Master 2 internships
- Quantification of axonal regeneration in drosophila:
The motivation of this project is the restoration of damaged neuronal circuits, for instance for patients suffering from neurodegenerative diseases. The Drosophila is an excellent model for studying axonal regeneration, because during metamorphosis (transition to the adult stage), a significant and stereotyped restructuring of its nervous system is observed. In this living organism, particularly well-suited to genetic manipulation, F. Besse team (C. Medioni) has recently characterized a population of neurons (ie Bursicon) whose axons regenerate reproducibly during metamorphosis. Any alteration in the axonal regeneration of these neurons, thus in adult neurons, results in wing unfolding defects in adult Drosophila.
The aim of this project is to build and assess a protocol to characterize regulators involved in axonal regeneration. For each targeted gene, the strength in alteration is assessed at the population level by evaluating the unfolding defects at the population level. Then, Bursicon neurons will be imaged in adults in two populations, one control and one with the inactive gene. Last, the localization of the regulator induced by the gene will be analyzed in a wild type population, also by imaging. Two challenges arise. First, (at least) 15 individuals are to be imaged in each population (gene-suppressed, control, wild-type), yielding a significant number of 3D images are to be analyzed. Second, an objective comparison of groups (control vs gene-suppressed for one gene, or across genes) is necessary, which cannot be reached by visual assessment.
The goal of this internship is to defined automated image processing methods to quantify the regulator localization as for the axonal arborization.
Requirements:
- Last year of master in computer sciences or applied mathematics (with interest in biology) or bioinformatics (with interest in image processing)
- Knowledge in image processing, preferably 3D
- Computer skills: programming (python), image processing/graphics libraries
- Written and spoken English
Practical information:
- This work takes place in a collaboration between IBV (C. Medioni) and Morpheme, a joint research team between INRIA, CNRS and the University of Nice Côte d’Azur.
- This internship is located in Sophia Antipolis (French Riviera).
- This internship is remunerated
- Duration: 6 months, expected start: early 2025
- To candidate, please send a curriculum vitae, referees coordinates and a motivation letter to
- Grégoire Malandain (Gregoire.Malandain@inria.fr)
- Use of Markov fields for naming ascidian embryos :
The aim of this internship is to use the Markov field framework to name ascidian embryo cells.
Developmental biology aims to understand the dynamics of tissue or organ formation within an organism. Many model organisms are used to study development. Most of these are animals, such as the nematode C. elegans or Drosophila. Each of these models has its own specificities, which can make it more suitable for observation and/or analysis. For example, the C. elegans nematode is transparent, making it ideal for observation under the light microscope, and it has a remarkable stereotyped development, with each adult having exactly the same number of somatic cells, each cell having a homologous cell in other individuals. In other words, each cell can be identified by a name (Sulston, 1983). This property not only enables cell comparisons in population studies, but also the aggregation of different cell types.
For vertebrates, mice are generally the model of choice. However, as close relatives of vertebrates, tunicates are also widely used as models for studying embryonic development, the ascidians being the largest group among the tunicates. The embryos of some tunicate species are transparent, making them suitable for microscopic observation. Moreover, the development of ascidian embryos (in the early stages) is stereotyped at the cellular level, as in the nematode C. Elegans, even from one species to another. This stereotypy has led to the development of nomenclatures (Kofold, 1893; Castle, 1896; Conklin, 1905) based on light microscopy observations, which provide unambiguous names for embryonic cells during the early stages of development.
In order to aggregate information (e.g. gene expression) by cell, it is therefore necessary to be able to “name” the cells of an embryo at any given time. This project aims to infer the cell names of an embryo (derived from a 3D image) from a database of already named developing embryos (derived from time series of 3D images) (Guignard, 2020). While the number of cells is representative of the stage of development, there is no exact correspondence between the cells of two embryos with the same number of cells, due to the heterochrony of divisions (the temporal order of cell divisions is different from one embryo to another). On the other hand, it is possible to quantify a similarity between the cells of two embryos (Malandain, 2024), with which we can estimate empirical naming probabilities for each cell. Furthermore, an embryo can be represented as a graph (cells being nodes, edges being adjacencies between cells), which naturally introduces joint probabilities between the names of two adjacent cells.
This project will investigate Markov fields in a Bayesian framework on the graph defined by the embryo under study. Likelihood and prior information (interactions between neighboring cells on the graph) will be inferred by statistical analysis of the annotated embryo database. The naming of individual cells will be obtained by optimizing the model using simulated annealing. To this end, the cell uniqueness constraint for a given name will be preserved by Kawasaki dynamics.
Key words: Markov field, Simulated annealing, Bayesian approach, Ascidian embryo
Prerequisite:
- Last year of Master’s degree in computer science or applied mathematics (with an interest in biology).
- Computer skills: programming (python)
- Written and spoken English
Practical information :
- This work is part of a collaboration between the CRBM (P. Lemaire) and Morpheme, a joint research team between INRIA, CNRS and the Université de Nice Côte d’Azur.
- This internship is located in Sophia Antipolis (Côte d’Azur, France).
- This internship is paid.
- Length of internship: 6 months, planned start: early 2025
- To apply, please send a curriculum vitae, contact details of referees and a covering letter to
- Xavier Descombes (Xavier.Descombes@inria.fr)
- Grégoire Malandain (Malandain@inria.fr)
PhD scholarships
PhD Thesis
Distillation for deep learning based processing of microscopy images
Developmental biology aims to better understand the morphogenesis. The latest microscopy techniques provide temporal sequences of 3D images with a high spatio-temporal resolution allowing to follow the embryo or organ development at sub-cellular scale [1]. The acquisition of such 3D+t series results in huge quantities of data for which manual analysis of a large number of images is not possible. Therefore, sophisticated image analysis tools have been developed in the recent years for this particular goal, for plant meristems [2], ascidians embryos [3] or sea urchin sea urchin embryos [4], in particular for series where the cell walls or membranes have been marked.
In this context, two difficulties have to be leveraged. On the one hand, in long series, cells become smaller and membranes express less fluorescence: this impairs the efficiency of the membrane-based segmentation method. On the other hand, cell division orientation may be linked to the cell nuclei position and/or orientation [5], thus we are also interested in investigating the nuclei morphology and dynamics in developing embryos.
The first difficulty can be addressed by augmenting the segmentation method with enhanced or segmented membrane images issued from advanced neural network (NN) methods (eg [6]). Concerning the second one, efficient NN-based methods already exist for nuclei segmentation [7], but suffer from a high computational cost [8] which make them unapplicable for temporal series of 3D images (eg. Cellpose typically requires more than 40 minutes per 3D image). The first objective of this PhD project will be to improve the computational efficiency of NN-based segmentation methods. To that end, the teacher-student paradigm within the distillation framework [9] is certainly the method of choice since it allows to decrease the number of coefficients of the target network (or “student” network), thus the cost of the inferring stage, while preserving the segmentation quality by learning from a reference (or “teacher”) network. We will elaborate on this paradigm and in particular we will investigate a student learning scheme from several independent teacher networks.
Imaging membranes and nuclei in the same embryos usually requires two fluorophores, excited by different wavelengths, which result in a series of two-channels images. However, multi-laser excitation increases the energy received by the embryo (thus may impair its normal development) or prevent to use one channel to image a protein of interest. Imaging both the membrane and the nuclei with the same wavelength addresses both limitations. The second objective of this PhD project will then to investigate whether a NN can separate the nuclei from the membranes in images where both have been imaged.
Requirements:
- Master in computer sciences or applied mathematics (with interest in biology), preferably with knowledge in deep learning.
- Knowledge in image processing, preferably 3D
- Computer skills: programming (python), image processing/graphics libraries
- Written and spoken English
Practical information:
- This work takes place in a collaboration between CRBM (P. Lemaire’s team) and Morpheme, a joint research team between INRIA, CNRS, INSERM and the University of Nice Côte d’Azur.
- This PhD is located in Sophia Antipolis (French Riviera).
- Gross salary: 2100 euros/month
- To candidate, please send a curriculum vitae, referees coordinates and a motivation letter to
- Grégoire Malandain (Gregoire.Malandain@inria.fr)
PhD positions
Diffusion model for super-resolution fluorescent microscopy inverse
problem
PhD proposal for 2025 (Duration: 3 years)
Supervisors: Laure Blanc-Feraud (blancf@i3s.unice.fr), Xavier Descombes (xavier.descombes@inria.fr),
Sebastien Schaub (sebastien.schaub@imev-mer.fr)
Host institution: MORPHEME research group (INRIA, CNRS, I3S, Sophia-Antipolis, France)
Collaboration: Luca Calatroni (Luca.calatroni@unige.it), Machine learning Genoa Center, Italy.
Context and PhD objectives
Conventional optical microscopy techniques, as confocal microscopes, are widely used in biology for cellular and sub-cellular structures investigation. However, their spatial resolution is limited by the light diffraction
phenomena and it is typically around 200nm in the transverse plane and 400nm in the optical axis.
Over the recent years, several super-resolution techniques have been developed to overcome this drawback.
Among them we focus on fluctuation of fluorescent molecules methods as they don’t need any specific materiel or fluorophore. The super-resolved image reconstruction is formalized in mathematical terms as an inverse problem with regularization. In the morpheme group we have proposed model-driven approach (e.g.,COL0RME [1]) as well as model-based data-driven approaches (e.g., GANs [2] or Plug& Play [3]).
A different and increasingly popular class of methods producing outstanding results in many applied fields is based on the use of modern generative learning approaches. Among them, diffusion models [4] have attracted the attention of several researchers working in the field of inverse problems due to their ability of combining variational inference approaches with the ability of neural networks to learn unknown posterior distributions distributions. Their use in the field of image microscopy, however, remains limited.
The purpose of this PhD thesis is to develop an inverse super-resolution method based on the use of diffusion model, inspired, e.g., from recent works in inverse problems as [5]. This approach will be developed with different regularization terms according to the biological structures of interest. The advantage will be the automatic estimation of model parameters and the possibility of associating uncertainties with the results.
Acquisitions with specific microscope views will also be developed in collaboration with the Laboratoire debiologie du d´eveloppement de Villefranche sur mer, enabling super-resolved 3D reconstruction.
Candidate profile
Master/engineer degree in computer science, applied mathematics, data science with background in image processing, imaging inverse problems, deep learning and optimisation. Good coding skills for numerical simulation (Pytorch, Python, MATLAB, …). A general interest in health and biology is welcome.
Practical information
MORPHEME research team is a joint research group between INRIA Sophia Antipolis Méditerranée., I3S
Lab (Universit´e Cote d’Azur and CNRS).
Remuneration:
Application procedure
Please send your CV, motivation letter, marks of the last two years of study and the name and e-mail
address of a contact for recommendation to Laure Blanc-F´eraud (blancf@i3s.unice.fr), Luca Calatroni (calatroni@i3s.unice.fr), Xavier Descombes (xavier.descombes@inria.fr)
