Kernel Analysis for Noisy Microarray Data

Microarray technique measures gene expression levels under various conditions, simultaneously. Microarray data are successfully analyzed by kernel methods for a variety of applications. A major drawback of microarray is technically error prone. To gain accurate analysis, we propose a method which produces a noise-tolerant kernel matrix. First of all, we devise a new distance function for microarray data. The distance function is robust to noise, but not metric. Although we need a kernel matrix to apply kernel methods, yet construction of a kernel matrix from the non-metric distances is not an easy task. Our algorithm for building a kernel matrix is based on a maximum-entropy method using constraints derived from the distances. Promising results are shown in classification and network inference. kernel methods, microarray data, partial distance, maximum entropy method, network inference