UC San Diego

Classification of gene expression profiles to distinguish one disease state from another is essential for the realization of personalized medicine. Recent approaches towards this problem use prior knowledge about interaction among bio-molecules to improve classification accuracy and the biological relevance of the predictive features. However, many such network-based methods do not significantly outperform their unconstrained counterparts in terms of sensitivity and specificity due to unexplained reasons. This behavior, observed across diverse datasets and methods, is a cause of concern as it implies that something is wrong with the data, the algorithms or both. This work focuses on understanding the reasons behind this problem through extensive simulation of gene expression profile to help the development of better classifiers in the future. We infer that when using networks whose interactions do not agree well with the patterns of gene expression, improvement in classification performance will not be significant. Because this improvement is dependent on the classifier also, future network-based methods need to understand their properties with respect to network noise and know the quality of actual network mapping to make meaningful inferences from the performance results