Question about adaptation of the model for new type of study material

[El-Castor]

Hello,

First of all, thank you for your very interesting work. I wanted to try your tools on my brachypodium embryo single cell data (plants). Surprisingly, it was able to annotate my different cell clusters very roughly using the seurat markers filtered as follows: avg_log2FC > 1.0 and p_val_adj < 0.05.

However, I have a few questions. If I've understood correctly, the 'tissuename' variable in the gptcelltype() command is important in guiding chatGPT to identify the tissue type, which will greatly influence its results. In fact, I used a custom gene annotation, '1G780903' for example, and despite this it was able to roughly identify certain cell types that could be expected in my tissue. My question is, is it possible to train it a little by providing it with a list of known cell markers for our study model, and/or also a pivot file of orthologs between different species with in addition a database of cell markers in different tissues and cell types for different species in order to have a more accurate and realistic annotation? If so, would you have any suggestions as to how this could be achieved?

Thank you in advance,

Best

[AYYYang]

Hello, I think maybe instead of having to train the model (which is very expensive), you could provide some additional information in the prompt (I am not sure if this is already a feature for GPTCellType).

What maybe you could do is printing out the prompt itself, see https://winnie09.github.io/Wenpin_Hou/pages/gptcelltype.html, search for ## [1] "Note: OpenAI API key not found: returning the prompt itself."

What you can do, for now, is manually adding additional information you have such as a list of known cell markers for our study model and paste the prompt into ChatGPT. For example:

## Identify cell types of human PBMC cells using the following markers separately for each
##  row. Only provide the cell type name. Do not show numbers before the name.
##  Some can be a mixture of multiple cell types. 
## I also have an additional list of known cell markers for my own study model: <<insert your data here>>
## 0:HLA-DPB1,HLA-DRB1,HLA-DPA1,HLA-DRA,HLA-DRB5,HLA-DQB1,LYZ,TYMP,HLA-DQA1,HLA-DMB
## 1:S100A8,TYMP,S100A9,LYZ,CST3,FCGRT,LST1,AIF1,TYROBP,IFITM3
## 2:HLA-DPB1,MS4A1,HLA-DQB1,HLA-DRB1,HLA-DRA,TCL1A,CD79A,CD79B,HLA-DPA1,HLA-DRB5

This simple trick of using prompt engineering should give you some results that fit your study better.

One thing you might want to consider whether you are comfortable sharing your data (especially if unpublished) with GPT.