Tutorial: Receiver Operator Characteristic scoring
The receiver operator characteristic (ROC) method is a fast, non-parametric alternative to the ORA and resampling methods for generating gene set scores from gene scores.
The ROC is a well-known method for evaluating rankings of items, in this case genes. The ranking in this case comes from the gene scores. A gene set will get a good ROC if many genes in the gene set are near the top of the list.
The score measured for each gene set is the area under the ROC curve, a value between 0 and 1. If the genes in the gene set are randomly distributed in the ranking, you would expect a value near 0.5. Values near 1 indicate the genes in the gene set are near the top of the list, while values near 0 indicate the genes in the gene set are near the bottom of the list. In principle both values near 0 and near 1 are statistically significant, but p-values reported by ermineJ are based on the assumption that only the top of the list is of interest (e.g., we’re not considering “under-representation analysis”).
Unlike the other methods in ermineJ other than the PRC method, the ROC uses only the ranks of the gene scores. That is, all it cares about is the ordering of items obtained by your gene scores (e.g., t-test or fold-change), but doesn’t use the information about the relative values of the scores.
P-values for this analysis are computed using algorithms described in Breslin et al., 2004*. For more information on the ROC, you could do worse than reading the Wikipedia page (http://en.wikipedia.org/wiki/Receiver_operator_characteristic).
When to use ROC scoring
Like other non-parametric techniques, using ranks costs some statistical power, but also makes fewer assumptions. Specifically, if you think the ordering of items in your data is more accurate than the actual p-values themselves, the ROC might be appropriate. The PRC method is similar in that it uses ranks, but puts more emphasis on genes in the set which are ranked very near the top. In contrast the ROC method looks at overall trends in the rankings.
If you have read the ORA page (please do), you will recognize the first 5 steps of the wizard. The only different one is the last step, where the only difference is no p-value threshold is needed.
There is another explanation of gene scores here.