一、為什麼 scRNA QC 不能直接套到 ST？
在 scRNA-seq，「一個 barcode = 一個細胞」是基本假設。在 ST 上這個假設只在 image-based 平台成立。對於 Visium 這類 55 µm spot，一個 spot 可能涵蓋 0、1、5、甚至 10 個細胞。所以：
- nFeature 高不一定是 doublet——可能只是 spot 裡細胞多。
- MT% 高不一定是死細胞——可能該區是肌肉/心臟組織。
- 過濾要先依組織區塊 (tissue region) 做分組，再下手。
另一個 ST 特有的問題：組織覆蓋 (tissue coverage)。capture array 上必有些 spot 落在組織外，這些必須先排除。Space Ranger / Loupe 會給 in_tissue 標籤。
In scRNA-seq, "one barcode = one cell" is foundational. That only holds for image-based ST platforms. For Visium-style 55 µm spots, one spot can contain 0, 1, 5, or even 10 cells. So:
- High nFeature isn't necessarily a doublet — the spot may simply contain more cells.
- High MT% isn't necessarily dead cells — the area may be muscle/cardiac tissue.
- Filter by tissue region, then apply thresholds.
An ST-specific issue: tissue coverage. Many spots on the capture array fall off-tissue and must be excluded first — Space Ranger / Loupe provides the in_tissue flag.
二、ST 的 QC 核心指標
in_tissue
spot 是否落在組織內。Space Ranger 自動偵測；務必檢查 H&E 影像確認沒有把組織誤判成背景。
Whether a spot is inside tissue. Space Ranger detects it automatically; always check the H&E image to confirm tissue isn't mis-classified as background.
nCount / nFeature
UMI 與基因數，跟 scRNA 一樣。但不要用單一閾值——應該分組織區塊看分佈。
UMI and gene counts, like scRNA. But don't use one global threshold — examine distributions per tissue region.
percent.mt
MT 比例。心肌/肝組織天然偏高，過濾前先看空間圖。Visium FFPE 因為用 probe，MT% 通常較低。
Mitochondrial fraction. Naturally elevated in cardiac/liver tissue — view the spatial map before filtering. Probe-based Visium FFPE typically has lower MT%.
percent.hb
血紅素基因比例。富血管組織偏高，是「該區是血液」而非「該 spot 壞」。
Hemoglobin fraction. Elevated in vascular regions — usually means "this region is blood-rich," not "this spot is bad."
局部離群
SpotSweeper 把每個 spot 跟自己的 k-nearest spatial neighbours (Visium 用 k = 36) 比較，標記出「跟周遭格格不入」的 spot。比 global threshold 更精準。
SpotSweeper compares every spot to its k-nearest spatial neighbors (k = 36 for Visium hex grid) and flags "out-of-context" spots — more accurate than global thresholds.
影像品質
H&E 染色不均、組織皺褶、氣泡都會造成局部偏差。務必目視確認影像，並可在 Loupe 圈出可疑區。
Uneven H&E staining, tissue folds, bubbles all create local artifacts. Always inspect the image, and use Loupe to mask suspicious regions if needed.
互動：global vs spatially-aware QC
下方左圖：spot 顏色深淺對應 nCount。拖動「全域閾值」會用單一截斷值；切換到「空間鄰域」則會比照局部鄰居。觀察哪一種會誤殺正常但低表達的組織區。
Left: spot color = nCount. Move the "global threshold" to use a single cutoff; switch to "spatial neighborhood" to compare each spot with its local neighbors. Notice which method mistakenly removes biologically low-expression regions.
灰：被剔除；紫：保留；底色色塊代表組織區塊
三、ST QC 決策流程
🌳 ST QC 流程
QC 範例
library(Seurat); library(SpotSweeper); library(SpatialExperiment) vis <- Load10X_Spatial("visium/") # 基本指標 / Basic metrics vis[["percent.mt"]] <- PercentageFeatureSet(vis, pattern = "^MT-") vis[["percent.hb"]] <- PercentageFeatureSet(vis, pattern = "^HB[AB]") SpatialFeaturePlot(vis, features = c("nCount_Spatial", "percent.mt")) # 全域簡易過濾（先寬鬆） / Lenient global filter vis <- subset(vis, subset = nCount_Spatial > 500 & percent.mt < 25) # SpotSweeper 局部離群（轉成 SpatialExperiment） / Local-outlier QC spe <- as.SingleCellExperiment(vis) |> as.SpatialExperiment() spe <- localOutliers(spe, metric = "sum", direction = "lower", n_neighbors = 36) spe <- spe[, !spe$local_outliers]
import scanpy as sc import squidpy as sq adata = sc.read_visium("visium/") adata.var_names_make_unique() # 標記 mt / hb 基因 / Tag mt & hb genes adata.var["mt"] = adata.var_names.str.startswith("MT-") adata.var["hb"] = adata.var_names.str.contains("^HB[AB]") sc.pp.calculate_qc_metrics(adata, qc_vars=["mt","hb"], inplace=True) # 空間視覺化 / Spatial QC view sq.pl.spatial_scatter(adata, color=["total_counts","pct_counts_mt"]) # 簡易過濾（in_tissue 已自動套用） / Simple filter sc.pp.filter_cells(adata, min_counts=500) adata = adata[adata.obs.pct_counts_mt < 25].copy()
📝 自我檢測
1. Visium 心臟切片整片 MT% 都偏高，最佳處置？
1. A Visium heart section has globally elevated MT%. Best response?
2. SpotSweeper 的核心優勢是？
2. The key advantage of SpotSweeper?
3. 為什麼「nFeature 高」在 Visium 不一定代表 doublet？
3. Why doesn't "high nFeature" in Visium always mean a doublet?