Single-cell Embedding

Single-cell Embedding is an AI Tool that creates numeric vectors for cells in an AnnData .h5ad dataset.

Supported backends are UCE 4-layer, Geneformer V1 10M, and scGPT Whole-human.

What it does

The tool reads a single-cell dataset and asks the selected foundation model to compute one embedding vector per cell. Similar cells should usually have more similar vectors, so the output can be useful for exploration, visualization, clustering, and downstream analysis.

This is different from annotation. Annotation gives cells labels. Embedding creates a representation that other tools can inspect.

Inputs

Input	Meaning
AI Model	Installed UCE 4-layer, Geneformer V1 10M, or scGPT Whole-human
AnnData file	.h5ad file from Data
Species	Organism matching the dataset
Batch size	Number of cells processed together
CSV preview rows	Number of cells exported to the lightweight CSV preview

The AnnData .X matrix should contain raw scRNA-seq counts. UCE expects gene symbols in var_names. Geneformer V1 is human-only and expects Ensembl IDs in var["ensembl_id"] or var_names. scGPT is human-only and expects gene symbols in var["gene_name"], var["feature_name"], or var_names.

Outputs

Output	Meaning
Embedded AnnData	Main .h5ad artifact with obsm["X_uce"] or obsm["X_geneformer"]
Embedding preview CSV	Bounded table with raw embedding dimensions and two preview PCA columns
Summary JSON	Counts, dimensions, settings, warnings, and provenance
Intermediate files	UCE processing artifacts when UCE is selected, marked as intermediate

Use the embedded AnnData file for downstream scientific work. The CSV preview is only meant for quick inspection. Its PCA is computed over the bounded exported rows, not over the whole dataset.

In pipelines

Use this tool when you want:

• h5ad -> single-cell embedding -> viewer/report;
• h5ad -> embedding -> downstream script/plugin;
• a reusable cell-embedding step before visualization.

In the pipeline editor, the AI Model dropdown only shows compatible installed models.

For the ready-made path, open Pipelines, choose Single-cell embedding and preview, select the AnnData file and an installed AI Model, and run. The preset connects both the embedded AnnData and its preview to the Single-cell Viewer. Pipeline outputs are registered in Data before the viewer step starts.

Reading the result

Check:

• cell count and gene count;
• embedding dimensions;
• warnings about input format;
• provenance;
• whether the species matches the dataset.

Embeddings are not a final biological conclusion. They are a representation to inspect, compare, and validate.