Generating Dynamic Structures Through Physics‐Based Sampling of Predicted Inter‐Residue Geometries

ABSTRACT Deep learning‐based methods, such as AlphaFold2, have revolutionized the prediction of static protein structures. However, modeling alternative conformations and dynamic structures remains an unsolved problem. Here, we present trRosettaX2‐Dynamics (trX2‐D), an innovative solution building on our CASP15 and CASP16 winning method, trRosettaX2. trX2‐D tackles this challenge by employing physics‐based iterative sampling of trRosettaX2's predicted inter‐residue geometric distributions. The model underwent pre‐training on high‐resolution X‐ray structures, followed by fine‐tuning on approximately 7000 dynamic NMR structures. This dual training regime significantly bolsters its capacity to predict alternative conformations and dynamic structures. At its core, trX2‐D employs a Transformer‐based neural network to initially predict a set of inter‐residue geometric constraints. These constraints are then iteratively sampled to generate dynamic structures, entirely circumventing the need for prior knowledge of native structural states. Extensive benchmarking across three distinct datasets—two focused on alternative conformations and one on dynamic structures—demonstrates trX2‐D's promising ability to predict alternative conformations and accurately capture structural dynamics. This work highlights the potential of integrating deep learning predictions with physics‐based sampling to advance the field of protein dynamic structure prediction.