Sensitivity analysis of gene ranking methods in phenotype prediction.

INTRODUCTION: It has become clear that noise generated during the assay and analytical processes has the ability to disrupt accurate interpretation of genomic studies. Not only does such noise impact the scientific validity and costs of studies, but when assessed in the context of clinically translatable indications such as phenotype prediction, it can lead to inaccurate conclusions that could ultimately impact patients. We applied a sequence of ranking methods to damp noise associated with microarray outputs, and then tested the utility of the approach in three disease indications using publically available datasets. MATERIALS AND METHODS: This study was performed in three phases. We first theoretically analyzed the effect of noise in phenotype prediction problems showing that it can be expressed as a modeling error that partially falsifies the pathways. Secondly, via synthetic modeling, we performed the sensitivity analysis for the main gene ranking methods to different types of noise. Finally, we studied the predictive accuracy of the gene lists provided by these ranking methods in synthetic data and in three different datasets related to cancer, rare and neurodegenerative diseases to better understand the translational aspects of our findings. RESULTS AND DISCUSSION: In the case of synthetic modeling, we showed that Fisher's Ratio (FR) was the most robust gene ranking method in terms of precision for all the types of noise at different levels. Significance Analysis of Microarrays (SAM) provided slightly lower performance and the rest of the methods (fold change, entropy and maximum percentile distance) were much less precise and accurate. The predictive accuracy of the smallest set of high discriminatory probes was similar for all the methods in the case of Gaussian and Log-Gaussian noise. In the case of class assignment noise, the predictive accuracy of SAM and FR is higher. Finally, for real datasets (Chronic Lymphocytic Leukemia, Inclusion Body Myositis and Amyotrophic Lateral Sclerosis) we found that FR and SAM provided the highest predictive accuracies with the smallest number of genes. Biological pathways were found with an expanded list of genes whose discriminatory power has been established via FR. CONCLUSIONS: We have shown that noise in expression data and class assignment partially falsifies the sets of discriminatory probes in phenotype prediction problems. FR and SAM better exploit the principle of parsimony and are able to find subsets with less number of high discriminatory genes. The predictive accuracy and the precision are two different metrics to select the important genes, since in the presence of noise the most predictive genes do not completely coincide with those that are related to the phenotype. Based on the synthetic results, FR and SAM are recommended to unravel the biological pathways that are involved in the disease development.