Multi-TGDR: An Extension of the Threshold Gradient Descent Regularization for Multiclass Classification of Microarray Experiments

• Activity Number: 367
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Mayte Suarez-Farinas*+ and Suyan Tian
• Companies: Rockefeller University and First Hospital of the Jilin University
• Keywords: TGDR ; classification ; microarray ; bagging ; regularization ; feature-selection

Abstract

While regularization methods has been successfully used in classification of high-throughput data, most algorithms do not address multiple classes and data from several independent experiments, both commonly encountered in practice. Here, we extend the Threshold Gradient Descent Regularization (TGDR) algorithm to the multi-class setting. When several microarray experiments are combined to build a classifier, a meta-analysis version of TGDR that combines the individual classifiers with the same model/structure but with study-varying parameters, can be used. Here, we propose an explicit method to estimate the overall coefficients of the Meta-TGDR selected biomarkers, formally allowing the class prediction for samples from an independent study. This feature, broaden the applicability of Meta-TGDR and allows a fairer comparison with TGDR. Using real-world applications, we compare the performance of multi-TDGR and compared with a pairwise-TDGR strategy. We evaluate the effect of batch/experiment adjustment on the performance of TGDR and Meta-TDGR classifiers. We also show that the addition of Bagging procedure to these methods improves stability while maintaining predictive performance.