Notes on Gene Expression Profiling by Microarrays – Clinical Implications

From Hofmann W (editor). 2006. Gene Expression Profiling by Microarrays, Clinical Implications. Cambridge University Press.

“Approximately 5-10% of all breast cancers are of hereditary origin, and two major breast cancer susceptibility genes have been identified to data, BRCA1 [19] and BRCA2 [20]. Although these high-penetrance syndromes account for a small proportion of all cancers, the identification of these genes and the investigation of their roles in breast cancer development and progression emphasize the genetic aspect of cancer in general, and provide a basis for extrapolating findings from ‘genetically defined’ studies to the investigation of the more common forms of sporadic disease.

“While mutation screening in the two known breast cancer susceptibility genes for hereditary breast cancer families, allowing mutation carriers to make informed decisions regarding surveillance and/or prophylactic approaches, has become commonplace at oncogenetic clinics across the world, the techniques used for screening are time-consuming and expensive. Studies of the histopathological features, genomic alterations and hormone receptor levels in these tumors support the notion that breast cancers caused by germline mutations in BRCA1 and BRCA2 differ from each other and from tumors not caused by mutations in these genes at a molecular level. While certain characteristics, such as medullary histology and ER negativity, are more common among BRCA1 derived breast tumors, which make up a somwhat homogeneous group, BRCA2 derived breast tumors and, to an even greater extent, non-BRCA1/2 (BRCAx) breast tumors constitute considerably more heterogeneous groups. Therefore, an alternative means of classifying BRCA1, BRCA2 and non-BRCA1/2 associated tumors would greatly facilitate the identification of patients carrying mutations in these genes. Moreover, comprehensive understanding of the underlying defects causing the development of hereditary breast cancers may greatly improve both treatment strategies and intervention options for the affected patients, and may give insights into breast cancer biology in general.” (p. 136-137)

“While Wang and coauthors, in concordance with van’t Veer and colleaugues, suggest an improvement in the risk assessment of breast cancer patients with the use of gene expression profiles compared to conventional criteria commonly used in Europe (St. Gallen, [42]) and the USA (NIH, [43]), the overlap in genes involved in the two predictors is minimal. Differences in techniques used and, maybe more importantly, patients included may account for a certain degree of these discrepancies; this demonstrates the need for transparency in terms of already published results as well as the implementation of well-designed, large, prospective studies aimed at answering very specific questions in terms of outcome prediction. While gene expression-based techniques lend promise to the improvement of individual patient care, these studies also illustrate the heterogeneity of the disease and that we have not yet reached a point where these applications can be readily incorporated into clinical practice.” (p. 149)

“Response to therapy: As mentioned, although ER status is predictive of response to hormonal treatment, the predictive value is not 100%. Moreover, there are, to date, no clinically useful predictive markers for a patient’s response to chemotherapy. In addition, all patients eligible for chemotherapy receive the same treatment although the average expected benefit is low; some patients fail to respond and suffer unnecessary toxicity, while others would benefit from a more severe treatment. Hence, selection of patients most likely to benefit from adjuvant systemic therapy would be a great advance in the clinical management of breast cancer. The assessment of one or a few individual markers has not been shown to be powerful enough to reveal the complex biology of clinical breast cancer and response to therapy. However, patterns of expression of larger numbers of genes could be successful in distinguishing between sensitive and resistant tumors.” (p. 149)

“As it has been illustrated by a few research groups, probably the most plausible approach for translating microarray-derived gene expression profiles into clinically applicabile routine assays is first to identify diagnostic or prognostic gene expression profiles consisting of a small number of genes using whole genome microarras (and fresh frozen tissues), and then validating the clinical efficacy of these genes in prospective studies using a simple and robust conventional assay such as RT-PCR and formalin-fixed paraffin-embedded tissue. Such an assay could be incorporated readily into clinical practice, and would also be the most cost-effective. While this strategy may improve and refine the stratification of patients greatly into available treatment alternatives, the greatest obstacle in the advancement of clinical oncology is the identification and development of novel, targeted drugs with optimal efficiency.” (p. 157)

“Microarray mRNA expression analysis is a powerful screening technique that tests a sample of 1 ugram total RNA with many thousands of parallel hybridizations. Even if each of these hybridizations is performed with precision and high reproducibility [48], the measures required to protect against false positives (random fluctuations exceeding the permitted threshold of variation) are draconic. For instance, on an array configuration with 45,000 probe sets and 20,000 expressed transcripts, the preset P-value used in differential expression analysis has to be P<10^-4 in order to have two or fewer false positives. Since microarray experiments are expensive, investigators avoid high numbers of repetition, often three times for well-defined cell preparations [48]. This means that, by accepting P<10^-4 in order to prevent false positives, one will lose most true positives as well, a situation that is not desired. We have learned from a number of related microarray experiments on the same biological object that linking the different data sets is a powerful way of guiding the investigator towards appropriate gene prioritization.

“Suppose beta-cell mRNA expression is measured in conditions A vs. B and we want to know which of the 20,000 transcripts called present are truly changed. When the frequency of true positives is 1% and the acceptance threshold P-value of each compared transcript is 1% and the acceptance threshold P-value of each compared transcript is 0.01, the differential expression analysis A vs. B would generate a data set of 200 true and 200 false positives. In terms of follow-up, this is a large project, as the obvious way to find out which are the true and which are false positives is to validate the transcripts one by one with other techniques like real-time RT-PCR. However, such analysis is an enormous task that would fill the time alloted to several Ph.D. students. But suppose that we test beta-cell mRNA expression a second time, now in conditions C vs. D. With a frequency of true positives of 1.5% and a threshold P-value of 0.01, we would now generate a second data set with 300 true and 200 false positives. When we accept as null hypothesis that the two comparisons bear no other than a random relationship in terms of behavior of RNA molecules, then the expected number of true and false positives exhibiting differential expression in A vs. B and C vs. D would be between 4 and 5.” (p. 203)