DESeq2 LRT test for multi-factor designs
1
0
Entering edit mode
mep19 • 0
@mep19-23552
Last seen 2.5 years ago

Hi!

I'm using DESeq2 on an experimental setup that assesses the effect of a treatment (control & heatwave) on 3 different genotypes (with replicates at the genotype level).

I have used Wald tests to assess differences in the treatment effect between pairs of genotypes (i.e., genotypeA vs. genotypeB, genotypeA vs. genotypeC, and genotypeB vs. genotypeC). I would also like to use an LRT to assess differences in the treatment effect between all 3 genotypes simultaneously, but i'm not 100% confident that i am using the right full/reduced models to do this.

One option is to use the full model with the interaction term, and reduced model with no interaction term:

ddsoption1 <- DESeqDataSetFromMatrix(countData = counts, colData = coldata, design= ~genotype + treatment + genotype:treatment)

LRT_test1 <- DESeq(ddsoption1, test="LRT", reduced= ~genotype + treatment)


Or, perhaps I should be specifying a full model with no interaction term and reduced model with just treatment:

ddsoption2 <- DESeqDataSetFromMatrix(countData= counts, colData= coldata, design= ~genotype + treatment)

LRT_test2 <- DESeq(ddsoption2, test="LRT", reduced= ~treatment)


Is anyone able to help clarify which model setup i should be using to address my question?

Thanks in advance for any help, its much appreciated :)

deseq2 • 555 views
2
Entering edit mode
@mikelove
Last seen 22 minutes ago
United States

Think about what you want your null model to be. You already know what the full model is (the one in the first chunk). Should the null be that the genotype effect is the same across all treatments (#1) or that there is no genotype effect (#2). Then based on your choice of null model you should use the first or second chunk but you should use the full model as specified in the first chunk.

0
Entering edit mode

How is the reduced model works lets say when i have only genotype as the factor in my design , i now understand when i have multiple things in my metadata how to implement the reduced model but confusion in case when I have a single condition or factor . The explanation that is written is genes that is reported is the comparison between full model and reduced model. So lets say i have 3 sample groups and one condition then what is being compared to what[I'm not able to explain in statistical vocabulary ]. It is straight forward in wald test where I know which is my reference but in case of LRT it is difficult to get it. Any help or suggestion would be really appreciated

0
Entering edit mode

For more understanding of likelihood ratio testing, I'd recommend consulting a local statistician.