Am 14.11.2025 um 11:00 spricht Dr. Helgi I. Ingólfsson (Lawrence Livermore National Laboratory, Kalifornien) zum Thema “Capturing lipid-protein interactions during RAS-RAF activation using machine learning-driven multiscale simulations”.

Abstract: 

Molecular dynamics (MD) simulations offer detailed insights into biological mechanisms. However, using MD to capturing conformational transitions between different protein states can be extremely challenging, especially given complex reaction coordinates, unknown, and/or rugged energy surfaces. To address these challenges, we extended the massively parallel Multiscale Machine-Learned Modeling Infrastructure (MuMMI) by incorporating a machine learning (ML)-based structure generator. Two sets of MD simulation ensembles are executed, representing the two endpoints we want to sample protein configurations between. The ML structure generator is trained on the endpoint ensembles and a MuMMI multiscale simulation is then run, where ML-generated structures are drawn from latent space paths between the two endpoint ensembles. For each selected structure, MD simulations are performed, lipids are pre-equilibrated using a macro model, and the systems are set up and run using coarse-grained simulations. These running simulations guide further structure selection, refining the sampling between the endpoints. The ML model is iteratively retrained with incoming simulation data, enhancing the latent space representations until sufficiently fine-grained paths between the two endpoint ensembles are achieved.

Using the extended MuMMI protein path sampling framework we mapped the RAS-RAF protein conformations from autoinhibition to activation. The autoinhibited RAF-14-3-3₂ complex binds RAS on the plasma membrane, and multiple protein-membrane engagements help loosen the autoinhibited state. After the release of RAF autoinhibition two RAF kinase domains associate forming a fully formed dimer complex.