TriGlue: a Biology-Inspired Generative Model for Generating Molecular Glue-Induced Ternary Complex
Yuliang Yan, Shuo Yan, Haochun Tang, Yiqin Sun, Enyan Dai
cs.LG, cs.AI
2026-07-24
TriGlue estimates the protein interface first, then jointly generates the glue molecule and docking pose via flow matching, beating pocket-given baselines on RMSD, drug-likeness, and docking score.
A molecular glue is a small-molecule drug that pulls an E3 ubiquitin ligase and a target protein together into a ternary complex, triggering the cell to degrade the target. That lets you drug proteins ordinary inhibitors cannot touch. But designing a new glue from scratch is hard: you must generate a novel small molecule, dock two proteins in the right relative pose, and ensure the molecule bridges a compatible protein interface, all when that interface does not exist until the complex forms. Structure-based generative drug design (AR, TargetDiff, PocketXMol) assumes a single protein with a predefined binding pocket; molecular glues have no pocket. DeepTernary predicts ternary structure for a known ligand but does not design one.
TriGlue mirrors the real biology: the glue induces the interface first, then the complex assembles. It splits the joint problem into two stages.
Stage one is interface estimation. An SE(3)-equivariant EGNN encodes each protein chain; cross-protein attention maps each chain to K virtual interface points, and a Gaussian-ellipsoid constraint pulls those points to match the true interface distribution. Equivariance means predictions do not depend on an arbitrary coordinate frame. On isolated monomers the interface is otherwise invisible, so this gives a compact, geometry-aware prior.
Stage two is interface-conditioned ternary flow matching. One multimodal conditional flow-matching framework jointly denoises three flows: a rigid-body transform flow (rotation on SO(3), translation in Euclidean space), a ligand-coordinate flow, and an atom-type flow on the probability simplex. The transform is initialized by differentiable Kabsch alignment of the two virtual interfaces, then refined iteratively, with AlphaFold's IPA as the latent encoder. Training uses 22,000 ternary complexes from TernaryDb, clustered by sequence similarity with 94 held-out test complexes at zero sequence overlap with training.
Every baseline (AR, TargetDiff, PocketXMol) was given a binding pocket; TriGlue was not, and still leads.
| Metric | PocketXMol | TriGlue |
| RMSD (avg) | 6.74 | 5.22 |
| QED avg (drug-likeness) | 0.69 | 0.79 |
| Vina dock (binding) | -7.65 | -8.85 |
On interface estimation TriGlue beats PeSTo and ScanNet by roughly an order of magnitude (W2 distance 12.65 vs 58). Against the only direct ternary competitor, DeepTernary, the DockQ-above-0.23 acceptance rate is 35.2% vs 30.8%. Removing interface correction drops average DockQ from 0.22 to 0.15, confirming the conditioning matters. Feeding generated glues to AlphaFold3 gives high pTM and ipTM for most families.
This is one of the few works to treat de novo molecular glue design as a generative task and to confront the unknown-interface difficulty head-on. For computational drug design, the two-stage estimate-then-condition recipe is transferable to other interface-induced drug problems. Code is open-sourced.
There is no dedicated limitations section, and several hard problems stand out. Every result is computational (Vina, DockQ, AlphaFold3) with no wet-lab validation: no synthesized molecule, no degradation assay, no binding measurement. The model is rigid-body only and assumes proteins do not change conformation on binding, whereas real glues such as CRBN allosteric remodeling involve backbone and side-chain flexibility, a modeling gap rather than an engineering one. Synthesizability (SA) is worse than PocketXMol, which the paper attributes to PocketXMol's roughly 100-million-molecule training set. The best-case DockQ ceiling is lower than DeepTernary's; TriGlue wins on average stability. The test set comes only from TernaryDb's own pipeline, with no external independent set.