Skip to Main content Skip to Navigation
Habilitation à diriger des recherches

Bioimage Informatics for Phenomics

Abstract : While we have the technologies and computational tools to analyze entire genomes, transcrip- tomes and proteomes, the computational description of phenotypes resulting from this molecular basis is still lagging behind. Yet, the quantitative description of the diverse aspects of the phe- nome is a prerequisite for understanding the complex genotype-phenotype relationships in living systems. High Content Screening (HCS) allows to systematically explore many different aspects of the phenome, in particular cell morphology, the dynamics of cellular behavior and the spatial dis- tribution of transcripts and proteins inside cells. Monitoring and analyzing the changes in these aspects upon perturbation by gene silencing or drug treatment have the potential to unravel the relationship between these cellular properties and the molecular mechanisms that regulate them. Similarly, the analysis of stained tissue slides allows to study architectural changes depending on disease related variables. Large-scale imaging approaches, such as HCS and histopathology, thus provide information that is complementary to information at the molecular level, traditionally studied in bioinformatics. In order to make best use of these challenging and complex large-scale image data sets, we need robust and sophisticated methods capable of integrating a large set of image features in order to reach a biologically meaningful description of the data. For this reason, computer vision is the method of choice for computational phenotyping. This manuscript describes my contributions to the field of computational phenotyping by com- puter vision. After an introduction to the field of Bioimage Informatics as well as some back- ground on High Content Screening, I will describe a number of different projects I have been working on over the last years exemplifying the different types of information that can be studied with images: (1) analysis of morphological phenotypes by supervised learning, (2) analysis of temporal information, in the form of phenotypic and spatial trajectories, (3) analysis of local- ization patterns, i.e. the spatial distributions of biomolecules inside cells and (4) analysis at the tissular scale. These methods have been applied to large-scale screens on cell division and migration, namely the first genome-wide screen by time-lapse microscopy in a human cell line. Some more recent applications include the study of the spatial aspects of gene expression, where we aim at un- derstanding the patterns according to which RNA localize inside cells, as well as the field of digital pathology, where we wish to predict clinical variables, such as outcome or response to treatment, from large stained images of diseased tissue. While often used for diagnostic purpose, histopathology data is also informative about cellular phenotypes and therefore allows to bridge the gap between phenotypic analysis at the cellular level and implications for disease (at the patient level). The most striking evolution in this field is the advent of deep learning, that has revolutionized computer vision over the last years. I will discuss the role deep learning is going to play in the near future with respect to the different applications mentioned above, and the methodological or conceptional developments that are most promising for each application in turn. Altogether, one of the major challenges in bioinformatics today is to establish relationships be- tween the molecular level (e.g. the level of genetic mutations or transcripts) and the level of an entire biological system (e.g. cell or even patient level) by analysis of large-scale omics data sets. These associations need to cross 9 orders of magnitude. The projects and methods I am showing in this manuscript will contribute to bridging this gap.
Complete list of metadatas

Cited literature [286 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/tel-02981391
Contributor : Thomas Walter <>
Submitted on : Tuesday, October 27, 2020 - 11:27:06 PM
Last modification on : Wednesday, January 20, 2021 - 2:58:01 PM

File

ThomasWalter_HDR_2020.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : tel-02981391, version 1

Citation

Thomas Walter. Bioimage Informatics for Phenomics. Bioinformatics [q-bio.QM]. Sorbonne Université, 2020. ⟨tel-02981391⟩

Share

Metrics

Record views

49

Files downloads

21