Establishment of predictive blood-based signatures in medical large scale genomic data sets

Abstract

Increasing data has led to tremendous success in discovering molecular biomarkers based on high throughput data. However, the translation of these so-called genomic signatures into clinical practice has been limited. The complexity and volume of genomic profiling requires heightened attention to robust design, methodological details, and avoidance of bias. During this thesis, novel strategies aimed at closing the gap from initially promising pilot studies to the clinical application of novel biomarkers are evaluated.<br /> First, a conventional process for genomic biomarker development comprising feature selection, algorithm and parameter optimization, and performance assessment was established. Using this approach, a RNA-stabilized whole blood diagnostic classifier for non-small cell lung cancer was built in a training set that can be used as a biomarker to discriminate between patients and control samples. Subsequently, this optimized classifier was successfully applied to two independent and blinded validation sets. Extensive permutation analysis using random feature lists supports the specificity of the established transcriptional classifier.<br /> Next, it was demonstrated that a combined approach of clinical trial simulation and adaptive learning strategies can be used to speed up biomarker development. As a model, genome-wide expression data derived from over 4,700 individuals in 37 studies addressing four clinical endpoints were used to assess over 1,800,000 classifiers. In addition to current approaches determining optimal classifiers within a defined study setting, randomized clinical trial simulation unequivocally uncovered the overall variance in the prediction performance of potential disease classifiers to predict the outcome of a large biomarker validation study from a pilot trial. Furthermore, most informative features were identified by feature ranking according to an individual classification performance score.<br /> Applying an adaptive learning strategy based on data extrapolation led to a datadriven prediction of the study size required for larger validation studies based on small pilot trials and an estimate of the expected statistical performance during validation. With these significant improvements, exceedingly robust and clinically applicable gene signatures for the diagnosis and detection of acute myeloid leukemia, active tuberculosis, HIV infection, and non-small cell lung cancer are established which could demonstrate disease-related enrichment of the obtained signatures and phenotype-related feature ranking.<br /> In further research, platform requirements for blood-based biomarker development were exemplarily examined for micro RNA expression profiling. The performance as well as the technical sample handling to provide reliable strategies for platform implementation in clinical applications were investigated.<br /> Overall, all introduced methods improve and accelerate the development of biomarker signatures for molecular diagnostics and can easily be extended to other high throughput data and other disease settings.