A new approach for filtering noise from high-density oligonucleotide microarray datasets

Although DNA microarrays are powerful tools for profiling gene expression, the dynamic range and the sheer number of signals produced require efficient procedures for distinguishing false positive results (noise) from changes in expression that are 'real' (independently reproducible). We have developed an approach to filter noise from datasets generated when high density oligonucleotide-based microarrays are used to compare two distinct RNA populations. First, we performed comparisons between chips hybridized with cRNAs prepared from an identical starting RNA population; an 'Increase' or 'Decrease' call in such a comparison was defined as a false positive. Plotting the average distribution of these false positive signal intensities across 18 such comparisons of nine independent RNA preparations allowed us to develop a series of noise-filtering lookup tables (LUTs). Using a database of 70 separate chip-to-chip comparisons between distinct RNA preparations prepared by different workers at different sites and at different times, we show that the LUTs can be used to predict the likelihood that a given transcript called Increased or Decreased in one comparison will again be called Increased or Decreased in a replicate comparison. Evidence is presented that this LUT-based scoring system provides greater predictive value for reproducible microarray results than imposition of arbitrary foldchange thresholds and accurately predicts which microarray-identified changes will be validated by independent assays such as quantitative real-time PCR.