Note

# CITE-seq analysis with totalVI¶

With totalVI, we can produce a joint latent representation of cells, denoised data for both protein and RNA, integrate datasets, and compute differential expression of RNA and protein. Here we demonstrate this functionality with an integrated analysis of PBMC10k and PBMC5k, datasets of peripheral blood mononuclear cells publicly available from 10X Genomics subset to the 14 shared proteins between them. The same pipeline would generally be used to analyze a single CITE-seq dataset.

If you use totalVI, please consider citing:

• Gayoso, A., Steier, Z., Lopez, R., Regier, J., Nazor, K. L., Streets, A., & Yosef, N. (2021). Joint probabilistic modeling of single-cell multi-omic data with totalVI. Nature Methods, 18(3), 272-282.

[1]:

import sys

#if branch is stable, will install via pypi, else will install from source
branch = "stable"

if IN_COLAB and branch == "stable":
!pip install --quiet scvi-tools[tutorials]
elif IN_COLAB and branch != "stable":
!pip install --quiet git+https://github.com/yoseflab/scvi-tools@\$branch#egg=scvi-tools[tutorials]


[2]:

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

import scvi
import scanpy as sc

sc.set_figure_params(figsize=(4, 4))

Global seed set to 0


This dataset was filtered as described in the totalVI manuscript (low quality cells, doublets, lowly expressed genes, etc.)

[3]:

adata = scvi.data.pbmcs_10x_cite_seq(run_setup_anndata=False)

INFO     Downloading file at data/pbmc_10k_protein_v3.h5ad

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

[4]:

sc.pp.highly_variable_genes(
n_top_genes=4000,
flavor="seurat_v3",
batch_key="batch",
subset=True,
layer="counts"
)

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.

[7]:

scvi.model.TOTALVI.setup_anndata(
protein_expression_obsm_key="protein_expression",
layer="counts",
batch_key="batch"
)

INFO     Using batches from adata.obs["batch"]
INFO     No label_key inputted, assuming all cells have same label
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.


## Prepare and run model¶

[8]:

vae = scvi.model.TOTALVI(adata, latent_distribution="normal")

[9]:

vae.train()

GPU available: True, used: True
TPU available: False, using: 0 TPU cores
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1,2]

Epoch 321/400:  80%|█████████████████████████████████████████████████████████████████                | 321/400 [05:16<01:17,  1.02it/s, loss=1.22e+03, v_num=1]Epoch   321: reducing learning rate of group 0 to 2.4000e-03.
Epoch 363/400:  91%|█████████████████████████████████████████████████████████████████████████▌       | 363/400 [05:57<00:37,  1.02s/it, loss=1.22e+03, v_num=1]Epoch   363: reducing learning rate of group 0 to 1.4400e-03.
Epoch 377/400:  94%|████████████████████████████████████████████████████████████████████████████▎    | 377/400 [06:12<00:22,  1.01it/s, loss=1.21e+03, v_num=1]

[10]:

plt.plot(vae.history["elbo_train"], label="train")
plt.plot(vae.history["elbo_validation"], label="validation")
plt.title("Negative ELBO over training epochs")
plt.ylim(1200, 1400)
plt.legend()

[10]:

<matplotlib.legend.Legend at 0x7f7254ffdac0>


## Analyze outputs¶

We use Scanpy for clustering and visualization after running totalVI. It’s also possible to save totalVI outputs for an R-based workflow. First, we store the totalVI outputs in the appropriate slots in AnnData.

[11]:

adata.obsm["X_totalVI"] = vae.get_latent_representation()

rna, protein = vae.get_normalized_expression(
n_samples=25,
return_mean=True,
transform_batch=["PBMC10k", "PBMC5k"]
)

n_samples=25,
return_mean=True,
transform_batch=["PBMC10k", "PBMC5k"]
)
parsed_protein_names = [p.split("_")[0] for p in adata.obsm["protein_expression"].columns]


Now we can compute clusters and visualize the latent space.

[12]:

sc.pp.neighbors(adata, use_rep="X_totalVI")

[13]:

sc.pl.umap(
color=["leiden_totalVI", "batch"],
frameon=False,
ncols=1,
)

/data/yosef2/users/valehvpa/miniconda3/envs/scvi-tools-dev/lib/python3.8/site-packages/anndata/_core/anndata.py:1220: FutureWarning: The inplace parameter in pandas.Categorical.reorder_categories is deprecated and will be removed in a future version. Removing unused categories will always return a new Categorical object.
c.reorder_categories(natsorted(c.categories), inplace=True)
... storing 'batch' as categorical


To visualize protein values on the umap, we make a temporary protein adata object. We have to copy over the umap from the original adata object.

[14]:

pro_adata = sc.AnnData(adata.obsm["protein_expression"].copy(), obs=adata.obs)
# Keep log normalized data in raw
# these are cleaner protein names -- "_TotalSeqB" removed


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

[15]:

names = adata.obsm["protein_foreground_prob"].columns
for p in names:


### Visualize denoised protein values¶

[16]:

sc.pl.umap(
gene_symbols="protein_names",
ncols=3,
vmax="p99",
use_raw=False,
frameon=False,
wspace=0.1
)


### Visualize probability of foreground¶

Here we visualize the probability of foreground for each protein and cell (projected on UMAP). Some proteins are easier to disentangle than others. Some proteins end up being “all background”. For example, CD15 does not appear to be captured well, when looking at the denoised values above we see little localization in the monocytes.

Note

While the foreground probability could theoretically be used to identify cell populations, we recommend using the denoised protein expression, which accounts for the foreground/background probability, but preserves the dynamic range of the protein measurements. Consequently, the denoised values are on the same scale as the raw data and it may be desirable to take a transformation like log or square root.

By viewing the foreground probability, we can get a feel for the types of cells in our dataset. For example, it’s very easy to see a population of monocytes based on the CD14 foregroud probability.

[17]:

sc.pl.umap(
color=["{}_fore_prob".format(p) for p in parsed_protein_names],
ncols=3,
color_map="cividis",
frameon=False,
wspace=0.1
)


## Differential expression¶

Here we do a one-vs-all DE test, where each cluster is tested against all cells not in that cluster. The results for each of the one-vs-all tests is concatenated into one DataFrame object. Inividual tests can be sliced using the “comparison” column. Genes and proteins are included in the same DataFrame.

Important

We do not recommend using totalVI denoised values in other differential expression tools, as denoised values are a summary of a random quantity. The totalVI DE test takes into account the full uncertainty of the denoised quantities.

[18]:

de_df = vae.differential_expression(
groupby="leiden_totalVI",
delta=0.5,
batch_correction=True
)

DE...: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 18/18 [01:17<00:00,  4.31s/it]

[18]:

proba_de proba_not_de bayes_factor scale1 scale2 pseudocounts delta lfc_mean lfc_median lfc_std ... raw_mean1 raw_mean2 non_zeros_proportion1 non_zeros_proportion2 raw_normalized_mean1 raw_normalized_mean2 is_de_fdr_0.05 comparison group1 group2
MARC1 0.9988 0.0012 6.724225 0.000125 0.000002 0.0 0.5 8.587596 8.736490 2.600155 ... 0.325505 0.004393 0.250515 0.004156 1.266389 0.013068 True 0 vs Rest 0 Rest
S100A12 0.9972 0.0028 5.875328 0.003628 0.000051 0.0 0.5 9.018895 9.217430 2.523524 ... 9.915121 0.147708 0.966213 0.047257 37.111893 0.446665 True 0 vs Rest 0 Rest
S100A8 0.9972 0.0028 5.875328 0.026917 0.000376 0.0 0.5 8.532126 8.781158 2.287990 ... 73.262054 1.203990 0.999588 0.115531 284.169373 3.517996 True 0 vs Rest 0 Rest
S100A9 0.9968 0.0032 5.741396 0.049121 0.000853 0.0 0.5 8.341504 8.646438 2.241871 ... 132.937378 2.829969 1.000000 0.178817 486.117737 7.608763 True 0 vs Rest 0 Rest
CLEC4D 0.9958 0.0042 5.468460 0.000113 0.000002 0.0 0.5 8.162704 8.291159 2.463758 ... 0.270705 0.003918 0.223321 0.003443 1.026796 0.012392 True 0 vs Rest 0 Rest

5 rows × 22 columns

Now we filter the results such that we retain features above a certain Bayes factor (which here is on the natural log scale) and genes with greater than 10% non-zero entries in the cluster of interest.

[19]:

filtered_pro = {}
filtered_rna = {}
for i, c in enumerate(cats):
cid = "{} vs Rest".format(c)
cell_type_df = de_df.loc[de_df.comparison == cid]
cell_type_df = cell_type_df.sort_values("lfc_median", ascending=False)

cell_type_df = cell_type_df[cell_type_df.lfc_median > 0]

pro_rows = cell_type_df.index.str.contains('TotalSeqB')
data_pro = cell_type_df.iloc[pro_rows]
data_pro = data_pro[data_pro["bayes_factor"] > 0.7]

data_rna = cell_type_df.iloc[~pro_rows]
data_rna = data_rna[data_rna["bayes_factor"] > 3]
data_rna = data_rna[data_rna["non_zeros_proportion1"] > 0.1]

filtered_pro[c] = data_pro.index.tolist()[:3]
filtered_rna[c] = data_rna.index.tolist()[:2]


We can also use general scanpy visualization functions

[20]:

sc.tl.dendrogram(adata, groupby="leiden_totalVI", use_rep="X_totalVI")

[21]:

sc.pl.dotplot(
filtered_rna,
groupby="leiden_totalVI",
dendrogram=True,
standard_scale="var",
swap_axes=True
)


Matrix plot displays totalVI denoised protein expression per leiden cluster.

[22]:

sc.pl.matrixplot(
groupby="leiden_totalVI",
gene_symbols="protein_names",
dendrogram=True,
swap_axes=True,
use_raw=False, # use totalVI denoised
cmap="Greens",
standard_scale="var"
)


This is a selection of some of the markers that turned up in the RNA DE test.

[23]:

sc.pl.umap(
color=[
"leiden_totalVI",
"IGHD",
"FCER1A",
"SCT",
"GZMH",
"NOG",
"FOXP3",
"CD8B",
"C1QA",
"SIGLEC1",
"XCL2",
"GZMK",
],
legend_loc="on data",
frameon=False,
ncols=3,
layer="denoised_rna",
wspace=0.1
)