scGPT

Use This Skill When

Use this skill for the local scGPT repository at /DATA/disk0/zhaosy/home/scGPT. It is the right choice when the task involves:

• preparing AnnData inputs with scGPT's own preprocessing pipeline
• matching genes to a pretrained scGPT vocabulary
• tokenizing binned expression inputs for transformer models
• extracting cell embeddings with pretrained checkpoints
• fine-tuning scGPT for integration or annotation-style downstream tasks
• understanding how scGPT expects binned values, special tokens, and batch labels
• working through scGPT tutorials such as integration, annotation, GRN, perturbation, or reference mapping

Do not use this skill for generic Scanpy work that does not depend on scGPT checkpoints or tokenization.

Start Here

1. Confirm the checkpoint directory contains args.json, vocab.json, and best_model.pt.
2. Decide whether the task is fine-tuning, embedding extraction, or tutorial-guided experimentation.
3. Run preprocessing before tokenization or embedding unless the input has already been prepared for scGPT.
4. Check vocabulary overlap before spending time on training or inference.

Choose A Path

Preprocess and bin

The core preprocessing path in this repo is scgpt.preprocess.Preprocessor. Typical steps include:

• filter genes by counts
• optionally filter cells
• normalize total counts
• optionally log1p transform
• subset highly variable genes
• bin values into discrete bins and store them in adata.layers["X_binned"]

Fine-tune for integration

The clearest end-to-end example in the local repo is examples/finetune_integration.py. It demonstrates:

• loading a dataset
• building str_batch and batch_id
• preprocessing and HVG selection
• matching checkpoint vocabulary
• tokenizing and padding batches
• training / evaluation for an integration workflow

If the user asks "how should I use scGPT on my AnnData?", this example is often the best starting point.

Extract cell embeddings

Use scgpt.tasks.cell_emb.embed_data(...) or related functions when the goal is to generate cell embeddings from a pretrained model directory.

This path:

• loads args.json, vocab.json, and best_model.pt
• filters genes to those found in the vocabulary
• builds the transformer model
• encodes cells and writes embeddings to adata.obsm["X_scGPT"]

Reference mapping or tutorial workflows

The local repo ships tutorials for:

• annotation
• integration
• multiomics
• GRN
• perturbation
• reference mapping

Use them when the user wants the project-supported path instead of building a custom pipeline from scratch.

Guardrails

• Do not assume arbitrary gene identifiers will work. scGPT inference depends on the checkpoint vocabulary.
• Do not skip binning if the target workflow expects X_binned.
• Do not mix raw, normalized, and logged layers casually; keep track of which layer is used at each stage.
• Do not assume all checkpoints use the same configuration; always read args.json.
• When extracting embeddings, verify the gene_col used to map genes into the vocabulary.

Read More Only If Needed

• Read references/local-usage.md for checkpoint expectations, preprocessing shape, and local entry points.
• Read references/workflow-notes.md for best starting paths and common mistakes.