Benchmarking Cofolding Methods for Molecular Glue Ternary Structure Prediction

Molecular glues (MGs) represent an emerging therapeutic paradigm capable of inducing or stabilizing protein–protein interactions (PPIs), with broad applications in creating neomorphic interactomes and targeted protein degradation. However, current discovery efforts remain largely confined to experimental screening, while in silico rational design of MGs remains a formidable challenge. A critical step toward rational design lies in accurate ternary complex modeling, which is less explored and highly challenging due to the involvement of small-molecule-induced de novo PPIs. Here, we tested the ability of recently developed cofolding models, including AlphaFold 3, Boltz-1, Chai-1, Protenix, and RoseTTAFold All-Atom. Although these models were not specifically trained on ternary complex structures, whether their capability to learn diverse interatomic interactions can generalize well to such ternary systems remains an open question. We systematically curated a data set, named MG-PDB, with 221 noncovalent MG-engaged ternary complexes. MGBench was further introduced as a comprehensive benchmark set, which comprises 88 ternary structures excluded from cofolding models' training data through rigorous time-based partitioning. Our benchmark results demonstrated that AlphaFold 3 achieved the best overall performance among cofolding methods, in terms of both PPI interface prediction (50.6% success rate) and MG–protein interaction recovery (32.9% success rate). However, our homology study showed that most of their successful predictions actually stemmed from memorization. Further analysis revealed three phenomena of current cofolding methods for MG ternary structure prediction. First, these methods struggle to accurately model large interaction interfaces. Second, their predictive accuracy is notably reduced for domain–domain complexes compared to domain–motif interactions. Lastly, they face specific challenges in modeling MG degrader complexes with sufficient accuracy. We showcased they relied on the existing interaction patterns and highlighted the need for further improvements in novel E3 ligase systems. These findings reveal fundamental gaps in existing methods to learn atomic-level interaction rules for MG-engaged ternary complex modeling. As fully open resources, MG-PDB and MGBench establish the essential benchmark for MG ternary complex modeling, providing the definitive standard for evaluating future cofolding methods.