Any given gene can occur more than once in a data set platform (e.g. microarray). This is due to the occurrence of two or more probes (or probe sets) that target the same gene. On most microarray designs we have looked at, somewhere in the neighborhood of 30 percent of the probes are “repeats”, though this depends on the design. We refer to these as “gene replicates”, because they provide replicate measures of the same gene. However, these “replicates” may not be equivalent: they may target different splice forms, or have different sensitivity or specificity. In some cases a probe set may not work at all and give very poor signals, while another probe for the same gene gives a robust signal.
In a gene set score analysis, it does not make sense to count each of these “replicates” independently when determining the size of a gene set: a gene set that consists of five replicates of the same gene should not be considered five genes in the statistical analysis. Thus in ermineJ a gene is only counted once. For the correlation analysis, each pair of genes is only counted once. This means that even if there are five occurrences of a gene on the array, it will only be counted once.
What remains to determine is how to summarize the results for the replicates: the five replicates in our example have to be distilled down to a single value.
ErmineJ offers two ways of dealing with this situation, one which is conservative and one which, while less conservative, might be sensible when there is uncertainty as to the reliability of individual replicates.
The conservative choice is referred to as “Mean“: the different occurrences of the gene are each given equal weight. For correlation analysis, the same rule applies but at the level of pairs of genes. Thus if there are two replicates of gene A and two replicates of gene B, a total of 4 comparisons is possible. The final contribution of the A-B comparison is 1/4 of the correlations measured between all the replicates. This method makes the most sense to use if your replicates are usually actual spot replicates, that is, the same sequence occuring multiple times on the array.
The less conservative choice is referred to as “Best“: the only score counted for a gene is the best one. Thus if a gene has two replicates with scores 3 and 4 (- log p values), then only the 4 is counted; with the mean method the final value would be 3.5. For correlation analysis, the best pairwise correlation is stored. In the A-B example give above, this means that the best of the four comparisons is kept. This method might make more sense if your “replicates” tend to target different transcripts or sets of transcripts
Which method should you use?
The choice of which method to use depends on how conservative you want to be, combined with knowledge you have about the platform design. If you consider gene “replicates” not to really be replicates, then using the “best” option might make sense. If your platform design just has exact replicate spots of the same sequence, then using “Mean” might be sensible.