Techn. Universiẗat, Diss., 2006--München.

À l'ère de la biologie computationnelle, l'annotation fonctionnelle reste un défi central. Les méthodes d’annotation récentes reposent sur l’hypothèse d’association par culpabilité et s’appuient sur l’intégration de données pour la recherche de partenaires fonctionnels. Cependant, la majorité de ces méthodes souffrent de l’hétérogénéité des données et du manque de spécificité du contexte biologique ce qui expliquerait un taux élevé de faux positifs parmi les prédictions. Ce travail de thèse développe une approche intégrative de données moléculaires contrôlant leur hétérogénéité pour annoter des gènes d’Arabidopsis thaliana impliqués dans la réponse aux stress. Les contributions majeures de cette thèse sont: (1) l'annotation fonctionnelle de groupes de gènes coexprimés par l'intégration de données omiques (2) la construction d'un réseau de corégulation par une analyse transversale des groupes coexprimés qui renforce les liens fonctionnels entre les gènes. (3) le développement d’une méthode d’apprentissage supervisé pour l’inférence de fonction centrée sur les termes de la GO Slim en contrôlant le FDR. En identifiant une règle de décision par terme, cette méthode a permis de prédire la fonction de 47 gènes partiellement annotés ou orphelins. / In the era of computational biology, functional annotation remains a major challenge. Recent annotation methods are based on the guilt by association assumption and rely on data integration to identify functional partners. However, most of these methods suffer from data heterogeneity and a lack of biological context specificity which would probably explain the high rate of false positives among predictions. This thesis develops an approach of molecular data integration controlling their heterogeneity in order to annotate Arabidopsis thaliana genes involved in stress response. The major contributions of this thesis are: (1) functional annotation of groups of co-expressed genes by omics data integration (2) the construction of a coregulatory gene network through a cross-analysis of the coexpressed groups strengthening the functional links between genes (3) the development of a supervised learning method for the inference of gene function centered on the GO Slim terms with a control of the FDR. By identifying a decision rule by term, this method was used to predict the function of 47 orphan or partially annotated genes.

Annotation is essential in machine learning. Building an accurate object detec-tion model requires a large, diverse dataset, which poses challenges due to thetime-consuming nature of manual annotation. This thesis was made in collabora-tion with Project Ngulia, which aims at developing technical solutions to protectand monitor wild animals. A contribution of this work was to integrate an effi-cient semi-automatic image annotation tool within the Ngulia system, with theaim of streamlining the annotation process and improving the employed objectdetection models. Through research into available annotation tools, a custom toolwas deemed the most cost-effective and flexible option. It utilizes object detec-tion model predictions as annotation suggestions, improving the efficiency of theannotation process. The efficiency was evaluated through a user test, with partic-ipants achieving an average reduction of approximately 2 seconds in annotationspeed when utilizing suggestions. This reduction was supported as statisticallysignificant through a one-way ANOVA test. Additionally, it was investigated which images should be prioritized for an-notation in order to obtain the the most accurate predictions. Different samplingmethods were investigated and compared. The performance of the obtained mod-els remained relatively consistent, although with the even distribution methodat top. This indicate that the choice of sampling method may not substantiallyimpact the accuracy of the model, as the performance of the methods was rela-tively comparable. Moreover, different methods of selecting training data in there-training process was compared. The difference in performance was consider-ately small, likely due to the limited and balanced data pool. The experimentsdid however indicate that incorporating previously seen data with unseen datacould be beneficial, and that a reduced dataset can be sufficient. However, furtherinvestigation is required to fully understand the extent of these benefits.

No

Chicken is an important non-mammalian vertebrate model organism for biomedical research, especially for vaccine production and the study of embryology and development. Chicken is also an important agricultural species and major food source for high-quality protein worldwide. In addition, chicken is an important model organism for comparative and evolution genomics. Exploitation of this genome as a biomedical model is hindered by its incomplete structural and functional annotation. This incomplete annotation makes it difficult for researchers to model their functional genomics datasets. Improving structural and functional annotation of the chicken genome will allow researchers to derive biological meaning from their functional genomics datasets. The objectives of this study were to identify proteins expressed in multiple chicken tissues, to functionally annotate experimentally confirmed proteins expressed in different chicken tissues, to quantify and assess the Gene Ontology (GO) annotation quality, and to facilitate functional annotation of microarray data. The results of this research have proven to be fundamental resource for improving the structural and functional annotation of chicken genome. Specifically, we have improved the structural annotation of the chicken genome by adding support to predicted proteins. In addition, we have improved the functional annotation of the chicken genome by assigning useful biological information to proteomics datasets and the whole genome chicken array. The Gene Ontology Annotation Quality (GAQ) and Array GO Mapper (AGOM) tools developed in this study will sustainably continue to facilitate functional modeling of chicken arrays and high-throughput experimental datasets from microarray and proteomics studies. The ultimate positive impact of these results is to facilitate the field of biomedical research with useful information for comparative biology, better understanding of chicken biological systems, diseases, drug discovery and eventually development of therapies.

Nous pr´esentons dans ce m´emoire une m´ethode d'annotation de tableaux guid´ee par les connaissances d'un domaine d'application formalis´ees dans une on- tologie. Apr`es avoir pr´esent´e le contexte applicatif et une ´etude bibliographique sur l'annotation s´emantique et l'extraction d'information, nous pr´esentons les diff´erentes ´etapes de notre syst`eme : annotation des cellules, des colonnes puis des relations repr´esent´ees par le tableau. Nous traitons diff´eremment les donn´ees selon qu'elles sont num´eriques ou symboliques. Nous commen¸cons par d´eterminer si une colonne d'un tableau contient des donn´ees num´eriques ou symboliques. Les donn´ees symboliques sont annot´ees avec les termes de l'ontologie, en utilisant une comparaison mot `a mot des termes employ´es dans le tableau avec ceux d´efinis dans l'ontologie. Les donn´ees num´eriques sont extraites, ainsi que les unit´es de mesure employ´ees, et compar´ees avec les unit´es et intervalles de valeurs possibles d´efinis dans l'ontologie pour les types de donn´ees num´eriques. Le type de donn´ees repr´esent´e par chaque colonne du tableau est alors d´etermin´e, en utilisant `a la fois le contenu de la colonne (deux m´ethodes diff´erentes sont employ´ees suivant que la colonne contient des donn´ees num´eriques ou symboliques) et le titre de la colonne. Une fois le type des colonnes reconnu, les relations s´emantiques repr´esent´ees par le tableau sont identifi´ees en utilisant `a la fois le titre du tableau et la signature du tableau, qui est compar´ee avec la signature des relations s´emantiques d´efinies dans l'ontologie. Les relations reconnues sont ensuite instanci´ees pour chaque ligne du tableau. Les annotations que nous manipulons sont floues, c'est-`a-dire qu'au lieu de faire un lien direct entre un ´el´ement du tableau et un ´el´ement de l'ontologie, nous proposons plusieurs valeurs possibles pour l'annotation, en as- sociant `a chaque valeur un degr´e repr´esentant la confiance que l'on accorde `a cette valeur. Les diff´erentes ´etapes de notre m´ethode d'annotation de tableaux ont ´et´e ´evalu´ees exp´erimentalement, en prenant comme domaine d'application la microbiologie alimentaire.

One of the key research topics of post-genome study is annotation of the gene with regards to specific function and biological processes. This can help us to understand the precise role that a gene or a group of genes carries. In this thesis, I developed techniques to automatically annotate genes on single gene and a group of genes levels. It is shown that these techniques improve our understanding of biological systems/diseases, and will aid drug discovery. In the first project, I attempted to achieve precise annotation for single genes. In the second and third projects, I performed annotations of a group of genes using pathway knowledge. I examined this problem from supervised and unsupervised learning aspects, respectively. The main contributions of the work are organized as follows: In gene annotation project, I built up an automated scheme to reconcile the term differences arising from the different automated annotation services. The method leaves less than 20% of the annotations for manual work. The generalization performance across other species is of a similar standard, again leaving less than 20% of the annotations for manual inspection. In addition, less than 10% of the results have different functions from EcoCyc results in E.coli genome annotation task. Overall, this method can significantly reduce human effort involvement (6 months’ work by several biologists for a bacterial genome) to resolve inconsistent gene annotations. Then I started from the current limitations of pathway analysis and presented a novel approach for pathway discovery. Enrichment analysis is the most popular approach to map gene expression profiling from genes to biological pathways. It is a powerful tool to identify pathways enriching of differentially expressed gene; however, it is unable to discover active/inhibitive pathways. In this study, I attempted to resolve this issue by integrative classification of KEGG and TF gene sets. I assumed that the pathways with good classification performance should be considered as the active/inhibitive pathways. Based on this hypothesis, I built up a generic approach to incorporate two types of biological data for active pathway discovery. The experimental results show that integration of transcription factor data boosts classification performance. In addition, this method identified relevant biological pathways, which are highly associated with tumour genesis and development. But they are ignored by Gene Set Enrichment Analysis, such as cancer pathway, inflammation and metabolic pathways. Furthermore, this method achieves comparable classification performance with the best-reported results. Lastly, I performed subtyping analysis of Rheumatoid Arthritis patients based on gene expression profiling. I revalidated the two clusters of patients based on two independent cohorts. The experimental results indicate that the subgroup structure does not correspond to the drug response status. In addition, I developed a pathway subtyping approach and achieved the same number of clusters as gene-level clustering results. The pathway clustering results show that one group of the patients has high proliferation and low inflammation response, while the other group has the reverse trend. It suggests that designing drugs with better trade-off between anti-inflammation and anti-proliferation for specific subgroup of patients may achieve better clinical outcomes.

No abstract available

L'annotation automatique d'images est un domaine du traitement d'images permettant d'associer automatiquement des mots-clés ou du texte à des images à partir de leur contenu afin de pouvoir ensuite rechercher des images par requête textuelle. L'annotation automatique d'images cherche à combler les lacunes des deux autres approches actuelles permettant la recherche d'images à partir de requête textuelle. La première consiste à annoter manuellement les images, ce qui n'est plus envisageable avec le nombre croissant d'images numériques, d'autant que différentes personnes annotent les images différemment. La seconde approche, adoptée par les moteurs de recherche d'images sur le web, est d'utiliser les mots de la page web contenant l'image comme annotation de cette image, avec l'inconvénient de ne pas prendre du tout en compte le contenu de l'image. Quelques systèmes d'annotation automatique d'images commencent à émerger, avec certaines limites : le nombre d'objets reconnus reste de l'ordre de 10 à 100, alors que les humains sont capables de reconnaître de l'ordre de 10000 objets ; les mots-clés générés comme annotation pour une image sont parfois en contradiction entre eux, par exemple "éléphant" et "ours polaire" peuvent être détectés dans une même image ; la base de données pour l'apprentissage des objets est construite manuellement. Les travaux effectués au cours de cette thèse visent à proposer des solutions à ces problèmes, d'une part en introduisant de la connaissance dans l'annotation automatique d'images, d'autre part en proposant un système complètement automatique, où notamment la base d'images pour l'apprentissage est construite automatiquement à partir des images du Web. Cette thèse est constituée de trois parties : La première partie concerne la catégorisation d'une image en fonction de son type (photo, carte, peinture, clipart) puis pour les photographies, on s'intéresse à savoir quel est le contexte de la scène photographiée : est-ce une photographie d'intérieur ou d'extérieur, une photographie prise de nuit ou de jour, une photographie de nature ou de ville ? Y a-t-il des visages dans la photo ? Y a-t-il du ciel, de l'herbe, de l'eau, de la neige, une route,

Given an unstructured collection of captioned images of cluttered scenes featuring a variety of objects, our goal is to simultaneously learn the names and appearances of the objects. Only a small fraction of local features within any given image are associated with a particular caption word, and captions may contain irrelevant words not associated with any image object. We propose a novel algorithm that uses the repetition of feature neighborhoods across training images and a measure of correspondence with caption words to learn meaningful feature configurations (representing named objects). We also introduce a graph-based appearance model that captures some of the structure of an object by encoding the spatial relationships among the local visual features. In an iterative procedure we use language (the words) to drive a perceptual grouping process that assembles an appearance model for a named object. We also exploit co-occurrences among appearance models to learn hierarchical appearance models. Results of applying our method to three data sets in a variety of conditions demonstrate that from complex, cluttered, real-world scenes with noisy captions, we can learn both the names and appearances of objects, resulting in a set of models invariant to translation, scale, orientation, occlusion, and minor changes in viewpoint or articulation. These named models, in turn, are used to automatically annotate new, uncaptioned images, thereby facilitating keyword-based image retrieval.