Systemic microRNA measurement, a useful tool for intelligent clinical decision support systems in bladder cancer

Abstract

Abstract Purpose: Urothelial carcinoma of the bladder (UC) is the 4th most common solid malignancy in men and 5th most common overall with an estimated 70,000 new cases of UC and over 14,000 deaths from the disease expected in 2010 in the United States. Although the majority of patients with invasive UC present without radiographic or clinical evidence of disease beyond the bladder, up to 56% of patients die from the result of occult metastasis not detected by current staging modalities. The potential of microRNAs (miRs) as novel tumor markers has been the focus of recent scrutiny because of their tissue specificity, stability, and association with clinicopathological parameters. Prognostic tools based on conventional clinical and pathologic staging can quantify the risk of death from UC, but their accuracy is imperfect due to the heterogeneous biologic behavior of tumors and/or patients with similar clinical and pathologic features. Use of biomarkers specific to the tumor and/or patient can provide prognostic utility over that available from routine clinical features. Data have emerged documenting altered systemic miRs expression across a spectrum of cancers, thus the potential to rapidly screen individuals by analyzing miRs from serum collected by venipuncture is minimally invasive and reproducible and may identify reliable tools for screening or early detection of UC. Our aim was to assess the expression of all known and predicted non-coding RNAs species, contrasting the miRs in the circulation of patients with superficial or invasive UCs or without any cancerous disease, to determine whether we can identify systemic miRs as screening tools for bladder cancer. Patients and Methods: Whole blood samples were prospectively collected from preoperative cancer (n=20) or non-cancerous patients (n=18), with similar age and sex distribution. Total RNA was isolated from plasma, and hybridized with MDACC custom-made non-coding RNAs (ncRNA) arrays, yielding 19200 measurement values per patient. We used the freely available biostatistics suite Bioconductor together with custom algorithms in R for all preprocessing steps and the biomarker optimization. The best biomarker combines two potentially contradictory features: high accuracy and low number of predictors. Several statistical and machine learning methods were tested for developing intelligent clinical decision support systems for diagnosis prediction: random forests of classification trees, nearest shrunken centroids, and logistic regression. For all of them, we applied appropriate techniques to estimate the generalization performance of the obtained classifiers (bootstrapping for the random forests, cross-validation for the other two methods). Results: Screening for differentially expressed miRs by applying t-tests and subsequently computing false discovery rates identified 79 miRs at that are likely to be systematically dysregulated (local FDR < 0.5) in serum of UC patients. None of these 79, however, suffices by itself to distinguish UC cases from control patients. This demonstrates the need for multivariate biomarkers that cleverly combine the expression levels of several miRs into a single prognostic figure. The three classification methods that we used to find such combinations were compared for their predictive accuracy. The logistic regression (LR), preceeded by a coarse unsupervised filtering of miRs, turned out to be the most accurate algorithm in this application scenario. In a leave-one-out cross-validation (LOO-CV) it was able to predict the presence or absence of bladder cancer correctly in 87% of cases, based on the expression levels of 568 miRs and the gender of the patients. We also applied LR to separate invasive BC from other cases (which could be either superficial BC or normal), where we achieved 92% LOO-CV accuracy. For the 3-way classification problem, the accuracy was 79%. Investigation of the computational results suggests that the analysis of a larger data set could likely lead to more accurate classifiers based on fewer miRs. We observe a relatively high prognostic utility of the gender, roughly on par with that of the most informative miRs, which is not surprising given the vast difference of incidences of UCs in males and females. Conclusion: These findings suggest that invasive UCs display specific systemic miR profiles, which could aid in discriminating among other pathologies usually encountered at similar age groups of patients. This finding is of notable clinical consequence and predicts unlimited potential for impacting clinical practice patterns by directing appropriate management of this disease and reducing death from UC. The use of reliable markers also has the potential for reducing the cost of health care delivery by improving and streamlining surveillance protocols, and by personalizing the therapy.