Estudio y aplicación de métodos supervisados para el análisis de patrones de expresión génica

Tesis doctoral inédita leída en la Universidad Autónoma de Madrid, Facultad de Ciencias, Departamento de Biología Molecular. Fecha de lectura: 14-03-2006DNA microarray technology has opened the possibility of measuring the expression of thousands of genes in a single experiment. This offers new possibilities but, at the same time, it poses new challenges at the different steps of the analysis of the generated data. In this thesis 1 study the use of machine learning methods for the analysis of microarray data. In the íirst part of this thesis I use the multilayer perceptron, a supervised method, for the prediction of gene function based on the expression promes of genes. I combine the multilayer perceptron with functional information included in the Munich Information Center for Protein Sequences (MIPS) functional catalog. I also study the characteristics of functional classes present in the database and conclude that the learning of catalogs like MIPS fiom the expression profiles is hard due to the inherent biological interconnection of processes. In the second part of this thesis I study the use of the perceptron and support vector machines (SVM) for the prediction of phenotypical classes based on their expression profiles. The phenotypical classes predicted here are cancer and healthy classes. I study the combination of both methods (perceptron and SVM) with ,he Self Organising 'Itee Algorithm (SOTA), a robust unsupervised neural ietwork based on Self Organising Maps (SOM) to improve the ability of the ,erceptron and SVM for the classiñcation task. Through the use of SOTA, ,he dimensionality of the data is previously reduced by clustering genes. In iubsequent steps clusters of genes, instead of genes themselves, are used. I :onclude that SVM are robust enough to handle high dimensional data sets tnd do not require previous clustering of the data. However, the perceptronbenefits from the initial compression of the data and the rate of trne positives increases strongly when working with cluster8 instead of genes. In addition to the classiílcation method, I propose a way for identifying the genes responsible for the dehition of the classes, obtaining relationships from genotype to phenotype. This second part of the method makes use of biological functional annotation, searching the (according to the perceptron weights) important clusters for Gene Ontology (GO) terms. Then a comparison with the rest of clusters is performed. The significant over- or under-representation of terms in the compared groups of gene clusters is used as a complementary criterion to determine the genes defining the classes and so, to increase the accuracy of the detection of important clusters. Finally, I use a supervised bi-clustering method, called signature alg* rithm to study the modular structure of a complex disease lüce cancer. I carry out a functional study of the obtained moduls to asses the results