HLCA reference model incompatible with scvi-tools > 1.0.3

[luisa-cusick]

Thank you for sharing this very useful tool!

I am following the tutorial for HLCA mapping: https://github.com/theislab/scarches/blob/hlca_tutorial_improvements/notebooks/hlca_map_classify.ipynb and I downloaded the most recent HLCA reference model: https://zenodo.org/record/7599104/files/HLCA_reference_model.zip.

I tried loading the model into an sca.models.SCANVI object with the provided command:

surgery_model = sca.models.SCANVI.load_query_data(
    adata_query,
    ref_model_dir,
    freeze_dropout=True,
)

This fails with scvi-tools versions > 1.0.3 because the model base has an added attribute starting in version 1.0.4. I was able to get this working with some changes to my environment, but it would be nice to be able to use the newest version of SCVI. Do you have an updated HLCA scANVI or scVI model?

Here's the full traceback:

in ArchesMixin.load_query_data(cls, adata, reference_model, inplace_subset_query_vars, accelerator, device, unfrozen, freeze_dropout, freeze_expression, freeze_decoder_first_layer, freeze_batchnorm_encoder, freeze_batchnorm_decoder, freeze_classifier)
    116 if _SETUP_ARGS_KEY not in registry:
    117     raise ValueError(
    118         "Saved model does not contain original setup inputs. "
    119         "Cannot load the original setup."
    120     )
--> 122 setup_method = getattr(cls, registry[_SETUP_METHOD_NAME])
    123 setup_method(
    124     adata,
    125     source_registry=registry,
   (...)
    128     **registry[_SETUP_ARGS_KEY],
    129 )
    131 model = _initialize_model(cls, adata, attr_dict)

KeyError: 'setup_method_name'

Thank you!

Luisa

[shabnamzaman]

I struggled with this endlessly, too. What finally worked for me was a SCANVI model from the Hugging face repo. I came across it in the documentation for scvi-tools. In short:

from scvi.hub import HubModel

save_dir = tempfile.TemporaryDirectory()
hubmodel = HubModel.pull_from_huggingface_hub(
    "scvi-tools/human-lung-cell-atlas-scanvi", cache_dir=save_dir.name
)
adata = hubmodel.adata
model = hubmodel.model

And then proceed as usual.

import scarches as sca

surgery_model = sca.models.SCANVI.load_query_data(
    adata_query,
    model,
    freeze_dropout=True,
)

But yes, I can confirm that the models downloaded by and/or referred to in the scArches tutorials are incompatible. My current environment has the following installed:

• scanpy version: 1.11.4
• scarches version: 0.6.1
• scvi-tools version: 1.4.0