HAL will be down for maintenance from Friday, June 10 at 4pm through Monday, June 13 at 9am. More information
Skip to Main content Skip to Navigation
Theses

Deep Learning pour l’identification de différences phénotypiques visuelles subtiles entre lignées neuronales, modèles de la maladie de Parkinson

Abstract : High-content screening has experienced a significant growth since the mid-2000s. This technology is of primary interest to the pharmaceutical industry as it allows in principle the discovery of therapeutic molecules for diseases whose molecular pathways are poorly identified. Until now, a measurable cell phenotype must first be identified in order to evaluate the effect of a compound library on it. From an image analysis point of view, treated cells are automatically detected in hundreds of thousands of images and measurements of descriptors allow to finely differentiate effective treatments from a negative control. In collaboration with the companies Ksilink and Sanofi, we have been confronted with a new type of high-content screen to identify compounds effective against Parkinson's disease. This one, presenting images of neurons, made analysis dependent on robust cell segmentation obsolete. In addition, using cell differentiation and genome editing techniques, the controls of the cell model are two neuron cultures, isogenic but for one mutation: the GS2019 mutation that induces the disease. Nevertheless, the heterogeneity of these complex cells and the fine differences between the two isogenic lines do not allow the human eye to identify two distinct phenotypes. In order to allow the automatic detection of differences between phenotypes, we have proposed to use deep learning approaches. This work was mainly divided into two steps. The first step consisted in identifying a network architecture capable of classifying neural images. We learned that neuron cultures show phenotype differences in a very heterogeneous and therefore not systematic way. The second step consisted in proposing methods to explain and interpret subtle differences in phenotype, to ensure that the screening is performed on the basis of a proven difference between phenotypes and not on the basis of a technical bias. Based on the premise that differences between neuronal phenotypes are difficult to visually apprehend between two different images due to the high natural variability of neurons, we propose the idea that transforming the same image from one phenotype to another may represent an interesting approach. Indeed, we show that it is possible to train antagonistic networks to transform an image of a neuron that does not carry the mutation into a neuron that carries it and vice versa. In this way, we have been able to highlight potential assay biases but also what we believe to be true morphological differences related to the pathological mutation that were previously invisible.
Document type :
Theses
Complete list of metadata

https://tel.archives-ouvertes.fr/tel-03537532
Contributor : Abes Star :  Contact
Submitted on : Thursday, January 20, 2022 - 2:33:08 PM
Last modification on : Thursday, March 17, 2022 - 10:08:43 AM
Long-term archiving on: : Thursday, April 21, 2022 - 7:07:51 PM

File

Champetier_2020_These.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03537532, version 1

Collections

Citation

Tiphaine Champetier. Deep Learning pour l’identification de différences phénotypiques visuelles subtiles entre lignées neuronales, modèles de la maladie de Parkinson. Bio-Informatique, Biologie Systémique [q-bio.QM]. Université Paris sciences et lettres, 2020. Français. ⟨NNT : 2020UPSLE037⟩. ⟨tel-03537532⟩

Share

Metrics

Record views

54

Files downloads

9