CellTypeAI: Automated cell identification for scRNA-seq using local generative-AI

CellTypeAI, a streamlined, scalable program developed to perform context-dependent cell identification in scRNA-seq datasets, which leverages locally-run LLMs with enhanced RAG methods and optional ensemble methods.

CellTypeAI builds upon local LLM hosting technol-ogies and integrates directly into scanpy-based scRNA-seq analysis pipelines to enable accurate cell type identification of pre-clustered scRNA-seq data

Table of Contents

• Installation
• Usage
• Features
• Contributing
• License

Installation

Dev version:

#Using pip:
pip install git+https://github.com/rhdaw/celltypeai

#Using uv: 
uv add git+https://github.com/rhdaw/celltypeai

# Navigate into the directory to install requirements from .toml
cd CellTypeAI
pip install .

# Or using uv:
uv pip install -r pyproject.toml

pip installation (to be confirmed)

pip install celltypeai
#or
uv add celltypeai

Usage

Ollama Requirements:

• Install Ollama (https://ollama.com/)
• Pull model of choice (https://ollama.com/search), i.e:

ollama pull phi4:14b

• Initialise ollama server via ollama serve in terminal / powershell

ollama serve

Running CellTypeAI

Following standard scRNA-seq processing with leiden OR louvain clustering.

import celltypeai as cta

cta.cell_annotator(
                "Human",
                "Intestine",
                adata,
                model="qwen3:32b",
                num_genes=200,
                n_iterations=3,
            )

On completion, annotations are stored in adata.obs["cell_type_ai"]

cell_annotator(adata, species, tissue, ...)

Takes a processed .h5ad object, extracts Differentially Expressed Genes (DEGs) from each cluster, and utilizes a local LLM via Ollama to define and annotate cell types.

Returns:

• adata (AnnData): The input object with a new .obs column containing the predicted cell types.

Contributing

Fork the repo, create a feature branch, and submit a pull request. Please add tests and update documentation for new features.

License

GNU GPL-2.0 license - see LICENSE file