GSVA ssGSEA vs. Broad implementation of ssGSEA
2
1
Entering edit mode
Bob ▴ 10
@6c962abc
Last seen 16 days ago
United States

Hello,

Both the GSVA module's implementation of ssGSEA and the Broad/GenePattern implementation (https://github.com/broadinstitute/ssGSEA2.0) are based on the papers from Barbie et al (1) and Abazeed et al (2), however, I cannot find a combination of parameters for which they produce identical results to each other.

For example, taking the random matrices a la the GSVA documentation:

set.seed(2023-11-09)

p <- 1000 ## number of genes
n <- 20    ## number of samples
ngs <- 10 ## number of gene sets
X <- matrix(rnorm(p*n), nrow=p,
            dimnames=list(paste0("g", 1:p), paste0("s", 1:n)))

gs <- as.list(sample(10:100, size=ngs, replace=TRUE))
gs <- lapply(gs, function(n, p)
  paste0("g", sample(1:p, size=n, replace=FALSE)), p)

names(gs) <- paste0("gs", 1:length(gs))

and running with the default parameters of GSVA's ssGSEA:

ssgseaPar <- GSVA::ssgseaParam(X, gs)
gsva_ssgsea <- GSVA::gsva(ssgseaPar, verbose=FALSE)

and running the broad implementation on the same data:

gct <- cmapR::GCT(mat=X)
gmt <- purrr::imap(gs, \(x, name) list(head=name, desc=name, len=length(x), entry=x))
cmapR::write_gct(gct, "X.gct", appenddim = FALSE)
cmapR::write_gmt(gmt, "gs.gmt")
ssGSEA2("X.gct", "compare/test01", "gs.gmt")
broad_ssgsea <- cmapR::parse_gctx("compare/test01-scores.gct")@mat

produces correlated but different results:

tibble::tibble(gsva=as.numeric(gsva_ssgsea), broad=as.numeric(broad_ssgsea)) |>
  ggpubr::ggscatter("broad", "gsva",  add = "reg.line", cor.coef = TRUE) +
  ggplot2::theme_bw()

(See plot below). The critical parameters for GSVA seem to be normalize and alpha, but there seem to be no analogs in the Broad implementation.

Is there a way to make these two implementations behave identically?

Thank you!

enter image description here

sessionInfo( )
R version 4.4.0 (2024-04-24)
Platform: aarch64-apple-darwin23.2.0
Running under: macOS Sonoma 14.6.1

Matrix products: default
BLAS:   /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib 
LAPACK: /Users/rzimmermann/.brew/Cellar/r/4.4.0_1/lib/R/lib/libRlapack.dylib;  LAPACK version 3.12.0

locale:
[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

time zone: America/New_York
tzcode source: internal

attached base packages:
[1] parallel  stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
 [1] magrittr_2.0.3    doParallel_1.0.17 iterators_1.0.14  foreach_1.5.2     verification_1.42 dtw_1.23-1        proxy_0.4-27      CircStats_0.2-6   MASS_7.3-61       boot_1.3-31       fields_16.3       viridisLite_0.4.2 spam_2.11-0       gtools_3.9.5      pacman_0.5.1     

loaded via a namespace (and not attached):
  [1] rstudioapi_0.16.0           jsonlite_1.8.9              magick_2.8.5                farver_2.1.2                ragg_1.3.3                  fs_1.6.4                    zlibbioc_1.50.0             vctrs_0.6.5                 memoise_2.0.1               rstatix_0.7.2               htmltools_0.5.8.1           S4Arrays_1.4.1              usethis_3.0.0               broom_1.0.7                 Rhdf5lib_1.26.0             Formula_1.2-5               SparseArray_1.4.8           rhdf5_2.48.0                htmlwidgets_1.6.4           cachem_1.1.0                mime_0.12                   lifecycle_1.0.4             pkgconfig_2.0.3             rsvd_1.0.5                  Matrix_1.7-0                R6_2.5.1                    fastmap_1.2.0               GenomeInfoDbData_1.2.12     MatrixGenerics_1.16.0       shiny_1.9.1                 digest_0.6.37               colorspace_2.1-1            AnnotationDbi_1.66.0        S4Vectors_0.42.1            rprojroot_2.0.4             irlba_2.3.5.1              
 [37] pkgload_1.4.0               textshaping_0.4.0           GenomicRanges_1.56.1        RSQLite_2.3.7               ggpubr_0.6.0                beachmat_2.20.0             labeling_0.4.3              cytolib_2.16.0              fansi_1.0.6                 mgcv_1.9-1                  httr_1.4.7                  abind_1.4-8                 compiler_4.4.0              remotes_2.5.0               bit64_4.5.2                 withr_3.0.1                 backports_1.5.0             BiocParallel_1.38.0         carData_3.0-5               DBI_1.2.3                   pkgbuild_1.4.4              HDF5Array_1.32.1            maps_3.4.2                  ggsignif_0.6.4              DelayedArray_0.30.1         sessioninfo_1.2.2           rjson_0.2.23                tools_4.4.0                 httpuv_1.6.15               glue_1.8.0                  nlme_3.1-166                rhdf5filters_1.16.0         promises_1.3.0              grid_4.4.0                  cmapR_1.16.0                generics_0.1.3             
 [73] gtable_0.3.5                tzdb_0.4.0                  tidyr_1.3.1                 hms_1.1.3                   car_3.1-3                   BiocSingular_1.20.0         ScaledMatrix_1.12.0         utf8_1.2.4                  XVector_0.44.0              BiocGenerics_0.50.0         pillar_1.9.0                stringr_1.5.1               vroom_1.6.5                 GSVA_1.53.4                 later_1.3.2                 splines_4.4.0               flowCore_2.16.0             dplyr_1.1.4                 lattice_0.22-6              bit_4.5.0                   annotate_1.82.0             RProtoBufLib_2.16.0         tidyselect_1.2.1            SingleCellExperiment_1.26.0 Biostrings_2.72.1           miniUI_0.1.1.1              IRanges_2.38.1              SummarizedExperiment_1.34.0 stats4_4.4.0                Biobase_2.64.0              devtools_2.4.5              matrixStats_1.4.1           stringi_1.8.4               UCSC.utils_1.0.0            codetools_0.2-20            tibble_3.2.1               
[109] BiocManager_1.30.25         graph_1.82.0                cli_3.6.3                   systemfonts_1.1.0           xtable_1.8-4                munsell_0.5.1               Rcpp_1.0.13                 GenomeInfoDb_1.40.1         png_0.1-8                   XML_3.99-0.17               ellipsis_0.3.2              ggplot2_3.5.1               readr_2.1.5                 blob_1.2.4                  dotCall64_1.1-1             profvis_0.4.0               urlchecker_1.0.1            sparseMatrixStats_1.16.0    SpatialExperiment_1.14.0    GSEABase_1.66.0             scales_1.3.0                purrr_1.0.2                 crayon_1.5.3                rlang_1.1.4                 KEGGREST_1.44.1
  1. Barbie, D. A., Tamayo, P., Boehm, J. S., Kim, S. Y., Susan, E., Dunn, I. F., . Hahn, W. C. (2010). Systematic RNA interference reveals that oncogenic KRAS- driven cancers require TBK1, Nature, 462(7269), 108-112. https://doi.org/10.1038/nature08460

  2. Abazeed, M. E., Adams, D. J., Hurov, K. E., Tamayo, P., Creighton, C. J., Sonkin, D., et al. (2013). Integrative Radiogenomic Profiling of Squamous Cell Lung Cancer. Cancer Research, 73(20), 6289-6298. http://doi.org/10.1158/0008-5472.CAN-13-1616

enrichment GSVA ssGSEA • 712 views
ADD COMMENT
0
Entering edit mode
Robert Castelo ★ 3.4k
@rcastelo
Last seen 2 days ago
Barcelona/Universitat Pompeu Fabra

Hi, I can only speak for the GSVA implementation of the ssGSEA algorithm, which as you already know, it corresponds to the algorithm described in the publication by Barbie et al. (2009) and which included these two parameters on the exponent of the random walk terms and whether the result should be normalized to the range of the resulting values. The README file of the repo at https://github.com/broadinstitute/ssGSEA2.0 corresponding to other ssGSEA algorithm you're referring to, ssGSEA 2.0, says "This is an updated version of the original ssGSEA R-implementation", so I don't see why we should expect both algorithms, the first version and version 2.0, to give exactly the same results if the authors themselves say that 2.0 is an "updated version" of the first ssGSEA algorithm.

I looked briefly at the code in https://github.com/broadinstitute/ssGSEA2.0/blob/master/src/ssGSEA2.0.R, where the ssGSEA2() function goes from line 49 to 1193, so it's difficult to see at a first glance where the two algorithms differ. If you would like to have an implementation of ssGSEA 2.0 as part of the GSVA package, please open a new issue at https://github.com/rcastelo/GSVA/issues describing why do you think GSVA should have such an implementation, given that there is already an available R implementation by the authors of the ssGSEA algorithm.

ADD COMMENT
0
Entering edit mode

Thanks for the prompt answer! I looked into the "original ssGSEA R-implementation" (https://github.com/GSEA-MSigDB/ssGSEA-gpmodule), unnormalized ssGSEA2 and unnormalized GSVA, and they seem to be near identical. The difference lies pretty much entirely in the normalization strategy. ssGSEA2 seems to use NES (which relies on permuting the gene sets), and GSVA ssGSEA uses the score range to normalize it. ssGSEA 1.0 does not have a post-projection normalization option. We do not need NES in particular.

I have one more question, though: I was wondering if Barbie et al 2009 proposed a rationale for their normalization. In the GSVA documentation, it says that this strategy is outlined on "(Barbie et al. 2009, online methods, pg. 2)". However I don't find this online methods in the current version (at least in front of the paywall). Do they justify it or just simply state that this is a way to normalize the scores?

Thanks again!

ADD REPLY
0
Entering edit mode

Thanks for your analysis of the differences between ssGSEA 2.0 and ssGSEA 1.0, as I said in my previous message, please file an issue in the GitHub repo if you would like to have version 2.0 as part of the GSVA package. Regarding how the normalization step is described in the original article, this is the sentence:

"Signature values were normalized using the entire set of 128 lung adenocarcinomas and 17 normal lung specimens."

located in the last page of the PDF version. If I recall correctly, because this happened more than 10 years ago, I looked up back then the Java code of the ssGSEA module in the GenePattern suite and found that normalization step done in the way GSVA is implementing it. You can email the authors of ssGSEA and I'm sure they'll be happy to provide you a PDF copy of their article.

ADD REPLY
0
Entering edit mode
@df6c68e9
Last seen 14 days ago
United States

The ssGSEA2.0 code permutes the sorted vector of values in order to calculate permutation ES, rather than permuting the gene labels. See line 741 of https://github.com/broadinstitute/ssGSEA2.0/blob/master/src/ssGSEA2.0.R where a vector of random indices is passed to gsea.score() and then used by that function to reorder the sorted vector of input values. This causes the permutation ES to be much nearer to 0 than they should be, which then leads to larger NES than expected. I think this is an error, and I have brought up the issue to the authors, so I would recommend running both methods with the weight parameter set to 0 for the sake of comparisons. This is the exponent to which the absolute values of the sample data are raised. I believe this is the same as the alpha parameter in GSVA::ssgseaParam().

ADD COMMENT

Login before adding your answer.

Traffic: 552 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6