Accurate diagnosis of cancer is of great importance due to the global increase in new cancer cases. Cancer researches show that diagnosis by using microarray gene expression data is more effective compared to the traditional methods. This study presents an extensive evaluation of a variant of Deep Belief Networks - Discriminative Deep Belief Networks (DDBN) - in cancer data analysis. This new neural network architecture consists Restricted Boltzman Machines in each layer. The network is trained in two phases; in the first phase the network weights take their initial values by unsupervised greedy layer-wise technique, and in the second phase the values of the network weights are fine-tuned by back propagation algorithm. We included the test results of the model that is conducted over microarray gene expression data of laryngeal, bladder and colorectal cancer. High dimensionality and imbalanced class distribution are two main problems inherent in the gene expression data. To deal with them, two preprocessing steps are applied; Information Gain for selection of predictive genes, and Synthetic Minority Over-Sampling Technique for oversampling the minority class samples. All the results are compared with the corresponding results of Support Vector Machines which has previously been proved to be robust by machine learning studies. In terms of average values DDBN has outperformed SVM in all metrics with accuracy, sensitivity and specificity values of 0.933, 0.950 and 0.905, respectively.