Our research develops and applies machine-learning methods for biomolecular structure prediction, spanning proteins, protein complexes, RNA and cell membranes. Describing biomolecules in their biological context is central to understanding molecular mechanism and underpins applications in biotechnology, immunology and drug discovery. We apply for Arrhenius to run two parts of this programme that fit AI-dedicated hardware poorly: large-scale RNA structure prediction, whose dominant stage cannot be GPU-accelerated, and multiscale modelling of cell membranes, whose production stage is classical molecular dynamics rather than machine learning.
RNA structure prediction. Deep-learning methods such as AlphaFold3 and Boltz now predict RNA structure, but accuracy remains far below that achieved for proteins, and our own work has mapped the limits of these methods (Ludaic and Elofsson, under revision at NAR). Two factors dominate: the shallow and hard-won homology information available for RNA, and the neglect of ions, which are essential to RNA folding. We are generating structural ensembles for the approximately 4,000 Rfam families that lack experimentally resolved structures, and will retrain Boltz-2 to account for ion-dependent RNA structure, building on our CASP16 work on blind prediction of water and ion ensembles around RNA (Kretsch et al., Proteins 2026). Unlike proteins, RNA homology detection requires secondary-structure-aware covariance models - Infernal cmsearch and cmalign against Rfam and RNAcentral - for which no GPU implementation exists. This CPU stage, not the GPU inference that follows it, currently limits the project.
Multiscale membrane modelling. Within CEMERAI (Centre for Modelling Life at the Membrane with AI; Stockholm University, University of Gothenburg and Umea University), we build predictive multiscale models of erythrocyte and mitochondrial membranes. Current structural biology and AI methods treat membrane proteins one at a time and rarely capture how proteins, lipids and glycans jointly shape membrane organisation. We extend our AlphaFold-based methods for complex prediction (Bryant et al., Nat Commun 2022; Burke et al., Nat Struct Mol Biol 2023) to protein complexes embedded in a lipid environment, integrate experimental constraints from proteomics, glycoproteomics and lipidomics generated by CEMERAI partners, and scale from individual complexes to the organisation of whole membrane patches. Coarse-grained and all-atom molecular dynamics with GROMACS provide both the physical description and data to validate models.
Both workstreams share a structure: a large CPU stage producing multi-terabyte intermediates, feeding a GPU stage that consumes them. Separating the two across sites would mean transferring these datasets between centres for every campaign, which is not practical. This is why we request CPU and GPU resources on one system.
Our deep-learning method development remains on Berzelius, which is dedicated to AI workloads: AlphaFold-based protein complex and antibody-antigen prediction, flexibility prediction flow-matching generative models, and physically informed all-atom generative modelling.
The group comprises six PhD students, two postdocs and two senior scientists, funded by Vetenskapsradet, the Knut and Alice Wallenberg Foundation, SciLifeLab/DDLS and SeRC. We collaborate with NBIS on improving structure-prediction workflows for the broader Swedish user community.