Presentation by Fodil Azzaz of his project to integrate the AFMB via ERC Starting Grant funding

Abstract

Membrane proteins are at the core of numerous pathological pathways, ranging from intoxication to cancer. Obtaining accurate structural models of these therapeutic targets is therefore a major challenge to understand the molecular mechanisms of disease and to design targeted treatments. While experimental structures and AI-based prediction (e.g AlphaFold) represent fundamental advances, they often capture the protein in an isolated state, outside its native membrane context. They fail to account for the membrane constraints that actively reshape the conformational landscape of proteins, an effect we have termed the epigenetic dimension of protein structure. To address this challenge, we present a new computational approach that combines all-atom molecular dynamics simulations with graph neural networks. Our method uses structures predicted by molecular dynamics simulation to train an AI model the fundamental relationship between the identity of surrounding lipids and the structural deformations they induce on the protein. The outcome is a tool capable of predicting the functional and physiological structure of a membrane protein in its native environment, thus providing a crucial advance for antidote design.