Modeling and analyzing single-cell multimodal data with deep parametric inference

Hu H, Feng Z, Lin H, et al., Modeling and analyzing single-cell multimodal data with deep parametric inference, Brief Bioinform, 2023 Jan 19;24(1):bbad005. The proliferation of single-cell multimodal sequencing technologies has enabled us to understand cellular heterogeneity with multiple views, providing novel and actionable biological insights into the disease-driving mechanisms. Here, we propose a comprehensive end-to-end single-cell multimodal data analysis framework named Deep Parametric Inference (DPI). The python packages, datasets and user-friendly manuals of DPI are freely available at https://github.com/studentiz/dpi.

The dpi framework works with scanpy and supports the following single-cell multimodal analyses

• Multimodal data integration
• Multimodal data noise reduction
• Cell clustering and visualization
• Reference and query cell types
• Cell state vector field visualization

Pip install

pip install dpi-sc

Datasets

The dataset participating in "Single-cell multimodal modeling with deep parametric inference" can be downloaded at DPI data warehouse

Tutorial

We use Peripheral Blood Mononuclear Cell (PBMC) dataset to demonstrate the process of DPI analysis of single cell multimodal data. The following code is recommended to run on a computer with more than 64G memory.

Import dependencies

import scanpy as sc
import dpi

Retina image output (optional)

%matplotlib inline
%config InlineBackend.figure_format = 'retina'

Load dataset

# The dataset can be downloaded from [Datasets] above.
sc_data = sc.read_h5ad("PBMC_COVID19_Healthy_Annotated.h5ad")

Set marker collection

rna_markers = ["CCR7", "CD19", "CD3E", "CD4"]
protein_markers = ["AB_CCR7", "AB_CD19", "AB_CD3", "AB_CD4"]

Preprocessing

dpi.preprocessing(sc_data)
dpi.normalize(sc_data, protein_expression_obsm_key="protein_expression")
sc_data.var_names_make_unique()
sc.pp.highly_variable_genes(
    sc_data,
    n_top_genes=3000,
    flavor="seurat_v3",
    subset=False
)
dpi.add_genes(sc_data, rna_markers)
sc_data = sc_data[:,sc_data.var["highly_variable"]]
dpi.scale(sc_data)

Prepare and run DPI model

Configure DPI model parameters

dpi.build_mix_model(sc_data, net_dim_rna_list=[512, 128], net_dim_pro_list=[128], net_dim_rna_mean=128, net_dim_pro_mean=128, net_dim_mix=128, lr=0.0001)

Run DPI model

dpi.fit(sc_data, batch_size=256)

Visualize the loss

dpi.loss_plot(sc_data)

Save DPI model (optional)

dpi.saveobj2file(sc_data, "COVID19PBMC_healthy.dpi")
#sc_data = dpi.loadobj("COVID19PBMC_healthy.dpi")

Visualize the latent space

Extract latent spaces

dpi.get_spaces(sc_data)

Visualize the spaces

dpi.space_plot(sc_data, "mm_parameter_space", color="green", kde=True, bins=30)
dpi.space_plot(sc_data, "rna_latent_space", color="orange", kde=True, bins=30)
dpi.space_plot(sc_data, "pro_latent_space", color="blue", kde=True, bins=30)

Preparation for downstream analysis

Extract features

dpi.get_features(sc_data)

Get denoised datas

dpi.get_denoised_rna(sc_data)
dpi.get_denoised_pro(sc_data)

Cell clustering and visualization

Cell clustering

sc.pp.neighbors(sc_data, use_rep="mix_features")
dpi.umap_run(sc_data, min_dist=0.4)
sc.tl.leiden(sc_data)

Cell cluster visualization

sc.pl.umap(sc_data, color="leiden")

Observe multimodal data markers

RNA markers

dpi.umap_plot(sc_data, featuretype="rna", color=rna_markers, ncols=2)
dpi.umap_plot(sc_data, featuretype="rna", color=rna_markers, ncols=2, layer="rna_denoised")

Protein markers

dpi.umap_plot(sc_data, featuretype="protein", color=protein_markers, ncols=2)
dpi.umap_plot(sc_data, featuretype="protein", color=protein_markers, ncols=2, layer="pro_denoised")

Reference and query

Reference objects need to be pre-set with cell labels.

sc.pl.umap(sc_data, color="initial_clustering", frameon=False, title="PBMC COVID19 Healthy labels")

Demonstrate reference and query capabilities with unannotated asymptomatic COVID-19 PBMCs.

# The dataset can be downloaded from [Datasets] above.
filepath = "COVID19_Asymptomatic.h5ad"
sc_data_COVID19_Asymptomatic = sc.read_h5ad(filepath)

Unannotated data also needs to be normalized.

dpi.normalize(sc_data_COVID19_Asymptomatic, protein_expression_obsm_key="protein_expression")

Referenced and queried objects require alignment features.

sc_data_COVID19_Asymptomatic = sc_data_COVID19_Asymptomatic[:,sc_data.var.index]

Run the automated annotation function.

dpi.annotate(sc_data, ref_labelname="initial_clustering", sc_data_COVID19_Asymptomatic)

Visualize the annotated object.

sc.pl.umap(sc_data_COVID19_Asymptomatic, color="labels", frameon=False, title="PBMC COVID19 Asymptomatic Annotated")

Cell state vector field

Simulate and visualize the cellular state when the CCR7 protein is amplified 2-fold.

dpi.cell_state_vector_field(sc_data, feature="AB_CCR7", amplitude=2, obs="initial_clustering", featuretype="protein")