Correlation and variation-based method for identifying reference genes from large datasets

Reference genes are assumed to be stably expressed under most circumstances. Previous studies have shown that identification of potential reference genes using common algorithms, such as NormFinder, geNorm, and BestKeeper, are not suitable for microarray-sized datasets. The aim of this study was to evaluate existing methods and develop methods for identifying reference genes from microarray datasets.

We evaluated the correlation between outputs from 7 published methods for identifying reference genes, including NormFinder, geNorm, and BestKeeper, using subsets of published microarray data. From these results, seven novel combinations of published methods for identifying reference genes were evaluated.

Our results showed that NormFinder's and geNorm's indices had high correlations (R(2) = 0.987, P < 0.0001), which is consistent with the findings of previous studies. However, NormFinder's and BestKeeper's indices (R(2) = 0.489, 0.01 < P < 0.05) and NormFinder's coefficient of variance (CV) suggested a lower correlation (R(2) = 0.483, 0.01 < P < 0.05). We developed two novel methods with high correlations with NormFinder (R(2) values of both methods were 0.796, P < 0.0001). In addition, computational times required by the two novel methods were linear with the size of the dataset.

Our findings suggested that both of our novel methods can be used as alternatives to NormFinder, geNorm, and BestKeeper for identifying reference genes from large datasets. These methods were implemented as a tool, OLIgonucleotide Variable Expression Ranker (OLIVER), which can be downloaded from http://sourceforge.net/projects/bactome/files/OLIVER/OLIVER_1.zip.