Error message & multivariate experimental design

[Marc van Oostrum]

Hi

I'm currently analysing an experiment with the following experimental design:

I have a number of combinations of (mouse) organs, cell types and two subcellular fractions (A and B) with A beeing a subcellular fraction and B beeing a control fraction -used as "background" to determine what is enriched in subcellular fraction A.

When I process the experiment in MSstats I define a condition as "organ-celltype-fraction" and as Bioreplicate I define the animal Nr. from which the organs originate. 

Sometimes many organs are collected from one animal, sometimes only one organ from one animal. There are no technical replicates, each sample is only represented by 1 MS run and the subcellular fraction are always paired - except for few conditions where we had to remove an MS run due to technical issues. Hence few "BioReplicate" animal numbers are only represented once. Every condition has 5-10 replicates.

groupComparison gives the following error:

Warning messages:
1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
   Hessian is numerically singular: parameters are not uniquely determined

 

...and I'm not sure what it means, could you eventually help me with this?

I figured that this might occur due to the fact that some "BioReplicates" only occur once - could that be and do I need to remove them?

A second question: is it not possible to define a multivariate experimental design in MSstats?

I think it would be useful for my particular experimental design to define interaction terms in the linear model in order to compare  differences in fractions A over B across different cell types and organs. 

Many thanks for your help!

Best regards

Marc

[Mateusz Staniak]

Hi,

this is just a warning, no need to worry about that.

Currently, it's not possible to modify the model that MSstats uses for hypothesis testing.

Kind regards,

Mateusz

[Marc van Oostrum]

Hi Mateusz

Thanks for your response.

 Regarding the interaction model  - in limma it is described that one can define an interaction model as follows (with their particular example used in the manual): "StrainMu:TreatmentS // (Mu.S-Mu.U)-(WT.S-WT.U) // Interaction"

In my experimental design this would correspond to: ("celltype-organ-X-Fraction-A" - "celltype-organ-X-Fraction-B") - ("celltype-organ-Y-Fraction-A" - "celltype-organ-Y-Fraction-B") with the animal Nr. representing the "BioReplicate".

If I pass such a matrix to contrast.matrix (using 2x "-1" and 2x "1" for the four conditions) the output seems to be what I'm interested in - judging by the resulting log2FC values. 

Is this valid or do you see any reason that speaks against doing this?

Many thanks for your help!

Best

Marc

[vitek...@gmail.com]

This looks good to me