[R] pairwise difference operator

[Marc Schwartz]

On Fri, 2004-07-30 at 18:30, Adaikalavan Ramasamy wrote:
> There was a BioConductor thread today where the poster wanted to find
> pairwise difference between columns of a matrix. I suggested the slow
> solution below, hoping that someone might suggest a faster and/or more
> elegant solution, but no other response.
> 
> I tried unsuccessfully with the apply() family. Searching the mailing
> list was not very fruitful either. The closest I got to was a cryptic
> chunk of code in pairwise.table().
> 
> Since I do use something similar myself occasionally, I am hoping
> someone from the R-help list can suggest alternatives or past threads.
> Thank you.
> 
> ### Code ###
> pairwise.difference <- function(m){
>   npairs  <- choose( ncol(m), 2 )
>   results <- matrix( NA, nc=npairs, nr=nrow(m) )
>   cnames  <- rep(NA, npairs)
>   if(is.null(colnames(m))) colnames(m) <- paste("col", 1:ncol(m), sep="")
>   
>   k <- 1
>   for(i in 1:ncol(m)){
>     for(j in 1:ncol(m)){
>       if(j <= i) next;
>       results[ ,k] <- m[ ,i] - m[ ,j]
>       cnames[k]    <- paste(colnames(m)[ c(i, j) ], collapse=".vs.")
>       k <- k + 1
>     }
>   }
>   
>   colnames(results) <- cnames
>   rownames(results) <- rownames(m)
>   return(results)
> }
> 
> ### Example using a matrix with 5 gene/row and 4 columns ###
> mat <- matrix( sample(1:20), nc=4 )
> colnames(mat) <- LETTERS[1:4]; rownames(mat) <- paste( "g", 1:5, sep="") 
> mat
>     A  B  C  D
> g1 10 16  3 15
> g2 18  5 12 19
> g3  7  4  8 13
> g4 14  2  6 11
> g5 17  1 20  9
> 
> pairwise.difference(mat)
>    A.vs.B A.vs.C A.vs.D B.vs.C B.vs.D C.vs.D
> g1     -6      7     -5     13      1    -12
> g2     13      6     -1     -7    -14     -7
> g3      3     -1     -6     -4     -9     -5
> g4     12      8      3     -4     -9     -5
> g5     16     -3      8    -19     -8     11


How about this:

I am taking advantage of the combinations() function in the 'gregmisc'
package to define the pairwise column combinations based upon the input
matrix colnames. Given that, perhaps Greg might want to add this
function to the package if it holds up to scrutiny. Additional error
checking would be required as I note below.

pairwise.diffs <- function(x)
{
  if(is.null(colnames(x)))
    colnames(x) <- 1:ncol(x)

  col.diffs <- combinations(ncol(x), 2, colnames(x))
  result <- x[, col.diffs[, 1]] - x[, col.diffs[, 2]]
  colnames(result) <- paste(col.diffs[, 1], ".vs.", col.diffs[, 2], 
                            sep = "")
  result
}

What I am essentially doing is creating the matrix 'col.diffs' to hold
the combinations of the colnames in matrix 'x'. If 'x' does not have
colnames, I set them to the column indices. Then in line 2, I do the
pairwise subtractions. Line 3 simply sets up the colnames in the result
as the combinations.

Note that the subtractions, as you have above, are the first column
minus the second column in the pairwise combinations. You would also
want to check for an input matrix of <3 columns, since the 'result' in
that case would be a vector, rather than a matrix. In that case, you
could add code to coerce 'result' to a matrix, or simply not allow
matrices with <3 columns.

So, using your example matrix above (different seed value):

> mat <- matrix(sample(1:20), nc=4)
> colnames(mat) <- LETTERS[1:4]
> rownames(mat) <- paste( "g", 1:5, sep="") 
> mat
    A  B  C  D
g1  1 17 13 10
g2 12  5  7 16
g3  2 19  6 14
g4 20  4 11  8
g5  3 15 18  9

> pairwise.diffs(mat)
   A.vs.B A.vs.C A.vs.D B.vs.C B.vs.D C.vs.D
g1    -16    -12     -9      4      7      3
g2      7      5     -4     -2    -11     -9
g3    -17     -4    -12     13      5     -8
g4     16      9     12     -7     -4      3
g5    -12    -15     -6     -3      6      9


HTH,

Marc Schwartz