I read your preprint with great interest. The inclusion of batch specific decoders seems interesting for the dataset integration problem.
I was wondering if you all had code to reproduce the figures in the paper. I am a developer of scvi-tools and was surprised to see the results in your preprint.
In particular, I ran scVI with the PBMC dataset on your docs website and achieve what appears to be substantially different results. See here:
I would also like to correct a misconception. The implementation of scVI does not actually encode the batch covariates by default (and as far as I'm aware has never done this). Thus, query data can be passed through a pre-trained encoder just like in SCALEX. For an example of this see here:
Hello,
I read your preprint with great interest. The inclusion of batch specific decoders seems interesting for the dataset integration problem.
I was wondering if you all had code to reproduce the figures in the paper. I am a developer of scvi-tools and was surprised to see the results in your preprint.
In particular, I ran scVI with the PBMC dataset on your docs website and achieve what appears to be substantially different results. See here:
https://colab.research.google.com/drive/1bOXBuOI-haWoBOJG8TeO1fsqSKVUwv3R?usp=sharing
I would also like to correct a misconception. The implementation of scVI does not actually encode the batch covariates by default (and as far as I'm aware has never done this). Thus, query data can be passed through a pre-trained encoder just like in SCALEX. For an example of this see here:
https://colab.research.google.com/drive/19AfhDo6kGA_BxPgC1CijjYnpIYQrLw_G?usp=sharing
Thanks!