roastgsa vignette (main)
1 May 2024
Contents
1 Installation
if (!require("BiocManager", quietly = TRUE))
install.packages("BiocManager")
BiocManager::install("roastgsa")2 Paper associated to R package
library( roastgsa )The R package roastgsa contains several functions to perform
gene set analysis, for both competitive and self-contained hypothesis
testing.
It follows the work by Gordon Smyth’s group on rotation based methods
for gene set analysis [1], code available in R through functions
roast and romer from the limma package [2].
They propose to reconsider the sample randomization approach based on permutations of GSEA [3] to the more general procedure scheme using rotations of the residual space of the data, which can be used even when the number of degrees of freedom left is very small.
In our understanding, the test statistics provided in romer, which
are all functions of the moderated t-statistics ranks, are too limited. In
this R package we propose to complete the romer functionality by providing
other statistics used in the GSA context [4]. We consider the KS based test
statistics in the GSEA and GSVA [5] as well as computationally more efficient
procedures such as restandardized statistics based on summary statistics
measures. The methodology is presented and compared using both simulated
and benchmarking data in the following publication:
roastgsa: a comparison of rotation-based scores for gene set enrichment analysis (2023).
We encourage anybody wanting to use the R package to first read the associated paper.
In this R package, we have also provided several tools to summarize
and visualize the roastgsa results. This vignette will show a work-flow
for gene set analysis with roastgsa using microarray data (available in
the GSEABenchmarkeR R package [6]).
3 Data: array experiment from GSEABenchmarkeR
We consider the fourth dataset available in the GSEABenchmarkeR R
package, which consists of a microarray study with 30 samples, 15 paired samples
corresponding to two different groups (that take values 0 and 1 respectively).
library(GSEABenchmarkeR)
geo2keggR <- loadEData("geo2kegg", nr.datasets=4)
geo2kegg <- maPreproc(list(geo2keggR[[4]]))
y <- assays(geo2kegg[[1]])$expr
covar <- colData(geo2kegg[[1]])
covar$GROUP <- as.factor(covar$GROUP)
covar$BLOCK <- as.factor(covar$BLOCK)The objective is to understand which KEGG processes might be affected
by such two genotyping conditions. These are made available through the
EnrichmentBrowser R package. Here we will consider gene sets of size
between 11 and 499.
#library(EnrichmentBrowser)
#kegg.hs <- getGenesets(org="hsa", db="kegg")
data(kegg.hs)
kegg.gs <- kegg.hs[which(sapply(kegg.hs,length)>10&sapply(kegg.hs,length)<500)]4 Gene set enrichment analysis using roastgsa
4.1 Data treatment
There are some important elements that should be taken into
consideration before undertaking the roastgsa. First, the expression
data should be approximately normally distributed, at least in their
univariate form. For RNA-seq data or any other form of counts data, prior
normalization, e.g., rlog or vst (DESeq2), zscoreDGE (edgeR) or voom
(limma), should be used. Besides, filtering for expression intensity
and variability could be considered, especially when the gene
variation across individuals is limited by the experimental coverage.
For microarray intensities, as it happens in the example of this vignette, we can perform a quantile normalization to balance out the libraries.
library(preprocessCore)
ynorm <- normalize.quantiles(y)
rownames(ynorm) <- rownames(y)
colnames(ynorm) <- colnames(y)
par(mfrow=c(1,2))
boxplot(y, las = 2)
boxplot(ynorm, las = 2)
y <- ynorm4.2 Model setting
Indicating the model to be fitted is essential for the roastgsa
implementation. This follows a similar strategy to the limma
specifications. With the attributes form, covar and contrast it can be
set the linear model and the specific estimated coefficient to be used
in roastgsa for hypothesis testing. The matrix design is found
automatically taking the form and covar parameters (see
below). The contrast can be in a vector object (length equal to the
number of columns in the matrix design), an integer with the column of
the term of interest in the matrix design, as well as the name of such column.
form = as.formula("~ BLOCK + GROUP")
design <- model.matrix(form, covar)
contrast = "GROUP1"Three parameters that define the roast hypothesis testing have to be
properly specified: -a- the number of rotations to draw the null
hypothesis (nrot); -b- the statistic to be used (set.statistic);
-c- the type of hypothesis (self.contained argument), competitive
(set to FALSE) or self-contained (set to TRUE).
Regarding the test statistics, the maxmean (directional) and the
absmean (nondirectional) are recommended to maximise the power, with
mean.rank being a good non-parametric alternative for a more robust
statistic if it is known that a few highly differentially expressed
genes might influence heavily the statistic calculation. For a more
“democratic” statistic, one that counterbalance both ends of
the distribution (negative and positive), we encourage using the
mean statistic.
Below, there is the standard roastgsa instruction (under competitive
testing) for maxmean and mean.rank statistics.
fit.maxmean.comp <- roastgsa(y, form = form, covar = covar,
contrast = contrast, index = kegg.gs, nrot = 500,
mccores = 1, set.statistic = "maxmean",
self.contained = FALSE, executation.info = FALSE)
f1 <- fit.maxmean.comp$res
rownames(f1) <- substr(rownames(f1),1,8)
head(f1)## total_genes measured_genes est nes pval adj.pval
## hsa05230 70 69 0.5425709 3.926355 0.001996008 0.05112851
## hsa04115 73 73 0.8053681 3.429170 0.001996008 0.05112851
## hsa04215 32 31 0.6703507 3.271317 0.001996008 0.05112851
## hsa04530 169 164 0.4898137 3.263415 0.001996008 0.05112851
## hsa04520 93 92 0.6180546 3.257475 0.001996008 0.05112851
## hsa05203 204 194 0.6093395 3.096194 0.001996008 0.05112851fit.meanrank <- roastgsa(y, form = form, covar = covar,
contrast = 2, index = kegg.gs, nrot = 500,
mccores = 1, set.statistic = "mean.rank",
self.contained = FALSE, executation.info = FALSE)
f2 <- fit.meanrank$res
rownames(f2) <- substr(rownames(f2),1,8)
head(f2)## total_genes measured_genes est pval.diff adj.pval.diff
## hsa04710 34 33 2526.621 0.005988024 0.9919553
## hsa04012 85 84 2213.738 0.013972056 0.9919553
## hsa00230 128 124 -1623.702 0.025948104 0.9919553
## hsa04725 113 111 -1226.869 0.041916168 0.9919553
## hsa05223 72 72 1850.736 0.061876248 0.9919553
## hsa05213 58 58 2278.138 0.065868263 0.9919553
## pval.mixed adj.pval.mixed
## hsa04710 0.463073852 0.7032453
## hsa04012 0.079840319 0.5970770
## hsa00230 0.323353293 0.6296880
## hsa04725 0.516966068 0.7442039
## hsa05223 0.053892216 0.5970770
## hsa05213 0.001996008 0.22155694.3 Visualization of the results
Several functions to summarize or visualize the results can be applied
to objects of class roastgsa, which are found as output of the
roastgsa function.
The plot function of a roastgsa object produces a general image
of the differential expression results within any tested gene set. If
type = "stats", it shows the ordered moderated t-statistics in
various formats, area for up- and down- expressed genes, barcode plot
for these ordered values and density. With the argument whplot it can be
selected the gene set of interest (either an integer with the ordered
position in the roastgsa output or the name of the gene set).
plot(fit.maxmean.comp, type ="stats", whplot = 2, gsainfo = TRUE,
cex.sub = 0.5, lwd = 2)
If the roastgsa statistic is mean.rank, the moderated-t statistic
centred ranks are printed instead.
plot(fit.meanrank, type ="stats", whplot = 1, gsainfo = TRUE,
cex.sub = 0.4, lwd = 2)
Even though the type = "GSEA" option is directly interpretable for
ksmean and ksmax statistics, we find it useful for seeing the behavior
in both Kolgomorov-Smirnov type scores and simple summary statistics:
plot(fit.maxmean.comp, type = "GSEA", whplot = 2, gsainfo = TRUE,
maintitle = "", cex.sub = 0.5, statistic = "mean")
plot(fit.meanrank, type = "GSEA", whplot = 1, gsainfo = TRUE,
maintitle = "", cex.sub = 0.5, statistic = "mean")
Another graphic that is proposed in this package to visualize the genes activity
within a gene set can be obtained through the function heatmaprgsa_hm.
The main intention is to illustrate the variation across samples for the
gene set of interest. We highly recommend the generation of this graphic,
or any other similar plot that shows sample variation for the tested gene sets.
Not only for showing which genes are activated but also as quality control to
detect samples that can be highly influential in the GSA analysis.
hm <- heatmaprgsa_hm(fit.maxmean.comp, y, intvar = "GROUP", whplot = 3,
toplot = TRUE, pathwaylevel = FALSE, mycol = c("orange","green",
"white"), sample2zero = FALSE)
The same sort of graphic can be drawn at pathway level, i.e.,
summarized (average) pathway activity, when pathwaylevel = TRUE. With the parameter whplot it can be specified which pathways
are included in the comparison.
hm2 <- heatmaprgsa_hm(fit.maxmean.comp, y, intvar = "GROUP", whplot = 1:50,
toplot = TRUE, pathwaylevel = TRUE, mycol = c("orange","green",
"white"), sample2zero = FALSE)
4.4 Effective signature size
Gene sets in standard databases might contain from moderately to highly correlated genes, even when the effect of the known covariates is adjusted a priori. The variance of summary statistics such as mean or maxmean increases with the intra-gene set correlation, meaning that the power for detecting gene set changes between experimental conditions depends on a combination of genewise effect sizes, signature size and intra-gene set correlation.
We define the effective signature size of a tested gene set by the total number of genes that are needed, if these were selected at random, to achieve the same summary statistic variance as that for the testing set. This can be viewed as a realistic measure of the total number of independent variables that contribute to the power of the test. To get an estimate of the effective signature size, sample variances of rotation scores for randomly generated sets of several sizes are compared to the observed rotation scores variance for the tested gene set:
## Computationally intensive
# varrot <- varrotrand(fit.maxmean.comp, y,
# testedsizes = c(3:30, seq(32,50, by=2), seq(55,200,by=5)), nrep = 200)A p-value that approximates the probability of obtaining a variance as extreme as the observed test variance in randomly selected sets of several sizes is computed:
# ploteffsignaturesize(fit.maxmean.comp, varrot, whplot = 1)4.5 Functions to present the results
The outcome obtained in roastgsa can be organized in a html table using
the htmlrgsa function. The reported file shows all numerical results that
can be found in printing an object of class roastgsa as well as the plots
obtained by plotStats, plotGSEA, heatmaprgsa_hm or
varrotrand + ploteffsignaturesize. A column with html-links for the
differential expression results at gene level can also be provided through the
argument geneDEhtmlfiles, which we find extremely useful for researchers to
understand which are the genes that drive the pathway activity.
#htmlrgsa(fit.maxmean.comp, htmlpath = getwd(), htmlname = "file.html",
# plotpath = paste0(getwd(),"/plotsroast/"), geneDEhtmlfiles = NULL,
# plotstats = TRUE, plotgsea = TRUE, indheatmap = TRUE,
# ploteffsize = TRUE, y = y, intvar = "GROUP", whplot = 1:50,
# mycol = c("orange","green","white"), varrot=varrot,
# sorttable = sorttable,dragtable = dragtable)Complete information on the usage of this function can be found in the manual.
4.6 Single sample GSA
To complement the outcome of the roastgsa, we encourage the usage of
single sample GSA visualization methods. These approaches are interesting for
assessing the variability of gene set effects at sample level (averaging out
gene wise coefficients) instead of at gene level.
The heatmaprgsa_hm instruction used above already provides some insight
of sample variation. Besides, in the roastgsa package we have implemented a
function for single sample gene set analysis based on z-scores [7].
We present two methods that correct for the overall structure in
the genome (method = "GScor" and method = "GSadj"), leading to competitive
testing, and one method that ignores the rest of the genes in the data
(method = "zscore"), self-contained testing. For small
sample sizes, we recommend using “GScor” whereas for sufficiently large
sample size (>50) “GSadj” is our preferred option.
ss1 <- ssGSA(y, obj=fit.maxmean.comp, method = c("GScor"))Visualizing the summarized gene set information at sample level is highly recommended to check the variation of the samples in the groups of interest.
plot(ss1, orderby = covar$GROUP, whplot = 1, col = as.numeric(covar$GROUP),
samplename = FALSE, pch =16, maintitle = "", ssgsaInfo = TRUE,
cex.sub = 0.8, xlab = "Group", ylab = "zscore - GS")
5 sessionInfo
sessionInfo()## R version 4.4.0 beta (2024-04-15 r86425)
## Platform: x86_64-pc-linux-gnu
## Running under: Ubuntu 22.04.4 LTS
##
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.19-bioc/R/lib/libRblas.so
## LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.10.0
##
## locale:
## [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
## [3] LC_TIME=en_US.UTF-8 LC_COLLATE=en_US.UTF-8
## [5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
## [7] LC_PAPER=en_US.UTF-8 LC_NAME=C
## [9] LC_ADDRESS=C LC_TELEPHONE=C
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
##
## time zone: America/New_York
## tzcode source: system (glibc)
##
## attached base packages:
## [1] stats4 stats graphics grDevices utils datasets methods
## [8] base
##
## other attached packages:
## [1] preprocessCore_1.66.0 hgu133plus2.db_3.13.0
## [3] org.Hs.eg.db_3.19.1 AnnotationDbi_1.66.0
## [5] GSEABenchmarkeR_1.24.0 SummarizedExperiment_1.34.0
## [7] GenomicRanges_1.56.0 GenomeInfoDb_1.40.0
## [9] IRanges_2.38.0 S4Vectors_0.42.0
## [11] MatrixGenerics_1.16.0 matrixStats_1.3.0
## [13] Biobase_2.64.0 BiocGenerics_0.50.0
## [15] roastgsa_1.2.0 knitr_1.46
## [17] BiocStyle_2.32.0
##
## loaded via a namespace (and not attached):
## [1] DBI_1.2.2 bitops_1.0-7
## [3] GSEABase_1.66.0 rlang_1.1.3
## [5] magrittr_2.0.3 compiler_4.4.0
## [7] RSQLite_2.3.6 png_0.1-8
## [9] vctrs_0.6.5 pkgconfig_2.0.3
## [11] crayon_1.5.2 fastmap_1.1.1
## [13] magick_2.8.3 dbplyr_2.5.0
## [15] XVector_0.44.0 labeling_0.4.3
## [17] caTools_1.18.2 utf8_1.2.4
## [19] rmarkdown_2.26 graph_1.82.0
## [21] UCSC.utils_1.0.0 KEGGgraph_1.64.0
## [23] tinytex_0.50 purrr_1.0.2
## [25] bit_4.0.5 xfun_0.43
## [27] zlibbioc_1.50.0 cachem_1.0.8
## [29] jsonlite_1.8.8 blob_1.2.4
## [31] highr_0.10 DelayedArray_0.30.0
## [33] BiocParallel_1.38.0 parallel_4.4.0
## [35] R6_2.5.1 bslib_0.7.0
## [37] RColorBrewer_1.1-3 limma_3.60.0
## [39] jquerylib_0.1.4 Rcpp_1.0.12
## [41] bookdown_0.39 Matrix_1.7-0
## [43] tidyselect_1.2.1 abind_1.4-5
## [45] yaml_2.3.8 gplots_3.1.3.1
## [47] codetools_0.2-20 curl_5.2.1
## [49] lattice_0.22-6 tibble_3.2.1
## [51] withr_3.0.0 KEGGREST_1.44.0
## [53] evaluate_0.23 BiocFileCache_2.12.0
## [55] Biostrings_2.72.0 pillar_1.9.0
## [57] BiocManager_1.30.22 filelock_1.0.3
## [59] KernSmooth_2.23-22 generics_0.1.3
## [61] RCurl_1.98-1.14 ggplot2_3.5.1
## [63] munsell_0.5.1 scales_1.3.0
## [65] gtools_3.9.5 xtable_1.8-4
## [67] glue_1.7.0 tools_4.4.0
## [69] annotate_1.82.0 XML_3.99-0.16.1
## [71] grid_4.4.0 colorspace_2.1-0
## [73] GenomeInfoDbData_1.2.12 cli_3.6.2
## [75] KEGGandMetacoreDzPathwaysGEO_1.23.0 fansi_1.0.6
## [77] S4Arrays_1.4.0 dplyr_1.1.4
## [79] Rgraphviz_2.48.0 gtable_0.3.5
## [81] sass_0.4.9 digest_0.6.35
## [83] SparseArray_1.4.0 farver_2.1.1
## [85] memoise_2.0.1 htmltools_0.5.8.1
## [87] lifecycle_1.0.4 httr_1.4.7
## [89] EnrichmentBrowser_2.34.0 KEGGdzPathwaysGEO_1.41.0
## [91] statmod_1.5.0 bit64_4.0.56 References
Appendix
[1] E. Lim, D. Wu, G. K. Smyth, M.-L. Asselin-Labat, F. Vaillant, and J. E. Visvader. ROAST: rotation gene set tests for complex microarray experiments. Bioinformatics, 26(17):2176-2182, 2010.
[2] M. E. Ritchie, B. Phipson, D. Wu, Y. Hu, C. W. Law, W. Shi, and G. K. Smyth. limma powers differential expression analyses for RNAsequencing and microarray studies. Nucleic acids research, 43(7):e47, 2015.
[3] A. Subramanian, P. Tamayo, V. K. Mootha, S. Mukherjee, B. L. Ebert, M. A. Gillette, A. Paulovich, S. L. Pomeroy, T. R. Golub, E. S. Lander, and J. P. Mesirov. Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Proceedings of the National Academy of Sciences, 102(43):15545-15550, 2005.
[4] B. Efron and R. Tibshirani. On testing the significance of sets of genes. The Annals of Applied Statistics, 1(1):107-129, 2007.
[5] S. Hanzelmann, R. Castelo, and J. Guinney. GSVA: gene set variation analysis for microarray and RNA-Seq data. 14(1):7, 2013.
[6] Geistlinger L, Csaba G, Santarelli M, Schiffer L, Ramos M, Zimmer R, Waldron L (2019). GSEABenchmarkeR: Reproducible GSEA Benchmarking. R package version 1.6.0, https://github.com/waldronlab/GSEABenchmarkeR.
[7] Caballe Mestres A, Berenguer Llergo A and Stephan-Otto Attolini C. Adjusting for systematic technical biases in risk assessment of gene signatures in transcriptomic cancer cohorts. bioRxiv (2018).