Cross-Platform Gene Signature Development

• Activity Number: 611
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Medical Devices and Diagnostics
• Author(s): Prasad Patil*+
• Companies: Johns Hopkins University
• Keywords: Gene signature ; personalized medicine ; cross-platform ; prediction

Abstract

One of the key obstacles in the path from genomic signature creation to use in the clinic is retraining on new microarray platforms. Most genomic markers are developed on one platform and eventually moved to a different platform for production use. Here we investigate the use of top-scoring pairs (TSPs) as a mechanism for cross-platform signature development. Top-scoring pairs are rank-based, using comparisons between the relative levels of multiple raw features as the basis for the predictions. We extend this approach to the creation of classifiers that are invariant to changes in platform and require minimal data pre-processing. Using 28 curated breast cancer microarray datasets spanning 12 different platforms, we show that models based on just five TSPs are comparable to the existing standard for predicting molecular subtypes (which currently require 50 genes). We are also able to merge datasets from different platforms and train models on the resulting "chimera", which vastly improves cross-platform accuracy. These models require no retraining or normalization, making them ideal for robust gene signature development.