Training a deep neural network to learn 3D division rules
- Type de publi. :
- Date de publi. : 04/09/2023
-
Auteurs :
Alexandre DurrmeyerJean-Christophe PalauquiPhilippe Andrey
-
Organismes :
Institut Jean-Pierre Bourgin - Sciences du végétal
Université Paris-Saclay
AgroParisTech
Institut Jean-Pierre Bourgin - Sciences du végétal
Université Paris-Saclay
AgroParisTech
Institut Jean-Pierre Bourgin - Sciences du végétal
Université Paris-Saclay
AgroParisTech
Résumé : Plant embryo development is an intricate process whereby a complex organism arises from a single cell. At the cellular scale, this process hinges on the fundamental mechanisms of cell division and growth. Cell division, in particular, plays a pivotal role by reshaping tissue properties like mechanical constraints and topology, thereby influencing morphology. Predicting the positioning of a new cell wall during division is thus crucial for understanding morphogenesis. Early attempts to predict cell division positioning relied on geometric approaches based on the shape of the mother cell. While geometry contributes to division plane positioning, other factors like chemical signaling, mechanical constraints, and tissue-scale optimizations also exert influence. However, deciphering how each factor impacts division remains challenging. Current models, whether based on geometry, stresses, or other properties, are typically rule-based, incorporating manually defined rules that may contain human biases or overlook certain aspects. It is then difficult to define whether the data does not fit any rule or if the right rule was simply not identified. In A. thaliana embryos, a geometric division rule successfully fit the observed divisions from stage 1C to stage 32C, with the exception of the stage 16C basal intern cells (Moukhtar et al. 2019, Laruelle et al. 2022). Do these latter cells follow a different rule from the previous stages, or is there an alternative common rule? To ensure that the proposed division rule aligns as closely as possible with observations, we propose a method to automatically learn predicting cell division plane positioning in a supervised framework. Leveraging U-Net3D, a deep learning architecture tailored for biological image segmentation, we train a model to approximate division rules based on an image dataset of before/after division pairs. This formulation transforms the prediction of division plane positioning into a more accessible segmentation task. Additionally, representing mother cells by their 3D image masks enables the model to approximate a strict geometric division rule from fully preserved geometric information. We conducted multiple experiments to assess the potential and limitations of this novel approach. We first evaluated the network's ability to learn a simple division rule using synthetic shapes like cuboids and ellipsoids. Subsequently, we tested the model's generalization capability across different shapes and its performance in approximating complex division rules. Finally, we applied the approach to real A. thaliana embryo cells spanning developmental stages 1C to 32C. We trained the model to learn stage-specific division rules and subsequently a universal rule applicable across all stages.
Fichiers liés :
Poster.pdf
Source