Contrasts For Gene Differentiaion

Question

[Bdh_Dtu]

I am helping a colleague automate some AffyMetrix gene chip analysis in R. This is based on the pipeline described by Gillespie et al 2010 ( http://www.biomedcentral.com/1756-0500/3/81 ).

During the analysis, contrasts will be calculated. The experiment can be comparing mutant-type expression with wild-type expression at each time step (m00-w00, m04-w04, m08-w08, ...) or comparing the timecourse of the single gene type, e.g. only wild-type (w00-w04, w04-w08, w08-w12, ...). In some cases we want both things, i.e. three sets in total: m0x-w0x, m0x-m0y, w0x-w0y.

This needs to be generalised as much as possible so that sensible results are obtained (and graphs outputted) with minimal user input.

My colleague has specifically said that he is interested in all contrasts, for example: w00-w04, w04-w08, and w00-w08. Does this make sense? Is it useful? I am not a biologist - my area is mathematics. Including the full range of contrasts causes trouble for the script.

Intuitively it seems that having e.g. w00-w04, w04-08 and w00-w08 is overkill since the third contrast can be derived from summing the first two: linear dependence. Am I right in believing that contrasts should be orthogonal to avoid redundancy?

Any advice on this? Thanks!

By the way, currently I generate the full set of contrasts as follows:

paste2 <- function(...) paste(...,sep="-")
X <- c("m00", "m04", "m08", "m12", "w00", "w04", "w08", "w12")
cons <- outer(X,X,FUN="paste2")
contrasts <- upperTriangle(cons,FALSE)
#mutant on mutant
MMc <- contrasts[grepl("m\\d*-m\\d*",contrasts)]
#mutant on wild
MWc <- contrasts[grepl("m\\d*-w\\d*",contrasts)]
#wild on wild
WWc <- contrasts[grepl("w\\d*-w\\d*",contrasts)]

To generate the reduced set of contrasts I will use:

genContrasts <- function(cons){
 tp <- unique(sub("(\\D*)\\d*", "\\1", cons))
 mat <- list()
 for (i in tp)
  { mat[[i]] <- cons[grep(i, cons)]
  }
 contrasts <- lapply(mat, function(x)paste(x[-length(x)],
                                           x[-1],
                                           sep="-"))
 if(length(mat)>1)
  { contrasts[["crossed"]] <- apply(do.call(cbind,mat), 1,
                                    function(x){paste(x,
                                                      collapse="-")})
  }
 return(contrasts)
}

# Example:

X <- c("m00", "m04", "m08", "m12", "w00", "w04", "w08", "w12")
genContrasts(X)  # returns a list of three
X <- c("m00", "m04", "m08", "m12")
genContrasts(X)  # returns a list of one

Answer 1

[Csgillespie]

I realise that this answer comes a bit late, but better late than never!

As you mentioned, you are going over board with your contrasts.

1. If you are doing every possible contrast, then you should probably adjust for multiple testing. However, adjusting isn't obvious or trivial in this case since the tests aren't independent.

2. As you mentioned, you really want to try and get orthogonal contrasts. In particular, x vs y, y vs z and x vs z is not really a great idea.

3. You mention that you want to automate this analysis. Why? Are you conducting so many experiments that this is a problem?

4. If you believe that w00 vs m00 is going to be different, then the next obvious contrast to try would be w00-w04 vs m00-m04. This is because you are (probably) interested in a difference at the second time point after you have adjusted for the first time point difference. I presume that your colleague hadn't considered this ;)

5. If you do need to automate the analysis, what about using the timecourse package for your initial test, then if that suggests that interesting things are going on, you could think about what contrasts would be interesting for that particular analysis.