luosc
commented
1 week ago Thank you for the note. My short answer is that I do not know. You are welcome to test CAST on your data. We are excited to see if it eventually works. I am happy to assist if you encounter any technical issues in this process. Our self-supervised learning scheme and models have succeeded with integrating translatome and transcriptome, we really wanted to test proteomics with transcriptomics and optimize based on that, but we did not have such data by the date of paper acceptance. In my recent review paper, I also mentioned that a common challenge in the field is dealing with weakly linked modalities. It is almost non-quantitative how robust the batch correction effects could be for a self supervised learning model. Alternatively, you can supply it with human input and expertise, e.g. a curated low-dimensional embedding that you see better correspondence between these modalities. We encourage users to test CAST using integrated embeddings (output from single-cell integration methods) instead of raw counts to see how it goes.
OMIC-coding
Excellent work, there's no doubt that CAST would be applied in various spatially resolved multi-omics data analysis tasks. However, I wanted to know if conducting an alignment between MALDI-MSI and Visium data using CAST is possible, as these two modalities don't share any common variables. Thanks a lot!