A comparison of rotation-based scores for gene set analysis

Gene-wise differential expression is usually the first major step in the statistical analysis of high-throughput data obtained from tech-niques such as microarrays or RNA-sequencing. The analysis at gene level is often complemented by the screening of the data in a broader biological context that considers as unit of analysis meaningful groups of genes that may have functions in certain biological processes. Among the vast number of publications about gene set analysis, the rotation test for gene set analysis, also referred by roast, is a general sample randomization approach that maintains the integrity of the intra-gene set correlation structure in defining the null distribution of the test. In this work we compare the performance of several enrichment score functions using such rotational approach for hypothesis testing. We find that computationally intensive measures based on Kolmogorov-Smirnov statistics fail to improve the rates of simpler measures of GSA like mean and maxmean scores. We also show the importance of ac-counting for the gene linear dependence structure of the testing set, which it is linked to the loss of effective signature size. In this regard, weighted statistics are introduced with the aim of maximizing the ef-fective signature size. These are found to out-power other usual scores in some simulations scenarios. The average of absolute values is found to be the most powerful score using both simulated and benchmarking data. All tools are available in the roastgsa R package.