interaction models DESeq2
1
0
Entering edit mode
aec ▴ 70
@aec-9409
Last seen 15 months ago

Dear all,

I have some doubts about the differences of the following interaction models:

~ind+grp+cnd+grp:cnd
~grp+grp:ind+grp:cnd
~grp+cnd+grp:ind+grp:cnd


After reading the DESeq2 manual and forums I still did not understand the differences.

My experiment design is:

    GROUP   CONDITION   ind.n
s1  CONTROL no_treat    p1
s2  CONTROL no_treat    p2
s3  CONTROL no_treat    p3
s4  CONTROL treat   p1
s5  CONTROL treat   p2
s6  CONTROL treat   p3
s7  KO  no_treat    p2
s8  KO  no_treat    p3
s9  KO  no_treat    p1
s10 KO  treat   p2
s11 KO  treat   p3
s12 KO  treat   p1


And I would like to know the differential response of the treatment between groups accounting for the patient effect.

deseq2 interaction model • 192 views
0
Entering edit mode
@mikelove
Last seen 1 day ago
United States

If the material in ?results and the vignette sections on interactions are not sufficient, I'd recommend discussing analysis plans with a statistician. I can only provide software support, but not statistical and experimental design support for DESeq2 users.

0
Entering edit mode

@Michael, I followed the example described in the DESeq2 documentation Group-specific condition effects, individuals nested within groups with this model ~grp+grp:ind+grp:cnd and obtained a bunch of DE genes for the interaction I am interested, the differential response to treatment between groups. However, applying the limma-voom duplicateCorr (patient as random effect) and the combined grp_cnd factor as is described in 9.7 section of the manual which should be an equivalent alternative I do not get any significant genes for the contrast (grp2cnd2-grp2cnd1)-(grp1cnd2-grp1cnd1), but the top DE genes with DESeq2 are also at the top positions of limma. Is the different calculation of the Pval affecting the results? Is limma more conservative than DESeq2?

0
Entering edit mode

“Is the different calculation of the Pval affecting the results?”

These are different methods which model the data differently. I would stick with one method because when you try multiple on your data now you are in a position to choose which method having seen the results of both.