TriGlue

Molecular glue degraders are small molecules that work by sticking two proteins together: they recruit an E3 ubiquitin ligase to a target protein, and the resulting ternary complex tags that target for proteasomal destruction. Because a glue needs no deep pocket on the target itself, the modality reaches proteins conventional inhibitors cannot — and for the same reason it is hard to design computationally. There is no pocket to dock into: the binding site exists only once the two proteins have met, and the interface creating it is unknown at the outset.

TriGlue recasts molecular glue design as ternary complex generation. Given the unbound structures of an E3 ligase and a target protein, it generates the glue molecule and the relative pose of the two proteins at once, mirroring the two-step mechanism of glue action by splitting the task into interface estimation and then interface-conditioned complex assembly. That decomposition is the central move: it hands the generator a geometric prior about where the ternary pocket will form before it places a single atom.

TriGlue was developed by researchers at the Hong Kong University of Science and Technology (Guangzhou), Jilin University, and the University of Sydney. It joins structure-conditioned generative design frameworks like MagicDock, which generates ligands against a fixed receptor, and DFMDock, which samples protein-protein docking poses, but unlike either it must solve both at once.

#Key Features

• Ternary complex formulation: Generation is posed over the whole E3 ligase, target, and glue system, so ligand generation, docking, and complex assembly are solved jointly.
• SE(3)-equivariant interface estimation: An equivariant module predicts a geometrically constrained protein-protein interface from unbound monomers, supplying the pocket that structure-based drug design normally assumes.
• Ellipsoid interface prior: Parametrizing the interface as a Gaussian ellipsoid compresses a diffuse surface patch into a compact, differentiable conditioning target.
• Coupled ternary flow matching: One network generates the glue's atom coordinates and atom types alongside the rigid-body transformation that assembles the two proteins.
• Generalizes to unseen pairs: The model is trained once and run at inference on new target and E3 ligase pairs rather than refit per target.

#Technical Details

The interface stage encodes both proteins as graphs with a four-layer EGNN at hidden dimension 128 and predicts K virtual interface points per protein by attention-weighted aggregation of residue coordinates, supervised against a Gaussian ellipsoid fit to interface residue Cα atoms within 8 Å. The second stage applies interface-conditioned flow matching to three coupled variables: ligand atom coordinates, atom types over a simplex, and the rigid-body transformation, with rotations interpolated along SO(3) geodesics. The denoiser stacks two invariant point attention blocks at hidden dimension 128 and trains with reparametrized conditional flow matching.

Training and evaluation use TernaryDB, 22,303 ternary complexes curated from the Protein Data Bank and split by sequence-similarity clustering to prevent leakage. TriGlue reports a ligand RMSD of 5.22 ± 1.41 Å, the lowest against AR, TargetDiff, and PocketXMol, with a mean Vina score of −6.23, QED of 0.67, and a high-affinity ratio of 0.79. On docking it reaches a 35.2% success rate at DockQ > 0.23 against DeepTernary's 30.8%, and is also compared with EquiDock and DiffDock-PP. Interface prediction outperforms PeSTo and ScanNet. Synthesizability trails PocketXMol, which the authors attribute to that model's larger training set.

#Applications

TriGlue is aimed at targeted protein degradation programs, where teams must choose which E3 ligase to pair with a given target and what chemistry could bridge them. From two monomer structures it returns both a candidate glue and the ternary geometry that glue would induce, giving medicinal chemists a structural hypothesis rather than an unanchored molecule, and a basis for triage when no plausible interface or valid bridging molecule emerges. Where SynPROTAC addresses bifunctional PROTAC degraders, TriGlue covers the monovalent glue side.

#Impact

Computational glue design has lagged behind small-molecule and biologic design because the ternary formulation resists the standard pocket-conditioned recipe. TriGlue's contribution is to show that a learned interface prior, placed between the unbound proteins and the generator, makes the joint problem tractable, improving on ternary docking and pocket-conditioned generation baselines alike. The work is a preprint awaiting peer review, and its evaluation is entirely computational with no experimental confirmation that generated glues induce degradation. Code is cited only through an anonymized double-blind review repository, and no pretrained weights, license, model card, or data card have been released.