Hi Christian.

Thanks for this.  That's great!

Can you send me a copy of your PNG file to my email?

I will add this to:
http://groups.google.com/group/aroma-affymetrix/web/redundancy-tests

Cheers,
Mark


>
> Dear Mark,
>
> I have now created the following function, which people can use to
> compute firma scores:
>
> # - - - - - - - - - - - - - - - - - - -
> firma <- function(m, method="rlm") {
> ## function to compute FIRMA scores (adapted from aroma.affymetrix)
> ## m:      dataframe containing normalized probe intensities (linear
> scale)
> ##         with 1.column containing the transcript_cluster_id
> ##         and 2. column containing the corresponding probeset_ids
> ## method: fitting model, one of "rlm" or "mdp"
>
>    ## convert expression levels to log2
>    y <- as.matrix(log2(m[,3:ncol(m)]));
>
>    ## dimensions
>    K <- nrow(y);   # number of probes
>    I <- ncol(y);   # number of arrays
>
>    ## fit log-additive model
>    if (method == "rlm") {
>       fit <- .Call("R_rlm_rma_default_model", y, 0, 1.345,
> PACKAGE="preprocessCore");
>    } else if (method == "mdp") {
>       mp <- medpolish(y, trace.iter=FALSE);
>       fit <- list(Estimates = c(mp$overall + mp$col,  mp$row),
>                   StdErrors = rep(0, length(c(mp$row, mp$col))));
>    } else {
>       stop("method must be <rlm, mdp>");
>    }#if
>
>    ## extract parameters
>    est <- fit$Estimates;
>    se  <- fit$StdErrors;
>
>    ## chip effects
>    beta <- est[1:I];
>
>    ## probe affinities
>    if (K == 1) {
>       ## if only one probe must have affinity=1 since sum constraint
>       alpha <- 0;
>    } else {
>       ## affinities sum to zero (on log scale)
>       alpha <- est[(I+1):length(est)];
>       alpha[length(alpha)] <- -sum(alpha[1:(length(alpha)-1)]);
>    }#if
>
>     ## estimates on the intensity scale
>     theta <- 2^beta;
>     phi   <- 2^alpha;
>
>    ## calculate residuals
>    phi   <- matrix(phi,   nrow=K, ncol=I, byrow=FALSE);
>    theta <- matrix(theta, nrow=K, ncol=I, byrow=TRUE);
>    yhat  <- phi *theta;
>    eps   <- (2^y)/yhat;  # rma uses y/yhat
>
>    ## estimate of standard error
>    u.mad <- mad(unlist(log2(eps)), center=0);
>
>    ## get probeset_ids
>    id <- unique(m[,2]);
>
>    ## firma scores
>    fs <- sapply(id,
> function(x){apply(log2(eps[which(m[,2]==x),,drop=F])/u.mad, 2,
> median)});
>    fs <- t(fs);
>    rownames(fs) <- id;
>
>    return(fs);
> }#firma
> # - - - - - - - - - - - - - - - - - - -
>
> Now I can apply this function to the normalized intensities of gene
> UNR:
>
>  > library(preprocessCore)
>  > score.mdp <- firma(unr, "mdp")
>  > score.rml <- firma(unr, "rlm")
>
> Then I can plot the firma scores for robust fitting ("rlm") compared
> to
> median-polish ("mdp"):
>
>  > png(file="firmaplots6.png", width=540, height=800)
>  > oldpar <- par(pty="m", mfcol=c(3, 2), mar=c(5,5,4,2))
>  > for (i in 1:6) {
>  >    plot(score.rml[,i], type="l", ylim=c(-3,5),
> main=colnames(score.rml)[i], xlab="Probeset", ylab="Score")
>  >    abline(h=0, lty=3)
>  >    lines(score.mdp[,i], lty=2)
>  > }
>  > par(oldpar)
>  > dev.off()
>
> The attached plots compare "rlm" (solid lines) and "mdp" (dashed
> lines)
> and should reflect the firma scores for heart and muscle shown in
> Figure
> 3B of Purdom et al.
> As you can see the difference between"rlm" and "mdp" is pretty small.
>
> Best regards
> Christian
>
> P.S.: It seems that I cannot attach the png-file.
>
>
> On Mar 18, 8:28 am, cstratowa <[email protected]
> ingelheim.com> wrote:
>> Dear Mark,
>>
>> Probably, there is not much difference, maybe I will check it.
>> I will let you know but it will take some time.
>>
>> Best regards
>> Christian
>>
>> On Mar 17, 11:46 am, Mark Robinson <[email protected]> wrote:
>>
>> > Hi Christian.
>>
>> > > However, in this respect I have also the following question:
>> > > How does using "median polish" compare to using
>> > > "R_rlm_rma_default_model"?
>> > > Are the final scores still of some use if you use medpol?
>>
>> > Short answer is I haven't investigated this too thoroughly.
>>
>> > But, my guess is that it wouldn't be "too" different.  That prediction
>>  
>> > is based on the fact that the chip effects are in the same ballpark,  
>> > as you can see from the Aroma_vs_Affy (Aroma=R_rlm_rma, Affy=medpol)  
>> > plot in the following thread:
>>
>> >http://groups.google.com/group/aroma-affymetrix/browse_thread/thread/...
>>
>> > But, I'd be interested to hear more details if you do look into it
>> more.
>>
>> > Cheers,
>> > Mark
>>
>> > > Best regards
>> > > Christian
>>
>> > > On Mar 16, 9:13 am, Mark Robinson <[email protected]> wrote:
>> > >> Hi Christian.
>>
>> > >>  From what I can tell looking at your code (rather quickly, i must
>> > >> admit), there will be 2 differences between aroma.affymetrix and
>> what
>> > >> you have:
>>
>> > >> 1. We use the 'preprocessCore' codebase for the robust fitting of
>> the
>> > >> linear model (... but maybe you are just using median polish as an
>> > >> illustration).  For example, you might try:
>>
>> > >> library(preprocessCore)
>> > >> f <- .Call("R_rlm_rma_default_model", log2(yTr), 0,
>> > >> 1.345,PACKAGE="preprocessCore")
>> > >> [... and piece together the alpha, beta, etc ...]
>>
>> > >> 2. The "estimate of standard error" is calculated genewise, over
>> > >> residuals from all probes/samples (i.e. u.mad should be a scalar  
>> > >> not a
>> > >> vector).
>>
>> > >> Hope that helps.
>> > >> Mark
>>
>> > >> On 16/03/2009, at 6:32 PM, cstratowa wrote:
>>
>> > >>> Dear all,
>>
>> > >>> After reading the FIRMA paper I would like to understand the
>> > >>> implementation, but this is not easy since the source code is hard
>>  
>> > >>> to
>> > >>> read. Nevertheless, I tried and would like to know if this is  
>> > >>> correct.
>>
>> > >>> According to the page on exon array analysis you do the following:
>>
>> > >>> I, fit a summary of the entire transcript
>> > >>>> plmTr <- ExonRmaPlm(csN, mergeGroups=TRUE)
>> > >>>> fit(plmTr, verbose=verbose)
>>
>> > >>> II, fit the FIRMA model for each exon
>> > >>>> firma <- FirmaModel(plmTr)
>> > >>>> fit(firma, verbose=verbose)
>>
>> > >>> However, I would like to understand the underlying source code.
>>
>> > >>> For this example let us assume that we have quantile-normalized
>> > >>> intensities yTr for a transcript  containing two exons:
>> > >>>> yTr
>> > >>>     HeartA   HeartB    HeartC  MuscleA  MuscleB  MuscleC
>> > >>> 1   5.74954   18.0296    2.50436   15.5857   26.1744   31.0075
>> > >>> 2   9.59819   23.0093   22.01120   70.1742   32.8408  102.0080
>> > >>> 3 114.50800   87.1742   70.34080  312.3410  266.1740  601.3410
>> > >>> 4  66.34080   52.0075   67.34080  184.1740  266.1740  147.0080
>> > >>> 5 210.17400  142.0080  173.34100  514.5080  659.1740  509.6740
>> > >>> 6 104.00800   84.3408   70.34080  333.5080  324.1740  231.0080
>> > >>> 7 194.00800  124.5080  234.00800  443.6740  767.5080  716.8410
>> > >>> 8 319.34100  282.6740  283.50800  656.0080  807.6740  954.6740
>>
>> > >>> Here rows 1:4 code for exon 1 and rows 5:8 code for exon 2.
>>
>> > >>> I, fit a summary of the entire transcript
>> > >>> To simplify issues I will fit the data using median polish:
>> > >>> # 1. fit median polish
>> > >>>> mp <- medpolish(log2(yTr))
>>
>> > >>> # 2. data set specific estimates (probe affinities)
>> > >>>> beta  <- mp$overall+mp$col
>> > >>>> thetaTr <- 2^beta
>>
>> > >>> # 3. array-specific estimates
>> > >>>> alpha <- mp$row
>> > >>>> alpha[length(alpha)] <- -sum(alpha[1:(length(alpha)-1)])
>> > >>>> phiTr <- 2^alpha
>>
>> > >>> II, fit FIRMA model for each exon
>> > >>> # 1. calculate residuals
>> > >>>> phi   <- matrix(phiTr, nrow=nrow(yTr), ncol=ncol(yTr))
>> > >>>> theta <- matrix(thetaTr, nrow=nrow(yTr), ncol=ncol(yTr),
>> > >>> byrow=TRUE)
>> > >>>> yhat  <- phi *theta
>> > >>>> eps   <- yTr/yhat    # rma uses y/yhat
>>
>> > >>> # 2. estimate of standard error
>> > >>>> u.mad <- apply(log2(eps), 2, mad, center=0)
>>
>> > >>> # 3. compute final score statisitc
>> > >>> # for 1. exon
>> > >>>> y1 <- log2(eps[1:4,])
>> > >>>> F1 <- apply(y1/u.mad, 2, median)
>> > >>>> F1
>> > >>>     HeartA      HeartB      HeartC     MuscleA     MuscleB
>> > >>> MuscleC
>> > >>> -0.89938777 -0.03792624 -0.69409936  0.11536565 -0.61385296
>> > >>> 1.08709568
>>
>> > >>> # for 2. exon
>> > >>>> y2 <- log2(eps[5:8,])
>> > >>>> F2 <- apply(y2/u.mad, 2, median)
>> > >>>> F2
>> > >>>     HeartA      HeartB      HeartC     MuscleA     MuscleB
>> > >>> MuscleC
>> > >>> -0.02899616 -1.64645153 -0.70048533 -0.39996057  0.02666064
>> > >>> -1.46657055
>>
>> > >>> Now my question is:
>> > >>> Is this calculation of the final score statistic F1 for exon 1 and
>>  
>> > >>> F2
>> > >>> for exon 2 correct?
>> > >>> Did I miss something?
>>
>> > >>> Best regards
>> > >>> Christian
>>
>> > >> ------------------------------
>> > >> Mark Robinson
>> > >> Epigenetics Laboratory, Garvan
>> > >> Bioinformatics Division, WEHI
>> > >> e: [email protected]
>> > >> e: [email protected]
>> > >> p: +61 (0)3 9345 2628
>> > >> f: +61 (0)3 9347 0852
>> > >> ------------------------------
>>
>> > ------------------------------
>> > Mark Robinson
>> > Epigenetics Laboratory, Garvan
>> > Bioinformatics Division, WEHI
>> > e: [email protected]
>> > e: [email protected]
>> > p: +61 (0)3 9345 2628
>> > f: +61 (0)3 9347 0852
>> > ------------------------------
> >
>



--~--~---------~--~----~------------~-------~--~----~
When reporting problems on aroma.affymetrix, make sure 1) to run the latest 
version of the package, 2) to report the output of sessionInfo() and 
traceback(), and 3) to post a complete code example.


You received this message because you are subscribed to the Google Groups 
"aroma.affymetrix" group.
To post to this group, send email to [email protected]
To unsubscribe from this group, send email to 
[email protected]
For more options, visit this group at 
http://groups.google.com/group/aroma-affymetrix?hl=en
-~----------~----~----~----~------~----~------~--~---

Reply via email to