The properties of training data set such as size, distribution and the number of attributes significantly contribute to the generalization error of a learning machine. A not well-distributed data set is prone to lead to a partial overfitting model. Two approaches proposed in this dissertation for the binary classification enhance useful data information by mining negative data. First, an error driven compensating hypothesis approach is based on Support Vector Machines (SVMs) with (1+k)-iteration learning, where the base learning hypothesis is iteratively compensated k times. This approach produces a new hypothesis on the new data set in which each label is a transformation of the label from the negative data set, further producing the positive and negative child data subsets in subsequent iterations. This procedure refines the base hypothesis by the k child hypotheses created in k iterations. A prediction method is also proposed to trace the relationship between negative subsets and testing data set by a vector similarity technique. Second, a statistical negative example learning approach based on theoretical analysis improves the performance of the base learning algorithm learner by creating one or two additional hypotheses audit and booster to mine the negative examples output from the learner. The learner employs a regular Support Vector Machine to classify main examples and recognize which examples are negative. The audit works on the negative training data created by learner to predict whether an instance is negative. However, the boosting learning booster is applied when audit does not have enough accuracy to judge learner correctly. Booster works on training data subsets with which learner and audit do not agree. The classifier for testing is the combination of learner, audit and booster. The classifier for testing a specific instance returns the learner's result if audit acknowledges learner's result or learner agrees with audit's judgment, otherwise returns the booster's result. The error of the classifier is decreased to O(e^2) comparing to the error O(e) of a base learning algorithm.

Cette thèse présente: 1) le développement d'une nouvelle approche pour trouver des associations directes entre des paires d'éléments liés indirectement à travers diverses caractéristiques communes, 2) l'utilisation de cette approche pour associer directement des fonctions biologiques aux domaines protéiques (ECDomainMiner et GODomainMiner) et pour découvrir des interactions domaine-domaine, et enfin 3) l'extension de cette approche pour annoter de manière complète à partir des domaines les structures et les séquences des protéines. Au total, 20 728 et 20 318 associations EC-Pfam et GO-Pfam non redondantes ont été découvertes, avec des F-mesures de plus de 0,95 par rapport à un ensemble de référence Gold Standard extrait d'une source d'associations connues (InterPro). Par rapport à environ 1500 associations déterminées manuellement dans InterPro, ECDomainMiner et GODomainMiner produisent une augmentation de 13 fois le nombre d'associations EC-Pfam et GO-Pfam disponibles. Ces associations domaine-fonction sont ensuite utilisées pour annoter des milliers de structures de protéines et des millions de séquences de protéines pour lesquelles leur composition de domaine est connue mais qui manquent actuellement d'annotations fonctionnelles. En utilisant des associations de domaines ayant acquis des annotations fonctionnelles inférées, et en tenant compte des informations de taxonomie, des milliers de règles d'annotation ont été générées automatiquement. Ensuite, ces règles ont été utilisées pour annoter des séquences de protéines dans la base de données TrEMBL / This thesis presents: 1) the development of a novel approach to find direct associations between pairs of elements linked indirectly through various common features, 2) the use of this approach to directly associate biological functions to protein domains (ECDomainMiner and GODomainMiner), and to discover domain-domain interactions, and finally 3) the extension of this approach to comprehensively annotate protein structures and sequences. ECDomainMiner and GODomainMiner are two applications to discover new associations between EC Numbers and GO terms to protein domains, respectively. They find a total of 20,728 and 20,318 non-redundant EC-Pfam and GO-Pfam associations, respectively, with F-measures of more than 0.95 with respect to a “Gold Standard” test set extracted from InterPro. Compared to around 1500 manually curated associations in InterPro, ECDomainMiner and GODomainMiner infer a 13-fold increase in the number of available EC-Pfam and GO-Pfam associations. These function-domain associations are then used to annotate thousands of protein structures and millions of protein sequences for which their domain composition is known but that currently lack experimental functional annotations. Using inferred function-domain associations and considering taxonomy information, thousands of annotation rules have automatically been generated. Then, these rules have been utilized to annotate millions of protein sequences in the TrEMBL database