scGPT Whole-human

scGPT is a single-cell foundation model built for single-cell and multi-omics data. In Liatir it is tracked as a preview AI Model for future embedding, integration, perturbation, and gene-network workflows.

What it does

scGPT learns representations of cells and genes from large single-cell datasets. Those representations can support tasks such as cell embedding, reference mapping, batch correction, and perturbation hypotheses.

Current status in Liatir

This model is visible as a preview. Liatir documents the model and keeps its metadata in the AI Model registry, but install and run controls are not enabled yet.

The next implementation step is a dedicated managed runtime box that can install the Python package, download a selected checkpoint, validate AnnData inputs, and write embeddings/provenance through the shared Liatir I/O contract.

Expected inputs

• AnnData .h5ad file.
• Expression matrix and gene metadata.
• Optional batch or cell metadata for integration workflows.

Expected outputs

• Cell embeddings.
• UMAP-ready tables or matrices.
• Optional batch-corrected representations.
• JSON/CSV summaries and provenance.

Hardware and installation

Small examples may load on CPU, but practical foundation-model workflows should use a GPU. scGPT is a heavy scientific runtime and will stay isolated from the core app.

Official source

• scGPT on GitHub