File size: 6,741 Bytes
fec1b45 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 | # CellVerse: Do Large Language Models Really Understand Cell Biology?
Official repository for the paper "CellVerse: Do Large Language Models Really Understand Cell Biology?"
🌟 For more details, please refer to the project page: [https://cellverse-cuhk.github.io](https://cellverse-cuhk.github.io/).
[[🌐 Webpage](https://cellverse-cuhk.github.io/)] [[📖 Paper](https://arxiv.org/pdf/2505.07865)] [[🤗 Huggingface Dataset](https://huggingface.co/datasets/Karl28/CellVerse)] [[🏆 Leaderboard](https://cellverse-cuhk.github.io/#leaderboard)]
## 💥 News
- **[2025.05.08]** 🚀 We release the ***test*** set of CellVerse at [[🤗 Huggingface Dataset]](https://huggingface.co/datasets/Karl28/CellVerse), and the [inference code](https://github.com/zfkarl/CellVerse/tree/master/evaluation)!
## 👀 About CellVerse
The capabilities of **Large Language Models (LLMs)** in **Cell Biology Understanding** remain insufficiently evaluated and understood. We investigate the differences between the traditional single-cell analysis paradigm and the language-centric paradigm, and find that the latter offers advantages in terms of **unification**, **user-friendliness**, and **interpretability**.
<p align="center">
<img src="figs/fig_illus.png" width="95%"> <br>
</p>
To this end, we introduce **CellVerse**, a language-centric single-cell analysis benchmark designed for an equitable and in-depth evaluation of LLMs. The whole dataset encompasses four types of single-cell multi-omics data (scRNA-seq, CITE-seq, ASAP-seq, and ASAP-seq data) and spans three sub-tasks: cell type annotation, drug response prediction, and perturbation analysis. This approach allows CellVerse to uncover both ***capabilities and limitations of current LLMs in understanding cell biology.***
<p align="center">
<img src="figs/fig_cellverse.png" width="95%"> <br>
</p>
In addition, we systematically evaluate the performance of 14 open-source and closed-source advanced LLMs on CellVerse.
## 🚀 Inference
We provide two formats for inference on CellVerse using APIs and [vLLM](https://github.com/vllm-project/vllm).
Inference using APIs:
```bash
python ./evaluation/infer_api.py \
--model_name "$Model" \
--openai_api_key "$API" \
--base_url "$URL" \
--output_path "./results/response.json" \
--dataset_path "./data/cta_scrna_full.json"
```
Inference using [vLLM](https://github.com/vllm-project/vllm):
```bash
python ./evaluation/infer_vllm.py \
--model_name "$Model" \
--output_path "./results/response.json" \
--dataset_path "./data/cta_scrna_full.json"
```
## 💪 Evaluation
After getting model responses (saved in `output_path`), you can extract the answer and calculate the metrics.
Here we provide an example of the extracted answer at `./results/ms_cta_response_deepseek_r1.json`.
## 🏆 Leaderboard
| **#** | **Model** | **Type** | **Source** | **Date** | **Avg** | **CTA (scRNA-seq)** | **CTA (CITE-seq)** | **CTA (ASAP-seq)** | **DRP** | **PSA** | **PDA** |
| ----- | -------------------- | -------- | ----------- | ---------- | -------- | ------- | ------- | ------- | ------- | ------- | ------- |
| 1 | **DeepSeek-R1 🥇** | Open | [Link](https://huggingface.co/deepseek-ai/DeepSeek-R1) | 2025-03 | 53.05 |42.38 | 58.29 | 28.00 | 50.00 | 76.67 | 62.96 |
| 2 | **GPT-4.1-mini 🥈** | Close | [Link](https://platform.openai.com/docs/models/gpt-4.1-mini) | 2025-04 | 52.24 |40.51 | 59.14 | 29.33 | 55.00 |68.33 | 61.11 |
| 3 | **GPT-4o 🥉** | Close | [Link](https://platform.openai.com/docs/models/gpt-4o) | 2024-11 | 49.45 |35.70 | 58.29 | 28.00 | 47.50 | 71.67 | 55.56 |
| 4 | **GPT-4.1** | Close | [Link](https://platform.openai.com/docs/models/gpt-4.1) | 2025-04 | 49.41 |37.83 | 61.43 | 28.22 |49.38 | 73.33 | 46.30 |
| 5 | **LLaMA-3.3-70B** | Open | [Link](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) | 2024-12 |45.86 | 32.75 |52.57 | 22.00 | 43.75 | 66.67 | 57.41 |
| 6 | **DeepSeek-V3** | Open | [Link](https://huggingface.co/deepseek-ai/DeepSeek-V3) | 2025-03 |45.53 | 37.57 | 57.14 | 27.11 | 50.63 | 76.67 | 24.07 |
| 7 | **Qwen-2.5-72B** | Open | [Link](https://huggingface.co/Qwen/Qwen2.5-72B) | 2024-09 | 43.66 |24.73 | 50.29 | 28.44 | 50.00 | 73.33 | 35.19 |
| 8 | **Qwen-2.5-32B** | Open | [Link](https://huggingface.co/Qwen/Qwen2.5-32B) | 2024-09 | 37.97 |22.46 | 48.86 | 23.11 | 45.63 | 76.67 | 11.11 |
| 9 | **Qwen-2.5-7B** | Open | [Link](https://huggingface.co/Qwen/Qwen2.5-7B) | 2024-09 | 35.16 |13.77 | 30.86 | 10.67 | 49.38 | 76.67 | 29.63 |
| 10 | **GPT-4o-mini** | Close | [Link](https://platform.openai.com/docs/models/gpt-4o-mini) | 2024-07 | 34.38 |23.93 | 48.57 | 16.89 | 43.75 | 41.67 | 31.48 |
| 11 | **GPT-4** | Close | [Link](https://platform.openai.com/docs/models/gpt-4) | 2023-06 | 18.57 |35.16 | 53.43 | 21.56 |1.25 | 0.00 | 0.00 |
| 12 | **C2S-Pythia-1B** | Open | [Link](https://huggingface.co/vandijklab/C2S-Scale-Pythia-1b-pt) | 2025-04 | 0.00| 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| 12 | **C2S-Pythia-410M** | Open | [Link](https://huggingface.co/vandijklab/C2S-Pythia-410m-diverse-single-and-multi-cell-tasks) | 2024-09 | 0.00| 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| 12 | **C2S-Pythia-160M** | Open | [Link](https://huggingface.co/vandijklab/pythia-160m-c2s) | 2024-02 | 0.00| 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
## :white_check_mark: Citation
If you find **CellVerse** useful for your research and applications, please kindly cite using this BibTeX:
```latex
@misc{zhang2025cellverselargelanguagemodels,
title={CellVerse: Do Large Language Models Really Understand Cell Biology?},
author={Fan Zhang and Tianyu Liu and Zhihong Zhu and Hao Wu and Haixin Wang and Donghao Zhou and Yefeng Zheng and Kun Wang and Xian Wu and Pheng-Ann Heng},
year={2025},
eprint={2505.07865},
archivePrefix={arXiv},
primaryClass={q-bio.QM},
url={https://arxiv.org/abs/2505.07865},
}
```
🔥 Please contact `zfkarl1998@gmail.com` if you would like to contribute to the leaderboard or have any problems.
|