Tech Notes

Q. What changes have been made to the error model in GeneSpring 6.1?

A. GeneSpring 6.1 includes error model changes that improve the detection of down-regulated differentially expressed genes when using the log-ratio interpretation.

The GeneSpring Cross-Gene Error Model works by fitting a smooth curve to replicate variability (or deviations from 1 in the case of no replicates) resulting in error estimates that are based on pooling variability information for genes at similar expression levels. The model is fit to the data in normalized ratio mode; when working in log mode, the error model estimates need to be converted. In previous versions of GeneSpring, this conversion was done using a form of the delta method, based on Taylor series expansion of the log transformation. While this often works well, it sometimes results in down-regulated genes being assigned large error estimates, relative to up-regulated genes.

GeneSpring 6.1 replaces the delta method calculation with one based on assuming that the normalized ratio data come from a log-normal distribution; that is, that expression values in log mode are approximately normal. Under this assumption, it is straighforward to convert error estimates between normalized ratio and log modes, and better balance between down- and up-regulated genes is achieved.

The following references provide details on the GeneSpring 6.1 error model:

1995 David M. Rocke and Stefan Lorenzato, "A Two-Component Model for Measurement Error in Analytical Chemistry," Technometrics, 37, 176-184.

2001 David M. Rocke and Blythe Durbin, "A Model for Measurement Errors for Gene Expression Arrays," Journal of Computational Biology, 8, 557-569.