Mots-Clés

Statistical modelling Deep Learning approaches Applications to Genetics

Description

Context
Statistical models for the dynamics of plant physiology aim to describe and analyze the process
underlying plant development. In its simplest formulation, the model is a non-linear mixedeffects
model in which a limited number of physiological parameters control the shape and rate
of the dynamic traits. These physiological parameters can in turn be described as random
variables whose distributions depend on a set (possibly high-dimensional) of explanatory
cofactors. The inference of such models is rather complex, as it often requires the use of
sampling techniques such as SAEM or MCEM [1,2].

Subject
The aim of this post-doc is twofold:
1. First, to develop an alternative inference strategy based on recent tools and techniques
from deep learning to fit state-of-the-art non-linear mixed-effect models for
physiological dynamics. To this end, the inference will be presented as an optimization
problem where the target is a neural network, and where the loss function will be
chosen according to the nature of the physilogical feature (continuous or discrete).
2. Second, more refined statistical models involving differential equations will be
considered. These differential equations will be incorporated into the neural network
model using physics-informed neural networks (PINN, [3]).
All developments will be implemented using libraries for high-performance numerical
computing and optimization (PyTorch, Jax).
The models will be applied to the data collected in the G2WAS ANR project. The data consist of
250 grape varieties that have been dynamically phenotyped during 3 weeks for vegetative
biomass production by imaging at the PhenoArch platform. Each variety underwent 3 different
hydric scenarios (from well watered to severe drought stress). A set of 60K genetic
markers will be used as explanatory variables in the statistical model to explain and predict the
dynamics of plant physiology.

Required skills
Phd in one of these domains: computer science, statistics or machine learning,
Proficiency in Python programming, experience in working with large datasets,
Experience in working with reproducible research methods (GitLab, versioning, testing
of code),
Fluency in written English.
Previous experience in working with biologists and/or basic knowledge in genetics is a plus.

Working environment
The Post-Doc position is funded by the ANR G2WAS Project. You will work in the SOLsTIS team
of the MIA Paris-Saclay unit, located at AgroParisTech (Palaiseau). The Post-Doctoral researcher
will be supervised by Tristan Mary-Huard, Laure Sansonnet and Julien Chiquet, in close
collaboration with Vincent Segura and Timothée Flutre for the plant physiology and genetics
aspects.
This research project corresponds to a 1 year position, the starting can be as early as
September 2023.

Contact
Interested candidates should apply by sending a CV and a motivation letter at
tristan.mary-huard@agroparistech.fr
laure.sansonnet@inrae.fr
julien.chiquet@inrae.fr

References
[1] Kuhn & Lavielle (2005). Maximum likelihood estimation in nonlinear mixed effects model,
Comput. Stat. and Data Analysis.
[2] Liu & Wu (2007). Simultaneous inference for semiparametric nonlinear mixed‐effects
models with covariate measurement errors and missing responses. Biometrics.
[3] Raissi, Perdikaris & Karniadakis (2019). Physics-informed neural networks: A deep learning
framework for solving forward and inverse problems involving nonlinear partial differential
equations. J. of Comput. Phys.