Question

DESeq2 design & model matrix not full rank - am I nesting correctly?

0

Entering edit mode

amye107 • 0

@amye107-14845

Last seen 6.3 years ago

Hello,

I have a multi-factor design RNA-seq experiment I am analyzing with DESeq2. I am having trouble with the design due to "model matrix not full rank". I have been trying to work my way through the advice from the vignette and previous posts on this issue...

Here is my colData:

	treatment	condition	tank	tank.nested
1	control	high	A	A
2	control	high	A	A
3	control	high	A	A
4	control	high	A	A
5	control	high	B	B
6	control	high	B	B
7	control	high	B	B
8	control	high	B	B
9	control	low	C	C
10	control	low	C	C
11	control	low	C	C
12	control	low	C	C
13	control	low	D	D
14	control	low	D	D
15	control	low	D	D
16	control	low	D	D
17	infected	high	E	A
18	infected	high	E	A
19	infected	high	E	A
20	infected	high	E	A
21	infected	high	F	B
22	infected	high	F	B
23	infected	high	F	B
24	infected	high	F	B
25	infected	low	G	C
26	infected	low	G	C
27	infected	low	G	C
28	infected	low	G	C
29	infected	low	H	D
30	infected	low	H	D
31	infected	low	H	D
32	infected	low	H	D

Essentially, I am interested in: 1) what are the DE genes between infected and control? and 2) do the DE genes found differ depending on condition?

"Tank" denotes the enclosure the animals were sampled from, and needs to be controlled for in analyses if possible. Trying to follow previous posts, I tried to nest tank within treatment (hence the tank.nested column above). However, I realise that tank A&B are still only within high, C&D only within low etc. so have not solved the issue.

Could anyone suggest how to set up my colData and/or design to resolve the error of model matrix not full rank? And which results() and contrast=list should allow me to get at the comparisons I am trying to make.

I am getting myself very confused - I am waiting on a meeting with a local statistician to help me work through it. But would really appreciated some help in the meantime to try to get my head around it! Many thanks in advance.

deseq2 model matrix experimental design multiple factor design • 1.4k views

ADD COMMENT • link updated 6.3 years ago by Michael Love 41k • written 6.3 years ago by amye107 • 0

score 1 · Answer 1 · 2018-01-23

1

Entering edit mode

Michael Love 41k

@mikelove

Last seen 50 minutes ago

United States

You can't answer #1 and control for tank using fixed effects models, because treatment and tank are confounded. And you can't do a differential analysis in #2 and control for tank using fixed effects models because again, the treatment, condition and treatment:condition are confounded with tank.

Instead, you can use the limma-voom software for RNA-seq, and the duplicateCorrelation() function to inform the model that samples from the same tank are correlated. It's an alternative method for controlling for sample correlations than using fixed effects.