Contents

1 Setup

Although gene set enrichment analysis (GSEA) has become an integral part of high-throughput gene expression data analysis, the assessment of enrichment methods remains rudimentary and ad hoc. In the absence of suitable gold standards, the evaluation is commonly restricted to selected data sets and biological reasoning on the relevance of resulting enriched gene sets. However, this is typically incomplete and biased towards a novel method being presented.

As the evaluation of GSEA methods is thus typically based on self-defined standards, Mitrea et al. (2013) identified the lack of gold standards for consistent assessment and comparison of enrichment methods as a major bottleneck. Furthermore, it is often cumbersome to reproduce existing assessments for additional methods, as this typically involves considerable effort of data processing and method collection.

Leveraging the representative and extendable collection of enrichment methods available in the EnrichmentBrowser package, the GSEABenchmarkeR package facilitates efficient execution of these methods on comprehensive real data compendia. The compendia are curated collections of microarray and RNA-seq datasets investigating human diseases (mostly specific cancer types), for which disease-relevant gene sets have been defined a priori.

Consistently applied to these datasets, enrichment methods can then be subjected to a systematic and reproducible assessment of (i) computational runtime, (ii) statistical significance, especially how the fraction of significant gene sets relates to the fraction of differentially expressed genes, and (iii) phenotype relevance, i.e. whether enrichment methods produce gene set rankings in which phenotype-relevant gene sets accumulate at the top.

In the following, we demonstrate how the package can be used to

We start by loading the package.

library(GSEABenchmarkeR)

2 Expression data sources

The GSEABenchmarkeR package implements a general interface for loading compendia of expression datasets. This includes

In the following, we describe both pre-defined compendia in more detail and also demonstrate how user-defined data can be incorporated.

2.1 Microarray compendium

Although RNA-seq (read count data) has become the de facto standard for transcriptomic profiling, it is important to know that many methods for differential expression and gene set enrichment analysis have been originally developed for microarray data (intensity measurements). However, differences in data distribution assumptions (microarray: quasi-normal, RNA-seq: negative binomial) have made adaptations in differential expression analysis and, to some extent also in gene set enrichment analysis, necessary.

Nevertheless, the comprehensive collection and curation of microarray data in online repositories such as GEO still represent a valuable resource. In particular, Tarca et al. (2012 and 2013) compiled 42 datasets from GEO, each investigating a human disease for which a specific KEGG pathway exists.

These pathways are accordingly defined as the target pathways for the various enrichment methods when applied to the respective datasets. For instance, methods are expected to rank the Alzheimer’s disease pathway close to the top when applied to GSE1297, a case-control study of Alzheimer’s disease.

Furthermore, Tarca et al. made these datasets available in the Bioconductor packages KEGGdzPathwaysGEO and KEGGandMetacoreDzPathwaysGEO.

The GSEABenchmarkeR package simplifies access to the compendium and allows to load it into the workspace via

geo2kegg <- loadEData("geo2kegg")
## Loading GEO2KEGG data compendium ...
names(geo2kegg)
##  [1] "GSE1297"            "GSE14762"           "GSE15471"          
##  [4] "GSE16515"           "GSE18842"           "GSE19188"          
##  [7] "GSE19728"           "GSE20153"           "GSE20291"          
## [10] "GSE21354"           "GSE3467"            "GSE3585"           
## [13] "GSE3678"            "GSE4107"            "GSE5281_EC"        
## [16] "GSE5281_HIP"        "GSE5281_VCX"        "GSE6956AA"         
## [19] "GSE6956C"           "GSE781"             "GSE8671"           
## [22] "GSE8762"            "GSE9348"            "GSE9476"           
## [25] "GSE1145"            "GSE11906"           "GSE14924_CD4"      
## [28] "GSE14924_CD8"       "GSE16759"           "GSE19420"          
## [31] "GSE20164"           "GSE22780"           "GSE23878"          
## [34] "GSE24739_G0"        "GSE24739_G1"        "GSE30153"          
## [37] "GSE32676"           "GSE38666_epithelia" "GSE38666_stroma"   
## [40] "GSE4183"            "GSE42057"           "GSE7305"

A specific dataset of the compendium can be obtained via

geo2kegg[[1]]
## ExpressionSet (storageMode: lockedEnvironment)
## assayData: 22283 features, 16 samples 
##   element names: exprs 
## protocolData: none
## phenoData
##   sampleNames: GSM21215 GSM21217 ... GSM21229 (16 total)
##   varLabels: Sample Group
##   varMetadata: labelDescription
## featureData: none
## experimentData: use 'experimentData(object)'
##   pubMedIds: 14769913 
## Annotation: hgu133a

which returns, in this example, an ExpressionSet (documented in the Biobase package) that contains expression levels of 22,283 probe sets measured for 16 patients.

To prepare the datasets for subsequent analysis, the GSEABenchmarkeR package provides the function maPreproc. The function invokes EnrichmentBrowser::probe2gene on each dataset to summarize expression levels for probes annotated to the same gene. Here, we apply the function to the first 5 datasets of the compendium.

geo2kegg <- maPreproc(geo2kegg[1:5])
## Summarizing probe level expression ...

Now,

geo2kegg[[1]]
## class: SummarizedExperiment 
## dim: 12993 16 
## metadata(5): experimentData annotation protocolData dataId dataType
## assays(1): exprs
## rownames(12993): 780 5982 ... 388796 100505915
## rowData names(0):
## colnames(16): GSM21215 GSM21217 ... GSM21213 GSM21229
## colData names(2): Sample GROUP

returns a SummarizedExperiment that contains the summarized expression values for 12,994 genes. Furthermore, sample groups are defined in the colData column GROUP, yielding here 7 cases (1) and 9 controls (0).

se <- geo2kegg[[1]]
table(se$GROUP)
## 
## 0 1 
## 9 7

Note: The maPreproc returns datasets consistently mapped to NCBI Entrez Gene IDs, which is compatible with most downstream applications. However, mapping to a different ID type such as ENSEMBL IDs or HGNC symbols can also be done using the function EnrichmentBrowser::idMap.

2.2 RNA-seq compendium

The Cancer Genome Atlas (TCGA) project performed a molecular investigation of various cancer types on an unprecedented scale including various genomic high-throughput technologies. In particular, transcriptomic profiling of the investigated cancer types has comprehensively been carried out with RNA-seq in tumor and adjacent normal tissue.

Among the various resources that redistribute TCGA data, Rahman et al. (2015) consistently preprocessed the RNA-seq data for 24 cancer types and made the data available in the GEO dataset GSE62944.

The GSE62944 compendium can be loaded using the loadEData function, which provides the datasets ready for subsequent differential expression and gene set enrichment analysis.

Here, we load the compendium into the workspace using only two of the datasets.

tcga <- loadEData("tcga", nr.datasets=2)
## Loading TCGA data compendium ...
## Cancer types with tumor samples:
## ACC, BLCA, BRCA, CESC, COAD, DLBC, GBM, HNSC, KICH, KIRC, KIRP, LAML, LGG, LIHC, LUAD, LUSC, OV, PRAD, READ, SKCM, STAD, THCA, UCEC, UCS
## Cancer types with adj. normal samples:
## BLCA, BRCA, CESC, CHOL, COAD, ESCA, GBM, HNSC, KICH, KIRC, KIRP, LIHC, LUAD, LUSC, PAAD, PCPG, PRAD, READ, SARC, SKCM, STAD, THCA, THYM, UCEC
## Cancer types with sufficient tumor and adj. normal samples:
## BLCA, BRCA
## Creating a SummarizedExperiment for each of them ...
## BLCA tumor: 414 adj.normal: 19
## BRCA tumor: 1119 adj.normal: 113
names(tcga)
## [1] "BLCA" "BRCA"

For example, the breast cancer dataset contains RNA-seq read counts for roughly 20,000 genes measured in 1,119 tumor (1) and 113 adjacent normal (0) samples.

brca <- tcga[[2]]
brca
## class: SummarizedExperiment 
## dim: 12425 1232 
## metadata(4): annotation annotation dataId dataType
## assays(1): exprs
## rownames(12425): 1 2 ... 23140 26009
## rowData names(0):
## colnames(1232): TCGA-A2-A0T1-01A-21R-A084-07
##   TCGA-A7-A3J0-01A-11R-A213-07 ... TCGA-BH-A0BJ-11A-23R-A089-07
##   TCGA-BH-A0H5-11A-62R-A115-07
## colData names(3): sample type GROUP
table(brca$GROUP)
## 
##    0    1 
##  113 1119

2.3 User-defined data compendium

With easy and fast access to the GEO2KEGG and TCGA compendia, enrichment methods can be directly applied and assessed on well-studied, standardized expression datasets. Nevertheless, benchmarking with the GSEABenchmarkeR package is designed to be extendable to additional datasets as well.

Therefore, the loadEData function also accepts a directory where datasets, preferably of class SummarizedExperiment, have been saved as RDS files.

data.dir <- system.file("extdata", package="GSEABenchmarkeR")
edat.dir <- file.path(data.dir, "myEData")
edat <- loadEData(edat.dir)
names(edat)
## [1] "GSE42057x" "GSE7305x"
edat[[1]]
## class: SummarizedExperiment 
## dim: 50 136 
## metadata(5): experimentData annotation protocolData dataType dataId
## assays(1): exprs
## rownames(50): 3310 7318 ... 123036 117157
## rowData names(0):
## colnames(136): GSM1031553 GSM1031554 ... GSM1031683 GSM1031684
## colData names(2): Sample GROUP

3 Differential expression

To perform differential expression (DE) analysis between sample groups for selected datasets of a compendium, the GSEABenchmarkeR package provides the function runDE.

The function invokes EnrichmentBrowser::deAna on each dataset, which contrasts the sample groups as defined in the GROUP variable.

Here, we apply the function to 5 datasets of the GEO2KEGG compendium.

geo2kegg <- runDE(geo2kegg, de.method="limma", padj.method="flexible")
rowData(geo2kegg[[1]], use.names=TRUE)
## DataFrame with 12993 rows and 4 columns
##                             FC          limma.STAT                 PVAL
##                      <numeric>           <numeric>            <numeric>
## 780          0.432192831595346    2.63744839097559     0.01730300183827
## 5982        0.0350299089030893   0.537756178430442    0.597727503573133
## 3310         0.167336184494154   0.265329510306092    0.793951756244996
## 7849         0.145052817880692    1.87117543916442   0.0786643936871739
## 2978        -0.118651543432762  -0.887354363016077    0.387291389663339
## ...                        ...                 ...                  ...
## 8484         0.366663378857667    4.19469372714013 0.000610626502727196
## 23779       0.0862641275103565    1.22041702308892    0.238994318711422
## 79583        0.064153205958666   0.344811902279699    0.734472195494329
## 388796    -0.00387318363530115 -0.0304937839209508     0.97602900704413
## 100505915 -0.00572267222179455 -0.0856380470050309    0.932756398371501
##                       ADJ.PVAL
##                      <numeric>
## 780           0.01730300183827
## 5982         0.597727503573133
## 3310         0.793951756244996
## 7849        0.0786643936871739
## 2978         0.387291389663339
## ...                        ...
## 8484      0.000610626502727196
## 23779        0.238994318711422
## 79583        0.734472195494329
## 388796        0.97602900704413
## 100505915    0.932756398371501

Note: DE studies typically report a gene as differentially expressed if the corresponding DE p-value, corrected for multiple testing, satisfies the chosen significance level. Enrichment methods that work directly on the list of DE genes are then substantially influenced by the multiple testing correction.

An example is the frequently used over-representation analysis (ORA), which assesses the overlap between the DE genes and a gene set under study based on the hypergeometric distribution (see the vignette of the EnrichmentBrowser package, Appendix A, for an introduction).

ORA is inapplicable if there are few genes satisfying the significance threshold, or if almost all genes are DE.

Using padj.method="flexible" accounts for these cases by applying multiple testing correction in dependence on the observed degree of differential expression:

Note that resulting \(p\)-values are not further used for assessing the statistical significance of DE genes within or between datasets. They are solely used to determine which genes are included in the analysis with ORA - where the flexible correction ensures that the fraction of included genes is roughly in the same order of magnitude across datasets. Alternative strategies could also be applied (such as taking a constant number of genes for each dataset or generally excluding ORA methods from the assessment).

4 Enrichment analysis

In the following, we demonstrate how to carry out enrichment analysis in a benchmark setup. Therefore, we use the collection of human KEGG gene sets as obtained with getGenesets from the EnrichmentBrowser package.

library(EnrichmentBrowser)
kegg.gs <- getGenesets(org="hsa", db="kegg")

At the core of applying a specific enrichment method to a single dataset is the runEA function, which delegates execution of the chosen method to either EnrichmentBrowser::sbea (set-based enrichment analysis) or EnrichmentBrowser::nbea (network-based enrichment analysis). In addition, it returns CPU time used and allows saving results for subsequent assessment.

Here, we carry out ORA on the first dataset of the GEO2KEGG compendium.

kegg.ora.res <- runEA(geo2kegg[[1]], method="ora", gs=kegg.gs, perm=0)
kegg.ora.res
## $runtime
## elapsed 
##    0.55 
## 
## $ranking
## DataFrame with 323 rows and 4 columns
##                                                     GENE.SET  NR.GENES
##                                                  <character> <numeric>
## 1                         hsa00190_Oxidative_phosphorylation       106
## 2                                 hsa05010_Alzheimer_disease       151
## 3                                 hsa05012_Parkinson_disease       115
## 4                                hsa05016_Huntington_disease       162
## 5                                     hsa04714_Thermogenesis       186
## ...                                                      ...       ...
## 319                                     hsa03040_Spliceosome       112
## 320       hsa01521_EGFR_tyrosine_kinase_inhibitor_resistance        77
## 321         hsa04914_Progesterone-mediated_oocyte_maturation        77
## 322                             hsa00232_Caffeine_metabolism         5
## 323 hsa00524_Neomycin,_kanamycin_and_gentamicin_biosynthesis         5
##     NR.SIG.GENES      PVAL
##        <numeric> <numeric>
## 1             55   1.8e-12
## 2             68  2.23e-11
## 3             56  3.12e-11
## 4             70  1.15e-10
## 5             67  1.47e-06
## ...          ...       ...
## 319           12     0.999
## 320            6         1
## 321            5         1
## 322            0         1
## 323            0         1

The function runEA can also be used to carry out several methods on multiple datasets. As an example, we carry out ORA and CAMERA on 5 datasets of the GEO2KEGG compendium saving the results in a temporary directory.

res.dir <- tempdir()
res <- runEA(geo2kegg, methods=c("ora", "camera"), 
                gs=kegg.gs, perm=0, save2file=TRUE, out.dir=res.dir)
res$ora[1:2]
## $GSE1297
## $GSE1297$runtime
## elapsed 
##    0.36 
## 
## $GSE1297$ranking
## DataFrame with 323 rows and 4 columns
##                                                     GENE.SET  NR.GENES
##                                                  <character> <numeric>
## 1                         hsa00190_Oxidative_phosphorylation       106
## 2                                 hsa05010_Alzheimer_disease       151
## 3                                 hsa05012_Parkinson_disease       115
## 4                                hsa05016_Huntington_disease       162
## 5                                     hsa04714_Thermogenesis       186
## ...                                                      ...       ...
## 319                                     hsa03040_Spliceosome       112
## 320       hsa01521_EGFR_tyrosine_kinase_inhibitor_resistance        77
## 321         hsa04914_Progesterone-mediated_oocyte_maturation        77
## 322                             hsa00232_Caffeine_metabolism         5
## 323 hsa00524_Neomycin,_kanamycin_and_gentamicin_biosynthesis         5
##     NR.SIG.GENES      PVAL
##        <numeric> <numeric>
## 1             55   1.8e-12
## 2             68  2.23e-11
## 3             56  3.12e-11
## 4             70  1.15e-10
## 5             67  1.47e-06
## ...          ...       ...
## 319           12     0.999
## 320            6         1
## 321            5         1
## 322            0         1
## 323            0         1
## 
## 
## $GSE14762
## $GSE14762$runtime
## elapsed 
##    0.34 
## 
## $GSE14762$ranking
## DataFrame with 324 rows and 4 columns
##                                                         GENE.SET  NR.GENES
##                                                      <character> <numeric>
## 1                       hsa05150_Staphylococcus_aureus_infection        51
## 2                                             hsa04145_Phagosome       144
## 3                                         hsa05140_Leishmaniasis        70
## 4                                     hsa05416_Viral_myocarditis        56
## 5                        hsa04514_Cell_adhesion_molecules_(CAMs)       138
## ...                                                          ...       ...
## 320              hsa00440_Phosphonate_and_phosphinate_metabolism         6
## 321                               hsa00750_Vitamin_B6_metabolism         6
## 322                                 hsa00232_Caffeine_metabolism         5
## 323 hsa00400_Phenylalanine,_tyrosine_and_tryptophan_biosynthesis         5
## 324              hsa00471_D-Glutamine_and_D-glutamate_metabolism         5
##     NR.SIG.GENES      PVAL
##        <numeric> <numeric>
## 1             27   3.3e-15
## 2             45  1.12e-13
## 3             25  1.64e-09
## 4             21  1.23e-08
## 5             34  1.06e-07
## ...          ...       ...
## 320            0         1
## 321            0         1
## 322            0         1
## 323            0         1
## 324            0         1

Note: saving the results to file is typically recommended when carrying out several methods on multiple datasets for subsequent assessment. This makes results, potentially obtained from time-consuming computations, persistent across R sessions. In case of unexpected errors, this also allows resumption from the point of failure.

5 Benchmarking

Once methods have been applied to a chosen benchmark compendium, they can be subjected to a comparative assessment of runtime, statistical significance, and phenotype relevance.

To demonstrate how each criterion can be evaluated, we consider the example of the previous section where we applied ORA and CAMERA on 5 datasets of the GEO2KEGG compendium.

However, note that this minimal example is used to illustrate the basic functionality in a time-saving manner - as generally intended in a vignette. To draw conclusions on the individual performance of both methods, a more comprehensive assessment, involving application to the full compendium, should be carried out.

5.1 Runtime

Runtime, i.e. CPU time used, is an important measure of the applicability of a method. For enrichment methods, runtime mainly depends on whether methods rely on permutation testing, and how computationally intensive recomputation of the respective statistic in each permutation is (see Figure 4 in Geistlinger et al., 2016).

To obtain the runtime from the application of ORA and CAMERA to 5 datasets of the GEO2KEGG compendium, we can use the readResults function as we have saved results to the indicated result directory in the above call of runEA.

ea.rtimes <- readResults(res.dir, names(geo2kegg), 
                            methods=c("ora", "camera"), type="runtime")
ea.rtimes
## $ora
##  GSE1297 GSE14762 GSE15471 GSE16515 GSE18842 
##     0.36     0.34     0.36     0.34     0.34 
## 
## $camera
##  GSE1297 GSE14762 GSE15471 GSE16515 GSE18842 
##     0.23     0.22     0.26     0.22     0.28

For visualization of assessment results, the bpPlot function can be used to create customized boxplots for specific benchmark criteria.

bpPlot(ea.rtimes, what="runtime")

As both methods are simple gene set tests without permutation, they are among the fastest in the field - with CAMERA being roughly twice as fast as ORA.

mean(ea.rtimes$ora) / mean(ea.rtimes$camera)
## [1] 1.438017

5.2 Fraction of significant gene sets

The statistical accuracy of the significance estimation in gene set tests has been repeatedly debated. For example, systematic inflation of statistical significance in ORA due to an unrealistic independence assumption between genes is well-documented (Goeman and Buehlmann, 2007). On the other hand, the permutation procedure incorporated in many gene set tests has been shown to be biased (Efron and Tibshirani, 2007), and also inaccurate if permutation \(p\)-values are reported as zero (Phipson and Smyth, 2010).

These shortcomings can lead to inappropriately large fractions of significant gene sets, and can considerably impair prioritization of gene sets in practice. It is therefore important to evaluate resulting fractions of significant gene sets in comparison to other methods and with respect to the fraction of differentially expressed genes as a baseline.

We use the readResults function to obtain the saved gene set rankings of ORA and CAMERA when applied to 5 datasets of the GEO2KEGG compendium (see above call of runEA).

ea.ranks <- readResults(res.dir, names(geo2kegg), 
                            methods=c("ora", "camera"), type="ranking")
lengths(ea.ranks)
##    ora camera 
##      5      5
ea.ranks$ora[1:2]
## $GSE1297
## DataFrame with 323 rows and 4 columns
##                                                     GENE.SET  NR.GENES
##                                                  <character> <numeric>
## 1                         hsa00190_Oxidative_phosphorylation       106
## 2                                 hsa05010_Alzheimer_disease       151
## 3                                 hsa05012_Parkinson_disease       115
## 4                                hsa05016_Huntington_disease       162
## 5                                     hsa04714_Thermogenesis       186
## ...                                                      ...       ...
## 319                                     hsa03040_Spliceosome       112
## 320       hsa01521_EGFR_tyrosine_kinase_inhibitor_resistance        77
## 321         hsa04914_Progesterone-mediated_oocyte_maturation        77
## 322                             hsa00232_Caffeine_metabolism         5
## 323 hsa00524_Neomycin,_kanamycin_and_gentamicin_biosynthesis         5
##     NR.SIG.GENES      PVAL
##        <numeric> <numeric>
## 1             55   1.8e-12
## 2             68  2.23e-11
## 3             56  3.12e-11
## 4             70  1.15e-10
## 5             67  1.47e-06
## ...          ...       ...
## 319           12     0.999
## 320            6         1
## 321            5         1
## 322            0         1
## 323            0         1
## 
## $GSE14762
## DataFrame with 324 rows and 4 columns
##                                                         GENE.SET  NR.GENES
##                                                      <character> <numeric>
## 1                       hsa05150_Staphylococcus_aureus_infection        51
## 2                                             hsa04145_Phagosome       144
## 3                                         hsa05140_Leishmaniasis        70
## 4                                     hsa05416_Viral_myocarditis        56
## 5                        hsa04514_Cell_adhesion_molecules_(CAMs)       138
## ...                                                          ...       ...
## 320              hsa00440_Phosphonate_and_phosphinate_metabolism         6
## 321                               hsa00750_Vitamin_B6_metabolism         6
## 322                                 hsa00232_Caffeine_metabolism         5
## 323 hsa00400_Phenylalanine,_tyrosine_and_tryptophan_biosynthesis         5
## 324              hsa00471_D-Glutamine_and_D-glutamate_metabolism         5
##     NR.SIG.GENES      PVAL
##        <numeric> <numeric>
## 1             27   3.3e-15
## 2             45  1.12e-13
## 3             25  1.64e-09
## 4             21  1.23e-08
## 5             34  1.06e-07
## ...          ...       ...
## 320            0         1
## 321            0         1
## 322            0         1
## 323            0         1
## 324            0         1

The evalNrSigSets calculates the percentage of significant gene sets given a significance level alpha and a multiple testing correction method padj. We can visualize assessment results as before using bpPlot, which demonstrates here that CAMERA produces substantially larger fractions of significant gene sets than ORA.

sig.sets <- evalNrSigSets(ea.ranks, alpha=0.05, padj="BH")
sig.sets
##                ora   camera
## GSE1297   2.476780 17.33746
## GSE14762 12.654321 34.87654
## GSE15471  3.395062 22.83951
## GSE16515  2.469136 17.59259
## GSE18842  1.234568 21.29630
bpPlot(sig.sets, what="sig.sets")

5.3 Phenotype relevance

As introduced above, Tarca et al. (2012 and 2013) also assigned a target pathway to each dataset of the GEO2KEGG compendium, which is considered highly-relevant for the respective phenotype investigated. However, the relation between dataset, investigated phenotype, and assigned target pathway is not always clear-cut. In addition, there is typically more than one pathway that is considered relevant for the investigated phenotype.

On the other hand, evaluations of published enrichment methods often conclude on phenotype relevance, if there is any association between top-ranked gene sets and the investigated phenotype.

A more systematic approach is used in the MalaCards database of human diseases. Here, relevance of GO and KEGG gene sets is summarized from (i) experimental evidence and (ii) co-citation with the respective disease in the literature.

5.3.1 MalaCards disease relevance rankings

The GSEABenchmarkeR package provides MalaCards relevance rankings for the diseases investigated in the datasets of the GEO2KEGG and TCGA compendia. Here, we load the relevance rankings for KEGG gene sets and demonstrate how they can be incorporated in the assessment of phenotype relevance.

We note that the relevance rankings contain different numbers of gene sets for different diseases, because only gene sets for which evidence/association with the respective disease has been found are listed in a ranking.

For demonstration, we inspect the relevance rankings for Alzheimer’s disease (ALZ) and breast cancer (BRCA) containing 57 and 142 gene sets, respectively.

mala.kegg.file <- file.path(data.dir, "malacards", "KEGG.rds")
mala.kegg <- readRDS(mala.kegg.file)
sapply(mala.kegg, nrow)
##  ACC  ALZ BLCA BRCA CESC CHOL  CML COAD  CRC  DCM DLBC DMND ESCA  GBM HNSC 
##    9   57   65  142   22   33   56   28  161   23   52   99   90   99   72 
## HUNT KICH KIRC KIRP LAML  LES  LGG LIHC LUAD LUSC MESO   OV PAAD PARK PCPG 
##   34    4    8    8  108   49   24   98   54   23    3   31   70   39   12 
## PDCO PRAD READ SARC SKCM STAD TGCT THCA THYM UCEC  UCS  UVM 
##   31   12    2   73   42   24   24   81   61   90   29   55
mala.kegg$ALZ
## DataFrame with 57 rows and 4 columns
##                                              TITLE REL.SCORE MATCHED.GENES
##                                        <character> <numeric>     <integer>
## hsa05010                        Alzheimers disease     84.12            28
## hsa04932 Non-alcoholic fatty liver disease (NAFLD)     84.12             7
## hsa04726                      Serotonergic synapse     49.19             8
## hsa04728                      Dopaminergic synapse     49.19             8
## hsa04713                     Circadian entrainment     49.19             5
## ...                                            ...       ...           ...
## hsa05310                                    Asthma      9.81             1
## hsa05416                         Viral myocarditis      9.81             2
## hsa05330                       Allograft rejection      9.81             1
## hsa05332                 Graft-versus-host disease      9.81             2
## hsa05321          Inflammatory bowel disease (IBD)      9.81             2
##          TOTAL.GENES
##            <integer>
## hsa05010         177
## hsa04932         160
## hsa04726         115
## hsa04728         130
## hsa04713          98
## ...              ...
## hsa05310          35
## hsa05416          64
## hsa05330          41
## hsa05332          45
## hsa05321          67
mala.kegg$BRCA
## DataFrame with 142 rows and 4 columns
##                            TITLE REL.SCORE MATCHED.GENES TOTAL.GENES
##                      <character> <numeric>     <integer>   <integer>
## hsa05210       Colorectal cancer     166.1            35          70
## hsa05213      Endometrial cancer     166.1            23          61
## hsa05221  Acute myeloid leukemia     166.1            16          60
## hsa05218                Melanoma     166.1            35          78
## hsa05215         Prostate cancer     166.1            40          95
## ...                          ...       ...           ...         ...
## hsa05020          Prion diseases     13.23             6          43
## hsa05144                 Malaria     11.34             6          55
## hsa05143 African trypanosomiasis     11.28             5          36
## hsa04720  Long-term potentiation     10.93             7          67
## hsa05134           Legionellosis     10.81             6          59

5.3.2 Mapping between dataset ID and disease code

To obtain the relevance ranking of the respective disease investigated when assessing results on a specific dataset, a mapping between dataset and investigated disease is required. The function readDataId2diseaseCodeMap reads such a mapping from a tabular text file and turns it into a named vector - where the elements correspond to the disease codes and the names to the dataset IDs.

Here, we read the mapping between GSE ID and disease code for the GEO2KEGG compendium.

d2d.file <- file.path(data.dir, "malacards", "GseId2Disease.txt")
d2d.map <- readDataId2diseaseCodeMap(d2d.file)
head(d2d.map)
##      GSE1145     GSE11906      GSE1297     GSE14762 GSE14924_CD4 
##        "DCM"       "PDCO"        "ALZ"       "KIRC"       "LAML" 
## GSE14924_CD8 
##       "LAML"

5.3.3 Relevance score of a gene set ranking

To evaluate the phenotype relevance of a gene set ranking obtained from the application of an enrichment method to an expression dataset, the function evalRelevance assesses whether the ranking accumulates phenotype-relevant gene sets (i.e. gene sets with high relevance scores) at the top. Therefore, the function first transforms the ranks from the enrichment analysis to weights - where the greater the weight of a gene set, the more it is ranked towards the top of the GSEA ranking. These weights are then multiplied by the corresponding relevance scores and summed.

Here, we use evalRelevance to assess whether ORA, when applied to the GSE1297 dataset, recovers Alzheimer-relevant KEGG pathways.

ea.ranks$ora$GSE1297
## DataFrame with 323 rows and 4 columns
##                                                     GENE.SET  NR.GENES
##                                                  <character> <numeric>
## 1                         hsa00190_Oxidative_phosphorylation       106
## 2                                 hsa05010_Alzheimer_disease       151
## 3                                 hsa05012_Parkinson_disease       115
## 4                                hsa05016_Huntington_disease       162
## 5                                     hsa04714_Thermogenesis       186
## ...                                                      ...       ...
## 319                                     hsa03040_Spliceosome       112
## 320       hsa01521_EGFR_tyrosine_kinase_inhibitor_resistance        77
## 321         hsa04914_Progesterone-mediated_oocyte_maturation        77
## 322                             hsa00232_Caffeine_metabolism         5
## 323 hsa00524_Neomycin,_kanamycin_and_gentamicin_biosynthesis         5
##     NR.SIG.GENES      PVAL
##        <numeric> <numeric>
## 1             55   1.8e-12
## 2             68  2.23e-11
## 3             56  3.12e-11
## 4             70  1.15e-10
## 5             67  1.47e-06
## ...          ...       ...
## 319           12     0.999
## 320            6         1
## 321            5         1
## 322            0         1
## 323            0         1
obs.score <- evalRelevance(ea.ranks$ora$GSE1297, mala.kegg$ALZ)
obs.score
## [1] 831.4956

5.3.4 Random relevance score distribution

To assess the significance of the observed relevance score of an enrichment method applied to a specific dataset, i.e. to assess how likely it is to observe a relevance score equal or greater than the one obtained, the function compRand repeatedly applies evalRelevance to randomly drawn gene set rankings.

For demonstration, we compute relevance scores for 50 random gene set rankings and calculate the p-value as for a permutation test. This demonstrates that the relevance score obtained from applying ORA to GSE1297 significantly exceeds random scores.

gs.names <- ea.ranks$ora$GSE1297$GENE.SET
gs.ids <- substring(gs.names, 1, 8)
rand.scores <- compRand(mala.kegg$ALZ, gs.ids, perm=50)
summary(rand.scores)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##   507.6   597.1   635.5   632.9   666.5   749.7
(sum(rand.scores >= obs.score) + 1) / 51
## [1] 0.01960784

5.3.5 Theoretical optimum

The observed relevance score can be used to compare phenotype relevance of two or more methods when applied to one particular dataset. However, as the number of gene sets in the relevance rankings differs between phenotypes (see above Section 5.3.1 MalaCards disease relevance rankings), comparison between datasets is not straightforward as resulting relevance scores scale differently.

Therefore, 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 can then be used to compare observed scores between datasets.

Here, we compute the optimal score for the Alzheimer relevance ranking, which indicates that the score observed for ORA, when applied to GSE1297, is about 68% of the optimal score.

opt.score <- compOpt(mala.kegg$ALZ, gs.ids)
opt.score
## [1] 1227.125
round(obs.score / opt.score * 100, digits=2)
## [1] 67.76

5.3.6 Cross-dataset relevance score distribution

Evaluation of phenotype relevance with evalRelevance can also be done for several methods applied across multiple datasets. This allows to assess whether certain enrichment methods tend to produce rankings of higher phenotype relevance than other methods when applied to a compendium of datasets. As explained in the previous section, observed relevance scores are always expressed in relation to the respective optimal score.

For demonstration, we use evalRelevance to evaluate phenotype relevance of the gene set rankings produced by ORA and CAMERA when applied to 5 datasets of the GEO2KEGG compendium. We can visualize assessment results as before using bpPlot, which demonstrates here that ORA tends to recover more phenotype-relevant gene sets than CAMERA.

all.kegg.res <- evalRelevance(ea.ranks, mala.kegg, d2d.map[names(geo2kegg)])
bpPlot(all.kegg.res, what="rel.sets")

6 Advanced

6.1 Caching

Preparing an expression data compendium for benchmarking of enrichment methods can be time-consuming. In case of the GEO2KEGG compendium, it requires to summarize probe level expression on gene level and to subsequently carry out differential expression analysis for each dataset.

To flexibly save and restore an already processed expression data compendium, we can use the cacheResource function which builds on functionality of the BiocFileCache package.

cacheResource(geo2kegg, rname="geo2kegg")

This adds the selected 5 datasets of the GEO2KEGG compendium (as processed throughout this vignette) to the cache, and allows to restore it at a later time via

geo2kegg <- loadEData("geo2kegg", cache=TRUE)
## Loading GEO2KEGG data compendium ...
names(geo2kegg)
## [1] "GSE1297"  "GSE14762" "GSE15471" "GSE16515" "GSE18842"

Note: to obtain the original unprocessed version of the compendium, set the cache argument of the loadEData function to FALSE.

To clear the cache (use with care):

cache.dir <- rappdirs::user_cache_dir("GSEABenchmarkeR")
bfc <- BiocFileCache::BiocFileCache(cache.dir)
BiocFileCache::removebfc(bfc)

6.2 Parallel computation

Leveraging functionality from BiocParallel, parallel computation of the functions maPreproc, runDE, and especially runEA, when applied to multiple datasets is straightforward. Internally, these functions call BiocParallel::bplapply, which triggers parallel computation as configured in the first element of BiocParallel::registered(). As a result, parallel computation is implicitly incorporated in the above calls of these functions when carried out on a multi-core machine. See the vignette of the BiocParallel package for an introduction.

Inspecting

BiocParallel::registered()
## $SnowParam
## class: SnowParam
##   bpisup: FALSE; bpnworkers: 4; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bptimeout: 2592000; bpprogressbar: FALSE; bpexportglobals: TRUE
##   bpRNGseed: 
##   bplogdir: NA
##   bpresultdir: NA
##   cluster type: SOCK
## 
## $SerialParam
## class: SerialParam
##   bpisup: TRUE; bpnworkers: 1; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bptimeout: 2592000; bpprogressbar: FALSE; bpexportglobals: TRUE
##   bplogdir: NA

shows that the execution uses a MulticoreParam per default (on Windows: a SnowParam), where the bpnworkers attribute indicates the number of cores involved in the computation.

To change the execution mode of functions provided in the GSEABenchmarkeR package, accordingly configured computation parameters of class BiocParallelParam can either directly be registered via BiocParallel::register, or supplied with the parallel argument of the respective function.

For demonstration, we configure here a BiocParallelParam to display a progress bar

bp.par <- BiocParallel::registered()[[1]]
BiocParallel::bpprogressbar(bp.par) <- TRUE

and supply runDE with the updated computation parameter.

geo2kegg <- runDE(geo2kegg, parallel=bp.par)
## 
  |                                                                         
  |                                                                   |   0%
  |                                                                         
  |=================                                                  |  25%
  |                                                                         
  |==================================                                 |  50%
  |                                                                         
  |==================================================                 |  75%
  |                                                                         
  |===================================================================| 100%

Users that would like to use distributed computation, on e.g. an institutional computer cluster, should consult the vignette of the BiocParallel package to similarly configure a BiocParallelParam for that purpose.