Distinct Gene Profiles for Tumor and Non-Tumor Tissue in the Head and Neck: An Analytical Approach.

In a study of genetic alterations, the Multiplex Ligation-dependent Probe Amplification (MLPA) assay was used to measure gain or loss of 113 gene-probes in tumor and non-tumor tissue samples collected from each of the 220 patients with squamous head and neck cancer (HNSCC). Conditional and marginal models were available; both models account for correlated data but have different aspects. The conditional logistic regression model was proposed to estimate the subject-specific risk of tumor based on the paired tumor and non-tumor data collection, which was in contrast with the marginal model to estimate population-average risk.The modeling process included rigorous variable selection, an initial multivariable model, a final model selection, and model validation. Genes with individual effect (p<0.01) were considered as candidates for the initial multivariable model for tumor. The final model included gene-probes with p<0.01 and estimations of odds ratios (OR) 95% Confidence Intervals (CIs) and the model's predictive ability, measured by the receiver operating characteristic curve (ROC). A 10-fold cross-validation was performed to validate the model. Of 113 gene-probes, using the conditional approach, 16 genes in 7 chromosomes, remained in the final multivariable model with p<0.01 and an ROC score of 0.94. The cross-validation showed ROC mean (SD) score of 0.96(0.04). The marginal model, in contrast, ended with 8 gene-probes and had an observed ROC of 0.81.
CONCLUSION: The conditional approach appears to be the model of choice when assessing gene-probe risks of subjects with paired data collection and fewer missing covariates, compared to the marginal approach. This multiple gene model demonstrated excellent ability to discriminate tumor from non-tumor, and supports its contribution to the pathogenesis of HNSCC as well as their potential utility for further markers of early tumor detection.