Machine learning, docking, or physics for structure prediction of ligand-induced ternary complexes

Proteolysis-targeting chimeras (PROTACs) and molecular glues promote targeted protein degradation by recruiting an E3 ligase to proteins of interest (POIs). An accurate 3D structure of the ternary complex formed by E3 ligase, ligand, and POI is central to the rational design of degraders. Elucidating this structure with crystallography or cryo-EM can be challenging due to conformational flexibility, dynamic protein-protein interactions, and high-dimensional binding landscapes. To facilitate structure-based design in the absence of an experimental structure, computational approaches have been proposed: (i) multistep methods involving traditional docking pipelines, and (ii) single-step methods with deep learning models to directly predict the complex structure. Multistep methods are limited by sampling complexity, accurate input structures, scoring accuracy, and computational cost, while single-step methods are faster but are constrained by training-data scarcity. Here, we examine recent advances and emerging tools in modeling ternary complexes, critically discuss their predictive power and limitations, and highlight remaining challenges. • Methods for predicting the structure of ternary complexes can be classified into two categories: multistep and single-step methods. • Multistep methods are computationally demanding and highly dependent on force fields and simulation settings. • Single-step methods are faster but often lack sufficient generalizability. • Physics-based complementary methods can potentially improve the quality of the predicted structures.