Interpretation and Transformation of Microarray Data

Wednesday, October 1, 2003 - 11:30am - 11:55am
Keller 3-180
Wolfgang Huber (Deutsches Krebsforschungszentrum (Cancer Research)(DKFZ))
Data from microarray experiments is often reported in the form of logarithmic ratios or logarithm-transformed intensities. This amounts to the assumption that an increase from, say, 1000 units to 2000 units has the same biological significance as one from 10000 to 20000. While this approach is useful for large intensities, it fails when the true level of expression of a gene in one of the conditions is small or zero. However, these genes may be biologically relevant, perhaps even the most relevant ones.

We derive a measure of differential expression that has comparable resolution across the whole dynamic range of expression. Mathematically, it can be expressed in terms of a variance stabilizing transformation. The measure coincides with the log-ratio in those cases where the latter is well-defined, and is a meaningful extrapolation in those cases where the log-ratio is unstable. The measure is closely related to the standardized log-ratio (moving-window z-score), but has more preferable mathematical and computational properties.

We present a parametric statistical model that leads to a robust estimator for the transformation parameters, as well as the between-array normalization parameters. In applications to several benchmark datasets, this approach compares favorably to other normalization algorithms.