How to create pre-trained models

[ccruizm]

Good day,

I would like to use your tools to compare predictions on my dataset but would like to create my own pre-trained reference. Could you please share how this can be done?

Thanks in advance!

[canergen]

Hi, Run the Colab notebook (either locally or on Colab) and change Process_Query call to (4000 hvg is a suggestion that works well in my hands but depends on the complexity of the cell-types in your dataset, you can increase it if it's necessary for other processing like PCA). If the cell-types in the reference are not named based on an ontology, use _cl_obofolder==False.

adata = Process_Query( query_adata, ref_adata, query_labels_key=query_labels_key, query_batch_key=query_batch_key, ref_labels_key=ref_labels_key, ref_batch_key=ref_batch_key, unknown_celltype_label=unknown_celltype_label, save_path_trained_models=output_model_fn, cl_obo_folder="./PopV/ontology/", prediction_mode="retrain", n_samples_per_label=n_samples_per_label, use_gpu=0, compute_embedding=True, hvg=4000, ).adata

[ccruizm]

Thank you very much for your speedy reply and the recommendations to run it in an independent reference dataset. Will test it on my data and see how it performs.

[ccruizm]

Good day!

I am trying to run the code as you suggested setting cl_obo_folder=False but I am getting this error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[20], line 4
      1 # https://github.com/YosefLab/PopV/issues/19
      2 from popv.preprocessing import Process_Query
----> 4 adata = Process_Query(
      5     query_adata,
      6     ref_adata,
      7     query_labels_key=query_labels_key,
      8     query_batch_key=query_batch_key,
      9     ref_labels_key=ref_labels_key,
     10     ref_batch_key=ref_batch_key,
     11     unknown_celltype_label=unknown_celltype_label,
     12     save_path_trained_models=output_model_fn,
     13     # cl_obo_folder="./PopV/ontology/",
     14     cl_obo_folder=False,
     15     prediction_mode="retrain",
     16     n_samples_per_label=n_samples_per_label,
     17     use_gpu=0,
     18     compute_embedding=True,
     19     hvg=5000,
     20 ).adata

File ~/miniconda3/envs/popv/lib/python3.8/site-packages/popv/preprocessing.py:197, in Process_Query.__init__(self, query_adata, ref_adata, ref_labels_key, ref_batch_key, query_labels_key, query_batch_key, query_layers_key, prediction_mode, cl_obo_folder, unknown_celltype_label, n_samples_per_label, pretrained_scvi_path, save_path_trained_models, hvg, use_gpu, compute_embedding, return_probabilities)
    195     self.nlp_emb_file = cl_obo_folder + "cl.ontology.nlp.emb"
    196 try:
--> 197     with open(self.cl_obo_file) if self.cl_obo_file else True:
    198         pass
    199 except FileNotFoundError:

AttributeError: __enter__

I created a dedicated conda env to install PopV, and I am using the latest version (0.2.2). What do you think the problem is?

Thanks in advance!

[canergen]

Thanks for the report. It's fixed in v0.3.1 (uploaded to PyPI). Additionally, PopV now includes harmony-pytorch integration and a faster version of KNN.