We have never tested this, but I think we need some code changes to make scvi work with categorical covariates and continuous covariates. What's going on, in my understanding, is this:
In the scvi model we have a batch latent variable $s_n$
This $s_n$, functionally, is used as an input that the encoder and decoder are conditioned on.
In the simple case, $s_n$ is a one-hot encoding of a single categorical "batch" variable. We have this implemented.
To get more complex, the scvi codebase allows you to (and I think the published CZI model does) use additional categorical (or continuous) covariates. They all become part of $s_n$, and are used to condition the encoder and decoder, but now $s_n$ no longer needs to be one-hot.
For example, let's say we have 4 datasets ("dataset_id") and 2 suspension types ("suspension_type"). Like this
cell dataset_id suspension_type
0 A whole_cell
1 A nucleus
2 B nucleus
3 B whole_cell
4 A whole_cell
5 B nucleus
...
then we could set up a model with batch="dataset_id" and categorical_covariates=["suspension_type"] and the $s_n$ values might look like this
where the first two entries are one-hot for "dataset_id" and the last two entries are one-hot for "suspension_type".
The trick here is how to handle the embedding of batch for the case where we want batch to be embedded. Would we embed only the "batch", i.e. "dataset_id" and let the rest still be one-hot? Would we embed that whole s_nb thing jointly? Would we embed each part separately, i.e. one embedding for "dataset_id" and another separate embedding for "suspension_type"?
We have never tested this, but I think we need some code changes to make scvi work with categorical covariates and continuous covariates. What's going on, in my understanding, is this:
In the scvi model we have a batch latent variable $s_n$
This $s_n$, functionally, is used as an input that the encoder and decoder are conditioned on.
In the simple case, $s_n$ is a one-hot encoding of a single categorical "batch" variable. We have this implemented.
To get more complex, the scvi codebase allows you to (and I think the published CZI model does) use additional categorical (or continuous) covariates. They all become part of $s_n$, and are used to condition the encoder and decoder, but now $s_n$ no longer needs to be one-hot.
For example, let's say we have 4 datasets ("dataset_id") and 2 suspension types ("suspension_type"). Like this
then we could set up a model with
batch="dataset_id"andcategorical_covariates=["suspension_type"]and the $s_n$ values might look like thiswhere the first two entries are one-hot for "dataset_id" and the last two entries are one-hot for "suspension_type".
The trick here is how to handle the embedding of batch for the case where we want batch to be embedded. Would we embed only the "batch", i.e. "dataset_id" and let the rest still be one-hot? Would we embed that whole
s_nbthing jointly? Would we embed each part separately, i.e. one embedding for "dataset_id" and another separate embedding for "suspension_type"?