Seurat: Assay to use for differential expression in an integrated data

Question

[firestar]

I have a single-cell RNASeq Seurat object integrated using sctransform.

When running FindMarkers(), which assay is to be used? RNA, SCT or integrated? I assume the slot is always "data".

To look at some real data. Running DGE.

f1 <- FindMarkers(obj,ident.1 = "micro-wt-lps-01",ident.2="micro-wt-lps-05",
                  group.by="cell_type_condition_cluster",slot="data",assay="RNA")
f2 <- FindMarkers(obj,ident.1 = "micro-wt-lps-01",ident.2="micro-wt-lps-05",
                  group.by="cell_type_condition_cluster",slot="data",assay="SCT")
f3 <- FindMarkers(obj,ident.1 = "micro-wt-lps-01",ident.2="micro-wt-lps-05",
                  group.by="cell_type_condition_cluster",slot="data",assay="integrated")

The results and DEGs are all different.

> head(f1)
p_val   avg_logFC pct.1 pct.2    p_val_adj
Mmp12  8.656001e-23 -17.9039715 0.001 0.111 1.417853e-18
Cxcl2  2.153860e-18 -14.1067209 0.002 0.111 3.528023e-14
Xylt1  6.461056e-17  -0.6137400 0.010 0.222 1.058321e-12
Rhebl1 5.547881e-15  -0.3580077 0.018 0.278 9.087429e-11
Lpl    1.582957e-14  57.4116369 0.047 0.444 2.592883e-10
Tgfb2  1.202214e-13  -1.4634266 0.007 0.167 1.969227e-09
> head(f2)
p_val  avg_logFC pct.1 pct.2    p_val_adj
Hp      7.286762e-30 -0.2764460 0.011 0.333 1.193572e-25
S100a11 9.700438e-29 -0.2758875 0.012 0.333 1.588932e-24
Ifitm2  9.700438e-29 -0.2758875 0.012 0.333 1.588932e-24
Ccr2    9.700438e-29 -0.2758875 0.012 0.333 1.588932e-24
Ahnak   9.700438e-29 -0.2758875 0.012 0.333 1.588932e-24
Cybb    1.039943e-27 -0.4562356 0.018 0.389 1.703426e-23
> head(f3)
p_val  avg_logFC pct.1 pct.2    p_val_adj
Gas2l3  3.419395e-11  1.7086466 0.126 0.944 1.025819e-07
Cxcr4   2.159168e-09 -0.3924339 0.022 0.500 6.477503e-06
Kif2c   3.049359e-09 -0.5152160 0.119 0.722 9.148077e-06
Rps18   6.295892e-09 -0.8022487 0.034 0.167 1.888768e-05
mt-Atp6 2.352245e-08 -0.5896836 0.167 0.778 7.056735e-05
Rps8    2.648383e-08 -0.6596474 0.018 0.167 7.945150e-05

I picked two genes which were found in all three results to show how the fold change varies.

> f1[c("Cxcl2","Lpl"),]
             p_val avg_logFC pct.1 pct.2    p_val_adj
Cxcl2 2.153860e-18 -14.10672 0.002 0.111 3.528023e-14
Lpl   1.582957e-14  57.41164 0.047 0.444 2.592883e-10
> f2[c("Cxcl2","Lpl"),]
             p_val  avg_logFC pct.1 pct.2    p_val_adj
Cxcl2 2.060274e-18 -0.3262467 0.002 0.111 3.374729e-14
Lpl   2.822407e-14 -1.3114768 0.038 0.389 4.623102e-10
> f3[c("Cxcl2","Lpl"),]
           p_val  avg_logFC pct.1 pct.2 p_val_adj
Cxcl2 0.00328671 -27.233063 0.050 0.167         1
Lpl   0.44693810  -5.408191 0.117 0.556         1

Answer 1

[rpolicastro]

You shouldn't use SCT and integrated counts for anything outside of dimension reduction and clustering. Before differential expression you can run NormalizeCounts on the RNA assay.

Relevant Seurat links:

• SCtransform and differential expression in v4
• Differential expression testing post-integration

[firestar]

Thanks for the reply and links. By NormalizeCounts, I guess you mean NormalizeData. Also, why normalised data? Shouldn't I be using raw counts with covariates?