fetching model.pt

Description of feature

Can you please share the complete configuration of folder and files containing within the folder/path containing "model.pt". Its throwing error! Please share a small tutorial for executing this in small dataset. Its a kind request! Otherwise it will be a real pity if people can't make use of such an amazing work. Thanks :)

[arunhpatil]

Hi,

Thanks for the great tool. I was able to work with the tutorial using marker genes.

However, I am now stuck at the reference data and algorith used. Where can we download the models (model.pt)? If we have to create them ourselves which one to use (scPoli and scANVI)?

URL: https://devsystemslab.github.io/HNOCA-tools/vignettes/atlas_mapping/
hnoca_model = scarches.models.scPoli.load(
    "scpoli_HNOCA/",
    adata=hnoca_adata,
)

and

Quote: We already trained a scANVI model for this.

URL: https://devsystemslab.github.io/HNOCA-tools/vignettes/reference_mapping/
# Load reference model
ref_model = scvi.model.SCANVI.load(
    "scarches_BraunLinnarsson/model.pt",
    adata=ref_adata,
)

One of the GitHub users commented to locate the files here: https://huggingface.co/ChelseaBright/human-neural-organoid-cell-atlas-scpoli/tree/main

However, hnoca_cleanedmeta.h5ad is ~19Gb, while the reffered link has the h5ad which is only ~750 Mb. Can you point out where to find these model.pt files? I think this have been a recurring question and I couldn't find any solutions. Can you please clarify?

Thank you,
Arun.

[zhen-fu]

I was also trying to figure out where to find a pretrained HNOCA reference model for query mapping.

I first tried the model uploaded by Chelsea Bright on Hugging Face. That repository contains a complete model bundle, but it appears to have been exported using the newer scvi-tools-style model format. I was not able to load that model directly with the older scPoli/scArches workflow used by the current HNOCA-tools setup.

What worked for me was using the scPoli model provided in the Theis lab's neural_organoid_atlas GitHub repository:

https://github.com/theislab/neural_organoid_atlas

Specifically, I used:

supplemental_files/scpoli_model_params/

which contains the older scPoli model files:

attr.pkl
model_params.pt
var_names.csv

The matching AnnData object is not included in that model directory. I downloaded the minimal HNOCA reference AnnData from the third Zenodo deposition for query-to-reference mapping:

https://zenodo.org/records/15004818

I used hnoca_minimal_for_mapping.h5ad as the reference AnnData.

The following worked for me:

import torch
import scanpy as sc
from scarches.models import scPoli
import hnoca.map as mapping

model_dir = (
    "/path/to/neural_organoid_atlas/"
    "supplemental_files/scpoli_model_params"
)

adata_mini = sc.read_h5ad(
    "/path/to/hnoca_minimal_for_mapping.h5ad"
)

ref_model = scPoli.load(
    model_dir,
    adata=adata_mini,
    map_location=torch.device("cpu"),
)

mapper = mapping.AtlasMapper(ref_model)

For my query dataset, I first prepared the features to match the HNOCA reference:

query_prepared = mapping.prepare_features(
    query_adata,
    ref_model,
)

Because the scPoli reference model was trained using batch as the condition key, the query AnnData also needs a batch column in .obs. For example:

query_prepared.obs["batch"] = "query_batch"

I then mapped the query with:

mapper.map_query(
    query_prepared,
    retrain="partial",
    max_epochs=100,
    batch_size=1024,
)

This seemed to work (at least I got it running).

One additional note: I needed an older scArches-compatible Python environment because the published scPoli model uses the older scArches serialization format. In my case, the working environment included approximately:

Python 3.9
scArches 0.6.1
scvi-tools 1.1.2
anndata 0.10.7
scanpy 1.10.1
numpy 1.26.4
pandas 1.5.3
hnoca 0.1.1

Hopefully this helps others who are also trying to locate and load the pretrained HNOCA reference model.

[thompsr]

Hello,
I found the model on the Archmap site (archmap.bio) and emailed them to ask for the model
It is here (scpoli model):
https://www.archmap.bio/#/references/atlases/65b3b52e0cae6d0c1e93340c

I then followed their tutorial using the scpoli option:
https://archmap-docu.readthedocs.io/en/latest/tutorials/map_local.html