{"id":11851,"date":"2026-01-10T09:14:55","date_gmt":"2026-01-10T08:14:55","guid":{"rendered":"https:\/\/fermi.univ-tlse3.fr\/?post_type=tribe_events&p=11851"},"modified":"2026-01-12T09:38:46","modified_gmt":"2026-01-12T08:38:46","slug":"a-neural-network-assisted-exploration-of-the-complex-landscape-of-biomolecules-with-first-principles-accuracy-jer-lai-kuo-lcar-lcpq-seminar-23-01-2026-11h","status":"publish","type":"tribe_events","link":"https:\/\/fermi.univ-tlse3.fr\/fr\/home\/event\/a-neural-network-assisted-exploration-of-the-complex-landscape-of-biomolecules-with-first-principles-accuracy-jer-lai-kuo-lcar-lcpq-seminar-23-01-2026-11h\/","title":{"rendered":"A Neural Network\u2013Assisted Exploration of the Complex Landscape of Biomolecules with First-Principles Accuracy. – (Jer-Lai Kuo\/ LCAR-LCPQ \/ Seminar). – 23\/01\/2026, 11H"},"content":{"rendered":"

S\u00e9minaire LCAR-LCPQ<\/span><\/p>\n

Jer-Lai Kuo, <\/strong><\/span> Director of the CECAM-TW node; Research Fellow, Institute of Atomic and Molecular Sciences, Academia Sinica<\/span><\/p>\n

Summary<\/strong> :
\nStructure of simple biomolecules (such as peptides and saccharides) are known to be flexible. While molecular dynamics with empirical force fields have gained plenty of success, however, it\u2019s applicability to chemical reaction is often limited. While First-Principle methods (such as DFT) are powerful \u2013 the simulation length is limited due to the high cost of DFT. In a few studies on mono-saccharides, we have proposed a self-adapting scheme to train NNP (neural network potential) to reach quantum chemistry accuracy of 1 kcal\/mol based on a multi-model approach. A typical structural searching algorithm is to begin with more efficient (but less accurate) methods such as empirical force field or semi-empirical modes (for example DFTB) to explore the energy scape of the flexible bio-molecules. For a given aldohexose (AH), we can find 500~100 structurally distinct conformers covering an energy range of ~ 150 kcal\/mol. We demonstrated that from these trajectories (even if some of them lead to the same local minima), we can efficiently generate a data set in the order of 40k to 50k \u2013 this energetically and structurally distinct data set can be used to train an NNP with an accuracy of 1 kJ\/mol in energy and 1 kJ\/mol\/A in gradient. A self-adapt scheme is to cover all kinds aldose and ketose with the same accuracy by adding a small fraction of data from new types of aldoses and ketoses. The NNP trained in this way is decent enough to perform molecular dynamics simulations that cover an energy range of more than 200 kJ\/mol.<\/p>\n