YDS-GlueFold: Surpassing AlphaFold 3-Type Models for Molecular Glue-Induced Ternary Complex Prediction

Molecular glues represent an innovative class of drugs that enable previously impossible protein-protein interactions, but their rational design remains challenging, a problem that accurate ternary complex modeling can significantly address. Here we present YDS-GlueFold, a novel computational approach that enhances AlphaFold 3-type models by incorporating guided diffusion during inference to accurately predict molecular glue-induced ternary complex structures. We demonstrate YDS-GlueFold’s capabilities across eight diverse test cases, including both E3 ligase-based systems (VHL, CRBN complexes with mTOR-FRB, NEK7, and VAV1-SH3c, and KBTBD4 complexes with HDAC1) and non-E3 ligase complexes (FKBP12 complexes with mTOR-FRB, BRD9 and QDPR). The model achieves remarkable accuracy with RMSD values consistently below 2.5 Å compared to experimental structures. Importantly, 7 out of 8 test cases involve protein-protein pairs that were not present in the AlphaFold 3 training set, providing a rigorous test of the model’s ability to generalize beyond its training data. Notably, in the FKBP12 case, YDS-GlueFold correctly predicts a novel interface configuration instead of defaulting to known interactions present in training data, further demonstrating true generalization rather than mere memorization. Our results suggest that strategic enhancement of the inference process through guided diffusion can significantly improve ternary complex prediction accuracy, potentially accelerating the development of molecular glue therapeutics for previously undruggable targets.