Liatir

Appearance

UCE 4-layer ​

UCE, or Universal Cell Embeddings, creates a numeric representation of each cell in a single-cell dataset. Those vectors can be used for comparison, clustering, visualization, and downstream analysis.

When to use it ​

Use UCE when you have an AnnData .h5ad file and want cell embeddings rather than cell-type labels.

For a first-pass annotation, start with CellTypist Local Annotation. For foundation-model embeddings, use UCE.

What Liatir installs ​

Liatir installs UCE as its own isolated AI Model runtime box:

  • a Python 3.10/3.11 virtual environment;
  • the official UCE source checkout pinned to a verified commit;
  • the UCE Python requirements;
  • the 4-layer model weights;
  • token and species mapping files;
  • protein embedding assets used by the model.

This runtime is separate from CellTypist and from every other AI Model.

Inputs ​

The direct runner and pipeline AI Tool expect:

  • an AnnData .h5ad file;
  • a species selection;
  • optional batch size and CSV preview size settings.

UCE expects the AnnData .X matrix to contain scRNA-seq counts. It also expects var_names to contain gene symbols, not Ensembl IDs.

Outputs ​

Liatir creates:

  • an embedded .h5ad file with embeddings stored in obsm["X_uce"];
  • a lightweight CSV preview of the first embedded cells;
  • a JSON summary with cell count, gene count, embedding size, settings, and warnings;
  • intermediate files produced by the UCE workflow, marked as intermediate artifacts in Results.

The .h5ad output is the main scientific artifact. The CSV is intentionally a preview so large datasets do not create huge duplicate embedding tables.

Hardware ​

The UCE model card lists GPU as the system requirement. Liatir allows CPU runs for small validation datasets, but they can be slow. For repeated or large single-cell runs, use a GPU-capable workstation.

Technical notes ​

Liatir currently exposes the official 4-layer UCE workflow. It pins the source checkout and installs the official package versions used by that workflow. The larger 33-layer workflow is not exposed yet.

Supported species in the first Liatir runner are:

  • human;
  • mouse;
  • frog;
  • zebrafish;
  • mouse lemur;
  • pig;
  • crab-eating macaque;
  • rhesus macaque.

Official source ​