2026/01/20 by Ayşe Berçin Barlas, Benoist Laurent, Ezgi Karaca +2 · 1 voice
Biochemistry, Genetics and Molecular Biology · Materials Science · #Protein Structure and Dynamics #Bioinformatics and Genomic Networks #Enzyme Structure and Function
paper · doi:10.1016/j.jmb.2026.169650
openalex publication_date 2026/01/20 · openalex created_date 2026/01/21 · openalex updated_date 2026/07/23
• The assessment of protein interaction predictions usually relies on a single reference structure (provided by the experiments), even though protein assemblies and protein-protein interfaces involve dynamic processes. • Integrating the dynamic properties of protein-protein interface models (which can be obtained by tools such as MD simulations) can lead to notable changes when ranking the quality of the docking predictions (see Prévost and Sacquin-Mora, Proteins 2021). • The Dynabench database provides all-atom MD trajectories (three replicas, each one 100 ns long) for over 200 protein complexes listed in the DB5.5. Thus granting access to the dynamics of experimental interfaces, which can then serve as references when investigating the predicted interfaces produced by docking tools. • These simulations also offer a resource to explore interfacial flexibility, train machine learning models, redefine accuracy metrics for model evaluation, and informe the design of protein interfaces. Protein–protein interactions are central to numerous cellular processes, including transport, signaling, and immune response. Structural modeling of protein assemblies typically relies on AlphaFold or docking methods, which produce structural models evaluated against a single experimental reference. While AlphaFold2 and its extension, AlphaFold-Multimer, have advanced complex prediction, they, and conventional docking tools, offer only static representations. However, flexibility at protein–protein interfaces is increasingly recognized as critical for function. To address this limitation, DynaBench provides a benchmark of interface dynamics in biologically relevant protein assemblies. We performed MD simulations for over 200 protein-protein complexes listed in the Docking Benchmark 5.5 ( https://zlab.umassmed.edu/benchmark/ ), generating three 100 ns long replicas per complex. All trajectories are now publicly available online ( http://www-lbt.ibpc.fr/DynaBench ) via the MDposit platform (INRIA node), which is part of the EU-funded Molecular Dynamics Data Bank (MDDB). These simulations offer a unique resource for exploring interfacial flexibility, training machine learning models, redefining accuracy metrics for model evaluation, and informing the design of protein interfaces.