Cathy Philippe

Les tumeurs cérébrales malignes sont la première cause de mortalité par cancer chez l’enfant avec une survie médiane de 12 à 14 mois et une survie globale à 5 ans de 20%, pour les gliomes de haut grade. Ce travail de thèse propose des méthodes innovantes pour l’analyse de blocs de données de génomiques, dans le but d’accroître les connaissances biologiques sur ces tumeurs. Les méthodes proposées étendent les travaux de Tenenhaus et al (2011), introduisant le cadre statistique général : Regularized Generalized Canonical Correlation Analysis (RGCCA). Dans un premier temps, nous étendons RGCCA à la gestion de données en grande dimension via une écriture duale de l’algorithme initial (KGCCA). Dans un deuxième temps, la problématique de la sélection de variables dans un contexte multi-Blocs est étudiée. Nous en proposons une solution avec la méthode SGCCA, qui pénalise la norme L1 des poids des composantes. Dans un troisième temps, nous nous intéressons à la nature des liens entre blocs avec deux autres adaptations. D’une part, la régression logistique multi-Blocs (multiblog) permet de prédire une variable binaire, comme la réponse à un traitement. D’autre part, le modèle de Cox multi-Blocs (multiblox) permet d’évaluer, par exemple, le risque instantané de rechute. Enfin, nous appliquons ces méthodes à l’analyse conjointe des données de transcriptome et d’aberrations du nombre de copies, acquises sur une cohorte de 53 jeunes patients avec un gliome de haut grade primaire. Les résultats sont décrits dans le dernier chapitre du manuscrit.

Tumour lesion segmentation is a key step to study and characterise cancer from MR neuroradiological images. Presently, numerous deep learning segmentation architectures have been shown to perform well on the specific tumour type they are trained on (e.g., glioblastoma in brain hemispheres). However, a high performing network heavily trained on a given tumour type may perform poorly on a rare tumour type for which no labelled cases allows training or transfer learning. Yet, because some visual similarities exist nevertheless between common and rare tumours, in the lesion and around it, one may split the problem into two steps: object detection and segmentation. For each step, trained networks on common lesions could be used on rare ones following a domain adaptation scheme without extra fine-tuning. This work proposes a resilient tumour lesion delineation strategy, based on the combination of established elementary networks that achieve detection and segmentation. Our strategy allowed us to achieve robust segmentation inference on a rare tumour located in an unseen tumour context region during training. As an example of a rare tumour, Diffuse Intrinsic Pontine Glioma (DIPG), we achieve an average dice score of 0.62 without further training or network architecture adaptation.

Language is a unique trait of the human species, of which the genetic architecture remains largely unknown. Through language disorders studies, many candidate genes were identified. However, such complex and multifactorial trait is unlikely to be driven by only few genes and case-control studies, suffering from a lack of power, struggle to uncover significant variants. In parallel, neuroimaging has significantly contributed to the understanding of structural and functional aspects of language in the human brain and the recent availability of large scale cohorts like UK Biobank have made possible to study language via image-derived endophenotypes in the general population. Because of its strong relationship with task-based fMRI activations and its easiness of acquisition, resting-state functional MRI have been more popularised, making it a good surrogate of functional neuronal processes. Taking advantage of such a synergistic system by aggregating effects across spatially distributed traits, we performed a multivariate genome-wide association study (mvGWAS) between genetic variations and resting-state functional connectivity (FC) of classical brain language areas in the inferior frontal (pars opercularis, triangularis and orbitalis), temporal and inferior parietal lobes (angular and supramarginal gyri), in 32,186 participants from UK Biobank. Twenty genomic loci were found associated with language FCs, out of which three were replicated in an independent replication sample. A locus in 3p11.1, regulating EPHA3 gene expression, is found associated with FCs of the semantic component of the language network, while a locus in 15q14, regulating THBS1 gene expression is found associated with FCs of the perceptualmotor language processing, bringing novel insights into the neurobiology of language.

The expansion of the cerebral cortex is one of the most distinctive changes in the evolution of the human brain. Cortical expansion and related increases in cortical folding may have contributed to emergence of our capacities for high-order cognitive abilities. Molecular analysis of humans, archaic hominins, and non-human primates has allowed identification of chromosomal regions showing evolutionary changes at different points of our phylogenetic history. In this study, we assessed the contributions of genomic annotations spanning 30 million years to human sulcal morphology measured via MRI in more than 18,000 participants from the UK Biobank. We found that variation within brain-expressed human gained enhancers, regulatory genetic elements that emerged since our last common ancestor with Old World monkeys, explained more trait heritability than expected for the left and right calloso-marginal posterior fissures and the right central sulcus. Intriguingly, these are sulci that have been previously linked to the evolution of locomotion in primates and later on bipedalism in our hominin ancestors.

Neuroimaging-genetics cohorts gather two types of data: brain imaging and genetic data. They allow the discovery of associations between genetic variants and brain imaging features. They are invaluable resources to study the influence of genetics and environment in the brain features variance observed in normal and pathological populations. This study presents a genome-wide haplotype analysis for 123 brain sulcus opening value (a measure of sulcal width) across the whole brain that include 16,304 subjects from UK Biobank. Using genetic maps, we defined 119,548 blocks of low recombination rate distributed along the 22 autosomal chromosomes and analyzed 1,051,316 haplotypes. To test associations between haplotypes and complex traits, we designed three statistical approaches. Two of them use a model that includes all the haplotypes for a single block, while the last approach considers each haplotype independently. All the statistics produced were assessed as rigorously as possible. Thanks to the rich imaging dataset at hand, we used resampling techniques to assess False Positive Rate for each statistical approach in a genome-wide and brain-wide context. The results on real data show that genome-wide haplotype analyses are more sensitive than single-SNP approach and account for local complex Linkage Disequilibrium (LD) structure, which makes genome-wide haplotype analysis an interesting and statistically sound alternative to the single-SNP counterpart.

Neuroimaging-genetics cohorts gather two types of data: brain imaging and genetic data. They allow the discovery of associations between genetic variants and brain imaging features. They are invaluable resources to study the influence of genetics and environment in the brain features variance observed in normal and pathological populations. This study presents a genome wide haplotype analysis for 123 brain sulcus opening value (a measure of sulcal width) across the whole brain that include 16,304 subjects from UK Biobank. Using genetic maps, we defined 119,548 blocks of low recombination rate distributed along the 22 autosomal chromosomes, and analyzed 1,051,316 haplotypes. To test associations between haplotypes and complex traits, we designed three statistical approaches. Two of them use a model that includes all the haplotypes for a sin gle block, while the last approach considers one model by haplotype. All the statistics produced were assessed as rigorously as possible. Thanks to the rich imaging dataset at hand, we used resampling techniques to assess False Positive Rate for each statistical approach in a genome-wide and brain-wide context. The results on real data show that genome-wide haplotype analyses are more sensitive than single-SNP approach and account for local complex Linkage Disequilibrium (LD) structure, which makes genome-wide haplotype analysis an interesting and statistically sound alternative to the single-SNP counterpart.

Extensive heterogeneity in autism spectrum disorder (ASD) has hindered the characterization of consistent biomarkers, which has led to widespread negative results. Isolating homogenized subtypes could provide insight into underlying biological mechanisms and an overall better understanding of ASD. A total of 1093 participants from the population-based “Healthy Brain Network” cohort (Child Mind Institute in the New York City area, USA) were selected based on score availability in behaviors relevant to ASD, aged 6–18 and IQ >= 70. All participants underwent an unsupervised clustering analysis on behavioral dimensions to reveal subgroups with ASD traits, identified by the presence of social deficits. Analysis revealed three socially impaired ASD traits subgroups: (1) high in emotionally dysfunctional traits, (2) high in ADHD-like traits, and (3) high in anxiety and depressive symptoms. 527 subjects had good quality structural MRI T1 data. Site effects on cortical features were adjusted using the ComBat method. Neuroimaging analyses compared cortical thickness, gyrification, and surface area, and were controlled for age, gender, and IQ, and corrected for multiple comparisons. Structural neuroimaging analyses contrasting one combined heterogeneous ASD traits group against controls did not yield any significant differences. Unique cortical signatures, however, were observed within each of the three individual ASD traits subgroups versus controls. These observations provide evidence of ASD traits subtypes, and confirm the necessity of applying dimensional approaches to extract meaningful differences, thus reducing heterogeneity and paving the way to better understanding ASD traits.

Summary: Regularized generalized canonical correlation analysis (RGCCA) is a general multiblock data analysis framework that encompasses several important multivariate analysis methods such as principal component analysis, partial least squares regression, and several versions of generalized canonical correlation analysis. In this article, we extend RGCCA to the case where at least one block has a tensor structure. This method is called multiway generalized canonical correlation analysis (MGCCA). Convergence properties of the MGCCA algorithm are studied, and computation of higher-level components are discussed. The usefulness of MGCCA is shown on simulation and on the analysis of a cognitive study in human infants using electroencephalography (EEG).

Identifying the genes that contribute to the variability in brain regions involved in language processing may shed light on the evolution of brain structures essential to the emergence of language in Homo sapiens. The superior temporal asymmetrical pit (STAP), which is not observed in chimpanzees, represents an ideal phenotype to investigate the genetic variations that support human communication. The left STAP depth was significantly associated with a predicted enhancer annotation located in the 14q23.1 locus, between DACT1 and KIAA0586, in the UK Biobank British discovery sample (N = 16 515). This association was replicated in the IMAGEN cohort (N = 1726) and the UK Biobank non-British validation sample (N = 2161). This genomic region was also associated to a lesser extent with the right STAP depth and the formation of sulcal interruptions, “plis de passage,” in the bilateral STAP but not with other structural brain MRI phenotypes, highlighting its notable association with the superior temporal regions. Diffusion MRI emphasized an association with the fractional anisotropy of the left auditory fibers of the corpus callosum and with networks involved in linguistic processing in resting-state functional MRI. Overall, this evidence demonstrates a specific relationship between this locus and the establishment of the superior temporal regions that support human communication.