evalRelevance: Evaluating phenotype relevance of gene set rankings

View source: R/benchmark.R

evalRelevanceR Documentation

Evaluating phenotype relevance of gene set rankings


This function evaluates gene set rankings obtained from the application of enrichment methods to multiple datasets - where each dataset investigates a certain phenotype such as a disease. Given pre-defined phenotype relevance scores for the gene sets, indicating how important a gene set is for the investigated phenotype (as e.g. judged by evidence from the literature), this allows to assess whether enrichment methods produce gene set rankings in which phenotype-relevant gene sets accumulate at the top.


  method = "wsum",
  top = 0,
  rel.thresh = 0,

compOpt(rel.ranks, gs.ids, data2pheno = NULL, top = 0)

compRand(rel.ranks, gs.ids, data2pheno = NULL, perm = 1000)



Enrichment analysis rankings. A list with an entry for each enrichment method applied. Each entry is a list that stores for each dataset analyzed the resulting gene set ranking, obtained from applying the respective method to the respective dataset. Resulting gene set rankings are assumed to be of class DataFrame in which gene sets (required column named GENE.SET) are ranked according to a ranking measure such as a gene set p-value (required column named PVAL). See gsRanking for an example.


Relevance score rankings. A list with an entry for each phenotype investigated. Each entry should be a DataFrame in which gene sets (rownames are assumed to be gene set IDs) are ranked according to a phenotype relevance score (required column REL.SCORE).


A named character vector where the names correspond to dataset IDs and the elements of the vector to the corresponding phenotypes investigated.


Character. Determines how the relevance score is summarized across the enrichment analysis ranking. Choose "wsum" (default) to compute a weighted sum of the relevance scores, "auc" to perform a ROC/AUC analysis, or "cor" to compute a correlation. This can also be a user-defined function for customized behaviors. See Details.


Integer. If top is non-zero, the evaluation will be restricted to the first top gene sets of each enrichment analysis ranking. Defaults to 0, which will then evaluate the full ranking. If used with method="auc", it defines the number of gene sets at the top of the relevance ranking that are considered relevant (true positives).


Numeric. Relevance score threshold. Restricts relevance score rankings (argument rel.ranks) to gene sets exceeding the threshold in the REL.SCORE column.


Additional arguments for computation of the relevance measure as defined by the method argument. This includes for method="wsum":

  • perc: Logical. Should observed scores be returned as-is or as a *perc*entage of the respective optimal score. Percentages of the optimal score are typically easier to interpret and are comparable between datasets / phenotypes. Defaults to TRUE.

  • rand: Logical. Should gene set rankings be randomized to assess how likely it is to observe a score equal or greater than the respective obtained score? Defaults to FALSE.


Character vector of gene set IDs on which enrichment analysis has been carried out.


Integer. Number of permutations if rand set to TRUE.


The function evalRelevance evaluates the similarity of a gene set ranking obtained from enrichment analysis and a gene set ranking based on phenotype relevance scores. Therefore, the function first transforms the ranks 'r' from the enrichment analysis to weights 'w' in [0,1] via w = 1 - r / N; where 'N' denotes the total number of gene sets on which the enrichment analysis has been carried out. These weights are then multiplied with the corresponding relevance scores and summed up.

The function compOpt applies evalRelevance to the theoretically optimal case in which the enrichment analysis ranking is identical to the relevance score ranking. The ratio between observed and optimal score is useful for comparing observed scores between datasets / phenotypes.

The function compRand repeatedly applies evalRelevance to random rankings obtained from placing the gene sets randomly along the ranking, thereby assessing how likely it is to observe a score equal or greater than the one obtained.

It is also possible to inspect other measures for summarizing the phenotype relevance, instead of calculating weighted relevance scores sums (argument method="wsum", default). One possibility is to treat the comparison of the EA ranking and the relevance ranking as a classification problem, and to compute standard classification performance measures such as the area under the ROC curve (method="auc"). However, this requires to divide the relevance rankings (argument rel.ranks) into relevant (true positives) and irrelevant (true negatives) gene sets using the top argument. Instead of method="auc", this can also be any other performance measure that the ROCR package (https://rocr.bioinf.mpi-sb.mpg.de) implements. For example, method="tnr" for calculation of the true negative rate. Although such classification performance measures are easy to interpret, the weighted sum has certain preferable properties such as avoiding thresholding and accounting for varying degrees of relevance in the relevance rankings.

It is also possible to compute a standard rank-based correlation measure such as Spearman's correlation (method="cor") to compare the similarity of the enrichment analysis rankings and the relevance rankings. However, this might not be optimal for a comparison of an EA ranking going over the full gene set vector against the typically much smaller vector of gene sets for which a relevance score is annotated. For this scenario, using rank correlation reduces the question to "does a subset of the EA ranking preserve the order of the relevance ranking"; although our question of interest is rather "is a subset of the relevant gene sets ranked highly in the EA ranking".


A numeric matrix (rows = datasets, columns = methods) storing in each cell the chosen relevance measure (score, AUC, cor) obtained from applying the respective enrichment method to the respective expression dataset.


Ludwig Geistlinger <Ludwig.Geistlinger@sph.cuny.edu>

See Also

runEA to apply enrichment methods to multiple datasets; readResults to read saved rankings as an input for the eval-functions;


    # (1) simulated setup: 1 enrichment method applied to 1 dataset

    # simulate gene set ranking
    ea.ranks <- EnrichmentBrowser::makeExampleData("ea.res")
    ea.ranks <- EnrichmentBrowser::gsRanking(ea.ranks, signif.only=FALSE)

    # simulated relevance score ranking
    rel.ranks <- ea.ranks
    rel.ranks[,2] <- runif(nrow(ea.ranks), min=1, max=100)
    colnames(rel.ranks)[2] <- "REL.SCORE"
    rownames(rel.ranks) <- rel.ranks[,"GENE.SET"]
    ind <- order(rel.ranks[,"REL.SCORE"], decreasing=TRUE)
    rel.ranks <- rel.ranks[ind,]

    # evaluate
    evalRelevance(ea.ranks, rel.ranks)    
    compOpt(rel.ranks, ea.ranks[,"GENE.SET"])
    compRand(rel.ranks, ea.ranks[,"GENE.SET"], perm=3)

    # (2) simulated setup: 2 methods & 2 datasets
    methods <- paste0("m", 1:2)
    data.ids <- paste0("d", 1:2)

    # simulate gene set rankings
    ea.ranks <- sapply(methods, function(m) 
                    r <- EnrichmentBrowser::makeExampleData("ea.res") 
                    r <- EnrichmentBrowser::gsRanking(r, signif.only=FALSE)
                }, simplify=FALSE),

    # simulate a mapping from datasets to disease codes
    d2d <- c("ALZ", "BRCA")
    names(d2d) <- data.ids

    # simulate relevance score rankings
    rel.ranks <- lapply(ea.ranks[[1]],
            rr[,2] <- runif(nrow(rr), min=1, max=100)
            colnames(rr)[2] <- "REL.SCORE"
            rownames(rr) <- rr[,"GENE.SET"]
            ind <- order(rr[,"REL.SCORE"], decreasing=TRUE)
            rr <- rr[ind,]
    names(rel.ranks) <- unname(d2d)

    # evaluate
    evalRelevance(ea.ranks, rel.ranks, d2d)

waldronlab/GSEABenchmarkeR documentation built on Nov. 3, 2023, 10:59 p.m.