Autoencoder gene expression reconstruction accuracy

Hi scvi folks,

We had a discussion in the lab, and we’re hoping some scvi-verse expert could chime in.

A naive question here: what reconstruction accuracy should one expect from scvi tools in terms of total variance explained? ballpark value, is it 1%, 10%, 90%?

Where reconstruction accuracy is : mean((X-X’)^2)/mean((X-mean(X))^2) where X is cell x gene input matrix to scvi and X’ is the output of the autoencoder.


Thanks for reaching out! We are working on designing a model criticism module for scvi-tools, and metrics like the one you describe could be a good addition. Do you have references describing the use of this metric?