Predicting prostate cancer progression with penalized logistic regression model based on co-expressed genes

The prediction of cancer progression is one of the most challenging problems in oncology. In this paper, we apply the penalized logistic model to microarray data in combination with co-expression genes to identify patients with prostate cancer progression. Compared with conventional methods, penalized logistic regression (PLR) has some advantages such as providing an estimate of the probability in classification label, genetic interpretation of regression coefficients, and short computation time. We employed the top score pair (TSP) approach to select genes for PLR. The TSP method was originally proposed for binary classification of phenotypes according to the relative expression of one gene pair. In the proposed algorithm of this paper, we first identified co-expressed TSP genes and then used PLR to the microarray data for predicting prostate cancer. We applied the framework to the microarray analysis on prostate cancer progression. We have identified three gene pairs associated with prostate cancer progression for PLR model. We compared our approach with the standard classification techniques such as support vector machines (SVMs), Lasso, and Fisher discriminative analysis (FDA). We found that our method yielded better performance in terms of classification and prediction. Furthermore, it has the advantages to provide the underlying probability of predicting the classification, robust biomarker genes and interpretable regression coefficients.