Hi all, I have a couple of questions regarding the tool. First, thank you for providing both the tool.
For multiple multiome samples, is it better to use Cell Ranger aggr without normalization to integrate modalities and samples, or to process each sample’s matrix individually?
I’d like to use raw data for MultiVI, perform QC in Seurat, and then integrate the MultiVI latent space embedding. Would this be the recommended approach?
Additionally, are there papers that have used MultiVI for integrating multiple multiome samples with the code shared or example? And finally, how can we best assess whether sample and modality integration has been successful, within this framework?
Yes, you would probably like to preserve raw counts in cellranger for MultiVI analysis
Yes, QC in Seurat can be used to filter out bad cells.
In terms of code for running MultiVI, I can refer you to this tutorial (use your own data):
I guess there are papers that do that as well, need to find them.
In terms of assessing the quality of the integration, you should use UMAPs and scib-metrics with the latent space you got from the previous analysis. An example (of different models) can be shown here:
You can easily apply this kind of code to use the MultiVI latent space. However, scbi-metrics is mainly a comparative method to compare models’ latent space integration quality.