Step 9: scRNA Integration — Spatial Transcriptomics Tutorial

任務分類

一、整合的三種任務

🏷️

Label transfer

把 scRNA 的 cell-type 標籤投射到 spot（per-spot top-1 / weighted）。Seurat FindTransferAnchors 是經典方案。

Project scRNA cell-type labels to spots (per-spot top-1 / weighted). Seurat FindTransferAnchors is the classic approach.

🗺️

Cell mapping

Tangram：用 deep learning 學一個機率矩陣，每顆 sc 細胞對應到某 voxel/spot 的機率。同時可做 gene imputation。

Tangram: deep learning estimates a probability matrix mapping each sc cell to a voxel/spot. Also imputes missing genes.

🎯

High-res placement

CytoSPACE：把 sc cell 一對一映射到 ST spot，達到單細胞解析的空間圖；對 noise tolerance 強。

CytoSPACE: one-to-one assignment of sc cells to ST spots — yields a single-cell spatial map; strong noise tolerance.

方法清單

二、主流工具

工具	任務	特色
Seurat anchors	label transfer	與 scRNA-seq 工作流無縫銜接、社群最大	Seamless with scRNA workflow, largest community
Tangram	cell mapping + imputation	PyTorch、可同時擴充基因覆蓋；MERFISH/STARmap/Visium 通用	PyTorch, also imputes missing genes; works on MERFISH/STARmap/Visium
CytoSPACE	1-to-1 placement	2023 Nat Biotech、優於前期方法；對 noise robust	2023 Nat Biotech; outperforms prior methods; noise-robust
SpaGE	gene imputation	針對 image-based panel 補齊未被測到的基因	Imputes off-panel genes for image-based data
STEM	transfer learning	2024 Comm Bio；用 deep transfer learning 同時做 mapping + 標籤遷移	2024 Comm Bio; deep transfer learning for mapping + label transfer

工作流程

三、整合的標準工作流程

清乾淨 sc reference

用對的組織、用對的分群粒度。太細的 cluster 可能跟 ST 解析度不匹配。

Use the right tissue and the right granularity. Too-fine clusters may not match ST resolution.

挑選共用基因

Tangram 建議用每個 cell-type 各 ~100 個 marker（共 ~1k 基因）做 training set。

Tangram recommends ~100 markers × cell-type (~1 k genes) for training.

訓練 / 對位

Tangram constraint mode（每個 spot 細胞數估值帶入），有助於 Visium。CytoSPACE 直接用 ILP optimizer。

For Tangram, "constrained" mode (with per-spot cell count) helps on Visium. CytoSPACE uses an ILP optimizer.

驗證

把 imputed marker 跟 ground-truth marker 比較；以及 spatial domain 對應細胞類型是否符合解剖學。

Compare imputed markers with ground-truth; check whether spatial domains correspond to expected cell types per anatomy.

程式碼

實作

# Seurat label transfer
anchors <- FindTransferAnchors(reference = sc, query = vis, normalization.method = "SCT")
preds <- TransferData(anchorset = anchors, refdata = sc$celltype,
                       prediction.assay = TRUE, weight.reduction = vis[["pca"]])
vis[["predictions"]] <- preds
SpatialFeaturePlot(vis, features = c("T_cell", "B_cell"))

import tangram as tg
# Tangram cell mapping (Visium)
markers = sc.tl.rank_genes_groups_df(adata_sc, group=None).head(1500)["names"].tolist()
tg.pp_adatas(adata_sc, adata_st, genes=markers)
ad_map = tg.map_cells_to_space(adata_sc, adata_st, mode="constrained",
                                density_prior="rna_count_based", num_epochs=500)
ad_proj = tg.project_genes(ad_map, adata_sc)  # 補齊基因

# CytoSPACE
import cytospace
out = cytospace.run(scRNA = sc_counts, ST = st_counts,
                     scLabels = sc_meta["celltype"])  # one-to-one assignment

📝 自我檢測

1. 下列哪個工具最適合「想要單細胞解析的空間圖」？

1. Which tool best fits "I want a single-cell-resolution spatial map"?

A. RCTDA. RCTD

B. cell2locationB. cell2location

C. CytoSPACEC. CytoSPACE

D. Moran's ID. Moran's I

2. Tangram 除了 cell mapping 還能做？

2. Beyond cell mapping, Tangram also performs?

A. 基因 imputation（補齊未被測到的基因）A. Gene imputation (filling in untested genes)

B. 自動 segmentationB. Automatic segmentation

C. 估算 ligand-receptorC. Ligand-receptor inference

D. 移除 batch effectD. Removes batch effects

3. Tangram 訓練基因怎麼挑最有效率？

3. How to pick training genes for Tangram efficiently?

A. 全部基因都丟進去A. Pass in all genes

B. 每個細胞類型各取 top markers (共 ~1k)B. Top markers per cell type (~1k total)

C. 隨機取 100 個C. Random 100 genes

D. 只用 housekeeping 基因D. Only housekeeping genes