AI-Based Prediction of PROTAC- and Molecular Glue-Mediated Ternary Complexes: A Comparative Evaluation of AlphaFold 3 and Boltz-2

Abstract

Proteolysis targeting chimeras (PROTACs) and molecular glues induce ligand-mediated ternary complexes between an E3 ubiquitin ligase and a protein of interest, but their in silico modeling remains challenging due to conformational flexibility and weak protein-protein interfaces. Recent diffusion-based AI structure prediction models enable the direct prediction of protein-ligand complexes. Here we benchmarked AlphaFold 3 and Boltz-2 for predicting PROTAC- and molecular glue-mediated ternary complexes using a reproducible evaluation workflow. We curated a dataset of 40 experimentally resolved complexes from the Protein Data Bank, including 25 PROTAC and 15 molecular glue systems. Structural accuracy was assessed using complex RMSD and DockQ scores relative to the corresponding crystal structures and compared to model-internal confidence metrics. Both models outperform other current approaches in both accuracy and runtime. Boltz-2 shows higher prediction accuracy assessed by complex RMSD and DockQ scores. Predictions are generally more accurate for VHL-based PROTACs than for CRBN-based PROTACs. Predictions for molecular glue complexes show good overall accuracy. Error analysis indicates that prediction failures predominantly arise from misoriented global arrangements and twisting in flexible ternary complexes, while individual protein and ligand structures are often accurately modeled. Limitations in the generalizability of the models could also be observed, especially for more recently released structures. These findings suggest that diffusion-based AlphaFold-type models show promise in the structure-based prediction of PROTAC- and molecular glue-mediated ternary complexes.