Adaptive Matrix Metrics for Attribute Dependence Analysis in Differential High-throughput Data

Data-driven metric adaptation is proposed for proteome analysis of 2D-gel electrophoretic plots aiming at identification of stress related proteins in two barley cultivars with different response towards different salt stress conditions. Gradient descent is applied to the ratio of intraand inter-class distance sums to optimize the matrix parameters of generalized Mahalanobis distances in order to separate the several hundred dimensional data of protein intensities in the transformation space. The resulting matrix contains mutual dependence of spots, explaining differential stress reactions and putative protein interactions. We present interesting results obtained by the new metric learning method that possesses general applicability in biomedical data analysis.