Our work aims to perform the feature selection step on Multiple Kernel Learning by optimizing the Kernel Target Alignment score. It begins by building feature-wise gaussian kernel functions. Then by a constrained linear combination of the feature-wise kernels, we aim to increase the Kernel Target Alignment to obtain a new optimized custom kernel. The linear combination results in a sparse solution where only few kernels survive to improve KTA and consequently a reduced feature subset is obtained. Reducing considerably the original gene set allow to study deeper the selected genes for clinical purposes. The higher the KTA obtained, the better the feature selection, since we want to build custom kernels to use them for classification purposes later. The final kernel after optimizing the KTA is built by a linear combination of ‘Ki’ kernels, each one associated to a μi coefficient. The μ vector is computed during the optimization process.Sociedad Argentina de Informática e Investigación Operativa