IntegrateLayers after regression out different variables

[phoebee-h]

Hi,

I am using Seurat (from v4 to v5) and have a question regarding IntegrateLayers with FastMNN/Harmony.

Why does the integration result (i.e., the batch correction step in Seurat v5) not change when I provide different PCA inputs?
I am confident that the PCA embeddings differ after regressing out different variables. For instance, when I use different PCA reductions for UMAP and t-SNE, the embeddings clearly show differences and are comparable.

However, during integration, the newly created reductions from FastMNN (e.g., "mnn_rgcount" and "mnn_rgcc") appear to be the same, even though they are based on different PCA inputs. Consequently, the downstream UMAP and t-SNE plots also look identical.

Am I missing a step in the batch correction process after regressing out variables?

obj[["RNA"]] <- split(obj[["RNA"]], f = obj$batch)

## [Regress out nCount] 
## ... Already run NormalizeData/FindVariableFeatures ... 
obj <- ScaleData(obj, vars.to.regress = "nCount_RNA", features = rownames(obj)) 
obj <- RunPCA(obj, features = VariableFeatures(obj),reduction.name="pca_rgcount", nfeatures.print = 10) 
obj <- IntegrateLayers( object = obj, method = FastMNNIntegration, orig.reduction = "pca_rgcount", new.reduction = "mnn_rgcount", verbose = FALSE) 
obj <- obj %>% 
RunTSNE(reduction = "mnn_rgcount", reduction.name = 'tsnergcount', reduction.key = 'tsnergcount_') %>% 
RunUMAP(reduction = "mnn_rgcount", reduction.name = 'umaprgcount', reduction.key = 'umaprgcount_', ) %>% 
FindNeighbors(reduction = "mnn_rgcount", dims = 1:20) %>% 
FindClusters(reduction = "mnn_rgcount", resolution = 0.4, save.SNN = TRUE, cluster.name="cnt_cluster") 


## [Regress out Cell Cycle]
## ... Already run NormalizeData/FindVariableFeatures ... 
obj <- ScaleData(obj, vars.to.regress = c("S.Score", "G2M.Score"), features = rownames(obj)) 
obj <- RunPCA(obj, features = VariableFeatures(obj),reduction.name="pca_rgcc", nfeatures.print = 10) 
obj <- IntegrateLayers( object = obj, method = FastMNNIntegration, orig.reduction = "pca_rgcc", new.reduction = "mnn_rgcc", verbose = FALSE) 
obj <- obj %>% 
RunTSNE(reduction = "mnn_rgcc", reduction.name = 'tsnergcc', reduction.key = 'tsnergcc_') %>% 
RunUMAP(reduction = "mnn_rgcc", reduction.name = 'umaprgcc', reduction.key = 'umaprgcc_', ) %>% 
FindNeighbors(reduction = "mnn_rgcc", dims = 1:20) %>% 
FindClusters(reduction = "mnn_rgcc", resolution = 0.4, save.SNN = TRUE, cluster.name="cc_cluster") 

[igrabski]

Hi, thanks for your question! This is happening because the FastMNNIntegration implementation computes its own PCA, so the reduction passed into it isn't actually what's used. You should see different results with other integration methods when changing the input reduction.