GSDB correction based on the previously described control hybridizations greatly reduces self–self variation in this dataset with variable label incorporations (Figure 1F). To ensure that the correction does not simply flatten all differential ratios regardless of whether they are true or false, a series of external RNA controls was spiked in as true positive differentials. The correction applied to the external controls was identical. For hybridizations with a high degree of GSDB (Figure 1G) or for hybridizations with moderate GSDB (Figure 1I), the correction works well, reducing the variation in self–selfs but without affecting spiked in true positive ratios (Figure 1H–J). This conclusion also holds for higher degrees of GSDB, as well as for experiments with low dye incorporation that initially do not seem to suffer from GSDB (Supplementary information). Note that alleviating GSDB by reducing the amount of incorporated dye has the disadvantage of lowering sensitivity (Supplementary information).

WeWe compared the performance of GASSCO with the latest available one, VERA (Kelley et al, 2008), as well as with simply averaging dye swaps (Figure 1K and L), using the data presented in Figure 1C. As both VERA and averaging result in a single merged dataset, we also averaged our individually GSDB-corrected hybridizations for the sake of comparison. The results show clearly that GASSCO outperforms both methods. Importantly, as GASSCO corrects hybridizations individually (Figure 1A–J), it has the additional advantage of conserving statistical power, which is lost by methods that merge. The new method is also computationally less challenging than VERA (500-fold faster in our tests). As the degree of GSDB is associated with labelling efficiency (Figure 1E), both simple merging of dye swaps and methods that do not take into account the slide dependency of dye bias will only work well if label incorporation is identical between dye-swap replicates. In this study, ‘control' of sample labelling efficiency (Figure 1E) was only achieved by performing a number of reactions under varying conditions, matching labelled samples that happened to have the same incorporation. In practice, it is too challenging to control the degree of label incorporation precisely enough, especially for prolonged projects, which is one reason why the degree of GSDB varies (see Figure 3 for other examples of this variation from independent studies).

Identical clustering results were also obtained with iGSDB estimates based on leave-one-out cross-validation (data not shown). To approximate the minimum number of dye-swapped slides needed to get trustworthy iGSDB estimates, we compared estimates derived from all combinations to the original that was based on all five pairs. Using just two pairs of hybridizations resulted in a correlation between the different iGSDB estimates always >0.94, suggesting that three pairs of dye swaps may generally suffice. Although such iGSDB estimates are slightly less accurate than those obtained from self versus self hybridizations, Figure 2 shows that the GSDB correction method can successfully be applied to smaller projects without self versus self control hybridizations.

GSDB_ij is the GSDB of gene i on slide j. iGSDB_i is the intrinsic gene-specific dye bias of gene i and F_j is the slide-dependent factor of slide j. The method consists of the following steps: estimation of the iGSDBs once for an entire project; estimation of the slide-dependent factor for each hybridization; and lastly application of the individual corrections. The full R source code, including detailed documentation and worked examples, is available in the dyebias package from www.holstegelab.nl/publications/margaritis_lijnzaad/ or through BioConductor (www.bioconductor.org).