Note

This page was generated from scarches_scvi_tools.ipynb. Interactive online version: Colab badge.

Reference mapping with scvi-tools

This tutorial covers the usage of the scArches method with SCVI, SCANVI, and TOTALVI.

This particular workflow is useful in the case where a model is trained on some data (called reference here) and new samples are received (called query). The goal is to analyze these samples in the context of the reference, by mapping the query cells to the same reference latent space. This workflow may also be used in the scarches package, but here we demonstrate using only scvi-tools.

Imports and scvi-tools installation (colab)

[1]:
import sys

# If True, will install via pypi, else will install from source
stable = True
IN_COLAB = "google.colab" in sys.modules

if IN_COLAB and stable:
    !pip install --quiet scvi-tools[tutorials]
elif IN_COLAB and not stable:
    !pip install --quiet --upgrade jsonschema
    !pip install --quiet git+https://github.com/yoseflab/scvi-tools@master#egg=scvi-tools[tutorials]

if IN_COLAB:
    !pip install --quiet scrublet
     |████████████████████████████████| 69 kB 2.8 MB/s
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
nbclient 0.5.4 requires jupyter-client>=6.1.5, but you have jupyter-client 5.3.5 which is incompatible.
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
    Preparing wheel metadata ... done
     |████████████████████████████████| 813 kB 7.2 MB/s
     |████████████████████████████████| 678 kB 61.0 MB/s
     |████████████████████████████████| 127 kB 60.8 MB/s
     |████████████████████████████████| 242 kB 50.4 MB/s
     |████████████████████████████████| 211 kB 47.8 MB/s
     |████████████████████████████████| 2.0 MB 34.9 MB/s
     |████████████████████████████████| 1.4 MB 54.1 MB/s
     |████████████████████████████████| 3.2 MB 46.7 MB/s
     |████████████████████████████████| 41 kB 117 kB/s
     |████████████████████████████████| 8.8 MB 33.4 MB/s
     |████████████████████████████████| 48 kB 5.9 MB/s
     |████████████████████████████████| 636 kB 47.9 MB/s
     |████████████████████████████████| 123 kB 51.1 MB/s
     |████████████████████████████████| 282 kB 56.5 MB/s
     |████████████████████████████████| 829 kB 51.3 MB/s
     |████████████████████████████████| 1.3 MB 45.7 MB/s
     |████████████████████████████████| 51 kB 6.6 MB/s
     |████████████████████████████████| 80 kB 9.3 MB/s
     |████████████████████████████████| 1.1 MB 53.4 MB/s
     |████████████████████████████████| 294 kB 46.7 MB/s
     |████████████████████████████████| 142 kB 66.0 MB/s
     |████████████████████████████████| 63 kB 2.2 MB/s
  Building wheel for scvi-tools (PEP 517) ... done
  Building wheel for docrep (setup.py) ... done
  Building wheel for loompy (setup.py) ... done
  Building wheel for future (setup.py) ... done
  Building wheel for umap-learn (setup.py) ... done
  Building wheel for pynndescent (setup.py) ... done
  Building wheel for numpy-groupies (setup.py) ... done
  Building wheel for sinfo (setup.py) ... done
     |████████████████████████████████| 646 kB 8.0 MB/s
  Building wheel for annoy (setup.py) ... done
[2]:
import warnings
warnings.simplefilter(action='ignore', category=FutureWarning)

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

import anndata
import scvi
import scanpy as sc

sc.set_figure_params(figsize=(4, 4))
scvi.settings.seed = 94705
Global seed set to 0
/usr/local/lib/python3.7/dist-packages/numba/np/ufunc/parallel.py:363: NumbaWarning: The TBB threading layer requires TBB version 2019.5 or later i.e., TBB_INTERFACE_VERSION >= 11005. Found TBB_INTERFACE_VERSION = 9107. The TBB threading layer is disabled.
  warnings.warn(problem)
Global seed set to 94705

Online update of SCVI

Here we use the pancreas dataset described in the scArches manuscript, that is also widely used to benchmark integration methods.

[3]:
import gdown
url = 'https://drive.google.com/uc?id=1ehxgfHTsMZXy6YzlFKGJOsBKQ5rrvMnd'
output = 'pancreas.h5ad'
gdown.download(url, output, quiet=False)
Downloading...
From: https://drive.google.com/uc?id=1ehxgfHTsMZXy6YzlFKGJOsBKQ5rrvMnd
To: /content/pancreas.h5ad
126MB [00:00, 136MB/s]
[3]:
'pancreas.h5ad'
[4]:
adata_all = sc.read("pancreas.h5ad")
adata = adata_all.raw.to_adata()
print(adata)
AnnData object with n_obs × n_vars = 15681 × 1000
    obs: 'batch', 'study', 'cell_type', 'size_factors'
[5]:
adata.obs.study.value_counts()
[5]:
Pancreas inDrop         8391
Pancreas SS2            2961
Pancreas CelSeq2        2426
Pancreas CelSeq         1271
Pancreas Fluidigm C1     632
Name: study, dtype: int64

We consider the SS2 and CelSeq2 samples as query, and all the others as reference.

[6]:
query = np.array([s in ["Pancreas SS2", "Pancreas CelSeq2"] for s in adata.obs.study])

adata_ref = adata[~query].copy()
adata_query = adata[query].copy()

Train reference

We train the reference using the standard SCVI workflow, except we add a few non-default parameters that were identified to work well with scArches.

[7]:
scvi.model.SCVI.setup_anndata(adata_ref, batch_key="study")
INFO     Using batches from adata.obs["study"]
INFO     No label_key inputted, assuming all cells have same label
INFO     Using data from adata.X
INFO     Successfully registered anndata object containing 10294 cells, 1000 vars, 3 batches,
         1 labels, and 0 proteins. Also registered 0 extra categorical covariates and 0 extra
         continuous covariates.
INFO     Please do not further modify adata until model is trained.
[8]:
arches_params = dict(
    use_layer_norm="both",
    use_batch_norm="none",
    encode_covariates=True,
    dropout_rate=0.2,
    n_layers=2,
)

vae_ref = scvi.model.SCVI(
    adata_ref,
    **arches_params
)
vae_ref.train()
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
Epoch 1/400:   0%|          | 0/400 [00:00<?, ?it/s]
/usr/local/lib/python3.7/dist-packages/scvi/distributions/_negative_binomial.py:435: UserWarning: The value argument must be within the support of the distribution
  UserWarning,
Epoch 400/400: 100%|██████████| 400/400 [05:47<00:00,  1.15it/s, loss=495, v_num=1]

Now we obtain the latent representation, and use Scanpy to visualize with UMAP.

[9]:
adata_ref.obsm["X_scVI"] = vae_ref.get_latent_representation()
sc.pp.neighbors(adata_ref, use_rep="X_scVI")
sc.tl.leiden(adata_ref)
sc.tl.umap(adata_ref)
[10]:
sc.pl.umap(
    adata_ref,
    color=["study", "cell_type"],
    frameon=False,
    ncols=1,
)
../../_images/tutorials_notebooks_scarches_scvi_tools_17_0.png

Online update with query

We can load a new model with the query data either using

  1. The saved reference model

  2. The instantiation of the reference model in memory

[11]:
# save the reference model
dir_path = "pancreas_model/"
vae_ref.save(dir_path, overwrite=True)
[12]:
# both are valid
vae_q = scvi.model.SCVI.load_query_data(
    adata_query,
    dir_path,
)
vae_q = scvi.model.SCVI.load_query_data(
    adata_query,
    vae_ref,
)
INFO     .obs[_scvi_labels] not found in target, assuming every cell is same category
INFO     Using data from adata.X
INFO     Registered keys:['X', 'batch_indices', 'labels']
INFO     Successfully registered anndata object containing 5387 cells, 1000 vars, 5 batches,
         1 labels, and 0 proteins. Also registered 0 extra categorical covariates and 0 extra
         continuous covariates.
INFO     Using data from adata.X
INFO     Registered keys:['X', 'batch_indices', 'labels']
INFO     Successfully registered anndata object containing 5387 cells, 1000 vars, 5 batches,
         1 labels, and 0 proteins. Also registered 0 extra categorical covariates and 0 extra
         continuous covariates.
/usr/local/lib/python3.7/dist-packages/scvi/model/base/_archesmixin.py:96: UserWarning: Query integration should be performed using models trained with version >= 0.8
  "Query integration should be performed using models trained with version >= 0.8"
/usr/local/lib/python3.7/dist-packages/scvi/model/base/_base_model.py:151: UserWarning: Make sure the registered X field in anndata contains unnormalized count data.
  "Make sure the registered X field in anndata contains unnormalized count data."

This is a typical SCVI object, and after training, can be used in any defined way.

For training the query data, we recommend using a weight_decay of 0.0. This ensures the latent representation of the reference cells will remain exactly the same if passing them through this new query model.

[13]:
vae_q.train(max_epochs=200, plan_kwargs=dict(weight_decay=0.0))
adata_query.obsm["X_scVI"] = vae_q.get_latent_representation()
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
Epoch 1/200:   0%|          | 0/200 [00:00<?, ?it/s]
/usr/local/lib/python3.7/dist-packages/scvi/distributions/_negative_binomial.py:435: UserWarning: The value argument must be within the support of the distribution
  UserWarning,
Epoch 200/200: 100%|██████████| 200/200 [01:25<00:00,  2.33it/s, loss=1.13e+03, v_num=1]
/usr/local/lib/python3.7/dist-packages/scvi/model/base/_base_model.py:151: UserWarning: Make sure the registered X field in anndata contains unnormalized count data.
  "Make sure the registered X field in anndata contains unnormalized count data."
[14]:
sc.pp.neighbors(adata_query, use_rep="X_scVI")
sc.tl.leiden(adata_query)
sc.tl.umap(adata_query)
[15]:
sc.pl.umap(
    adata_query,
    color=["study", "cell_type"],
    frameon=False,
    ncols=1,
)
../../_images/tutorials_notebooks_scarches_scvi_tools_24_0.png

Visualize reference and query

[16]:
adata_full = adata_query.concatenate(adata_ref)

The concatenated object has the latent representations of both reference and query, but we are also able to reobtain these values using the query model.

[17]:
adata_full.obsm["X_scVI"] = vae_q.get_latent_representation(adata_full)
INFO     Input adata not setup with scvi. attempting to transfer anndata setup
INFO     Using data from adata.X
INFO     Registered keys:['X', 'batch_indices', 'labels']
INFO     Successfully registered anndata object containing 15681 cells, 1000 vars, 5 batches,
         1 labels, and 0 proteins. Also registered 0 extra categorical covariates and 0 extra
         continuous covariates.
/usr/local/lib/python3.7/dist-packages/scvi/model/base/_base_model.py:151: UserWarning: Make sure the registered X field in anndata contains unnormalized count data.
  "Make sure the registered X field in anndata contains unnormalized count data."
[18]:
sc.pp.neighbors(adata_full, use_rep="X_scVI")
sc.tl.leiden(adata_full)
sc.tl.umap(adata_full)
[19]:
sc.pl.umap(
    adata_full,
    color=["study", "cell_type"],
    frameon=False,
    ncols=1,
)
... storing 'study' as categorical
../../_images/tutorials_notebooks_scarches_scvi_tools_30_1.png

Online update of SCANVI

We’ll use the same Pancreas dataset, this time we set it up such that we register that the dataset has labels.

The advantage of SCANVI is that we’ll be able to predict the cell type labels of the query dataset. In the case of SCVI, a separate classifier (e.g., nearest-neighbor, random forest, etc.) would have to be trained on the reference latent space.

[20]:
scvi.model.SCANVI.setup_anndata(adata_ref, batch_key="study", labels_key="cell_type")
INFO     Using batches from adata.obs["study"]
INFO     Using labels from adata.obs["cell_type"]
INFO     Using data from adata.X
INFO     Successfully registered anndata object containing 10294 cells, 1000 vars, 3 batches,
         8 labels, and 0 proteins. Also registered 0 extra categorical covariates and 0 extra
         continuous covariates.
INFO     Please do not further modify adata until model is trained.

Train reference

SCANVI tends to perform better in situations where it has been initialized using a pre-trained SCVI model. In this case, we will use vae_ref that we have already trained above. In other words, a typical SCANVI workflow will be:

scvi_model = SCVI(adata_ref, **arches_params)
scvi_model.train()
scanvi_model = SCANVI.from_scvi_model(scvi_model, "Unknown")
scanvi_model.train()

Applying this workflow in the context of this tutorial:

[21]:
# unlabeled category does not exist in adata.obs[labels_key]
# so all cells are treated as labeled
vae_ref_scan = scvi.model.SCANVI.from_scvi_model(
    vae_ref,
    unlabeled_category="Unknown",
)
[22]:
vae_ref_scan.train(max_epochs=20, n_samples_per_label=100)
INFO     Training for 20 epochs.
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
Epoch 1/20:   0%|          | 0/20 [00:00<?, ?it/s]
/usr/local/lib/python3.7/dist-packages/scvi/distributions/_negative_binomial.py:435: UserWarning: The value argument must be within the support of the distribution
  UserWarning,
Epoch 20/20: 100%|██████████| 20/20 [00:38<00:00,  1.92s/it, loss=555, v_num=1]
[23]:
adata_ref.obsm["X_scANVI"] = vae_ref_scan.get_latent_representation()
sc.pp.neighbors(adata_ref, use_rep="X_scANVI")
sc.tl.leiden(adata_ref)
sc.tl.umap(adata_ref)
[24]:
sc.pl.umap(
    adata_ref,
    color=["study", "cell_type"],
    frameon=False,
    ncols=1,
)
../../_images/tutorials_notebooks_scarches_scvi_tools_39_0.png

Online update with query

[25]:
dir_path_scan = "pancreas_model_scanvi/"
vae_ref_scan.save(dir_path_scan, overwrite=True)

SCANVI is a bit trickier. Because the query data has the true cell type labels, we need to do something a bit hacky and tell the model which cells are labeled and which are unlabeled. We will make this cleaner in a future release.

[26]:
vae_q = scvi.model.SCANVI.load_query_data(
    adata_query,
    dir_path_scan,
)
vae_q._unlabeled_indices = np.arange(adata_query.n_obs)
vae_q._labeled_indices = []
INFO     Using data from adata.X
INFO     Registered keys:['X', 'batch_indices', 'labels']
INFO     Successfully registered anndata object containing 5387 cells, 1000 vars, 5 batches,
         8 labels, and 0 proteins. Also registered 0 extra categorical covariates and 0 extra
         continuous covariates.
/usr/local/lib/python3.7/dist-packages/scvi/model/base/_archesmixin.py:96: UserWarning: Query integration should be performed using models trained with version >= 0.8
  "Query integration should be performed using models trained with version >= 0.8"
/usr/local/lib/python3.7/dist-packages/scvi/model/base/_base_model.py:151: UserWarning: Make sure the registered X field in anndata contains unnormalized count data.
  "Make sure the registered X field in anndata contains unnormalized count data."
[27]:
vae_q.train(
    max_epochs=100,
    plan_kwargs=dict(weight_decay=0.0),
    check_val_every_n_epoch=10,
)
INFO     Training for 100 epochs.
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
Epoch 1/100:   0%|          | 0/100 [00:00<?, ?it/s]
/usr/local/lib/python3.7/dist-packages/scvi/distributions/_negative_binomial.py:435: UserWarning: The value argument must be within the support of the distribution
  UserWarning,
Epoch 100/100: 100%|██████████| 100/100 [01:28<00:00,  1.13it/s, loss=1.2e+03, v_num=1]
[28]:
adata_query.obsm["X_scANVI"] = vae_q.get_latent_representation()
adata_query.obs["predictions"] = vae_q.predict()
/usr/local/lib/python3.7/dist-packages/scvi/model/base/_base_model.py:151: UserWarning: Make sure the registered X field in anndata contains unnormalized count data.
  "Make sure the registered X field in anndata contains unnormalized count data."
[29]:
df = adata_query.obs.groupby(["cell_type", "predictions"]).size().unstack(fill_value=0)
norm_df = df / df.sum(axis=0)

plt.figure(figsize=(8, 8))
_ = plt.pcolor(norm_df)
_ = plt.xticks(np.arange(0.5, len(df.columns), 1), df.columns, rotation=90)
_ = plt.yticks(np.arange(0.5, len(df.index), 1), df.index)
plt.xlabel("Predicted")
plt.ylabel("Observed")


[29]:
Text(0, 0.5, 'Observed')
../../_images/tutorials_notebooks_scarches_scvi_tools_46_1.png

Analyze reference and query

[30]:
adata_full = adata_query.concatenate(adata_ref)

This just makes a column in the anndata corresponding to if the data come from the reference or query sets.

[31]:
adata_full.obs.batch.cat.rename_categories(["Query", "Reference"], inplace=True)
[32]:
full_predictions = vae_q.predict(adata_full)
print("Acc: {}".format(np.mean(full_predictions == adata_full.obs.cell_type)))

# adata_full.obs["predictions"] = ["Known"] * adata_ref.n_obs + list(adata_query.obs["predictions"].to_numpy())
adata_full.obs["predictions"] = full_predictions
INFO     Input adata not setup with scvi. attempting to transfer anndata setup
INFO     Using data from adata.X
INFO     Registered keys:['X', 'batch_indices', 'labels']
INFO     Successfully registered anndata object containing 15681 cells, 1000 vars, 5 batches,
         8 labels, and 0 proteins. Also registered 0 extra categorical covariates and 0 extra
         continuous covariates.
/usr/local/lib/python3.7/dist-packages/scvi/model/base/_base_model.py:151: UserWarning: Make sure the registered X field in anndata contains unnormalized count data.
  "Make sure the registered X field in anndata contains unnormalized count data."
Acc: 0.88278808749442
[33]:
sc.pp.neighbors(adata_full, use_rep="X_scANVI")
sc.tl.leiden(adata_full)
sc.tl.umap(adata_full)
[34]:
sc.pl.umap(
    adata_full,
    color=["study", "cell_type"],
    frameon=False,
    ncols=1,
)
... storing 'study' as categorical
... storing 'predictions' as categorical
../../_images/tutorials_notebooks_scarches_scvi_tools_53_1.png
[35]:
ax = sc.pl.umap(
    adata_full,
    frameon=False,
    show=False,
)
sc.pl.umap(
    adata_full[:adata_query.n_obs],
    color=["predictions"],
    frameon=False,
    title="Query predictions",
    ax=ax,
    alpha=0.7
)

ax = sc.pl.umap(
    adata_full,
    frameon=False,
    show=False,
)
sc.pl.umap(
    adata_full[:adata_query.n_obs],
    color=["cell_type"],
    frameon=False,
    title="Query observed cell types",
    ax=ax,
    alpha=0.7
)
Trying to set attribute `.uns` of view, copying.
../../_images/tutorials_notebooks_scarches_scvi_tools_54_1.png
../../_images/tutorials_notebooks_scarches_scvi_tools_54_2.png

Online update of TOTALVI

This workflow works very similarly for TOTALVI. Here we demonstrate how to build a CITE-seq reference and use scRNA-seq only data as the query.

Assemble data

For totalVI, we will treat two CITE-seq PBMC datasets from 10X Genomics as the reference. These datasets were already filtered for outliers like doublets, as described in the totalVI manuscript. There are 14 proteins in the reference.

[36]:
adata_ref = scvi.data.pbmcs_10x_cite_seq(run_setup_anndata=False)
INFO     Downloading file at data/pbmc_10k_protein_v3.h5ad
Downloading...: 24938it [00:00, 89917.06it/s]
INFO     Downloading file at data/pbmc_5k_protein_v3.h5ad
Downloading...: 100%|██████████| 18295/18295.0 [00:00<00:00, 80924.84it/s]
Observation names are not unique. To make them unique, call `.obs_names_make_unique`.

In general, there will be some necessary data wrangling. For example, we need to provide totalVI with some protein data – and when it’s all zeros, totalVI identifies that the protein data is missing in this “batch”.

It could have also been the case that only some of the protein data was missing, in which case we would add zeros for each of the missing proteins.

[37]:
adata_query = scvi.data.dataset_10x("pbmc_10k_v3")
adata_query.obs["batch"] = "PBMC 10k (RNA only)"
# put matrix of zeros for protein expression (considered missing)
pro_exp = adata_ref.obsm["protein_expression"]
data = np.zeros((adata_query.n_obs, pro_exp.shape[1]))
adata_query.obsm["protein_expression"] = pd.DataFrame(columns=pro_exp.columns, index=adata_query.obs_names, data = data)
INFO     Downloading file at data/10X/pbmc_10k_v3/filtered_feature_bc_matrix.h5
Downloading...: 37492it [00:01, 28105.76it/s]
Variable names are not unique. To make them unique, call `.var_names_make_unique`.
Variable names are not unique. To make them unique, call `.var_names_make_unique`.

We do some light QC filtering on the query dataset (doublets, mitochondrial, etc.)

[38]:
import scrublet as scr
scrub = scr.Scrublet(adata_query.X)
doublet_scores, predicted_doublets = scrub.scrub_doublets()
adata_query = adata_query[~predicted_doublets].copy()

adata_query.var['mt'] = adata_query.var_names.str.startswith('MT-')  # annotate the group of mitochondrial genes as 'mt'
sc.pp.calculate_qc_metrics(adata_query, qc_vars=['mt'], percent_top=None, log1p=False, inplace=True)
adata_query = adata_query[adata_query.obs.pct_counts_mt < 15, :].copy()
Preprocessing...
Simulating doublets...
Embedding transcriptomes using PCA...
Calculating doublet scores...
Automatically set threshold at doublet score = 0.33
Detected doublet rate = 4.7%
Estimated detectable doublet fraction = 55.3%
Overall doublet rate:
        Expected   = 10.0%
        Estimated  = 8.6%
Elapsed time: 19.2 seconds

Now to concatenate the objects, which intersects the genes properly.

[39]:
adata_full = anndata.concat([adata_ref, adata_query])
Observation names are not unique. To make them unique, call `.obs_names_make_unique`.

And split them back up into reference and query (but now genes are properly aligned between objects).

[40]:
adata_ref = adata_full[np.logical_or(adata_full.obs.batch == "PBMC5k", adata_full.obs.batch == "PBMC10k")].copy()
adata_query = adata_full[adata_full.obs.batch == "PBMC 10k (RNA only)"].copy()

Observation names are not unique. To make them unique, call `.obs_names_make_unique`.

We run gene selection on the reference, because that’s all that will be avaialble to us at first.

[41]:
sc.pp.highly_variable_genes(
    adata_ref,
    n_top_genes=4000,
    flavor="seurat_v3",
    batch_key="batch",
    subset=True,
)
Observation names are not unique. To make them unique, call `.obs_names_make_unique`.
Observation names are not unique. To make them unique, call `.obs_names_make_unique`.

Finally, we use these selected genes for the query dataset as well.

[42]:
adata_query = adata_query[:, adata_ref.var_names].copy()

Train reference

[43]:
scvi.model.TOTALVI.setup_anndata(
    adata_ref,
    batch_key="batch",
    protein_expression_obsm_key="protein_expression"
)
INFO     Using batches from adata.obs["batch"]
INFO     No label_key inputted, assuming all cells have same label
INFO     Using data from adata.X
INFO     Using protein expression from adata.obsm['protein_expression']
INFO     Using protein names from columns of adata.obsm['protein_expression']
INFO     Successfully registered anndata object containing 10849 cells, 4000 vars, 2 batches,
         1 labels, and 14 proteins. Also registered 0 extra categorical covariates and 0
         extra continuous covariates.
INFO     Please do not further modify adata until model is trained.
[44]:
arches_params = dict(
    use_layer_norm="both",
    use_batch_norm="none",
)
vae_ref = scvi.model.TOTALVI(
    adata_ref,
    **arches_params
)
[45]:
vae_ref.train()
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
Epoch 337/400:  84%|████████▍ | 337/400 [05:38<01:03,  1.01s/it, loss=1.24e+03, v_num=1]Epoch   337: reducing learning rate of group 0 to 2.4000e-03.
Epoch 376/400:  94%|█████████▍| 376/400 [06:18<00:23,  1.00it/s, loss=1.22e+03, v_num=1]Epoch   376: reducing learning rate of group 0 to 1.4400e-03.
Epoch 400/400: 100%|██████████| 400/400 [06:42<00:00,  1.01s/it, loss=1.23e+03, v_num=1]
[46]:
adata_ref.obsm["X_totalVI"] = vae_ref.get_latent_representation()
sc.pp.neighbors(adata_ref, use_rep="X_totalVI")
sc.tl.umap(adata_ref, min_dist=0.4)
[47]:
sc.pl.umap(
    adata_ref,
    color=["batch"],
    frameon=False,
    ncols=1,
    title="Reference"
)
... storing 'batch' as categorical
../../_images/tutorials_notebooks_scarches_scvi_tools_77_1.png
[48]:
dir_path = "saved_model/"
vae_ref.save(dir_path, overwrite=True)

Online update with query

[49]:
vae_q = scvi.model.TOTALVI.load_query_data(
    adata_query,
    dir_path,
    freeze_expression=True
)
INFO     .obs[_scvi_labels] not found in target, assuming every cell is same category
INFO     Found batches with missing protein expression
INFO     Using data from adata.X
INFO     Registered keys:['X', 'batch_indices', 'labels', 'protein_expression']
INFO     Successfully registered anndata object containing 10153 cells, 4000 vars, 3 batches,
         1 labels, and 14 proteins. Also registered 0 extra categorical covariates and 0
         extra continuous covariates.
/usr/local/lib/python3.7/dist-packages/scvi/model/base/_archesmixin.py:96: UserWarning: Query integration should be performed using models trained with version >= 0.8
  "Query integration should be performed using models trained with version >= 0.8"
[50]:
vae_q.train(200, plan_kwargs=dict(weight_decay=0.0))
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
Epoch 200/200: 100%|██████████| 200/200 [04:16<00:00,  1.28s/it, loss=749, v_num=1]
[51]:
adata_query.obsm["X_totalVI"] = vae_q.get_latent_representation()
sc.pp.neighbors(adata_query, use_rep="X_totalVI")
sc.tl.umap(adata_query, min_dist=0.4)

Impute protein data for query and visualize

Now that we have updated with the query, we can impute the proteins that were observed in the reference, using the transform_batch parameter.

[52]:
_, imputed_proteins = vae_q.get_normalized_expression(
    adata_query,
    n_samples=25,
    return_mean=True,
    transform_batch=["PBMC10k", "PBMC5k"],
)

Very quickly we can identify the major expected subpopulations of B cells, CD4 T cells, CD8 T cells, monocytes, etc.

[53]:
adata_query.obs = pd.concat([adata_query.obs, imputed_proteins], axis=1)

sc.pl.umap(
    adata_query,
    color=imputed_proteins.columns,
    frameon=False,
    ncols=3,
)
... storing 'batch' as categorical
../../_images/tutorials_notebooks_scarches_scvi_tools_87_1.png

Visualize reference and query

[54]:
adata_full_new = adata_query.concatenate(adata_ref, batch_key="none")
Observation names are not unique. To make them unique, call `.obs_names_make_unique`.
Observation names are not unique. To make them unique, call `.obs_names_make_unique`.
Observation names are not unique. To make them unique, call `.obs_names_make_unique`.
[55]:
adata_full_new.obsm["X_totalVI"] = vae_q.get_latent_representation(adata_full_new)
sc.pp.neighbors(adata_full_new, use_rep="X_totalVI")
sc.tl.umap(adata_full_new, min_dist=0.3)
INFO     Input adata not setup with scvi. attempting to transfer anndata setup
INFO     Found batches with missing protein expression
INFO     Using data from adata.X
INFO     Registered keys:['X', 'batch_indices', 'labels', 'protein_expression']
INFO     Successfully registered anndata object containing 21002 cells, 4000 vars, 3 batches,
         1 labels, and 14 proteins. Also registered 0 extra categorical covariates and 0
         extra continuous covariates.
[56]:
_, imputed_proteins_all = vae_q.get_normalized_expression(
    adata_full_new,
    n_samples=25,
    return_mean=True,
    transform_batch=["PBMC10k", "PBMC5k"],
)

for i, p in enumerate(imputed_proteins_all.columns):
    adata_full_new.obs[p] = imputed_proteins_all[p].to_numpy().copy()
[57]:
perm_inds = np.random.permutation(np.arange(adata_full_new.n_obs))
sc.pl.umap(
    adata_full_new[perm_inds],
    color=["batch"],
    frameon=False,
    ncols=1,
    title="Reference and query"
)
/usr/local/lib/python3.7/dist-packages/anndata/_core/anndata.py:1229: ImplicitModificationWarning: Initializing view as actual.
  "Initializing view as actual.", ImplicitModificationWarning
Trying to set attribute `.obs` of view, copying.
Observation names are not unique. To make them unique, call `.obs_names_make_unique`.
Observation names are not unique. To make them unique, call `.obs_names_make_unique`.
... storing 'batch' as categorical
../../_images/tutorials_notebooks_scarches_scvi_tools_92_1.png
[58]:
ax = sc.pl.umap(
    adata_full_new,
    color="batch",
    groups=["PBMC 10k (RNA only)"],
    frameon=False,
    ncols=1,
    title="Reference and query",
    alpha=0.4
)
... storing 'batch' as categorical
../../_images/tutorials_notebooks_scarches_scvi_tools_93_1.png
[59]:
sc.pl.umap(
    adata_full_new,
    color=imputed_proteins_all.columns,
    frameon=False,
    ncols=3,
    vmax="p99"
)
../../_images/tutorials_notebooks_scarches_scvi_tools_94_0.png