CellReasoner-32B

CellReasoner: A reasoning-enhanced large language model for cell type annotation 🧬🧠

<div align="center">

📄 Paper | 💻 GitHub

</div>

---

---

📌 Table of Contents

• 📖 CellReasoner: A reasoning-enhanced large language model for cell type annotation 🧬🧠
  • 📌 Table of Contents
  • 🔬 Key Highlights
  • 🔑 Key Results
  • 🧠 Model Zoo
  • 🏋️‍♂️ Training
  • 🚀 Usage
  • 📚 Citation

---

🔬 Key Highlights

• Only a few expert-level reasoning samples are needed to activate reasoning in a 7B LLM.
• CellReasoner achieves expert-level interpretability and zero-/few-shot generalization.
• Demonstrated superior performance across various scRNA-seq and scATAC-seq datasets.
• Compatible with marker-by-marker annotation, ontology mapping, and biological reasoning.

🧠 Less data, more reasoning: CellReasoner achieves accurate, interpretable, and scalable cell annotation with minimal supervision.

---

🔑 Key Results

PDAC dataset

---

PBMC3K dataset

---

🧠 Model Zoo

Our CellReasoner models are available on Hugging Face 🤗:

---

🏋️‍♂️ Training

We use the LLaMA-Factory framework for fine-tuning. It offers a flexible and efficient pipeline for supervised fine-tuning, LoRA, and multi-stage training strategies.

---

🚀 Usage

🛠️ Step 1: Prepare Conda Environment

Make sure you have a working conda environment with the necessary dependencies installed. We recommend:

conda create -n cellreasoner python=3.11
conda activate cellreasoner
pip install -r requirements.txt

---

🧪 Step 2: Preprocess Input Data

If your input is in Seurat .rds format, use the R preprocessing script:

Rscript s01.process_rds.R ./demo_data/pbmc_demo.rds ./output/ data/ranked_hvg.list

If your input is in AnnData .h5ad format, use the Python script:

python s01.process_h5ad.py \
    --input_file ./demo_data/pbmc_demo.h5ad \
    --output_path ./output_h5ad \
    --ranked_hvg_list ./data/ranked_hvg.list

Both pipelines will generate the following output files:

output/
├── pbmc_demo.h5
└── pbmc_demo.meta.csv

---

🧱 Step 3: Build Dataset for CellReasoner

Build the model input file using:

python s02.build_dataset.py \
    --h5_path ./output/pbmc_demo.h5 \
    --output_path ./output/ \
    --meta_file_path ./output/pbmc_demo.meta.csv

If your metadata includes cell type labels (for scoring), specify the column name:

python s02.build_dataset.py \
    --h5_path ./output/pbmc_demo.h5 \
    --output_path ./output/ \
    --meta_file_path ./output/pbmc_demo.meta.csv \
    --cell_type_column "seurat_annotations"

This will generate:

output/
└── pbmc_demo_for_CellReasoner.json

---

🤖 Step 4: Run Inference with CellReasoner

python s03.inference.py \
    --model "CellReasoner-7B" \
    --output_path "./output" \
    --input_json "./output/pbmc_demo_for_CellReasoner.json" \
    --batch_size 2

Result:

output/
└── pbmc_demo_CellReasoner_result.csv

---

📊 Evaluation and Reasoning Visualization

To compute scores, generate plots, or view reasoning outputs, refer to:

s03.inference.ipynb

Citation

@article {Cao2025.05.20.655112,
	author = {Cao, Guangshuo and Shen, Yi and Wu, Jianghong and Chao, Haoyu and Chen, Ming and Chen, Dijun},
	title = {CellReasoner: A reasoning-enhanced large language model for cell type annotation},
	elocation-id = {2025.05.20.655112},
	year = {2025},
	doi = {10.1101/2025.05.20.655112},
	URL = {https://www.biorxiv.org/content/early/2025/05/26/2025.05.20.655112},
	eprint = {https://www.biorxiv.org/content/early/2025/05/26/2025.05.20.655112.full.pdf},
	journal = {bioRxiv}
}

---