Fig. 4.

Inter- and intralab experimental reproducibility. (A) Expression change estimates derived from our physically grounded model are consistent across intralab, interlab, and dye-swap replicates. To test this, we used published quality-control data following the design scheme described in ref. 4. Each of these four experiments was replicated five times, in three different labs (see SI Text). The experimental design included a number of technical replicates in which the same sample is labeled with different agents, such that R_i = R⁰. Because they are measurements of the absence of expression change, these dye-swap experiments are useful for assessing the specificity of the data and the model used to extract them. We estimated expression changes for four models and computed the correlation over all pairs of dye-swap chips (light gray, Fig. S3). (B) The mean correlation between R_i predicted from our physically grounded model is 0.10, lower than those derived from the statistical (Eq. 2) model, the lowess-corrected statistical model, and the MicroArray Quality Control (MAQC)-reported expression changes. The mean correlation of R_i estimates derived from chips comparing distinct commercially derived samples (R_i ≠ R⁰) was comparable for all models, but the signal-to-noise ratio (S/N) for the physically grounded model is substantially higher than for the other models. The combination of these two observations leads us to conclude that the statistical power of our model substantially surpasses existing methodologies.