Feature selection is an important step in gene expression data analysis. However, many feature selection methods exist and a costly experimentation is usually needed to determine the most suitable one for a given problem. This paper presents the application of gradient boosting and neural network techniques for the construction of metamodels that can recommend rankings of {feature selection - classification} algorithm pairs for new gene expression classification problems. Results in a corpus of 60 public data sets show the superiority of these techniques in producing more useful rankings in relation to classical metamodels.