How to adapt the fgseaL function to perform rapidGSEA computation of gene ranks across 9 different phenotype labels?

Question

I wish to adapt the r language function fgseaL, https://github.com/ctlab/fgsea , to perform rapidGSEA, https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-016-1244-x , computation of inter-class deviation per gene and the subsequent gene rank sorting operation on 9 different phenotype labels as illustrated in the diagram immediately below:

I thought of applying the R-language rank() function on the Expression Data Matrix D. If that is not correct, what sequence of R language commands should we apply to the Expression Data Matrix D to calculate a key value sorted deviation measure across 8 labeled human leukemia groups and a healthy labeled normal control group prior to running fgseaL?

I show below how fgseaL finds the correlation matrix between the R language variable , mat, which corresponds to the Expression Data Matrix D and the R language variable , labels , which is a vector of gene phenotype labels

    tmatSc <- scale(t(mat))
    labelsSc <- scale(labels)[, 1]

    minSize <- max(minSize, 1)

    pathwaysFiltered <- lapply(pathways, function(p) { as.vector(na.omit(fmatch(p, rownames(mat)))) })
    pathwaysSizes <- sapply(pathwaysFiltered, length)

    toKeep <- which(minSize <= pathwaysSizes & pathwaysSizes <= maxSize)
    m <- length(toKeep)

    if (m == 0) {
        return(data.table(pathway=character(),
                          pval=numeric(),
                          padj=numeric(),
                          ES=numeric(),
                          NES=numeric(),
                          nMoreExtreme=numeric(),
                          size=integer(),
                          leadingEdge=list()))
    }

    pathwaysFiltered <- pathwaysFiltered[toKeep]
    pathwaysSizes <- pathwaysSizes[toKeep]

    corRanks <- var(tmatSc, labelsSc)[,1]
    ranksOrder <- order(corRanks, decreasing=T)
    ranksOrderInv <- invPerm(ranksOrder)
    stats <- corRanks[ranksOrder]

    pathwaysReordered <- lapply(pathwaysFiltered, function(x) ranksOrderInv[x])

    gseaStatRes <- do.call(rbind,
                           lapply(pathwaysReordered, calcGseaStat,
                                  stats=stats,
                                  returnLeadingEdge=TRUE))

I found a problem with the algorithm shown immediately below.
correcttest <- data.frame(names = row.names(normal))
correcttest <- cbind(correcttest3, normal)
correcttest <- cbind(correcttest3, ALL3m)
rownames(correcttest) <- correcttest$names
correcttest$names <- NULL
correctlabelnormal <- rep(0:0, 73)
correctlabelALL3m <- rep(1:1, 122)
correctlabel <- as.vector(c(correctlabelnormal,correctlabelALL3m))
s <- apply(correcttest, 1, function(x) coef(lm(x~correctlabel))[2])
o <- rank(s)
o <- max(o) - o + 1
res <- fgseaL(df,o,correctlabel,nperm = 2000,minSize = 1, maxSize=50000)
empty data table (0 rows) of 8 columns:   pathway,pval,padj,ES,NES,nMoreExtreme,size

I found the binary phenotype labeled group fgseaL test results below looked satisfactory.    
correcttest <- data.frame(names = row.names(normal))
correcttest <- cbind(correcttest3, normal)
correcttest <- cbind(correcttest3, ALL3m)
rownames(correcttest) <- correcttest$names
correcttest$names <- NULL
correctlabelnormal <- rep(0:0, 73)
correctlabelALL3m <- rep(1:1, 122)
correctlabel <- as.vector(c(correctlabelnormal,correctlabelALL3m))
fgseaL(df,correcttest,correctlabel,nperm = 2000,minSize = 1, maxSize=50000)
       pathway        pval        padj         ES       NES nMoreExtreme  size
1: Gene.Symbol 0.003940887 0.003940887 -0.2460126 -1.180009            3 45714
                                 leadingEdge
1: AKIRIN2,LRRC20,HSPA5,HSPA5,DTWD2,ZFYVE28,

Thank you for your consideration.

Answer 1

You rank the fit coefficient rather than the original score matrix. So, given a score matrix, D:

D = matrix(c(22,20,9,8,46,22,18,10,3,18,3,29,2,1,5,45,43,47,17,5,14,44,21,36), byrow=T, ncol=6)
cl = c(0,0,0,1,1,1)
s = apply(m, 1, function(x) coef(lm(x~cl))[2]) # [1]
o = rank(s)
o = max(o) - o + 1 # [2]

o is then the rank of each row.

[1] This fits each row as a linear model of cl and extracts the cl coefficient.

[2] This converts the ranking to be the same as shown in the figure. I don't know if this is important, but I would assume so given how GSEA methods tend to work.

Answer 2

I finally found another answer to my question. Please read this great article in May 12 2017 BioMed Central (BMC) Bioinformatics article titled Ranking metrics in gene set enrichment analysis: do they matter?.

Also, please read this blog , Diving into Genetics and Genomics: Gene Set Enrichment Analysis (GSEA) explained.

After reading these two articles, my choice for the best rapidGSEA local ranking measure is Minimum Significant Difference (i.e., MSD), because it has the best overall false positive rate (i.e., FPR) for larger sample sizes.

Finally, it is important to realize that fgseaL's phenotype labeling can hypothetically be emulated by GSEA with either one of it's sixteen possible ranking metrics or a custom ranking metric.

Answer 3

This weekend, I  installed the R language packages rapidGSEA and the Broad Institute GSEA , the Nvidia CUDA Toolkit 6.5-- Custom installation option  and MINGW makefiles on my son's Windows 8.1 notebook computer. This step was very tricky to accomplish because I had to maneuver around the "NVIDIA Installer cannot continue. The NVIDIA graphics driver is not compatible with this version of Windows" installation error message which prevents the NVIDIA compilers from being installed correctly. My objective is to run rapidGSEA without the CUDA enhancements since I am not using a computer with a NVIDIA GPU. I did this step because rapidGSEA and Broad Institute GSEA have a significantly richer set of gene set enrichment analytics than fgseaL as shown below:

The GSEA method takes five default arguments and three optional arguments GSEA <- function(exprsData, labelList, geneSets, numPermutations, metricString, dumpFileName="", checkInput=TRUE, doublePrecision=FALSE) {...}

exprsData, labelList and geneSets refer to the data obtained in the previous section. numPermutations denotes the number of permutations in the resampling test, metricString denotes the local ranking measure (one of the following):

naive_diff_of_classes
naive_ratio_of_classes
naive_log2_ratio_of_classes
stable_diff_of_classes
stable_ratio_of_classes
stable_log2_ratio_of_classes
onepass_signal2noise
onepass_t_test
twopass_signal2noise
twopass_t_test
stable_signal2noise
stable_t_test
overkill_signal2noise
overkill_t_test