Alterlab Scgpt

scGPT (single-cell foundation model)

Overview

scGPT (Cui et al., Nature Methods 2024; bowang-lab/scGPT) is a transformer foundation model pretrained on tens of millions of cells. It provides zero-shot and fine-tuned cell-type annotation, gene and cell embeddings, batch integration, and gene-regulatory / perturbation inference — all operating on AnnData (.h5ad) objects.

Its niche vs. the existing single-cell skills: scGPT is the pretrained-transformer route. For probabilistic latent-variable models use alterlab-scvi-tools; for the conventional Scanpy analysis pipeline use alterlab-scanpy; scGPT complements both.

When to Use This Skill

Use this skill when the user wants to:

• Annotate cell types with a pretrained foundation model (zero-shot or fine-tuned).
• Generate scGPT embeddings for cells or genes.
• Integrate batches using the transformer's representation.
• Run zero-shot inference / transfer to a new dataset without training from scratch.

Does NOT Trigger

Core Capabilities

1. Zero-shot cell embedding & annotation

pythonCopy code
import scanpy as sc
from scgpt.tasks import embed_data

adata = sc.read_h5ad("cells.h5ad")
adata = embed_data(
    adata,
    model_dir="checkpoints/scGPT_human",   # downloaded checkpoint folder
    gene_col="feature_name",               # column in adata.var holding gene symbols
    batch_size=64,
    device="cuda",
)
# cell embeddings land in adata.obsm["X_scGPT"]

gene_col must name an adata.var column of gene symbols matching the checkpoint's vocabulary (pass "index" to use var_names); symbols that miss the vocab are dropped, so check how many genes survive before trusting the embedding. Set use_fast_transformer=False when flash-attn is not installed. Zero-shot mode maps a new dataset onto scGPT's learned space without training — fast triage of cell identities. Fine-tuning on labeled reference data improves accuracy on a specific tissue.

2. Embeddings for downstream analysis

Produce cell embeddings (for clustering/visualization) or gene embeddings (for gene-network/similarity analysis). Feed embeddings back into a Scanpy neighbors/UMAP workflow.

3. Batch integration

Use the model representation to integrate across batches/donors, comparable in role to scVI-based integration but from the pretrained-transformer paradigm.

4. Environment, GPU, and dispatch

scGPT's last release is 0.2.4 (March 2025) and its pins have not moved since: it requires scvi-tools<1.0 and scanpy<2.0, and imports torchtext, whose own last release was 0.18.0 in April 2024. That stack will not co-install with a current scverse environment, so give scGPT its own venv and move data between environments as .h5ad files rather than trying to satisfy both sets of pins at once. Say so plainly when a user expects it to drop into their existing environment.

Checkpoints (whole-human, continual-pretrained, organ-specific) are downloaded by hand from the Drive links in the upstream README — there is no from_pretrained downloader. The whole-human model is the default choice; the continual-pretrained one is aimed at zero-shot cell-embedding tasks.

scGPT needs a GPU for realistic dataset sizes; fine-tuning is heavy. Dispatch fine-tuning / large inference via alterlab-remote-compute (submit → poll → harvest). Keep the AnnData I/O consistent with alterlab-anndata.

Because the model has been static for over a year while the foundation-model field has not, treat its published benchmarks as a 2024 snapshot: report the checkpoint name and date with any result, and where annotation accuracy matters, sanity-check the labels against marker-gene evidence (alterlab-scanpy) rather than accepting them unverified.

Resources

• references/scgpt_usage.md — install/pinning, checkpoints, embed/annotate/fine-tune calls, scverse integration, and paradigm comparison. Loaded on demand.

Part of the AlterLab Academic Skills suite.