massive GST (version 20230328)

In massiveGST: Competitive Gene Sets Test with the Mann-Whitney-Wilcoxon Test

This vignette shows how to use the massiveGST package.

Installation

Get the stable release from CRAN

install.packages("massiveGST")

On going bug correction and future improvements are in github.

if (!requireNamespace("BiocManager", quietly = TRUE)) install.packages("BiocManager")
BiocManager::install("stefanoMP/massiveGST")

Setup

suppressPackageStartupMessages(library(massiveGST, quietly = TRUE))

getting gene-profile from a file

In any cases, the names of the list have to match the gene names in any of their coding (gene-symbol, entrez, ensembl, etc.). To explain some details of the functions, we assume that positive values of the gene-profile are up-regulated in the treatment group, while the negative ones are up-regulated in the control group. In other words, the positive genes are associated with the treatment group, while the others are associated with the control samples.

Here, we consider the working gene-profile from Frattini et al, 2018, stored as external data in the package.

fname <- system.file("extdata", package="massiveGST")
fname <- file.path(fname, "pre_ranked_list.txt")
geneProfile <- get_geneProfile(fname)
class(geneProfile)
head(geneProfile)
tail(geneProfile)

getting gene-sets from msigdbr

msigdbr package is an R distribution of the most recent release of the MSigDB collection of gene-sets. The function get_geneSets_from_msigdbr is a wrapper allowing the creation of a data structure for following gene-set enrichment analysis. Function parameters are those of the msigdbr function, except for 'what', allowing to choose the coding of the gene names.

This example extracts gene-sets where the gene-names are the gene symbols. Other possibilities have been specified in the 'man' fo the functions, and essentially they are the names of the columns of the data frame retrieved by the msigdbr function.

geneSets <- get_geneSets_from_msigdbr(category = "H", what = "gene_symbol")
class(geneSets)
head(names(geneSets))

Optionally, the gene-sets collection can be stored in gmt formatted file

fname <- file.path(tempdir(), "hallmarks.gmt")
write_geneSets_to_gmt(geneSets, fileName = fname)

and retrieved with the code

fname <- file.path(tempdir(), "hallmarks.gmt")
tmp <- get_geneSets_from_local_files(fname)

class(geneSets)
head(names(geneSets))

Running the analysis with alternative = "two.sided"

system.time({ans <- massiveGST(geneProfile, geneSets, alternative = "two.sided")})
class(ans)
ans[1:6,]

The analysis result is essentially a data frame; the first class 'mGST' allows hooking other methods to handle the output.

The results can be saved in tab-separated value (tsv) format or an XLS format file.

fname <- file.path(tempdir(), "massiveGST_results.tsv")
save_as_tsv(ans, file_name = fname)

fname <- file.path(tempdir(), "massiveGST_results.xls")
save_as_xls(ans, file_name = fname)

The summary method.

The summary function allows customizing the result of the analysis.

summary(ans)[1:10,]

With it:

• The ordering of the table can be chosen.

summary(ans, order_by = "NES")[1:10,]
summary(ans, order_by = "p.value")[1:10,]
summary(ans, order_by = "bonferroni")[1:10,]

• Some columns can be removed.

(tmp <- summary(ans, order_by = "p.value", cols_to_remove = c("BH.value", "B.value"))[1:10,])

• The formattable package can be invoked to get a more readable table.

summary(ans, as.formattable = TRUE)
summary(ans, order_by = "p.value", cols_to_remove = c("BH.value", "B.value"), as.formattable = TRUE)

The value of the summary method is invisible, but it is a data frame.

tmp

Trimming the table fo results

Three functions allow to remove rows from the results table as need. The output of these function can be formatted with the summary method.

The first function removes the non significant gene-sets. By default, a 5\% level of significance is applied to BH.values (Benijamini and Hockberg adjustment of the p.values)

summary(cut_by_significance(ans), as.formattable = TRUE)

As a toy example, ...

summary(cut_by_significance(ans, level_of_significance = 0.01, where = "bonferroni"),
        cols_to_remove = c("BH.value", "NES", "size"), 
        order_by = "logit2NES",
        as.formattable = TRUE)

The functions cut_by_NES and cut_by_logit2NES remove the rows having a NES/logit2NES below a given threshold. They are equivalent, in fact

[logit2NES = \log_2 \frac{NES}{1-NES},] and back [NES = \frac{2^{logit2NES}}{1+2^{logit2NES}}.]

The advantage of representing the Normalized Enrichment Score ($NES$) as $logit2NES$ is that this last is a signed value: positive means association with the genes of the treatment group, while a negative value signals the association with the control samples. The default thresholds are set to 0.6 and 0.58 for $NES$ and $logit2NES$. These values say that the probability of association of the gene-set with the treatment group is 1.5 higher than the case of association with the control group.

Trimming the table of results according to the $NES$/$logit2NES$ means giving much more attention to the descriptive interpretation of the NES as a measure of strongness of association. Given a gene-set, the $NES$ is the percentile rank associated with the gene-set, seen as a single value (the average of the ranks), in the universe of the genes outside the gene-set.

tmp <- cut_by_significance(ans)
summary(cut_by_logit2NES(tmp), as.formattable = TRUE, order_by = "NES")
summary(cut_by_NES(tmp), as.formattable = TRUE, order_by = "NES")

Plotting the analysis

The results can be displayed two-way: a bar plot and a network graph. The corresponding functions provide a graphical rendering of the table in input in both cases. Then, eventually, the table has to be trimmed according to significance or NES.

plot(ans)

A meaningful display follows.

plot(cut_by_significance(ans), top = 30)

Here, the maximum number of bars has been restricted to 30.

The horizontal axis (signed-NES) is a linear transformation of the NES: [signed\mbox-NES = 2\cdot NES - 1] necessary to a) signal the direction of the association, and b) to bound the bars between -1.0 and 1.0

The network plot needs to specify 'as.network = TRUE' and provide the gene-sets collection.

plot(cut_by_significance(ans), gene_sets = geneSets, as.network = TRUE)

The similarity $S(A, B)$ between two gene-sets $A$ and $B$ comes from the convex combination [S(A,B)= \epsilon \cdot \delta_1(A, B)+ (1-\epsilon)\cdot \delta_0(A, B),] with $0\leq \epsilon \leq 1$, and $\delta_0$ is the Jaccard similarity, while $\delta_1$ is the overlap index. $logit2NES$ controls the color of the balls: red for those with positive values (associated with the treatment), green otherwise. $actualSize$ controls the dimension of the balls.

A more intensive analysis: playing with more than 10,000 gene-sets

Here, we show a more serious analysis. A larger collection of gene-sets have be considered.

The nature of the gene profile included in the package is strictly according to the treatment versus control logic. In this case, 9 fusion FGFR3-TACC3 positive samples have been compared to 535 other samples. The interest is in the treatment, and then the analysis requires the alternative hypothesis "greater".

system.time({C5BP_gs <- get_geneSets_from_msigdbr(category = "C5", 
                                                  subcategory = "BP", 
                                                  what = "gene_symbol")})

system.time({C5MF_gs <- get_geneSets_from_msigdbr(category = "C5", 
                                                  subcategory = "MF", 
                                                  what = "gene_symbol")})


system.time({C5CC_gs <- get_geneSets_from_msigdbr(category = "C5", 
                                                  subcategory = "CC", 
                                                  what = "gene_symbol")})

system.time({H_gs <- get_geneSets_from_msigdbr(category = "H", 
                                               what = "gene_symbol")})

# merging gene-sets collections
geneSets <- c(C5MF_gs, C5CC_gs, C5BP_gs, H_gs)
length(geneSets)

# running the analysis
system.time({ans <- massiveGST(geneProfile, geneSets, 
                               alternative = "greater")})

# removing non significant results
ans <- cut_by_significance(ans, 
                           level_of_significance = 0.05, 
                           where = "bonferroni")

# Tabular results
summary(ans, as.formattable = TRUE, cols_to_remove = "BH.value")

# Saving the table
fname <- file.path(tempdir(), "massiveGST_results.tsv")
save_as_xls(ans, file_name = fname)

# Inspecting the network of gene-sets
plot(ans, gene_sets = geneSets, as.network = TRUE)

Over-representation (Fisher's exact) test

Since version 1.2, a new function has been add to implement the overepresentation test (Fisher's exact text) that is integrated with both the tabular and graphical function of this package.

Getting the gene sets and the gene-profile

geneSets <- get_geneSets_from_msigdbr(category = "C5", subcategory = "CC", what = "gene_symbol")

fname <- system.file("extdata", package="massiveGST")
fname <- file.path(fname, "pre_ranked_list.txt")
geneProfile <- get_geneProfile(fname)

To mimic a set of significant genes coming from a differential expression procedure, we get the first 500 genes in the pre-ranked list,

geneList <- names(head(geneProfile, 500))

and run the enrichment analysis

ans <- massiveORT(geneList, geneSets)

The function massiveORT essentially is a wrapper to the function fisher.test in charge to 1) arrange the input to feed fisher.test in sequence for each gene set, 2) arrange the output in a data frame compatible with the other function of the package, and 3) compute the universe of genes for the analysis.

By default, the universe of genes necessary to the Fisher's test is computed as the collection of genes included at lest once in any gene-set. The function allows to consider a different universe. Se help file.

The tabular result comes from

summary(cut_by_significance(ans), as.formattable = TRUE)

The graphical output is from

plot(cut_by_significance(ans))

and

plot(cut_by_significance(ans), gene_sets = geneSets, as.network = TRUE)

Session info

sessionInfo()

massiveGST documentation built on March 31, 2023, 8:59 p.m.