Comparing gene set analysis methods on single-nucleotide polymorphism data from Genetic Analysis Workshop 16
Tintle, N. L., Borchers, B., Brown, M., & Bekmetjev, A., Comparing gene set analysis methods on single-nucleotide polymorphism data from Genetic Analysis Workshop 16, 2009, BMC proceedings, 3(S7), 96.
The authors write in their abstract that the purpose of this study was to confirm that GSEA and FET are not optimal for the analysis of SNP data when compared with the SUMSTAT.
The following methods were compared:
- Gene Set Enrichtment Analysis (GSEA)
- Fisher's Exact Test (FET)
- SUMSTAT (the sum of the test statistics for the individual genes)
- SUMQ (the sum of the squared test statistics), also termed SAM-GS in the paper
GSEA, SUMSTAT and SUMQ uses 1000 randomly selected gene sets containing the same number of genes as the set of interest as reference.
3 Study outcomes
3.1 Outcome O1
- More gene sets were found by SUMSTAT compared to FET and GSEA.
- There is a large overlap between the considered methods
Outcome O1 is presented as Venn diagram in Figure 1 of the original publication.
3.2 Outcome O2
- All methods have similar numbers of false-positives (i.e. similar numbers of identified gene sets in case there are no associated genes)
- Most gene sets containting associated SNPs are found by SUMSTAT (compared to the other methods)
- All approaches (except FET with cutoff=13.8) fail for the setting "pseudo gene set with 1-2 strongly associated SNPs
- The evaluated approaches seem to perform in the following order: SUMSTAT > SUMQ > GSEA > FET 5.9 > FET 9.2 > FET 13.8 > FET 18.4
Outcome O2 is presented as Table 1 in the original publication.
3.3 Further outcomes
The overall best method (SUMSTAT) was also applied to the experimental data in order to interpret the data.
4 Study design and evidence level
4.1 General aspects
You can describe general design aspects here. The study designs for describing specific outcomes are listed in the following subsections:
4.2 Design for Outcome O1
- The outcome was generated for a rather large experimental data from the so-called Framingham Heart Study.
The data set contained 957 samples, thereof 157 ever had heart disease and 167 ever had diabites.
- The cutoff for FET has been motivated but seems aribtrary to some extent
- Only a singel combination of configuration parameters was considered.
- The two phenotypes (heart disease and diabetis) were jointly analyzed.
- 706 gene sets were analyzed for each of the two phenotypes (1412 gene sets in total)
- One cutoff for FET has been considered
- The tradeoff between true positives and false positives (or sensitivty vs. specificity) in general depends on the cutoff. Since only a single cutoff was investigated and the methods predict different numbers of gene sets, it is difficult to interpret this result.
4.3 Design for Outcome O2
- For this outcome, simulated data for 666 vs. 210 samples was generated.
- The size of this data set is similar as the experimental data (O1), but not the same.
- 2000 SNPs were assumed as weakly assiciated with the phenotype, 19 were assumed as strongly associated
- Real gene sets and pseudo-gene sets were considerd
- 4 cutoffs for FET were evaluated
- The following regulation was simulated:
- real gene sets without regulated genes
- real gene sets with some weakly (1-9) regulated genes
- real gene sets with many weakly (10+) regulated genes
- real gene sets with one to two strongly associated genes (no weakly associated genes)
- pseudo-gene sets without strongly genes
- pseudo-gene sets with a large numbers of weakly associated genes
- pseudo-gene sets with some strongly genes
- It is counterintuitive that the proportion of identified gene sets does not monotonically depend on the chosen cutoff for FET
- There is no distinct evaluation in terms of true positives and false positives. Only the "percent of sets found as significant" are provided.
- For FET, extrapolation to smaller cutoffs seems to possibly outperform the other approaches. However, sensitivity and specificity should be considered.
- There is no graphical depiction of the outcome
5 Further comments and aspects
- The authors conclude that in agreement with other publications, GSEA and FET are less powerful.
- In the paper, there is no discussion about sensitivity vs. specificity
The list of cited or related literature is placed here.