entangle, so a single broken cell–cell bond no longer split the entire organism.

Our bounded question is smaller: does the published association between longer cells and larger clusters have the same direction in every independently evolved anaerobic lineage?

What was fixed before calculation

The protocol was published before the confirmatory calculation. It fixed:

  • the public source repository and exact upstream commit;
  • the SHA-256 identity of the source workbook;
  • the `Fig1e` cluster-radius and `Fig2d` cell-aspect-ratio sheets;
  • PA1–PA5 as the five biological inference units;
  • a within-line Spearman correlation between aspect ratio and log cluster radius;
  • success only if all five population-level correlations were positive;
  • three sensitivity analyses and a separate exploratory timing comparison.

There are 65 population-by-time rows. PA2 has no published cluster-radius mean at transfer 400, so that value remains missing rather than being interpolated. The final calculation contains 64 paired means. Thousands of underlying cells and clusters are not counted as thousands of independent replicates.

What we found

The frozen criterion passed. Every lineage showed a positive within-line association:

  • PA1: Spearman ρ = 0.956;
  • PA2: 0.958;
  • PA3: 0.923;
  • PA4: 0.956;
  • PA5: 0.967.

The median was 0.956. Under the protocol's deliberately limited model in which five independent population-level signs are equally likely to point either way, five positive signs correspond to `1/2⁵ = 0.03125`. That number describes directional convergence across these five lines. It is not a universal p-value for the evolution of multicellularity.

The prespecified checks agreed. After removing the shared day-zero ancestor, all five Spearman coefficients remained above 0.90. Pearson correlations with log radius ranged from 0.893 to 0.981, and Kendall's tau-b ranged from 0.821 to 0.872.

The result that makes the story less simple

We also compared the 50-transfer interval with the largest rise in cell aspect ratio with the interval containing the largest increase in log cluster radius. Their order was not consistent. The size jump came earlier in PA1, PA2, and PA5; the shape jump came earlier in PA3 and PA4.

That coarse comparison cannot identify a universal transition point. Fifty-transfer sampling, accumulated mutations, correlated time trends, and changes in the selection protocol all complicate temporal interpretation. The strong association is compatible with the source paper's mechanical experiments, but the summary trajectories alone cannot prove that elongation caused the size rise.

The v1.0 test: genome duplication helps, but is not enough

Tong and colleagues supplied two complementary forms of evidence. First, they engineered diploid and tetraploid versions of PA and PM snowflake-yeast backgrounds. The v1.0 analysis aggregates the 39,329 segmented cells and 17,227 24-hour clusters within each of four biological replicate strains per group before inference. Tetraploidy increased mean 24-hour cluster radius from 20.157 to 27.299 µm in PA and from 25.890 to 42.897 µm in PM. Each exact one-sided replicate-level permutation test has probability `1/70 = 0.0143`; both remain below 0.05 after Holm correction (`0.0286`). Cell volume and aspect ratio increased in the same direction as secondary outcomes.

Second, the public longitudinal tables follow PA1–PA5 and PM1–PM5 through transfer 1,000. Every evolved population-time G1 peak is at least 3.489N, yet the treatments diverge sharply in size. At transfer 1,000, mean PA radius is 400.074 µm and mean PM radius is 45.160 µm—a ratio of 8.859. All five PA lines exceed 300 µm, while the largest PM line is 53.226 µm. Genome duplication therefore provides an immediate size-related advantage but is not sufficient for the later macroscopic phenotype in these treatments.

The intervention and the longitudinal comparison answer different questions. The first supports an immediate effect in engineered strains. The second rejects a simple sufficiency explanation; it does not identify which later change completed the transition. The protocol was frozen after a source-compatibility audit and before the v1 analysis was implemented, so it is a prospective implementation contract rather than a blinded preregistration.

Two further pieces of the story

Later MuLTEE work makes the transition more interesting rather than more linear.

First, a small-cluster lineage and a large-cluster lineage evolved from one ancestor and remained together for roughly 4,300 generations. Their coexistence was maintained by a trade-off between rapid growth and survival during settling, mediated by dissolved oxygen. A new level of biological individuality created room for ecological diversification.

Second, macroscopic clusters appear able to move nutrients without a circulatory system. Their own metabolism creates density differences that drive fluid through the cluster. Above a size threshold, those flows can relieve limits expected from diffusion alone. Larger size can therefore create a physical process that makes still larger size possible.

These findings motivate a future coupled-threshold hypothesis: genomic size amplification and an entangling cellular geometry may need to coincide before robust macroscopic bodies emerge. This is a testable proposal, not a measured v1.0 threshold. The selected public tables do not contain a joined, quantitative time-resolved entanglement endpoint. A future protocol must acquire such an endpoint before inspecting the combined outcome.

Try it yourself

The interactive Snowflake Evolution Lab lets you replay each population from transfer 0 to 600, switch between the published cluster-size and cell-shape trajectories, and compare all metabolic treatments at day 600. The new genome explorer adds the engineered 2N/4N contrast, all 16 chromosome copy numbers, G1 peak and morphology through transfer 1,000, and the PA/PM sufficiency comparison. The animated cluster is deliberately labelled as an explanatory geometry model. Solid links are permanent chitinous parent–daughter junctions; dashed amber links are steric contacts rather than bonds; red gaps are modelled fractures. The packing and fracture views isolate those mechanisms, while the colony view combines them. These cues help you reason about the proposed mechanism, but they are not microscopy, fitted biological simulation, measured entanglement, or an unobserved ancestral reconstruction.

Stop at the boundary

prevalence claim.

  • These are five selected laboratory populations, not a random sample of evolutionary histories.
  • The reanalysis uses the original authors' published means and is not an independent replication.
  • Association through time does not isolate elongation from selection, mutations, or other changes.
  • Four engineered strains per group support a narrow intervention estimate, not a population-wide
  • The available joined tables do not measure a time-resolved entanglement endpoint.
  • Snowflake yeast are undifferentiated clonal clusters, not miniature animals or plants.
  • Nothing here shows that historical multicellular lineages followed the same route.

Inspect and reproduce