Integrative analysis in Seurat v5

Reference

• Integrative analysis in Seurat v5 • Seurat (satijalab.org)

Introduction

• 单细胞测序数据集的整合，例如跨实验批次、donor或条件的整合，通常是scRNA-seq工作流程中的重要一步。整合分析可以帮助匹配数据集之间的共享细胞类型和状态，这可以提高统计能力，最重要的是，有助于跨数据集进行准确的比较分析。【选择合适整合方法，去除批次效应带来的细胞差异，关注细胞在生物学上的真实分群、差异表达】

• Seurat v5使用IntegrateLayers功能实现了简化的整合分析。目前支持五种方法。这些方法中的每一种都在低维空间中执行集成，并返回降维（即integrated.xxx），该降维旨在跨批次共同嵌入共享细胞类型。

1. Anchor-based CCA integration (method=CCAIntegration)
2. Anchor-based RPCA integration (method=RPCAIntegration)
3. Harmony (method=HarmonyIntegration)
4. FastMNN (method= FastMNNIntegration)
5. scVI (method=scVIIntegration)

选一种方法整合 IntegrateLayers

obj <- IntegrateLayers(
  object = obj, method = CCAIntegration,
  orig.reduction = "pca", new.reduction = "integrated.cca",
  verbose = FALSE
)

obj <- IntegrateLayers(
  object = obj, method = RPCAIntegration,
  orig.reduction = "pca", new.reduction = "integrated.rpca",
  verbose = FALSE
)

obj <- IntegrateLayers(
  object = obj, method = HarmonyIntegration,
  orig.reduction = "pca", new.reduction = "harmony",
  verbose = FALSE
)

obj <- IntegrateLayers(
  object = obj, method = FastMNNIntegration,
  new.reduction = "integrated.mnn",
  verbose = FALSE
)

# For example, scVI integration requires `reticulate` which can be installed from CRAN (`install.packages("reticulate")`) as well as `scvi-tools` and its dependencies installed in a conda environment.
#Please see scVI installation instructions [here](https://docs.scvi-tools.org/en/stable/installation.html).

obj <- IntegrateLayers(
  object = obj, method = scVIIntegration,
  new.reduction = "integrated.scvi",
  conda_env = "../miniconda3/envs/scvi-env", verbose = FALSE
)

选一种方法，可视化+聚类

• FindNeighbors(),FindClusters(),RunUMAP()

## CCA--------------------------------------------------
obj <- FindNeighbors(obj, reduction = "integrated.cca", dims = 1:30)
obj <- FindClusters(obj, resolution = 2, cluster.name = "cca_clusters")
obj <- RunUMAP(obj, reduction = "integrated.cca", dims = 1:30, reduction.name = "umap.cca")
p1 <- DimPlot(
  obj,
  reduction = "umap.cca",
  group.by = c("Method", "predicted.celltype.l2", "cca_clusters"),
  combine = FALSE, label.size = 2
)

## SCVI--------------------------------------------------
obj <- FindNeighbors(obj, reduction = "integrated.scvi", dims = 1:30)
obj <- FindClusters(obj, resolution = 2, cluster.name = "scvi_clusters")
obj <- RunUMAP(obj, reduction = "integrated.scvi", dims = 1:30, reduction.name = "umap.scvi")
p2 <- DimPlot(
  obj,
  reduction = "umap.scvi",
  group.by = c("Method", "predicted.celltype.l2", "scvi_clusters"),
  combine = FALSE, label.size = 2
)

wrap_plots(c(p1, p2), ncol = 2, byrow = F)

两种整合方法 比较

• 在选择方法时，主要考虑聚类中保留的生物学信息。【看不同聚类marker gene 的特异性 ：聚类是不是一个类型聚在一起，而不是一个批次聚在一起（marker gene 在好几类都高表达）】

比较整合结果--Marker Gene

p1 <- VlnPlot(
  obj,
  features = "rna_CD8A", group.by = "unintegrated_clusters"
) + NoLegend() + ggtitle("CD8A - Unintegrated Clusters")
p2 <- VlnPlot(
  obj, "rna_CD8A",
  group.by = "cca_clusters"
) + NoLegend() + ggtitle("CD8A - CCA Clusters")
p3 <- VlnPlot(
  obj, "rna_CD8A",
  group.by = "scvi_clusters"
) + NoLegend() + ggtitle("CD8A - scVI Clusters")
p1 | p2 | p3

maker gene violinplot

• 看看CCA整合后的聚类，在其他整合中的分布 ：

obj <- RunUMAP(obj, reduction = "integrated.rpca", dims = 1:30, reduction.name = "umap.rpca")
p4 <- DimPlot(obj, reduction = "umap.unintegrated", group.by = c("cca_clusters"))
p5 <- DimPlot(obj, reduction = "umap.rpca", group.by = c("cca_clusters"))
p6 <- DimPlot(obj, reduction = "umap.scvi", group.by = c("cca_clusters"))
p4 | p5 | p6

cca_cluster check

将选择的整合后的结果作为新layer进行分析

Seurat v5 assays store data in layers. These layers can store raw, un-normalized counts (layer='counts'), normalized data (layer='data'), or z-scored/variance-stabilized data (layer='scale.data').

obj <- JoinLayers(obj)
obj

## An object of class Seurat 
## 35789 features across 10434 samples within 5 assays 
## Active assay: RNA (33694 features, 2000 variable features)
##  3 layers present: data, counts, scale.data
##  4 other assays present: prediction.score.celltype.l1, prediction.score.celltype.l2, prediction.score.celltype.l3, mnn.reconstructed
##  12 dimensional reductions calculated: integrated_dr, ref.umap, pca, umap.unintegrated, integrated.cca, integrated.rpca, harmony, integrated.mnn, integrated.scvi, umap.cca, umap.scvi, umap.rpca

• SCT normalization +整合 示例：

options(future.globals.maxSize = 3e+09)
obj <- SCTransform(obj)
obj <- RunPCA(obj, npcs = 30, verbose = F)
obj <- IntegrateLayers(
  object = obj,
  method = RPCAIntegration,
  normalization.method = "SCT",
  verbose = F
)
obj <- FindNeighbors(obj, dims = 1:30, reduction = "integrated.dr")
obj <- FindClusters(obj, resolution = 2)