168_单细胞空间deconvolution¶

import cell2location
import scanpy as sc
import numpy as np

# 1. 准备单细胞参考数据
adata_ref = sc.read_h5ad("scRNA_reference.h5ad")
# 需要: adata_ref.obs['cell_type'] - 细胞类型注释

# 2. 计算参考细胞类型的基因表达签名
from cell2location.models import RegressionModel

# 设置参考模型
RegressionModel.setup_anndata(adata_ref, labels_key="cell_type")
mod_ref = RegressionModel(adata_ref)
mod_ref.train(max_epochs=250, use_gpu=True)

# 导出签名
adata_ref = mod_ref.export_posterior(adata_ref)
inf_aver = adata_ref.varm["means_per_cluster_mu_fg"]

# 3. 准备空间数据
adata_vis = sc.read_h5ad("visium_data.h5ad")

# 4. 运行cell2location
from cell2location.models import Cell2location

Cell2location.setup_anndata(adata_vis)
mod = Cell2location(
    adata_vis,
    cell_state_df=inf_aver,
    N_cells_per_location=10,    # 每个spot的预期细胞数
    detection_alpha=20          # 检测灵敏度超参数
)

mod.train(max_epochs=30000, batch_size=None, train_size=1, use_gpu=True)

# 5. 提取结果
adata_vis = mod.export_posterior(adata_vis)

# 每个spot中各细胞类型的绝对丰度
cell_abundances = adata_vis.obsm["q05_cell_abundance_w_sf"]

# 6. 可视化
mod.plot_spatial_QC_across_batches()
cell2location.utils.plot_spatial(
    adata_vis, color=["T_cell", "Macrophage", "Cancer"],
    img_key="hires", spot_size=1.3
)

library(spacexr)

# 1. 准备参考数据
reference <- Reference(
  counts = sc_counts,           # 基因×细胞的counts矩阵
  cell_types = sc_cell_types,   # 细胞类型向量
  nUMI = colSums(sc_counts)     # 每个细胞的UMI数
)

# 2. 准备空间数据
coords <- data.frame(
  x = spatial_coordinates$x,
  y = spatial_coordinates$y,
  row.names = colnames(spatial_counts)
)
puck <- SpatialRNA(coords, spatial_counts, colSums(spatial_counts))

# 3. 运行RCTD
myRCTD <- create.RCTD(puck, reference, max_cores = 8)
myRCTD <- run.RCTD(myRCTD, doublet_mode = "full")
# doublet_mode: "full"=每个spot多细胞类型; "doublet"=最多2种

# 4. 提取结果
results <- myRCTD@results
# 每个spot的细胞类型权重
weights <- as.matrix(results$weights)
# 归一化为比例
proportions <- sweep(weights, 1, rowSums(weights), "/")

# 5. 可视化
library(ggplot2)
plot_data <- cbind(coords, proportions)
ggplot(plot_data, aes(x = x, y = y, color = T_cell)) +
  geom_point(size = 1.5) +
  scale_color_viridis_c() +
  theme_void() +
  labs(title = "T cell abundance (RCTD)")

import tangram as tg
import scanpy as sc

# 1. 准备数据
adata_sc = sc.read_h5ad("scRNA.h5ad")      # 单细胞
adata_sp = sc.read_h5ad("spatial.h5ad")     # 空间

# 2. 找到共有的高变基因
sc.pp.highly_variable_genes(adata_sc, n_top_genes=2000)
markers = adata_sc.var_names[adata_sc.var.highly_variable]

# 3. 预处理
tg.pp_adatas(adata_sc, adata_sp, genes=markers)

# 4. 训练映射模型
adata_map = tg.map_cells_to_space(
    adata_sc, adata_sp,
    mode="cells",           # "cells" or "clusters"
    density_prior="rna_count_based",
    num_epochs=500,
    device="cuda:0"
)

# 5. 投影细胞类型到空间
tg.project_cell_annotations(adata_map, adata_sp, annotation="cell_type")

# 6. 可视化
sc.pl.spatial(adata_sp, color=["T_cell", "Macrophage"], spot_size=1.5)

# === 环境安装 ===
pip install cell2location tangram-sc scvi-tools
# R: install.packages("spacexr")

# === cell2location完整流程 ===
python run_cell2location.py --sc_ref reference.h5ad --spatial visium.h5ad \
  --cell_type_key celltype --n_cells_per_spot 10 --output results/

# === RCTD ===
Rscript run_rctd.R --sc_counts sc_counts.rds --spatial spatial.rds \
  --cell_types celltypes.csv --output rctd_results/