Development of bioinformatics tools to predict disease predisposition from Next Generation Sequencing (NGS) data.

The sequencing of the human genome has opened up completely new avenues in research and the notion of personalized medicine has become common. DNA Sequencing technology has evolved by several orders of magnitude, coming into the range of $1,000 for a complete human genome. The promise of identifying genetic variants that influence our lifestyles and make us susceptible to diseases is now becoming reality. However, genome interpretation remains one the most challenging problems of modern biology. The focus of my PhD project is the development of bioinformatics tools to predict diseases predisposition from sequencing data. Several of these methods have been tested in the context of the Critical Assessment of Genome Interpretation (CAGI), always achieving good prediction performances. During my PhD project I faced the complete spectrum of challenges to be address in order to translate the sequencing revolution into clinical practice. One of the biggest problem when dealing with sequencing data is the interpretation of variants pathogenic effect. Dozens of bioinformatics tools have been created to separate mutations that could be involved in a pathogenic phenotype from neutral variants. In this context the problem of benchmarking is critical, as prediction performance are usually tested on different sets of variants, making the comparison among these tools impossible. To address this problem I performed a blinded comparison of pathogenicity predictors in the context of CAGI, realizing the most complete performance assessment among all the iterations of this collaborative experiment. Another challenge that needs to be address to realize the personalized medicine revolution is the phenotype prediction. During my PhD I had the opportunity to develop several methods for the complex phenotype prediction from targeted enrichment and exome sequencing data. In this context challenges like misinterpretation or overinterpretation of variants pathogenicity have emerged, like in the case of phenotype prediction from the Hopkins Clinical Panel. In addition, other complementary issues of phenotype predictions, like the possible presence of incidental findings have to be considered. Ad hoc prediction strategies have been defined while facing with different kinds of sequencing data. A clear example is the case of Crohn’s disease risk prediction. Always in the context of the CAGI experiment, three iterations of this prediction challenge have been run so far. Analysis of datasets revealed how population structure and bias in data preparation and sequencing could affect prediction performance, leading to inflated results. For this reason a completely new prediction strategy has been defined for the last edition of the Crohn’s disease challenge, exploiting data from Genome Wide Association Studies and Protein Protein Interaction network, to address the problem of missing heritability. Good prediction performance have been achieved, especially for individuals with an extreme predicted risk score. Last, my work has been focused on the prediction of a health related trait: the blood group phenotype. The accuracy of serological tests is very poor for minor blood groups or weak phenotypes. Blood groups incompatibilities can be harmful for critical individuals like oncohematological patients. BOOGIE exploits haplotype tables, and the nearest neighbor algorithm to identify the correct phenotype of a patient. The accuracy of our method has been tested in ABO and RhD systems achieving good results. In addition, our analyses paved the way for a further increase in performance, moving towards a prediction system that in the future could become a real alternative to wet lab experiments