bowang-lab/scGPT; PyPI scgpt 0.2.4, unchanged since 2025-03) under uv run python. Its old pins (scvi-tools<1.0, scanpy<2.0, torchtext) conflict with a current scverse stack, so use a dedicated environment, not one shared with alterlab-scanpy/alterlab-scvi-tools. Checkpoints (GB-scale) download manually from the repo's Drive links; a CUDA GPU is strongly recommended. I/O is AnnData (.h5ad); dispatch heavy fine-tuning via alterlab-remote-compute.\"\nmetadata:\nskill-author: AlterLab\nversion: \"1.1.0\"\nlast_updated: \"2026-09-23\"scGPT (Cui et al., Nature Methods 2024; bowang-lab/scGPT) is a transformer foundation\nmodel pretrained on tens of millions of cells. It provides zero-shot and fine-tuned\ncell-type annotation, gene and cell embeddings, batch integration, and\ngene-regulatory / perturbation inference — all operating on AnnData (.h5ad) objects.
Its niche vs. the existing single-cell skills: scGPT is the pretrained-transformer route.\nFor probabilistic latent-variable models use alterlab-scvi-tools; for the conventional\nScanpy analysis pipeline use alterlab-scanpy; scGPT complements both.
Use this skill when the user wants to:
\n| Scenario | \nUse instead | \n
|---|---|
| Probabilistic integration / latent model (scVI, scANVI) | \nalterlab-scvi-tools | \n
| Standard QC → cluster → UMAP → differential expression | \nalterlab-scanpy | \n
Read/write/wrangle the .h5ad data structure itself | \nalterlab-anndata | \n
| RNA velocity | \nalterlab-scvelo | \n
| Protein (not single-cell) language models | \nalterlab-esm | \n
import scanpy as sc\nfrom scgpt.tasks import embed_data\n\nadata = sc.read_h5ad(\"cells.h5ad\")\nadata = embed_data(\n adata,\n model_dir=\"checkpoints/scGPT_human\", # downloaded checkpoint folder\n gene_col=\"feature_name\", # column in adata.var holding gene symbols\n batch_size=64,\n device=\"cuda\",\n)\n# cell embeddings land in adata.obsm[\"X_scGPT\"]\ngene_col must name an adata.var column of gene symbols matching the checkpoint's vocabulary\n(pass \"index\" to use var_names); symbols that miss the vocab are dropped, so check how many\ngenes survive before trusting the embedding. Set use_fast_transformer=False when flash-attn\nis not installed. Zero-shot mode maps a new dataset onto scGPT's learned space without training\n— fast triage of cell identities. Fine-tuning on labeled reference data improves accuracy on a\nspecific tissue.
Produce cell embeddings (for clustering/visualization) or gene embeddings (for\ngene-network/similarity analysis). Feed embeddings back into a Scanpy neighbors/UMAP workflow.
\nUse the model representation to integrate across batches/donors, comparable in role to\nscVI-based integration but from the pretrained-transformer paradigm.
\nscGPT's last release is 0.2.4 (March 2025) and its pins have not moved since: it requires\nscvi-tools<1.0 and scanpy<2.0, and imports torchtext, whose own last release was 0.18.0\nin April 2024. That stack will not co-install with a current scverse environment, so give scGPT\nits own venv and move data between environments as .h5ad files rather than trying to satisfy\nboth sets of pins at once. Say so plainly when a user expects it to drop into their existing\nenvironment.
Checkpoints (whole-human, continual-pretrained, organ-specific) are downloaded by hand from the\nDrive links in the upstream README — there is no from_pretrained downloader. The\nwhole-human model is the default choice; the continual-pretrained one is aimed at zero-shot\ncell-embedding tasks.
scGPT needs a GPU for realistic dataset sizes; fine-tuning is heavy. Dispatch fine-tuning /\nlarge inference via alterlab-remote-compute (submit → poll → harvest). Keep the AnnData I/O\nconsistent with alterlab-anndata.
Because the model has been static for over a year while the foundation-model field has not,\ntreat its published benchmarks as a 2024 snapshot: report the checkpoint name and date with any\nresult, and where annotation accuracy matters, sanity-check the labels against marker-gene\nevidence (alterlab-scanpy) rather than accepting them unverified.
references/scgpt_usage.md — install/pinning, checkpoints, embed/annotate/fine-tune calls,\nscverse integration, and paradigm comparison. Loaded on demand.Part of the AlterLab Academic Skills suite.
\n","files":[{"path":"evals/evals.json","sizeBytes":2971,"isText":true},{"path":"references/scgpt_usage.md","sizeBytes":2951,"isText":true},{"path":"SKILL.md","sizeBytes":5843,"isText":true}],"reviewScore":null,"reviewSummary":null,"trust":{"provenance":"trusted-source-unreviewed","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow.","bodySource":null},"bodyLocked":false,"purchaseUrl":null,"sourceUrl":null,"report":{"provenance":"trusted-source-unreviewed","screen":{"ran":true,"outcome":"clean","suspicious":0,"notes":0,"hiddenCharacters":false},"virusScan":{"engine":"clamav","status":"clean","scannedAt":"2026-09-23T18:57:56.506436Z","sha256":"2D4072429202F644B873821B181D5C6548CCB4EF1B10FFDADEADC62F0CF3260A","sizeBytes":5547},"review":null,"source":{"repositoryUrl":"https://github.com/AlterLab-IEU/AlterLab-Academic-Skills","path":"skills/bioinformatics/alterlab-scgpt","license":"MIT","commit":"e4836c08a20da195a11f30f203a8cf23ec30aa95","subtreeSha":"F01CC27B35BC1B7945BCE9659B3EC3A3B571436723459CC074361DC8C576CE7F","lastSyncedAt":"2026-09-23T18:56:52.297238Z"},"reviewedAt":"2026-09-23T18:59:50.030869Z","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow."},"install":[{"target":"skills-cli","command":"npx skills add https://github.com/AlterLab-IEU/AlterLab-Academic-Skills/tree/main/skills/bioinformatics/alterlab-scgpt"},{"target":"claude-code","command":"claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install alterlab-ieu-alterlab-academic-skills@llmmart"},{"target":"git","command":"git clone https://github.com/AlterLab-IEU/AlterLab-Academic-Skills.git"}]}