TencentAILabHealthcare / ACTER

ACTER: the Antigen-oriented Classification of T cells by joint Embedding of the T cell Receptor sequence and the gene expression matrix

Configurationt

Server

• CPU(s): 64
• Main Memory: 251GiB

Package installation

conda create -n VDJ-clustering
conda activate VDJ-clustering

conda install scanpy mnnpy scrublet leidenalg

Run

import os

import matplotlib.pyplot as plt
import pandas as pd
import scanpy as sc
import VDJclustering as vdjc

# Merge gex and vdj from multiple samples
#   1st parameter: sample configuration tsvfile. The first 4 columns are mandatory: sample, gex_data, gex_data_type and vdj_data. The following columns are user optional and they are used to specify sample categories.
#   2nd paramter: output folder. It must be empty.
sample_config = "input.tsv"
out_path = "./"
adata = vdjc.read_gex_vdj(sample_config, out_path)
# We could also start with previous results
# adata = sc.read_h5ad(os.path.join(out_path, "merged_gex_vdj.h5ad"))

# Show cell numbers in different categories
vdjc.pp.show_sample_category(adata, "input_short.tsv")

# If you want to select tumor samples, then
adata = adata[adata.obs["tissue"] == "tumor"]
adata.obs

# Quality check
adata = vdjc.pp.gex_calculate_qc_metrics(adata, "mouse")

# filter out cells that have less than 200 genes expressed
# filter out genes that are detected in less than 3 cells
# filter out cells with percent mito >= 0.2
# filter out cells with percent ribo <= 0.05
adata = vdjc.pp.gex_filter(adata, min_genes=200, min_cells=3, 20, 5)

# normalize to depth 10 000
# logaritmize
adata = vdjc.pp.gex_norm_logtrans(adata, counts_per_cell_after=1e4)

# Detect highly variable genes
adata = vdjc.pp.highly_variable_genes_all(adata)
adata = vdjc.pp.highly_variable_genes_sample(adata)
# User could specify the threshold of selecting highly variable genes
# For example, select highly variable genes that are common in at least half of the samples
threshold = len(set(adata.obs['sample']))/2
var_select = adata.var.highly_variable_nbatches >= threshold
var_genes = var_select.index[var_select]
print(f"Detect {len(var_genes)} highly variable genes")

# Integrate GEX (count) and VDJ (Acthley factor)
adata_gex_vdj = vdjc.pp.generate_gex_vdj(adata, var_genes)

# Bach correct using MNN
corr_data = vdjc.pp.batch_correct_mnn(adata_gex_vdj, sample_config)

# Z-score transformation: give each feature (GEX and VDJ) a similar weight 
corr_data_scaled = vdjc.pp.zscore_transform(corr_data)

# Calculate PCA and neighborhood graph
corr_data_scaled = vdjc.tl.pca_neighbors(corr_data_scaled)

# Plot UMAP
vdj.pl.umap(corr_data_scaled, color="sample", title="MNN Corrected umap")

# Clustering
vdj.tl.leiden(corr_data_scaled, resolution=0.4, key_added="leiden_0.4")

# Filter cells in clusters
# In order to speed up the process, you could skip the alignment step and recomputation of Atchley Factors in function vdjc.tl.centers_corrs. Therefore, you could use the other function vdjc.tl.centers_corrs_noalign.
cen, cor = vdjc.tl.centers_corrs(corr_data_scaled, "leiden_0.4")
corr_data_scaled_filt = vdjc.tl.filter(corr_data_scaled, key="leiden_0.4", thr=0.9, centers=cen, corrs=cor)

# Filter clusters
cen, cor = vdjc.tl.centers_corrs(corr_data_scaled_filt, "leiden_0.4")
corr_data_scaled_filt = vdjc.tl.filter(corr_data_scaled_filt, key="leiden_0.4", thr=0.9, centers=cen, corrs=cor)