Single Cell Foundation Model Geneformer

Choose A Path

Tokenization

Use TranscriptomeTokenizer first for almost every Geneformer workflow. This step converts raw-count .loom or .h5ad data into tokenized datasets used by the downstream APIs.

Geneformer expects:

• row attribute ensembl_id
• cell attribute n_counts

Optional metadata can be passed through during tokenization.

Classification

Use Classifier for:

• cell state classification
• cell type annotation
• gene classification tasks

The input is a tokenized Geneformer .dataset object, not raw AnnData.

Embedding extraction

Use EmbExtractor when the task is to:

• extract CLS, cell, or gene embeddings
• plot or inspect cell embeddings
• generate state embeddings for later perturbation analysis

In silico perturbation

Use InSilicoPerturber for zero-shot or model-based perturbation analyses such as:

• deleting or shifting genes
• modeling start and goal cell states
• ranking perturbations by movement toward a desired cell state

This is one of Geneformer's defining workflows and should be treated as more than ordinary classifier inference.

Guardrails

• Do not pass feature-selected matrices into the tokenizer; the docs expect raw counts without feature selection.
• Do not use gene symbols where the tokenizer expects ensembl_id.
• Do not confuse tokenized .dataset files with AnnData objects.
• Do not skip tokenization and jump directly to classifiers or perturbation APIs.
• For perturbation tasks, be explicit about model type, embedding mode, and target cell states.

Official Workflow Surface

Read More Only If Needed

• For operational usage and required input fields, read references/workflows.md.
• For official source locations and model-specific notes, read references/sources-and-notes.md.