scGPT Whole-human
scGPT is a single-cell foundation model built for single-cell and multi-omics data. Liatir currently exposes the Whole-human checkpoint for local cell embeddings.
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
scGPT is published as signed Runtime Boxes for macOS arm64 Metal, Linux/Windows x86_64 CPU, Linux x86_64 CUDA 12.9 and Windows x86_64 CUDA 12.8. Liatir installs the complete target-specific environment and records the exact box release in Results provenance.
Expected inputs
- AnnData
.h5adfile. - Expression matrix and gene metadata.
- Optional batch or cell metadata for integration workflows.
Expected outputs
- Embedded AnnData with cell embeddings in
obsm["X_scgpt"]. - A bounded CSV preview for visualization.
- JSON summary, warnings and Runtime Box provenance.
Hardware and installation
Use only the exact published target compatible with the host. Windows CUDA is CUDA 12.8 and Linux CUDA is CUDA 12.9; support for those targets does not imply other CUDA or driver combinations. The model and its dependencies live inside the Runtime Box and are never added to the base app or system Python.