Presented at:
2005 Meeting of the Federation of Clinical Immunology Societies, Boston, MA, May 12-16, 2005.
NUSE and RLE: Quality assessment of oligonucleotide microarray data to quantify systemic variation
Collin F, Asare AL, Kolchinsky SA, Speed TP, Seyfert-Margolis VL
Immune Tolerance Network, Bethesda, MD; University of California, Berkeley, CA.
Background: Measurement of differential RNA expression in multicenter clinical trials requires special attention to the quality in sample preparation. Sample collection and handling can adversely affect results; therefore, quality metrics are required to detect and assess the potential errors induced by these factors. We compared the reliability of newly devised quality metrics derived from fitted statistical models of probe level data from high-density oligonucleotide microarrays and compared them to standard GeneChip® microarray quality metrics.
Method: We devised metrics using the Robust Multichip Analysis (RMA) process for deriving probe set summaries from GeneChip® microarrays. The first metric, Normalized Unscaled Standard Error (NUSE), provides a measure of relative chip quality derived from the residuals from the RMA model. The second metric, the Relative Log Expression (RLE), is an absolute metric that gauges variability of expression measures by summarizing the distribution of relative log expressions within a set of microarrays against a reference set. The RLE summaries are sensitive to technical sources of variability that are large compared to biological variation. These metrics were compared to standard quality metrics: Percent Present calls, GAPDH 3'/5', Background, and Scaling Factor. Two clinical trial sample sets were assessed: 368 microarrays from a Type I diabetes trial and 350 arrays from a ragweed allergy trial. A set of standard normal human control samples were sent blinded within patient sets during the course of ITN clinical trials and were used as the reference set against which RLE assessments were made.
Result: NUSE and RLE metrics detected systematic variation within certain participant sets that were not detectable using the standard Affymetrix quality metrics. Elevated GAPDH 3'/5' ratios typically cited as an indicator of poor quality RNA showed no relationship to quality when applying the NUSE and RLE (cor, 0.09). Hybridization/washing artifacts were easily visualized by plotting NUSE residuals. While not always true, Percent Present calls provided the closest approximation to NUSE and RLE; in cases of extremely low Percent Present, NUSE and RLE are adversely affected (cor, -0.50). In both trials in which these metrics were applied, we identified chips within a participant time series that required exclusion from the analysis that would not have been discovered otherwise.
Conclusion : In differentiating NUSE from RLE, NUSE values have no units and can only be used to assess the relative quality of arrays within an analysis set; RLE summaries provide a measure of reproducibility of gene expression data that can be compared across batches, experiments, or trials. Reflecting variability in expression measures, these proposed metrics provide a better basis for judging quality compared to standard metrics.