Liatir

Appearance

Your first analysis ​

This path creates a local single-cell embedding, opens its preview, and keeps the produced AnnData available for another tool or pipeline.

Before you start ​

You need:

  • a compatible AnnData .h5ad file;
  • enough free disk space for one supported AI Model;
  • a macOS arm64 machine for the currently complete local Beta path.

Geneformer and scGPT expect human data. UCE supports additional species but has different gene-identifier requirements. Review the model page before interpreting an output.

Run the no-code workflow ​

  1. Open Data and add the .h5ad file. Liatir records its path without copying or uploading it.
  2. Open AI Models, choose a model compatible with your machine and select Install. This is an explicit network operation and may use substantial disk space.
  3. Open Pipelines and choose Single-cell embedding and preview.
  4. Select the AnnData input and installed AI Model, then run the pipeline.
  5. Follow the analysis in Jobs. You can navigate elsewhere while it runs.
  6. Open the final entry in Results and inspect the single-cell preview, warnings and provenance.
  7. Add the embedded AnnData to Data or connect it to a downstream pipeline step. Use the AnnData output for scientific work; the bounded CSV/PCA view is only a preview.

The direct Single-cell Embedding AI Tool uses the same input contract and finalizer as the pipeline step.

Try a Nextflow workflow ​

If Nextflow and Java are already installed on a supported execution host, save the workflow under Tools → External Workflows. The same saved definition can run by itself or appear in a Liatir pipeline. Declare outputs explicitly so they become Liatir artifacts that can feed later tools without searching the Nextflow work directory.

See External Workflows for platform requirements, reproducible revisions, cancellation and resume behavior.

What to record ​

Before sharing a result, keep the Liatir Result and check:

  • input identity and scientific compatibility;
  • model/runtime or Nextflow/Java versions;
  • parameters, warnings and exit status;
  • output digest and lineage;
  • whether the preview represents the whole dataset or a bounded sample.

An embedding is not a cell annotation or a clinical conclusion.