AI-driven protein simulations from Berlin: TRR186 researcher contributes to two landmark studies

clementi nature chemistry
Date

Over the summer of 2025, TRR186 scientist Prof. Cecilia Clementi (Freie Universität Berlin) contributed to two high-profile publications in Nature Chemistry and Science, bringing breakthroughs in protein simulations.

Advancing Protein Simulation with Machine Learning

    In Nature Chemistry, Clementi and an international team introduce CGSchNet, a machine-learned coarse-grained model for proteins. The approach trains a graph neural network on thousands of all-atom molecular dynamics simulations to learn effective interactions between coarse-grained beads. CGSchNet can reproduce complex folding and unfolding dynamics at a fraction of the computational cost of all-atom simulations, while still capturing intermediate and metastable states that are crucial for protein function and misfolding, such as the formation of amyloid, which are pathological protein aggregates that appear in cases of Alzheimer’s disease, for example. The model also generalizes to proteins outside its training set and can estimate relative folding free energies of protein mutants, making it a promising tool for protein engineering and drug discovery.

“This work is the first to demonstrate that deep learning can lead to a simulation system that approximates all-atom protein simulations without explicitly modeling solvent or atomic detail” 

Prof. Cecilia Clementi.

Understanding How Proteins Function with Artificial Intelligence

    In Science, Clementi and colleagues from Microsoft Research AI for Science and Freie Universität Berlin report BioEmu, a generative deep learning system that emulates the equilibrium behavior of proteins. BioEmu can generate thousands of statistically independent protein structures per hour on a single GPU, integrating over 200 milliseconds of molecular dynamics simulations with experimental data. This allows the model to predict structural ensembles, thermodynamic properties, and stability changes with an accuracy that approaches laboratory measurements, while revealing hidden binding pockets, domain motions, and local unfolding events that are central to understanding protein function and drug design. 

“BioEmu provides a scalable method to model protein function at the genomic scale”

Prof. Cecilia Clementi.

Impact and Significance

Together, CGSchNet and BioEmu provide fast ways to explore how proteins move, fold, and bind ligands with near all-atom accuracy at a fraction of the computational cost of traditional methods. These studies are expected to have significant implications for understanding proteins dynamics and drug discovery. For TRR186, this means concrete new capabilities to:

  • Follow conformational changes in receptors and membrane protein complexes by efficiently sampling large-scale motions and state transitions.

  • Reveal “hidden” binding pockets by generating rare conformations and local unfolding events that are invisible in single static structures.

  • Connect atomistic and mesoscale models by translating detailed simulations into coarse-grained descriptions of protein assemblies that remain faithful to their thermodynamics.

These advances also highlight the strength of the Berlin ecosystem in AI-driven molecular science, with TRR186 at its core.

About Prof. Cecilia Clementi

Professor Cecilia Clementi is a theoretical and computational biophysicist. Within the TRR186, she leads project A12 "Machine-learned coarse-grained molecular dynamics of molecular complexes." She has previously conducted research as an Einstein Visiting Fellow at the Collaborative Research Centers “Investigation of Membranes – Molecular Mechanisms and Cellular Functions” and “Scale Cascades in Complex Systems” at Freie Universität Berlin. She is also the first scientist to be permanently recruited to work in Berlin following her support as an Einstein Visiting Fellow. Before moving to Berlin in 2020, Clementi was a professor of chemistry and physics at Rice University in Houston, Texas.

Further information: