Benchmarking Cofolding Methods for Molecular Glue Ternary Structure Prediction.

Area of Science:

• Computational biology
• Structural biology
• Drug discovery

Background:

• Molecular glues (MGs) offer a novel therapeutic approach by modulating protein-protein interactions (PPIs) for targeted protein degradation.
• Rational design of MGs is hindered by the challenge of accurately modeling ternary complexes, especially de novo interactions induced by small molecules.

Purpose of the Study:

• To evaluate the performance of state-of-the-art cofolding models in predicting the structures of MG-engaged ternary complexes.
• To establish comprehensive benchmark datasets (MG-PDB and MGBench) for assessing computational methods in this domain.

Main Methods:

• Systematic curation of the MG-PDB dataset comprising 221 noncovalent MG-ternary complexes.
• Creation of the MGBench benchmark set using time-based partitioning to exclude training data.
• Benchmarking AlphaFold 3, Boltz-1, Chai-1, Protenix, and RoseTTAFold All-Atom on ternary complex prediction.

Main Results:

• AlphaFold 3 demonstrated the highest performance among tested cofolding models for PPI interface prediction (50.6%) and MG-protein interaction recovery (32.9%).
• Model performance was largely attributed to memorization rather than generalization, with significant challenges in predicting large interfaces, domain-domain interactions, and MG degrader complexes.
• Existing models struggle with novel E3 ligase systems, indicating reliance on known interaction patterns.

Conclusions:

• Current cofolding models exhibit fundamental limitations in accurately modeling atomic-level interactions within MG-engaged ternary complexes.
• The developed MG-PDB and MGBench resources provide essential benchmarks for advancing computational methods in molecular glue ternary complex modeling.
• Further advancements are crucial for the rational design of molecular glues, particularly for novel therapeutic targets.