Feature selection algorithms for CGH data

Comparative Genomic Hybridization (CGH), when combined with microarray technology, measures copy number alterations (gains or losses of DNA segments) of a large number of genomic intervals. Selecting a small number of discriminative intervals from these genomic intervals is an important problem, known as feature selection, for accurate classification of CGH data. An important aspect of CGH data is that consecutive intervals are highly correlated This paper considers the problem of feature selection for multiclass CGH data. There are two phases in our approach. The first phase reduces the number of features by using Support Vector Machines. We develop a novel approach to find the important intervals using the classifiers that are generated. In this approach, K classifiers are built to learn each of the K different cancer types. The second phase reduces the features found from the first phase further by using a novel dynamic programming algorithm. This algorithm removes redundant proximate features by exploiting the fact that consecutive intervals in CGH data are highly correlated. We also propose a systematic way to select CGH samples from a large CGH dataset at predefined similarities and sizes. These samples can serve as benchmarks for research on CGH data. The experimental results on these datasets demonstrate that our approach can reduce the number of features by a significant amount without reducing the predictive accuracy of classification.