Project description
Predicting interactions of proteins by machine learning
Conventionally, the 3D structure of a protein is defined using experimental methods such as X-ray crystallography and cryo-electron microscopy. Recent advances in AI offer the potential to predict protein structure from its amino acid sequence with great accuracy. This is not the case, however, with protein dynamics and protein interactions that are central for predicting protein function. Funded by the Marie Skłodowska-Curie Actions programme, the MARMOTTE project aims to develop methods for predicting the interaction of a dynamic molecular structure. This will provide important insight into protein complexes and facilitate faster and more advanced drug discovery.
Objective
This year has seen a breakthrough in structural bioinformatics - deep learning-based methods, most notably Google DeepMind's AlphaFold2, have demonstrated near-experimental accuracy of protein structure predictions. However, even the best protein structure prediction methods do not automatically provide knowledge about protein dynamics and protein interactions, which is often essential to understand or predict the biological functions of proteins. Those functions are performed via intermolecular interactions, and such interactions almost always involve conformational changes of engaged partners. The problem of modeling dynamic protein structures and their complexes is still largely unsolved - this project aims to significantly contribute towards its future solution by exploring the link between computational geometry, statistical physics, and machine learning. The postdoctoral researcher will develop novel methods that: given a dynamic (moving) molecular structure, efficiently compute tessellation-derived contact areas; given a starting structure and its tessellation-derived contacts areas, predict (using a graph neural network) how the interatomic contact areas will change upon motion; given a protein complex model generated by docking, use the predicted statistical properties of the contact areas to estimate (using a graph neural network) the protein-protein binding energy score. If successfully developed, such methods will provide unique data about the dynamics of tessellation-derived interatomic contact areas. Most importantly, they will provide effective dynamics-aware scores for assessing and ranking structural models of protein complexes.
Fields of science
- natural sciencesbiological sciencesbiochemistrybiomoleculesproteins
- natural sciencesmathematicspure mathematicsgeometry
- natural sciencescomputer and information sciencesartificial intelligencemachine learning
- natural sciencescomputer and information sciencesartificial intelligencecomputational intelligence
Programme(s)
- HORIZON.1.2 - Marie Skłodowska-Curie Actions (MSCA) Main Programme
Funding Scheme
HORIZON-AG-UN - HORIZON Unit GrantCoordinator
75794 Paris
France