Biomarker lists stability in genomic studies: analysis and improvement by prior biological knowledge integration into the learning process

The analysis of high-throughput sequencing, microarray and mass spectrometry data has been demonstrated extremely helpful for the identification of those genes and proteins, called biomarkers, helpful for answering to both diagnostic/prognostic and functional questions. In this context, robustness of the results is critical both to understand the biological mechanisms underlying diseases and to gain sufficient reliability for clinical/pharmaceutical applications. Recently, different studies have proved that the lists of identified biomarkers are poorly reproducible, making the validation of biomarkers as robust predictors of a disease a still open issue. The reasons of these differences are referable to both data dimensions (few subjects with respect to the number of features) and heterogeneity of complex diseases, characterized by alterations of multiple regulatory pathways and of the interplay between different genes and the environment. Typically in an experimental design, data to analyze come from different subjects and different phenotypes (e.g. normal and pathological). The most widely used methodologies for the identification of significant genes related to a disease from microarray data are based on computing differential gene expression between different phenotypes by univariate statistical tests. Such approach provides information on the effect of specific genes as independent features, whereas it is now recognized that the interplay among weakly up/down regulated genes, although not significantly differentially expressed, might be extremely important to characterize a disease status. Machine learning algorithms are, in principle, able to identify multivariate nonlinear combinations of features and have thus the possibility to select a more complete set of experimentally relevant features. In this context, supervised classification methods are often used to select biomarkers, and different methods, like discriminant analysis, random forests and support vector machines among others, have been used, especially in cancer studies. Although high accuracy is often achieved in classification approaches, the reproducibility of biomarker lists still remains an open issue, since many possible sets of biological features (i.e. genes or proteins) can be considered equally relevant in terms of prediction, thus it is in principle possible to have a lack of stability even by achieving the best accuracy. This thesis represents a study of several computational aspects related to biomarker discovery in genomic studies: from the classification and feature selection strategies to the type and the reliability of the biological information used, proposing new approaches able to cope with the problem of the reproducibility of biomarker lists. The study has highlighted that, although reasonable and comparable classification accuracy can be achieved by different methods, further developments are necessary to achieve robust biomarker lists stability, because of the high number of features and the high correlation among them. In particular, this thesis proposes two different approaches to improve biomarker lists stability by using prior information related to biological interplay and functional correlation among the analyzed features. Both approaches were able to improve biomarker selection. The first approach, using prior information to divide the application of the method into different subproblems, improves results interpretability and offers an alternative way to assess lists reproducibility. The second, integrating prior information in the kernel function of the learning algorithm, improves lists stability. Finally, the interpretability of results is strongly affected by the quality of the biological information available and the analysis of the heterogeneities performed in the Gene Ontology database has revealed the importance of providing new methods able to verify the reliability of the biological properties which are assigned to a specific feature, discriminating missing or less specific information from possible inconsistencies among the annotations. These aspects will be more and more deepened in the future, as the new sequencing technologies will monitor an increasing number of features and the number of functional annotations from genomic databases will considerably grow in the next years.