Breakthrough in protein simulation: Berlin researchers revolutionize method!

Breakthrough in protein simulation: Berlin researchers revolutionize method!

On July 18, 2025, an international research team led by Prof. Dr. Cecilia Clementi from the Free University of Berlin a groundbreaking article in the magazineNature Chemistry, which is causing a stir in the scientific world. The article presents a new model that enables precise and efficient simulation of proteins - significantly faster than conventional molecular dynamics simulations.

The challenge of realistically depicting protein folding and dynamics has existed for over 50 years. The researchers used deep learning to develop a system that approximates all-atom protein simulations. The newly developed model, CGSchNet, uses a graph neural network to learn effective interactions between particles. These innovative approaches open up promising applications in the development of drugs and antibody therapies for cancer and other diseases.

Advances in Protein Simulation

The model focuses on the dynamic folding process of proteins, including the intermediate states that play a role in misfoldings such as amyloids. It simulates transitions between folded states that are crucial for the function of proteins. A significant advance of this model is the ability to accurately predict long-lived states of folded, unfolded, and disordered proteins.

One of the most notable features of the CGSchNet model is the ability to predict the relative stability of folded protein mutants, which previous methods have not been able to achieve. Prof. Dr. Cecilia Clementi, who was previously professor of chemistry and physics at Rice University in Houston, Texas, is strengthening research in theoretical and computational biophysics at the Free University of Berlin. Their work is supported by the Einstein Professorships, which support Berlin universities and the Charité in appointment or residency negotiations.

Data and models in detail

The study used comprehensive data sets. The benchmark dataset contains 1,262 targets and covers a wide range of protein structures, including 717 single domain proteins, 257 newly published targets, and 288 targets from CASP 8–14. These data sets were prepared to remove redundancies and adjust sequence identity to a cutoff of <30%.

The Human Protome dataset includes 20,595 proteins and enables differentiated analysis of single-domain and multidomain proteins. Of these, 19,512 proteins were predicted, which is potentially of great importance for structure prediction. In addition, the D-I-TASSER pipeline, which uses a hybrid approach to proteon structure prediction, was integrated into the research.

This pipeline performs all steps from deep MSA generation to threading template identification to structural refinement and model evaluation. The use of state-of-the-art algorithms and technologies such as DeepPotential and AttentionPotential illustrates the innovative character of the work.

The results of the study are transferable to a variety of proteins outside of the specific training data set, which underlines the relevance of the model for biochemical research. Application of the methods developed in the study could have far-reaching implications for the future of protein research and drug development. Further information is available in the detailed publication Nature to find.