Code for masked-reconstruction

[kevinfleisher]

Hello,

I have seen and used the tutorial for 'Zero shot inference' with the scGPT-spatial model. I was able to integrate a dataset into the model latent space, but I am wondering whether there will be a tutorial or functions on your Github to do masked reconstruction using these embeddings as your training paradigm suggests in the preprint. Is it already available to use somewhere?

Thanks

[ChloeXWang]

Hi @kevinfleisher,

Thanks for trying out scGPT-spatial! You can find the neighborhood-based masked prediction implementation in model.py. We will release a finetuning tutorial soon.

[kevinfleisher]

Hello, Thank you so much for the response. I see that the model was trained to do masked predictions on genes, but I am having some difficulty getting the code to run the way I would like it to. Basically I have some gene expression vectors and I am wanting to use your trained model to mask out certain genes and, using the embeddings from the model, reconstruct their expressions. Will you provide some way to do this with a function call from your model? Is this part of the fine-tuning tutorial you mentioned? If so that is great, but a function call using your pretrained model would really be fantastic. Maybe I could use some help figuring out a path to using the model and its attributes and methods to do just that on some test data I have. Thank you! Kevin Fleisher On Feb 12, 2025, at 5:22 PM, ChloeXWang ***@***.***> wrote: Hi @kevinfleisher<https://github.com/kevinfleisher>, Thanks for trying out scGPT-spatial! You can find the neighborhood-based masked prediction implementation in model.py. We will release a finetuning tutorial soon. — Reply to this email directly, view it on GitHub<#2 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/A2AOSDELP7NWTVAEB6PQXND2PPXT3AVCNFSM6AAAAABW4AQ5QGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNJVGIYTQMJSHE>. You are receiving this because you were mentioned.Message ID: ***@***.***>