Link to the subject

Context

Macromolecular assemblies are large complexes of proteins involved in most cellular processes. The field of structural biology aims to decipher their molecular structure. In particular, the goal is to identify and localize the proteins inside the structure, in order to infer the function of the macromolecular assemblies in the cellular processes.

While electron microscopy has been the technique of choice in structural biology for decades, recent progress in super-resolution [8] and expansion microscopy [3,6] have made possible to observe macromolecular assemblies with another category of imaging modalities called fluorescence imaging. Fluorescence techniques are able to label efficiently specific proteins in large complexes such as the centriole [5].

This emerging context has motivated the development of single particle reconstruction (SPR) methods for fluorescence microscopy. SPR consists in reconstructing the structure of a particle from acquisitions of a large number of replicates observed in random positions. The combination of the complementary information of all the views yields a reconstructed particle with higher resolution. This approach has become very popular and promising in recent years [2,4,7,1].

Subject

This internship addresses the first step of the SPR pipeline, that is the detection and extraction of the particles in the acquired fluorescence images, called particle picking. This is a crucial step, since the number and the quality of the picked particles have a decisive impact on the final reconstruction. Multiple challenges have to be addressed to design an efficient particle picking method. The first issue is the resolution anisotropy of the images: the blur is much stronger in the axial direction Z of the microscope than in the XY plane. The second issue is the presence of parasitic structures in the images such as microtubule filaments, which can influence the detection. Finally, the density of particles can be locally high (in particular because of the presence of couples of centrioles very close to each other), which makes more difficult to isolate indivudual particles. In practice, the particle picking in 3D images is often performed manually. This a tedious task prone to bias, which limits the number of particles and reliability of the picking.

We have recently proposed a new particle picking method based on a semi-automatic approach [9]: a few particles are manually selected in a restricted region of interest, and they are used to train a classifier to automatically pick the particles in the rest of the images. The challenge is to deal with a very low number of training particles (typically a few tens) to reduce as much as possible the user interaction.

The goal of this internship is to start from this baseline picking method, improve it, and compare it with existing state-of-the art approaches. Specifically, a study will be conducted regarding the model architecture, from neural network design to tailored loss improvements. Several popular deep-learned detection methods will be compared. Finally, the possibility to apply our method to another microscopy acquisition, namely cryo-electron tomography should also be considered. The developed code must be of high quality in order to be easily shared and used by biologist. The internship should result in the redaction of a research article.

Working environment

The intern will be a member of the IMAGeS team in the ICube laboratory in Illkirch. The internship will begin between January and May 2025, for a period of approximately 6 months.
Supervisors: Luc Vedrenne (PhD Student, luc.vedrenne@etu.unistra.fr), Denis Fortun (CNRS researcher, dfortun@unistra.fr), Etienne Baudrier (Assistant Professor, baudrier@unistra.fr).

Profile of the candidate

We are seeking a motivated M2-level student enrolled in a program specializing in computer science, machine learning and deep learning. Proficiency in the Python programming language is required. A knowledge of the PyTorch framework and an interest in biology and microscopy would be advantageous but are not required. The successful candidate will work closely with our team, benefiting from collaboration with biologists.

Application

Send a CV and a short description of your motivation, as well as the transcript of marks for the past 2 years to Luc Vedrenne(luc.vedrenne@etu.unistra.fr), Denis Fortun (dfortun@unistra.fr), and Etienne Baudrier(baudrier@unistra.fr).

References