Feature selection is one of the most important research topics in high dimensional array data analysis. We propose a two-way filtering based method that utilizes a pair of statistics coupled with rigorous cross-validation to identify the most informative features from different types of distributions. We evaluate the utility of the proposed adaptive feature selection method on six MicroArray Quality Control Phase II (MAQC-II) datasets. The results show that our method yields models with significantly fewer features and can achieve comparable or superior classification performance compared to models generated from other feature selection methods, suggesting high quality feature selection.