Use Geneformer to predict cell types
Hi,
I have a couple of datasets from different tissues and would like to generate a first cell type annotation of them with Geneformer. Is there a possibility to do it right now with a fine-tuned model? or do I have to fine-tune my own?
Thanks for the great tool, Christina!
Thank you for your question! We provide the trunk of the CELLxGENE multitask fine-tuned model here. We would suggest you start with this model, that is already tuned to recognizing cell types, and then fine-tune with labels that correspond to the cell types you'd like to annotate. There are limitations in the CELLxGENE cell type annotations, so if you take this approach you can ensure the possible labels for prediction match what you would be interested in annotating. If you start with this model as the pretrained weights as opposed to the initial pretrained one, your weights will already be tuned to distinguishing cell types so the fine-tuning can likely be fairly minimal, just to give the model a name for the given region in the embedding space.