Description
Date depot: 28 mars 2025
Titre: Geometric Machine Learning for Enhanced Modeling of Elongation and Theoretical Inquiries
Directrice de thèse:
Elodie LAINE (LCQB)
Encadrant :
Grégoire SERGEANT-PERTHUIS (LCQB)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Sciences de l’information et sciences du vivant
Resumé: The ribosome, a huge macromolecular complex comprised of RNAs and proteins, is responsible for translating messenger RNA (mRNA) into proteins. The movement of the ribosome along the mRNA, known as elongation, is crucial to this process, enabling the sequential attachment of amino acids to the protein undergoing biosynthesis. Yet, the associated conformational changes and the roles played by the ribosomal proteins remain largely unknown. Physics-based methods such as molecular dynamics simulations provide a means for fine-grained investigation of these questions. However, they remain limited in terms of space and time scales and are thus not adapted, in practice, to study such big and complex systems as the ribosome. In this project, we propose to leverage geometric deep learning, generalising Graph Neural Networks (GNN) to heterogenous structures, and Markov Random Fields (MRFs) to accelerate molecular dynamics simulations, using the elongation process as a case study. Our original approach will rely on the introduction of topological priors formalised as sheaves over partially ordered sets and defining coarse-grained models of specific regions of the ribosome that interact through overlapping sub-parts. We will shed light on the elongation process and provide an interpretable description of the learned dynamics by assessing the impact of the topology of these sheaves on the quality of inference and learning and by establishing a classification of critical points of the associated losses.