Statistical integration of p-values for enhancing discovery of radiotoxicity gene signatures

The aim of this study is to verify the effectiveness of a statistical integrative approach for merging expression data sets on the level of p-values. The data consist of two independent sets of expression levels for breast cancer patient lymphocyte tissue. The samples represent two groups: sensitive and resistant to the impact of ionizing radiation. Three approaches for integrating information derived from the two experiments to select a radiosensitivity gene signature were investigated: restrictive, non-statistical and integrative. Signature validity was assessed by verifying data separability using the support vector machine procedure and a logistic regression model selected using the likelihood ratio test to account for regularization. We demonstrated the value of additional information retained in the statistical method of gene selection based on combined p-values, which as the only logistic regression model in its optimal setting attained 100 % separability (AUC 86.2 % restrictive and 85.6 % non-statistical). Moreover, for the best support vector machine classifier, our p-value combination approach outperformed the restrictive and Arraymining methods (AUC 96.7 % versus 87.9 % and 94.6 % respectively). Further functional validation by signaling pathway research proved conjointly the biological relevance of the supreme gene signature.