Add a function of cell types annotation with local LLMs

[fredsamhaak]

Dear authors of omicverse,

First of all, thanks for providing such a useful tool!
I've added a function to annotate cell types using local LLMs (e.g. Qwen/Qwen2-7B-Instruct). This might be a useful feature for those who has local computing resources.

Feel free to contact me if you have any questions.

All the best,
He

[Starlitnightly]

Hi, Can you provide a full tutorial? In markdown format.

Best Regards.
Zehua

[fredsamhaak]

Hi @Starlitnightly, here is the tutorial. It is quite simple so I just put it directly below:

=====
Before starting with the cell type annotation, ensure you have preprocessed your data accordingly. The data preprocessing steps should provide you with a dictionary or a pandas DataFrame containing clusters and their respective gene markers.

To use ov.single.gptcelltype_local function to annotate cell types, you can simply set the model_name parameter:

# Example dictionary of gene markers for clusters
all_markers={
    'cluster1':['CD3D','CD3E'],
    'cluster2':['MS4A1']
}

model_name = 'path/to/your/local/model'  # e.g. '~/models/Qwen2-7B-Instruct'

# Annotate cell types
result = gptcelltype_local(all_markers, tissuename='PBMC', speciename='human', 
                           model_name=model_name, topgenenumber=5)
result

Note that you may encounter network problems that prevent you from downloading the model. In this case, please refer to https://zhuanlan.zhihu.com/p/663712983.

=====

Please feel free to contact me if you have other questions.

All the best,
He

[Starlitnightly]

I understand what you're saying, but is it possible for the model to be recognised for weights straight after it's been downloaded? Doesn't it require any other additional configuration?

Also, for local models, ollama is a good choice and compatible with our api.

Zehua

[fredsamhaak]

Yes, the model can be recognized for weights immediately after download without needing additional configuration. Here is an example:

Thanks for reminding me of ollama and I will keep an eye on it.

All the best,
He

[Starlitnightly]

Yes, the model can be recognized for weights immediately after download without needing additional configuration. Here is an example:

Thanks for reminding me of `ollama` and I will keep an eye on it.

All the best, He

Our tutorial is at https://github.com/Starlitnightly/omicverse/blob/master/omicverse_guide/docs/Tutorials-single/t_gptanno.ipynb. You can add a section on local LLM using at the end of the tutorial.

Zehua