Versions

Scale versions

Every refit that changes the item bank or the pool gets a version. Abilities are quoted with the version they were placed on; the current representation scale is v3, the perturbation scales are the 2026-10-20 fits.

2026-10-26

Renamed Virtual Cell Bench; foundation models on Tabula Sapiens

The site and report are now called Virtual Cell Bench. The seven foundation models are placed on the Tabula Sapiens panel, again with difficulties held fixed: Geneformer V2 +0.99, TranscriptFormer +0.94, UCE +0.81, scGPT +0.76, STATE +0.69, CellPLM +0.48, scPRINT large +0.17, against +0.98 for PCA-128D. CellPLM is also placed on the Parse PBMC panel (+2.27). The pooled scale still uses seven panels for the foundation models; the eight-panel placement follows when TranscriptFormer on Parse finishes.

2026-10-21

Foundation models placed on the curated panels

Seven pretrained single-cell models (scGPT, Geneformer V2, UCE, STATE, scPRINT large, CellPLM, TranscriptFormer) embed every cell of the eight curated panels and are read by the same frozen probe. They are placed with item difficulties held fixed, so no published ability moved; the boards mark them anchored. They are also placed on the pooled 14-panel scale from the seven panels finished so far (Tabula Sapiens to follow): scGPT +1.07, Geneformer +1.06, TranscriptFormer +1.02, UCE +0.88, STATE +0.76, CellPLM +0.37, scPRINT large −0.20, against +0.72 for the best simple model. The boards now show a featured set by default, one or two settings per simple family plus every learned model; "show all" reveals the full pool that defines the scale.

2026-10-20

Perturbation: unseen-combination scale with learned respondents

The context split becomes the primary perturbation setting. A second pooled scale, pooled_context, covers the five multi-context screens (Tahoe-100M, Parse cytokines, sciPlex3, McFarland, Frangieh), 71,897 conditions. New respondents: scVI counterfactual, scGen, and two MLPs from item features to a 50-dimensional PCA of the responses. MLP-256 leads at +4.44; scVI counterfactual sits at −0.08 and scGen at −0.42 after the Tahoe rerun.

2026-10-19

Representation v3: fourteen panels

The eight curated panels plus the six scaled-up scBaseCount human panels (human200, every filter-passing sample up to 200 per tissue). 424,478 items, 370,341 informative, 52,890 passed by nobody. Top of the scale: Nyström-256D +0.72, scVI-30D +0.72 and +0.71, PCA-128D +0.69. The 60-sample scBaseCount runs stay in the archive but leave the pooled scale.

2026-10-19

scBaseCount human200 panels and their pooled scale

More samples do not make single-tissue panels easier, because extra samples bring extra label vocabularies; blood is the exception. scVI-30D moves to the top of the scBaseCount-only pooled scale and wins outright on blood and bone marrow.

2026-10-16

scBaseCount panels, 60 samples per tissue

Eleven panels across human and mouse, 197,166 items on their own pooled scale. Six to 29% of items are passed by nobody, mouse worse than human: label mismatch across submissions, not rare biology. The mouse twelve-tissue panel is degenerate and excluded. scVI added to the six human panels; the human-only pooled scale puts it within 0.02 logits of PCA-64D.

2026-10-15 / 16

Representation v1 and the Parse cytokine screen

Six CELLxGENE panels and two Parse panels join Tabula Sapiens v2 on an eight-panel pooled scale of 146,428 items, with item identifiers namespaced by panel (v0 had double-counted 7,458 cells). Component B gains the Parse cytokine screen, 13,967 conditions over 169 contexts; the pooled unseen-perturbation scale grows to 15 screens.

2026-10-14

v0: first fits

Component A: Tabula Sapiens v2, 20,026 cells, 27 classical embeddings plus three scVI settings. Component B: thirteen scPerturb screens, 20,225 conditions. Ladders are monotone, floors last, and the mean shift is the best or second-best respondent on raw Pearson in 9 of 13 screens.

Naming

Pooled representation scales are repr_pooled_vN; each panel's own fit is repr_<panel>. Perturbation fits live under one run, pert_pool_v0, with pooled scales pooled (unseen perturbation) and pooled_context (unseen combination) and one fit per screen and split. Run outputs, item tables and Rasch reports are in the repository under runs/.