18 Uncertainties of a functional enrichment analysis
This section provides a summary of the paper by Wünsch et al. (2023), which explores uncertainties inherent in functional enrichment analysis. The study critically examines the sources of variability and challenges in this analytical approach, offering insights into improving its reliability and robustness.
18.0.1 Types of FEA
Functional enrichment analysis (FEA) typically involves one of over representation analysis (ORA), gene set enrichment analysis (GSEA) also known as functional class scoring (FCS), and Pathway Topology (PT).
- ORA
- ORA methods are the least complex among the three approaches of FEA.
- ORA methods requires a list of differentially expressed genes that are already analysed in differential expression analysis.
- The background population, the universe, can be a more general set of genes, like those in the human genome or more specific from those observed in an experiment.
- A contingency table is created and the null distribution is modelled using the hypergeometric distribution.
- FCS
- FCS methods aim to aggregate the values of the gene-level statistics (ranks) into gene set-level statistic (enrichment score, ES).
- FCS can be classified as one of FCS I, those that take the expression data as input or FCS II that take a pre-ranked list of genes as input. With the latter, the information of the conditions (phenotypes) of the samples is lost, as such phenotype permutation cannot be performed leaving the choice of null hypothesis to gene set permutation.
- PT
- PT additionally models interactions between the genes. This approach generally scores considerably lower in terms of popularity in the reference database.
18.0.2 Considerations
- Pre-filter expression data: Exclude lowly expressed genes to improve statistical power.
- Handle gene IDs carefully: Convert gene IDs to the required format and remove any duplicates.
- Normalise expression data: Address sample-specific biases to enable fair comparisons between samples.
- Use appropriate methods for differential expression analysis: Recommended methods include limma (voom), DESeq2, and edgeR.
- Select suitable gene-level statistics: For FCS II, choose metrics like moderated t-statistic to rank genes meaningfully.
- Adjust for multiple testing: Ensure your analysis includes a correction for multiple hypothesis testing. Some methods require manual adjustments.
- Choose gene set databases based on biological context: Ensure that the database aligns with the research question and the experimental system.
18.0.3 Recommendation
-
Awareness of Uncertainties:
- Recognise uncertainties in methods, parameter choices, and data preprocessing when conducting Gene Set Analysis (GSA).
- Understand that the method’s name alone does not capture the full analysis pipeline.
Clearly document all analysis choices, including methods, parameters, and preprocessing steps.
Select methods, parameters, and preprocessing steps before starting the analysis to minimise bias.
-
Set Technical Parameters:
- Fix technical parameters like the random seed and number of permutations before running the analysis to ensure reproducibility.
- Avoid adjusting these parameters to obtain favourable results.
-
Avoid Cherry-picking:
- Refrain from selectively reporting results based on favourable outcomes, as this can lead to over-optimistic and non-reproducible findings.
- Avoid excessive tweaking of the analysis strategy to fit the data post hoc.
Use different pipelines or parameter configurations as part of sensitivity analysis to check the consistency of results.
Share complete analysis workflows, including code and documentation, to allow others to replicate the findings accurately.