GluePlex: harnessing a co-folding model to predict molecular-glueinduced CRBN-neosubstrate ternary complex structures
Year: 2026
Type: preprint
Abstract: Molecular glue degraders act through a single induced geometry, and predicting that geometry in silico is the step that would allow degrader design to become rational rather than empirical. However, current co-folding performance on structurally novel neosubstrates remains uncertain. Here, we assembled 52 cereblon (CRBN) target–compound pairs with published ternary structures and stratified them by training-set exposure. On the 13 novel pairs, plain Boltz-2 co-folding produced an acceptable pose in 12 of 39 independent runs, with a median DockQ (protein-interface accuracy score) of 0.01. GluePlex combines ligand-free, restrained protein–protein docking with two passes of template-guided Boltz-2 co-folding. It increased success to 23 of 39 runs, with a median DockQ of 0.43 and 19 rank-1 successes among them. Ablation analysis, which tests the contribution of pipeline components by removing or replacing them, attributed most of this gain to docking and template transfer. Despite this improvement, ligand accuracy remained limited: the CRBN-binding warhead had a median root-mean-square deviation (RMSD) of 0.7 Å, compared with 2.8 Å for the distal moiety, whose placement remained inaccurate even when the protein interface was correct. Exploratory ligand truncations rescued some complexes but worsened others. The parent model’s target–ligand confidence helped identify cases likely to benefit, although these comparisons assume that truncation preserves the parent’s ternary geometry. Changing the docked conformer affected individual compounds without improving overall success, supporting the use of multiple conformers. Across configurations and truncations, an acceptable pose was recovered for 12 of 13 novel families while the remaining 1 can be recovered by plain co-folding, but no single configuration solved them all. In general, GluePlex improves prediction of novel CRBN ternary complexes, while model ranking and distal ligand placement remain key limitations.
Source: Research Square
Authors: Jin Wang, Hanfeng Lin, Haiyang Zheng
Institution: Baylor College of Medicine
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225821626
Cites:
Cited by:
Related to: