Feature selection is a key component in microarray data analysis. This is due to the fact that microarray datasets consists of features that are far exceed the number of instances. High dimensional data are also known to contain significant amount of noise and irrelevant variables that do not contribute to classification tasks and may even hinder classification performance. In this paper, a feature selection method which consists of two stages is proposed. At the first step, feature selection is done through a stacked Restricted Boltzmann Machines by means of comparing the error between reconstructed data and the original data. The next stage will use Partial Least Square to extract synthesis features from the previously selected features that will be then used for classification. The performance of the proposed method is done through the classification of ten microarray datasets that are widely used. The proposed model is able to out perform state-of-the-art in 2 datasets, namely 82.11% for GLIOMA and 72.39% for Breast datasets.