Regensburg 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

PLV: Plenarvorträge

PLV VIII

PLV VIII: Plenarvortrag

Donnerstag, 20. März 2025, 08:30–09:15, H1

Learning how biomolecules move and undergo chemical reactions — •Frauke Gräter — Max Planck Institute for Polymer Research, Mainz, Germany

Life is biochemistry in action. While molecular simulations of systems as complex as whole cells are now within reach, predicting chemical reactivity on relevant time and length scales remains a challenge. I will present our recent work towards bringing action * here: chemistry * to classical atomistic simulations and molecular design through machine learning.

We substitute costly quantum mechanical calculations with a graph neural network-based emulator which predicts reaction rates without explicitly modelling the reaction. To deal with the chemistries arising from the such reactions, we have developed a framework to parametrize a classical force field. GRAPPA leverages graph attention, is highly accurate, and can be easily fine-tuned. Our ML-based simulations can predict cascades of chemical reactions amidst the 'jiggling and wiggling' of biomolecules at an efficiency close to classical simulations.

Finally, I will demonstrate how we harness a flow-matching model based on geometric algebra and trained on Molecular Dynamics simulations to design novel proteins with tailored flexibilities. Our method generates conformational ensembles of unseen proteins without the need to run costly Molecular Dynamics simulations, and paves the way for generating novel proteins with biochemical functions that rely on molecular motions.

Keywords: machine learning; molecular dynamics; protein design; graph neural network; force field