Handling Technical and Biological Replicates

[agave82]

Hi, I have an experimental design as follows:

WT_1 WT_2 WT3 WT_4 KO_1 KO_2 KO_3 KO_4

*_1 and *_2 are biological replicates while *_3 and *_4 are technical replicates.

Please, can you explain how to handle technical and biological replicates in DEP?
I understand that I can utilise the option type="all" to make pairwise comparisons:
data_diff_all_contrasts <- test_diff(data_imp, type = "all")
I hav e done that but would that be enough to take into account the replicates or do I need add the design_formula option in the code?
If design_formula needs to be include please, can you explain how to do in my case of experimental design?
Thank you very much

[ojziff]

I have this same query: How do we compare groups of biological replicates?

In the vignette, there is only an example of how to compare groups of technical replicates (Ubi4_vs_Ubi6, Ubi4_vs_Ctrl, Ubi4_vs_Ubi1, Ubi6_vs_Ctrl, Ubi6_vs_Ubi1, Ctrl_vs_Ubi1) but how do you compare all Ubi samples with Ctrl? We could treat biological replicates (Ubi1, Ubi4, Ubi6) in the same way as technical replicates by giving them a replicate number 1 to 9 and the same condition name, but this is not ideal as limma is unaware that each biological replicate is in triplicate.

label	condition	replicate	technical_replicate
Ubi4_1	Ubi	1	1
Ubi4_2	Ubi	2	2
Ubi4_3	Ubi	3	3
Ubi6_1	Ubi	4	1
Ubi6_2	Ubi	5	2
Ubi6_3	Ubi	6	3
Ubi1_1	Ubi	7	1
Ubi1_2	Ubi	8	2
Ubi1_3	Ubi	9	3
Ctrl_1	Ctrl	1	1
Ctrl_2	Ctrl	2	2
Ctrl_3	Ctrl	3	3

I saw in a related issue #11 (comment) that you could treat the technical replicate as a batch effect and adjust for the technical replicate in the design with ~0 + technical_replicate + condition.

Is there any better solutions @arnesmits @adomingues @agave82 ?

Many thanks,
Oliver

[adomingues]

To my knowledge I think you covered most of it @ojziff.

The only other solution I can think of, and I have seen done / suggested, is to average the technical replicates before doing the differential analysis with DEP. It's a rough approach, but it the replicates are well correlated it should be fine.