Cross-Platform Analysis with Binarized Gene Expression Data

With widespread use of microarray technology as a potential diagnostics tool, the comparison of results obtained from the use of different platforms is of interest. When inference methods are designed using data collected using a particular platform, they are unlikely to work directly on measurements taken from a different type of array. We report on this cross-platform transfer problem, and show that working with transcriptome representations at binary numerical precision, similar to the gene expression bar code method, helps circumvent the variability across platforms in several cancer classification tasks. We compare our approach with a recent machine learning method specifically designed for shifting distributions, i.e., problems in which the training and testing data are not drawn from identical probability distributions, and show superior performance in three of the four problems in which we could directly compare.