Note

This page was generated from MultiVI_tutorial.ipynb. Interactive online version: . Some tutorial content may look better in light mode.

# Joint analysis of paired and unpaired multiomic data with MultiVI#

MultiVI is used for the joint analysis of scRNA and scATAC-seq datasets that were jointly profiled (multiomic / paired) and single-modality datasets (only scRNA or only scATAC). MultiVI uses the paired data as an anchor to align and merge the latent spaces learned from each individual modality.

This tutorial walks through how to read multiomic data, create a joint object with paired and unpaired data, set-up and train a MultiVI model, visualize the resulting latent space, and run differential analyses.

[1]:

!pip install --quiet scvi-colab
from scvi_colab import install
install()


## Data acquisition#

First we download a sample multiome dataset from 10X. We’ll use this throughout this tutorial. Importantly, MultiVI assumes that there are shared features between the datasets. This is trivial for gene expression datasets, which generally use the same set of genes as features. For ATAC-seq peaks, this is less trivial, and often requires preprocessing steps with other tools to get all datasets to use a shared set of peaks. That can be achieved with tools like SnapATAC, ArchR, and CellRanger in the case of 10X data.

Important

MultiVI requires the datasets to use shared features. scATAC-seq datasets need to be processed to use a shared set of peaks.

[ ]:

!wget https://cf.10xgenomics.com/samples/cell-arc/2.0.0/pbmc_unsorted_10k/pbmc_unsorted_10k_filtered_feature_bc_matrix.tar.gz
!sudo tar -xvf pbmc_unsorted_10k_filtered_feature_bc_matrix.tar.gz
!gunzip -f filtered_feature_bc_matrix/*.gz

[2]:

import scvi
import numpy as np
import scanpy as sc

scvi.settings.seed = 420

%config InlineBackend.print_figure_kwargs={'facecolor' : "w"}
%config InlineBackend.figure_format='retina'

Global seed set to 0
Global seed set to 420


## Data Processing#

Next, we’ll read the data into an Anndata object. We’ll then split the data to three datasets, and remove some modality-specific data from parts of the dataset, to demonstrate how MultIVI mixes multimodal and single-modal data. The data has 12012 cells, we’ll use 4004 for each dataset.

Reading the data into an AnnData object can be done with the read_10x_multiome function:

[3]:

# read multiomic data

Variable names are not unique. To make them unique, call .var_names_make_unique.


We can now split the dataset to three datasets, and remove a modality from two of them, so we’ll have equal-sized (4004 cells in each) datasets with multiome, expression-only, and accessibility-only observations:

[4]:

# split to three datasets by modality (RNA, ATAC, Multiome), and corrupt data
# by remove some data to create single-modality data
n = 4004


We can then use the organize_multiome_anndatas function to orgnize these three datasets into a single Multiome dataset. This function sorts and orders the data from the multi-modal and modality-specific AnnDatas into a single AnnData (aligning the features, padding missing modalities with 0s, etc).

[5]:

# We can now use the organizing method from scvi to concatenate these anndata


Note that organize_multiome_anndatas adds an annotation to the cells to indicate which modality they originate from:

[6]:

adata_mvi.obs

[6]:

batch_id modality
CCGCTAAAGGGCCATC_paired 1 paired
CCGCTAAAGTCTTGAA_paired 1 paired
CCGCTAAAGTTAGACC_paired 1 paired
CCGCTAAAGTTCCCAC_paired 1 paired
CCGCTAAAGTTTGCGG_paired 1 paired
... ... ...
TTTGTTGGTACGCGCA_accessibility 1 accessibility
TTTGTTGGTATTTGCC_accessibility 1 accessibility
TTTGTTGGTGATTACG_accessibility 1 accessibility
TTTGTTGGTTTCAGGA_accessibility 1 accessibility
TTTGTTGGTTTCCACG_accessibility 1 accessibility

12012 rows × 2 columns

Important

MultiVI requires the features to be ordered so that genes appear before genomic regions. This must be enforced by the user.

MultiVI requires the features to be ordered, such that genes appear before genomic regions. In this case this is already the case, but it’s always good to verify:

[7]:

adata_mvi = adata_mvi[:, adata_mvi.var["modality"].argsort()].copy()

[7]:

ID modality chr start end
MIR1302-2HG ENSG00000243485 Gene Expression chr1 29553 30267
AL391261.2 ENSG00000258847 Gene Expression chr14 66004522 66004523
FUT8-AS1 ENSG00000276116 Gene Expression chr14 65412689 65412690
FUT8 ENSG00000033170 Gene Expression chr14 65410591 65413008
AL355076.2 ENSG00000258760 Gene Expression chr14 65302679 65318790
... ... ... ... ... ...
chr15:101277030-101277907 chr15:101277030-101277907 Peaks chr15 101277030 101277907
chr15:101257856-101258771 chr15:101257856-101258771 Peaks chr15 101257856 101258771
chr15:101251516-101252373 chr15:101251516-101252373 Peaks chr15 101251516 101252373
chr15:101397608-101398445 chr15:101397608-101398445 Peaks chr15 101397608 101398445
KI270713.1:36924-37836 KI270713.1:36924-37836 Peaks KI270713.1 36924 37836

148458 rows × 5 columns

We also filter features to remove those that appear in fewer than 1% of the cells

[8]:

print(adata_mvi.shape)

(12012, 148458)
(12012, 80878)


## Setup and Training MultiVI#

We can now set up and train the MultiVI model!

First, we need to setup the Anndata object using the setup_anndata function. At this point we specify any batch annotation that the model would account for. Importantly, the main batch annotation, specific by batch_key, should correspond to the modality of the cells.

Other batch annotations (e.g if there are multiple ATAC batches) should be provided using the categorical_covariate_keys.

The actual values of categorical covariates (include batch_key) are not important, as long as they are different for different samples. I.e it is not important to call the expression-only samples “expression”, as long as they are called something different than the multi-modal and accessibility-only samples.

Important

MultiVI requires the main batch annotation to correspond to the modality of the samples. Other batch annotation, such as in the case of multiple RNA-only batches, can be specified using categorical_covariate_keys.

[9]:

scvi.model.MULTIVI.setup_anndata(adata_mvi, batch_key='modality')


When creating the object, we need to specify how many of the features are genes, and how many are genomic regions. This is so MultiVI can determine the exact architecture for each modality.

[10]:

mvi = scvi.model.MULTIVI(
)
mvi.view_anndata_setup()

Anndata setup with scvi-tools version 0.15.0b0.

Setup via MULTIVI.setup_anndata with arguments:

{
│   'layer': None,
│   'batch_key': 'modality',
│   'labels_key': None,
│   'size_factor_key': None,
│   'categorical_covariate_keys': None,
│   'continuous_covariate_keys': None
}

         Summary Statistics
┏━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┓
┃     Summary Stat Key     ┃ Value ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━┩
│         n_cells          │ 12012 │
│          n_vars          │ 80878 │
│         n_batch          │   3   │
│         n_labels         │   1   │
│ n_extra_categorical_covs │   0   │
│ n_extra_continuous_covs  │   0   │
└──────────────────────────┴───────┘

               Data Registry
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Registry Key ┃    scvi-tools Location    ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
└──────────────┴───────────────────────────┘

                     batch State Registry
┏━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┓
┃    Source Location    ┃  Categories   ┃ scvi-tools Encoding ┃
┡━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━┩
│ adata.obs['modality'] │ accessibility │          0          │
│                       │  expression   │          1          │
│                       │    paired     │          2          │
└───────────────────────┴───────────────┴─────────────────────┘

                     labels State Registry
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┓
┃      Source Location      ┃ Categories ┃ scvi-tools Encoding ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━┩
│ adata.obs['_scvi_labels'] │     0      │          0          │
└───────────────────────────┴────────────┴─────────────────────┘

[11]:

mvi.train()

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

Epoch 1/500:   0%|          | 0/500 [00:00<?, ?it/s]

/data/yosef2/users/jhong/miniconda3/envs/v15/lib/python3.9/site-packages/scvi/module/_multivae.py:567: UserWarning: The use of x.T on tensors of dimension other than 2 to reverse their shape is deprecated and it will throw an error in a future release. Consider x.mT to transpose batches of matricesor x.permute(*torch.arange(x.ndim - 1, -1, -1)) to reverse the dimensions of a tensor. (Triggered internally at  /opt/conda/conda-bld/pytorch_1645690191318/work/aten/src/ATen/native/TensorShape.cpp:2402.)
/data/yosef2/users/jhong/miniconda3/envs/v15/lib/python3.9/site-packages/scvi/distributions/_negative_binomial.py:56: UserWarning: Specified kernel cache directory could not be created! This disables kernel caching. Specified directory is /home/eecs/jjhong922/.cache/torch/kernels. This warning will appear only once per process. (Triggered internally at  /opt/conda/conda-bld/pytorch_1645690191318/work/aten/src/ATen/native/cuda/jit_utils.cpp:860.)
+ torch.lgamma(x + theta)

Epoch 244/500:  49%|████▉     | 244/500 [36:12<37:59,  8.90s/it, loss=6.68e+03, v_num=1]
Monitored metric elbo_validation did not improve in the last 45 records. Best score: 14214.882. Signaling Trainer to stop.


## Save and Load MultiVI models#

Saving and loading models is similar to all other scvi-tools models, and is very straight forward:

[12]:

mvi.save("trained_multivi")

[14]:

mvi = scvi.model.MULTIVI.load("trained_multivi", adata=adata_mvi)


## Extracting and visualizing the latent space#

We can now use the get_latent_representation to get the latent space from the trained model, and visualize it using scanpy functions:

[15]:

adata_mvi.obsm["MultiVI_latent"] = mvi.get_latent_representation()

/data/yosef2/users/jhong/miniconda3/envs/v15/lib/python3.9/site-packages/anndata/_core/anndata.py:1228: FutureWarning: The inplace parameter in pandas.Categorical.reorder_categories is deprecated and will be removed in a future version. Reordering categories will always return a new Categorical object.
c.reorder_categories(natsorted(c.categories), inplace=True)
... storing 'modality' as categorical
/data/yosef2/users/jhong/miniconda3/envs/v15/lib/python3.9/site-packages/anndata/_core/anndata.py:1228: FutureWarning: The inplace parameter in pandas.Categorical.reorder_categories is deprecated and will be removed in a future version. Reordering categories will always return a new Categorical object.
c.reorder_categories(natsorted(c.categories), inplace=True)
... storing 'modality' as categorical
/data/yosef2/users/jhong/miniconda3/envs/v15/lib/python3.9/site-packages/anndata/_core/anndata.py:1228: FutureWarning: The inplace parameter in pandas.Categorical.reorder_categories is deprecated and will be removed in a future version. Reordering categories will always return a new Categorical object.
c.reorder_categories(natsorted(c.categories), inplace=True)
... storing 'chr' as categorical


## Impute missing modalities#

In a well-mixed space, MultiVI can seamlessly impute the missing modalities for single-modality cells. First, imputing expression and accessibility is done with get_normalized_expression and get_accessibility_estimates, respectively.

We’ll demonstrate this by imputing gene expression for all cells in the dataset (including those that are ATAC-only cells):

[16]:

imputed_expression = mvi.get_normalized_expression()


We can demonstrate this on some known marker genes:

First, T-cell marker CD3.

[21]:

gene_idx = np.where(adata_mvi.var.index == "CD3G")[0]


Next, NK-Cell marker gene NCAM1 (CD56):

[22]:

gene_idx = np.where(adata_mvi.var.index == "NCAM1")[0]

[23]:

gene_idx = np.where(adata_mvi.var.index == "CD19")[0]