PMC:514539 / 5973-10835 JSONTXT

{"project":"2_test","denotations":[{"id":"15307894-5279523-8142006","span":{"begin":1905,"end":1907},"obj":"5279523"},{"id":"15307894-10647931-8142007","span":{"begin":3088,"end":3090},"obj":"10647931"}],"text":"Discussion\nOur initial results are compelling in that they suggest that we can expect biomarkers of high clinical significance for subsets of patients to be important for distinct subsets of patients. This also suggests that clinical validation of the utility of biomarkers should examine panels of expression biomarkers, not individual biomarkers. Disruption of genomic function via these patterns cannot be studied in the population level biomarker framework for the simple reason that methods that compare, say, group means, will find no difference between the sample groups if the number of case samples found in the two tails are even approximately equal. This is a sensible approach even from within the framework of population-based hypothesis testing, because the PPST test can be expected to be more robust to one or two outliers that might mislead simple parametric tests. Note that a number of genes are 'nearly significant' under the t-test but are strongly significant under the PPST test for the AB/BA patterns (e.g., Table 2).\nOur re-analysis of two independently generated data sets on astrocytoma progression demonstrates the utility of extending analysis to include a search for genes that are differentially expressed in a subset of patients. Of the tests examined, the parametric t-test showed the least internal consistency, while the PPST exhibited the highest internal consistency in identifying progression markers. Comparison to the non-parametric t-tests demonstrates that PPST is most similar to the nonparameteric t-test, but is more self-consistent. While the ABA test showed the least internal consistency across populations, it also exhibited low overlap with any other test, so the genes reported are unique and tend not to be found by others tests, matching expectations.\nOur results are consistent with Knudsen's 'two-hit' hypothesis on the genomic etiologies of cancer [49] with some insight into the diversity of genomic pathologies (functional 'hits') that may be relevant in patient populations. Studies of differential gene expression – and its role in the etiology of cancer and its responses to treatment – should seek these types of genes in addition to population-wide biomarkers, because they represent a subset of the genes that are expressed differentially in a significant subset of cancer patients. We recommend a major shift in perspective on the study on gene expression dysregulation away from the study of 'tumor populations' – which do not exist – toward the study of genomic pathologies in individual patients. For example, tumor subtypes are typically characterized by morphological characters, and these classifications may conflict with important expressotype subtypes that do not follow classical morphological tumor classes. Imposition of these subtypes on a study design may interfere with identifying expressotypes that provide high diagnostic, prognostic and theranostic value to the individual – and sets of individuals with similar expressotypes. This view is also consistent with the Hanahan-Weinberg model of oncogenesis [50], which envisions multiple possible mechanistic strategies to the acquisition of characteristics and capabilities of cancers including self-sufficiency in growth signals, insensitivity to anti-growth signals, tissue invasion \u0026 metastasis, limitless replicative potential, sustained angiogenesis and evasion of apoptosis. We also expect that individual cancers in different patients will be found to have evolved unique sets of solutions to each of these problems. Current prevailing methods for finding differentially expressed genes such as fold-change and t-tests do not allow for such complexities.\nOur comparison of the methods (Table 3) highlights the uniqueness of the ABA test. It is an extension of the PPST test; it specifically focuses on genes that are differentially expressed in subsets of patients. This ability is extremely important in search of genes with expression patterns that correlate with drug response. The ABA and the two-tailed t-test are not comparable because the ABA test allows us to find genes that the t-test specifically cannot (genes that are simultaneously overexpressed in some patients while underexpressed in others). Such test will have high variance (leading to a low t-test score) and low mean difference, and will thus not be significant. The PPST and the ABA tests extend our abilities beyond the t-test. Other improvements or even superior alternatives to these tests may be possible. The performance of these tests and all tests described to date for the AB type patterns and now for ABA patterns should be compared using extensive numerical simulations and cross-validation. Developments are needed to determine how best to select a threshold to allow deliberate control of the false positive and false negative error rates."}