Genomes, genes and what we found

We joined public tardigrade genomes and expression data on one reference, found the condensin subunits a published survey had reported missing, and measured how much tardigrade a genome AI was trained on. Findings here are preliminary until written up; each says what it rests on.

Everything known about what switches what — every gene-regulatory relationship recorded for tardigrades. Eight arrows, two disconnected pieces, and only two of them rest on anyone perturbing anything.

Two developmental datasets on one genome

We put two measurements of gene activity through H. exemplaris development on one genome: Levin et al. 2016, hour by hour in embryos, and Yoshida et al. 2017, across embryo and adult stages. They used different transcript sets, so their results couldn't be compared gene by gene. We aligned Levin's 12,734 assembled transcripts to the nHd_3.1 genome: 12,724 align (median 99.5% identity) and land on 9,069 genes. We joined them with Yoshida's gene-level counts, and lifted everything onto the chromosome-length Hi-C assembly.

The two datasets agree: across 8,541 shared genes, average expression correlates with Spearman ρ = 0.52, against 0.008 when gene labels are shuffled. That gives an hourly-plus-daily developmental expression atlas on one set of gene models.

Joining them also settled Levin's time axis, which the paper states two ways: the series runs about 98 hours, not 49.5 (see Development).

Tardigrades aren't missing condensin

We found every condensin subunit in the H. exemplaris proteome, including two divergent copies of CAP-H2, the one BLAST could not find. Condensin complexes fold chromosomes. In a survey of genome folding across the tree of life, Hoencamp et al. 2021 had found only the two core subunits in the tardigrade, SMC2 and SMC4, searching older genome assemblies, and wrote that this "will need to be investigated further".

In the 2017 nHd_3.1 proteome, with reciprocal best hits against human proteins:

That left condensin II without its kleisin, CAP-H2, which BLAST couldn't find. Kleisins change fast, so we built a profile from CAP-H2 in 33 other animals and searched with it. Each tardigrade has two CAP-H2 genes (in H. exemplaris, BV898_16921 and BV898_15594). Each carries both of CAP-H2's signature domains, matches the CAP-H2 profile and not the CAP-H one, and nothing turns up when the proteome is shuffled. The same profile finds human CAP-H2 with overwhelming confidence, and a CAP-H profile finds exactly the known CAP-H in each tardigrade.

Through development, the two CAP-H2 genes rise and fall with the other condensin II subunits: high in eggs in the first two days and in adults, low in between. So condensin II looks complete. Hoencamp et al.'s type I folding in tardigrades may still hold, but not because condensin II is missing. Credit where it's due: van Hooff et al. 2025 found the two Ramazzottius CAP-H2 genes first, in their survey of these complexes across eukaryotes; their sequences are identical to ours. Their survey didn't include H. exemplaris or Paramacrobiotus, and database annotation had already labelled three of the proteins "CAP-H2 C-terminal domain-containing". What we added: the genes in H. exemplaris, the species behind the 2021 question; the answer to that question; and the tests below.

A tree of 68 kleisins places all six tardigrade genes inside CAP-H2, as one tardigrade branch. The two copies come from one duplication: the same copies keep the same neighbouring genes across species, which dates it to before Hypsibius and Paramacrobiotus split, deep in their order. In single embryos, every condensin II subunit is switched on for the first ~46 hours and off after ~47 hours, while condensin I's kleisin stays on. We also redid Hoencamp et al.'s statistical test with the tardigrade moved to "complete". Their link between condensin II and genome folding still holds (p = 0.019, from 0.007), and the tardigrade becomes one of seven species with a complete condensin II and Rabl-like folding. Preprint: Meow-Meow 2026, doi:10.5281/zenodo.22832510 (all versions; version 2 adds the structure models); code and data on GitHub.

What the clasp looks like

A clasp only folds when it grips the ring, so we predicted each tardigrade CAP-H2 end together with its own species' ring protein, using Boltz-2 on a laptop GPU. The C-terminal end of the clasp is known to grip the head of SMC4. Five of the six tardigrade copies are predicted to do the same with confidence (ipTM 0.76–0.81; human 0.83). The controls behave: offered the other ring protein, SMC2, or scrambled into a random sequence of the same letters, the clasps fall to 0.19–0.59. The clasp's other end holds the neck of SMC2. With a short piece of the neck it gave no confident prediction; given the full head and neck, the b copies of H. exemplaris and R. varieornatus grip it in every sample (ipTM 0.62–0.69, human 0.57–0.61), on the same stretch of the neck as the human clasp, while scrambled clasps score 0.11–0.19. So in these two species copy b is predicted to grip both ring proteins, in separately modelled contacts. H. exemplaris a and Paramacrobiotus b still give no signal at this end.

We also connect this to an imaging result neither paper links it to: H. exemplaris chromosomes stay individualised through interphase in early embryos, each in its own compartment (Papell et al. 2026). Condensin II is what drives that kind of territory in other animals. Hoencamp's Hi-C came from whole animals, so one idea worth testing is that the architecture changes with stage. This is a hypothesis, not a result.

How much tardigrade does Evo 2 know?

We checked which tardigrade sequence actually went into Evo 2, a large DNA language model trained on OpenGenome2, a public collection of genomes across life. Its data card lists five tardigrade genomes.

• The reference genome is absent. The H. exemplaris reference, nHd_3.1, isn't in the training data. The only Hypsibius genome is an older assembly from a commercial culture (Sciento), known for bacterial contamination.
• Fragmented genomes lose most of their sequence. Every contig under 10 kb was dropped, leaving 35% of P. richtersi and 14% of I. dastychi.
• Gene-focused data see one tardigrade. The mRNA set holds 19,905 tardigrade sequences, all from P. metropolitanus, the only one with a RefSeq annotation. The genic-window set has 2,481 windows from it and 37 from Ramazzottius, whose annotation lists genes but no mRNAs, and none from the others.

So what Evo 2 "knows" about tardigrades is mostly one species, and what counted was annotation format, not biology. Small, fragmented and unannotated genomes of other groups are likely hollowed out the same way.

DNA methylation, from data already public

We're reading DNA methylation straight from public long-read (PacBio HiFi) files, whose methylation calls nobody has looked at, to test a 2022 prediction from gene content that tardigrades lost it. Methylated and unmethylated animals from the same pipeline serve as controls. Work in progress: early results agree with the prediction, and the main catch is that many tardigrade libraries were amplified, which erases methylation.