Improving flexible structure fitting into cryo-EM maps using multiple conformers generated by AlphaFold 2
Title: Improving flexible structure fitting into cryo-EM maps using multiple conformers generated by AlphaFold 2
SNIC Project: Berzelius-2022-212
Project Type: LiU Berzelius
Principal Investigator: Erik Lindahl <erik.lindahl@scilifelab.se>
Affiliation: Kungliga Tekniska högskolan
Duration: 2022-11-01 – 2023-05-01
Classification: 10603
Homepage: http://www.biophysics.se
Keywords:

Abstract

The resolution revolution has increasingly enabled single-particle cryo-EM reconstructions of previously inacces- sible systems, including large membrane protein complexes that constitute a disproportionate share of drug targets. Advanced techniques for model reconstruction use molecular dynamics simulations to refine atomistic models into cryo-EM maps [Igaev et al. eLife 2019, Blau et al. bioRxiv 2022]. However, proteins with multiple functional states remain a challenge for simulations-based refinement methods, since an initial model may be constructed in a conformation that differs significantly from the functional state captured in cryo-EM density. AlphaFold2 (AF2) has recently demonstrated the capability of generating an ensemble of protein states covering the conformational landscape of protein function [Del Alamo et al. eLife 2022]. Reducing the depth of the input multiple sequence alignments by stochastic subsampling as well as sequence clustering [Wayment-Steele et al. bioRxiv 2022] can lead to the generation of accurate models in multiple conformations. AF2-generated ensemble of protein states can be used to provide multiple inputs for density-guided simulation pipelines to reduce the initial model bias and hence improve fitting accuracy. This project will focus on the improvement of simulation-based density fitting techniques using currently released deep learning tools. The capability of the AlphaFold2 conformers package [Del Alamo et al. eLife 2022] to generate multiple valid input models for density-guided simulations and their impact on fitting results will be assessed. To test our approach we will use a diverse range of proteins that have multiple functional states and possess biomedical interests, including amino acid transporters, G-protein coupled receptors (CGPCR, FZD7, dopamine receptors etc.), ion channels, and several soluble cytosolic systems. While the AF2 will be run on Berzelius, MD simulations will be performed in our local cluster, Dardel, and EuroHPC platforms where we already have allocations. My group has vast experience in the development of the cryo-EM analysis tool RELION [Kimanius et al. eLife 2016] which has significantly improved reconstruction resolution. Furthermore, we have recently developed a density- guided simulations tool that fits structures via Bayes’ approach [Blau et al. bioRxiv 2022]. We also have a long track record in the field of molecular modeling where we develop the Gromacs MD toolkit [Abraham et al. SoftwareX 2015]. In the field of membrane proteins, the group has significant expertise in molecular modeling and experimental cryo-EM structure solving and has published numerous works on the conformational transitions and allosteric mod- ulation of ion- channels/transporters [Kim et al. Nature 2020, Bergh et al. eLife 2021]. The proposed integration of deep learning methods with MD-based fitting could help us take one step further to improve structure building quality. In total, we are requesting 19,800 GPU-hours of allocations and default storage space.