scReGAT 0.0.13

A GAT-based computational framework to predict long-range gene regulatory relationships

scReGAT

A GAT-based computational framework to predict long-range gene regulatory relationships.

Installation

install pytorch

Refer to: https://pytorch.org/get-started/previous-versions/

Example code:

conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.7 -c pytorch -c nvidia

install pyG

Refer to: https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html

Example code:

pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.0.0+cu117.html

install scReGAT

pip install scregat

Usage

1. Prepare scReGAT inputs

From Seuratbject

• Suppose SeuratObj_ATAC and SeuratObj_RNA are your SeuratObjects of scATAC-seq and scRNA-seq.
• Note that SeuratObj_ATAC must have two metadata columns named "celltype" and "celltype_rna", and SeuratObj_RNA must have a metadata column named "celltype".

library(Seurat)
library(SeuratDisk)
# save ATAC.h5ad
scregat_ATAC <- CreateSeuratObject(SeuratObj_ATAC@assays$ATAC@counts,meta.data = [e-mail])
scregat_ATAC@assays$[e-mail]$name <- rownames(scregat_ATAC@assays$[e-mail])
SaveH5Seurat(scregat_ATAC, filename = 'ATAC.h5Seurat',overwrite = T)
Convert(ATAC.h5Seurat, dest = "h5ad",overwrite = T)
# save RNA.h5ad
scregat_RNA <- CreateSeuratObject(SeuratObj_RNA@assays$RNA@counts,meta.data = [e-mail])
scregat_RNA@assays$[e-mail]$name = rownames(scregat_RNA@assays$[e-mail])
SaveH5Seurat(scregat_RNA, filename = 'RNA.h5Seurat',overwrite = T)
Convert(RNA.h5Seurat, dest = "h5ad",overwrite = T)

From SingleCellExperiment

• Suppose sce is your SingleCellExperiment Object of scATAC-seq.
• Note that sce must have two metadata columns named "celltype" and "celltype_rna".

library(Seurat)
library(Signac)
library(SingleCellExperiment)
counts <- assay(sce)
rownames(counts) <- GRangesToString(rowRanges(sce))
scregat_ATAC <- CreateSeuratObject(counts,meta.data = colData(sce))
scregat_ATAC@assays$[e-mail]$name <- rownames(scregat_ATAC@assays$[e-mail])
SaveH5Seurat(scregat_ATAC, filename = 'ATAC.h5Seurat',overwrite = T)
Convert(ATAC.h5Seurat, dest = "h5ad",overwrite = T)

Randomly select a subset of cells

set.seed(123)
scregat_ATAC <- subset(scregat_ATAC,cells = sample(colnames(scregat_ATAC),20000))

2. Run scReGAT

Import packages

import os
import pandas as pd
import torch
import anndata as ad
from torch_geometric.loader import DataLoader
import scanpy as sc
from scregat.data_process import prepare_model_input,sum_counts,plot_edge
from scregat.model import train_scregat, explain_model_ig

Set file path

SAMPLE = 'SF11857'
path_demo = '/home/txm/txmdata/scREGION/run/NatCan_GBM/' + SAMPLE + '/run_scReGAT/'
dir_scReGAT = 'scReGAT'
atac_file = 'ATAC.h5ad'

Read in scRNA-seq data

RNA_h5ad_file = "/home/txm/txmdata/scREGION/run/NatCan_GBM/" + SAMPLE + "/run_scReGAT/RNA.h5ad"
adata_rna = sc.read_h5ad(RNA_h5ad_file)
adata_rna.obs['celltype'] = adata_rna.obs['celltype'].astype('object')
df_rna = sum_counts(adata_rna,by = 'celltype',marker_gene_num=200)

Prepare model input

dataset_atac, dataset_graph = prepare_model_input(
    path_data_root = path_demo, 
    file_atac = atac_file, 
    df_rna_celltype = df_rna,
    path_eqtl = '/home/txm/txmdata/scREGION/scReGAT/v0.0.0/data/all_tissue_SNP_Gene.txt',
    hg19tohg38 = False, min_percent = 0.01)

Train the gene expression prediction model

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model_scReGAT, test_dataset = \
    train_scregat(path_demo, dataset_atac, dir_scReGAT, device,
                  hidwidth=16, numhead=8, num_epoch=20, learning_rate=1e-3,
                  split_prop = 0.6,print_process = True)

Predict regulatory intensity scores (RISs) by attribution method

• Load well-trained model

path_model = os.path.join(path_demo, dir_scReGAT)
file_model = os.path.join(path_model,'Model_batch_size_16_hidwidth_16_numhead_8_lr_0.001_numepoch_20_split_0.6.pt')
model_scReGAT = torch.load(file_model)

• Write predicted RIS matrix to "file_weight"

file_weight = os.path.join(path_model, "mat_RIS.tsv")
df_mat_RIS = explain_model_ig(
    dataset_atac, model_scReGAT, device, test_dataset, file_weight,print_process = True)

Visualize the cells profiled by the RIS matrix

plot_edge(df_mat_RIS, dataset_atac)