PaperlayerView exact revision on arXiv

Source-linked AI summary

TriGlue: a Biology-Inspired Generative Model for Generating Molecular Glue-Induced Ternary Complex

Yuliang Yan, Shuo Yan, Haochun Tang, Yiqin Sun, Enyan Dai

arXiv:2607.22143v2cs.LGcs.AI

TL;DR

Molecular glue design must coordinate ligand generation, protein–protein docking, and ternary complex assembly despite an unknown interface. TriGlue estimates that interface, then jointly generates the glue and complex pose, producing chemically valid molecules and plausible ternary complexes.

  • Problem

    Molecular glue design lacks computational methods that jointly address the unknown ligand-mediated interface, small-molecule generation, and protein docking required for ternary complex formation.

  • Method

    TriGlue estimates a geometrically constrained protein–protein interface, then uses interface-conditioned flow matching to jointly generate the molecular glue and rigid-body transformation.

  • Results

    Extensive experiments show that TriGlue generates chemically valid molecules and produces plausible ternary complexes, with the lowest average ligand-placement RMSD of 5.22 among tested baselines.

  • Takeaways & Limitations

    TriGlue demonstrates the potential of biology-inspired generative modeling for molecular glue discovery by integrating interface estimation with ternary complex generation.

Abstract

from arXiv · show

Molecular glue degraders have emerged as a promising strategy for targeted protein degradation by inducing ternary complex formation between an E3 ubiquitin ligase and a target protein. Despite their therapeutic potential, computational design of molecular glues remains largely unexplored. Unlike conventional structure-based drug design, molecular glue design is governed by the unknown protein-protein interface and requires the simultaneous modeling of ligand generation, protein-protein docking, and ternary complex assembly. In this work, we formulate molecular glue design as a ternary complex generation problem and propose a biology-inspired generative framework, TriGlue. Motivated by the mechanism of molecular glue action, we decompose ternary complex generation into two coupled stages: interface estimation and interface-conditioned complex generation. First, we develop an SE(3)-equivariant interface estimation module that predicts a geometrically constrained protein-protein interface from unbound monomer structures. Second, we introduce an interface-conditioned ternary flow matching network that jointly generates the molecular glue and predicts the rigid-body transformation required to assemble the ternary complex. Extensive experiments demonstrate that TriGlue generates chemically valid molecules and produces plausible ternary complexes, which highlight the potential of biology-inspired generative modeling for accelerating molecular glue discovery. Our code is available at https://github.com/yuliangyan0807/molecular-glue-design.

1 Introduction

Molecular glue design aims to generate a small molecule that simultaneously induces an E3 ligase–target interface, a novel ligand, and a coherent ternary complex. TriGlue addresses this problem by coupling interface estimation with interface-conditioned ternary generation, producing valid molecules and plausible complexes.

  • Motivation and challenges: Only a few hundred molecular glues have been identified, while traditional discovery relies on expensive, labor-intensive experimental screening.AI-based drug-design methods are not suitable because they typically model binary protein–ligand binding with a predefined pocket.
  • Motivation and challenges: Molecular glue design jointly optimizes the molecule-induced E3 ligase–target interface, the small molecule, and the proteins’ rigid-body transformation.These components must be coordinated to form a productive, coherent ternary complex.
  • TriGlue framework: An SE(3)-equivariant module predicts a geometrically constrained, biologically meaningful protein–protein interface from two monomer structures.The predicted interface provides a structural prior for subsequent ternary complex generation.
  • TriGlue framework: An interface-conditioned ternary flow matching network jointly generates the molecular glue and predicts the rigid-body rotation and translation needed for assembly.The module directly couples ligand generation with ternary complex positioning.
  • TriGlue framework: TriGlue decomposes ternary complex generation into interface estimation and interface-conditioned ternary flow generation.This biology-inspired formulation organizes the coupled design problem into two stages.
  • Results: Extensive experiments demonstrate that TriGlue generates valid small molecules and plausible ternary complexes.These results support the framework’s effectiveness for the proposed molecular glue design problem.

2 Preliminaries

Molecular glue design is framed as dynamic ternary-complex formation in which a ligand-mediated protein–protein interface drives cooperative assembly. Formally, the task is posed as generative modeling over receptor, target, and ligand structures represented through residue-level protein graphs.

  • Mechanistic Overview: Molecular glues reshape one protein’s surface to recruit a second partner, converting transient protein encounters into stable ternary assemblies through cooperative interactions.They can modulate protein function, localization, or stability.
  • Molecular Glue-induced Protein-protein Interface: The ligand-mediated protein–protein interface bridges complementary surface patches from both proteins and serves as a biological prior for ternary-complex formation.Occupying the interfacial region reduces the complex’s tendency to dissociate.
  • Dynamic Ternary Complex Assembly: Ternary assembly couples ligand placement, protein docking, and cooperative stabilization while unbound proteins undergo rigid-body rearrangements.This dynamic process motivates jointly modeling ligand structure and protein motion.
  • Formal Problem Definition: The design problem is formulated generatively as C = (R, M, T), with receptor R, target T, and small molecule M represented using residue-level protein graphs.Graph nodes correspond to residue identities and Cα coordinates, with edges constructed by k-nearest neighbors.

3 Method

TriGlue decomposes ternary complex generation into interface estimation and interface-conditioned complex generation, mirroring molecular glue-induced assembly. It estimates a geometry-aware virtual interface, then jointly generates the molecular glue and protein docking transformation using multimodality flow matching.

  • Framework: TriGlue factorizes ternary complex generation into two coupled modules: interface estimation and interface-conditioned complex generation.The decomposition matches molecular-glue action and converts joint modeling into two tractable sub-tasks.
  • Interface Estimation: The interface estimator infers a Global Virtual Interface from two monomer structures and aligns it with an ellipsoid-parameterized physical interface.EGNN encodes invariant features and SE(3)-equivariant coordinates; cross-protein attention aggregates residues into K virtual interface points.
  • Interface Estimation: The ellipsoid-parameterized interface captures the interaction region’s three-dimensional position and directional spread for downstream generative conditioning.Alignment minimizes discrepancies between the virtual interface’s first- and second-order moments and those of the ground-truth interface.
  • Interface-Conditioned Complex Generation: The interface-conditioned generator jointly models ligand conformation, ligand atom types, and rigid-body docking through three specialized flows.Rigid Transformation Flow predicts protein rotation and translation, Coordinate Flow generates ligand atom coordinates, and Atom Type Flow predicts chemical identities.
  • Interface-Conditioned Complex Generation: Rigid-body docking is initialized with a Differentiable Kabsch transformation from virtual interfaces and refined iteratively with residual Lie-algebra and translation updates.The refinement conditions on the current transformation, noisy input, and both virtual interfaces.

4 Experiment

TriGlue is evaluated on leakage-controlled TernaryDB splits for molecular generation, ternary-complex reconstruction, interface estimation, and interface-conditioning ablations. It achieves accurate ligand placement, outperforms interface-estimation baselines, and shows that interface modeling—especially correction—is important for docking performance.

  • Dataset: TernaryDB provides 22,303 curated ternary complexes, with MMseqs2-based sequence clustering and exclusion of clusters overlapping known molecular glue complexes to prevent leakage.The resulting validation and test splits are non-overlapping.
  • Evaluation Metrics: TriGlue is evaluated using affinity, molecular-property, ligand-pose, and protein-docking metrics, including Vina, Aff, QED, SA, RMSD, DockQ, and DockQ success rate.DockQ success is defined as DockQ > 0.23.
  • RQ1: Molecular Generation and Docking: TriGlue achieves the lowest average ligand-placement RMSD of 5.22 despite lacking binding-pocket information, while baselines receive pocket information.The comparison includes AR, TargetDiff, and PocketXMol for molecular generation, and DeepTernary, EquiDock, and DiffDock-PP for docking.
  • RQ2: Interface Estimation: TriGlue consistently outperforms PeSTo and ScanNet across JSD, Wasserstein-2 distance, Center-L2, and LogDet error for ellipsoid interface estimation on the held-out test split.Residue-level interface predictions from the methods are converted into ellipsoids before evaluation.
  • RQ3: Interface Conditioning: Removing either the ellipsoid prior or interface correction weakens unbound-target docking, while removing correction lowers average DockQ from 0.22 to 0.15.The ablation compares TriGlue with TriGlue w/o elli and TriGlue w/o corr using RMSD win rate and DockQ.

5 Related Work

This section situates TriGlue within advances in flow matching and deep learning for protein design. Flow matching extends from Euclidean generative modeling to structured non-Euclidean spaces, while protein design has diversified following AlphaFold.

  • Flow Matching: Flow Matching learns a continuous-time vector field transporting samples from a simple prior to the data distribution.The passage describes strong performance in Euclidean domains, including image, video, and robotics applications.
  • Flow Matching: Euclidean flow matching has been generalized to non-Euclidean spaces such as Riemannian manifolds for structured-object modeling.
  • Deep Learning for Protein Design: AlphaFold’s highly accurate protein structure prediction catalyzed broad generative approaches for protein design.Examples include RFDiffusion for enzyme design, SE(3)-Diffusion for protein backbone generation, Boltzgen for binder design, and methods for peptide design.

6 Conclusion … B Ellipsoid Construction

The paper frames molecular glue discovery as a generative modeling problem requiring coordinated ligand generation, protein–protein docking, and ternary complex assembly under unknown interface conditions. It addresses this challenge with TriGlue, a biology-inspired framework that decomposes ternary complex generation into interface estimation and interface-conditioned generation.

  • 6 Conclusion: TriGlue investigates molecular glue design from a generative modeling perspective.
  • 6 Conclusion: Molecular glue discovery requires coordinated modeling of ligand generation, protein–protein docking, and ternary complex assembly.
  • 6 Conclusion: This challenge arises under unknown protein–protein interface conditions.
  • 6 Conclusion: The proposed framework is named TriGlue.
  • 6 Conclusion: TriGlue is biology-inspired.
  • 6 Conclusion: TriGlue decomposes ternary complex generation into interface estimation and interface-conditioned generation.

C Architecture 17 … E Sample Algorithm

TriGlue’s implementation combines interface-aware geometric encoding, conditional flow matching, and iterative sampling to generate molecular glues and assemble ternary complexes. The framework represents interfaces compactly, predicts rigid transformations and ligand structures, and refines noisy states into final complexes.

  • A Terminology: Molecular glues modulate protein–protein interactions by enhancing affinity, either inducing novel interactions or stabilizing existing ones.They are classified as type I or type II according to whether they stabilize non-native or native interactions.
  • B Ellipsoid Construction: TriGlue represents an interface residue set with a Gaussian ellipsoid N(µ, Σ), whose centroid captures location and covariance eigenstructure captures spatial extent and principal orientation.The ellipsoid provides the compact geometric representation used for interface alignment and evaluation.
  • C Architecture: The latent ternary encoder uses Invariant Point Attention to jointly process scalar features, pairwise relations, and 3D coordinates with SE(3)-invariant predictions.Protein residue features combine amino-acid embeddings, local-frame heavy-atom coordinates, and backbone dihedrals, while ligand atoms use learned atom-type embeddings.
  • C Architecture: The architecture separately predicts rigid-body updates, ligand coordinates, and atom types conditioned on protein context and interface geometry.The rigid transformation block predicts residual rotation and translation updates, the coordinate block performs ligand–ligand and ligand–protein message passing, and the atom-type block samples from a predicted distribution.
  • D Flow Matching: TriGlue uses three conditional flows: Euclidean CFM for ligand coordinates, simplex CFM for atom types, and Riemannian CFM for rigid transformations.Rotations follow geodesics on SO(3), while translations use Euclidean flow.
  • D Flow Matching: The model is trained with reparameterized CFM objectives that directly predict target data and compute the conditional vector field implicitly from reconstructed targets.The paper explicitly states that these reparameterized objectives train the model.
  • E Sample Algorithm: Sampling begins by estimating virtual interfaces and applying differentiable Kabsch alignment to initialize an interface-consistent target-protein transformation.The corrected target coordinates then condition subsequent ternary-complex generation.
  • E Sample Algorithm: Starting from noisy rotations, translations, ligand coordinates, and atom types, the ternary flow network iteratively denoises these states before decoding the final ternary complex.Rotations use SO(3) geodesic updates, whereas translations, ligand coordinates, and atom types are refined with Euler steps over N iterations.

F Proofs · G Experimental Details

The proof establishes that TriGlue’s virtual interface points and ellipsoid parameters are SE(3)-equivariant. This follows from invariant attention weights and equivariant coordinate transformations under rigid-body transformations.

  • F Proofs: The ellipsoid parameters (ˆµ, ˆΣ) defined in Sec. 3.2 are also SE(3)-equivariant.The proof treats these parameters together with the virtual interface points.
  • F Proofs: Under a rigid-body transformation (R, t) ∈SE(3), all residue coordinates transform as i = Rzi + t.This transformation provides the basis for the equivariance proof.
  • F Proofs: Attention weights remain unchanged under SE(3) because they are computed from EGNN invariant features.This invariance allows the virtual interface points to transform consistently with the coordinates.
  • F Proofs: The virtual interface points yk are therefore SE(3)-equivariant.The proof derives this conclusion from the unchanged attention weights and transformed residue coordinates.
  • F Proofs: Centered interface coordinates satisfy k −ˆµ′ = R(yk −ˆµ), yielding equivariant transformation of (yk, ˆµ, ˆΣ).The centered-coordinate relation completes the argument for the ellipsoid parameters.

G.1 Dataset Details · G.2 Baseline Details · G.3 Training Details

The experiments use a TernaryDB1-derived molecular-glue ternary-complex split and evaluate generation, ternary structure reconstruction, and protein-protein docking against task-specific baselines. Training uses a two-stage framework with specified optimization, flow-matching, data-splitting, and compute settings.

  • G.1 Dataset Details: Each dataset sample contains a target protein, an E3 ligase, and a small-molecule ligand extracted from an experimentally resolved ternary complex.Protein lengths use resolved Cα residues from curated PDB files, while ligand descriptors are computed from per-complex SDF files with RDKit; QED denotes quantitative estimate of drug-like.
  • G.1 Dataset Details: The dataset is a molecular-glue-induced ternary-complex split derived from TernaryDB1.Table 4 summarizes training and held-out test splits using Median (Minimum, Maximum) numerical descriptors.
  • G.2 Baseline Details: Because no existing method evaluates both tasks jointly, baselines are selected separately for molecular design and ternary-complex structure modeling.The framework performs de novo molecular glue generation and reconstructs the unbound target protein into the ternary-complex coordinate frame.
  • G.2 Baseline Details: AR2, TargetDiff, and PocketXMol assess de novo small-molecule generation conditioned on protein binding pockets.These three methods represent structure-based drug design baselines for evaluating generated molecular glues.
  • G.2 Baseline Details: DeepTernary docks the unbound protein into the ternary-complex frame, while EquiDock and DiffDock-PP evaluate rigid-body protein-protein docking from unbound structures.DeepTernary is described as the first model designed for ternary complex structure prediction.
  • G.3 Training Details: The framework has interface estimation and interface-conditioned ternary flow generation stages, optimized with Adam at an initial learning rate of 5 × 10^-4.The interface model trains for 1200 epochs with batch size 8 on 8 GPUs, whereas the ternary flow model trains for 800 epochs with batch size 2 and gradient accumulation of 32.
  • G.3 Training Details: Flow matching uses Euclidean translation flows with σ = 1.0, Riemannian rotation flows on SO(3), and simplex atom types with K = 5.0 across 25 categories.Atom identities combine elemental type, hybridization, and aromaticity; sampling uses 50 iterative denoising steps.

G.4 Evaluation Metric Details · G.5 Ellipsoid Interface Prediction Metric Details

The evaluation measures molecular generation, ternary-complex reconstruction, and ellipsoid interface prediction using binding, physicochemical, structural, docking, and distributional metrics. These metrics specify how ligand quality, complex accuracy, and predicted interface geometry are quantified and interpreted.

  • G.4 Evaluation Metric Details: Molecular generation is evaluated by binding affinity and physicochemical properties, while ternary-complex reconstruction assesses ligand pose and protein docking accuracy.Binding uses AutoDock Vina-based scores and Aff; physicochemical evaluation includes QED and SA, while reconstruction uses RMSD and DockQ.
  • G.4 Evaluation Metric Details: Ternary-complex reconstruction uses ligand-coordinate RMSD and protein-protein DockQ, additionally reporting the success rate for DockQ greater than 0.23.Lower RMSD indicates closer agreement with the reference ligand pose, while DockQ summarizes protein docking quality.
  • G.4 Evaluation Metric Details: AutoDock Vina evaluation reports Vina score, Vina min after local energy minimization, and Vina dock after full re-docking; lower values indicate stronger predicted binding.Vina score evaluates the generated pose directly, Vina min reduces sensitivity to local geometric errors, and Vina dock estimates the best score achievable under docking.
  • G.4 Evaluation Metric Details: QED summarizes drug-likeness from molecular descriptors and ranges from 0 to 1, with larger values indicating more favourable drug-like properties.Descriptors include molecular weight, lipophilicity, topological polar surface area, hydrogen bonding, aromatic rings, rotatable bonds, and structural alerts.
  • G.4 Evaluation Metric Details: Synthetic accessibility combines fragment contributions, complexity penalties, and corrections; under the reported normalized convention, larger SA values indicate easier synthesis.Complexity penalties reflect features such as large rings, complex ring systems, stereochemical complexity, and large molecular size.
  • G.4 Evaluation Metric Details: Aff measures the proportion of generated molecules whose Vina dock score is stronger than the reference ligand for the same target, with higher values preferred.Because lower Vina scores denote stronger binding, Aff compares each generated molecule’s score against the corresponding reference score.
  • G.5 Ellipsoid Interface Prediction Metric Details: Ellipsoid interface prediction represents each interface as a Gaussian N(µ, Σ) and evaluates center, scale, joint geometry, and distributional discrepancies.Metrics include Center-L2, LogDet Error, Wasserstein-2 distance, and Jensen–Shannon divergence; JSD is estimated by Monte Carlo because the mixture is not a single Gaussian.

H Group Analysis · I Broader Impacts

TriGlue’s generated ternary complexes are evaluated with AlphaFold3 and receive structurally reasonable confidence for most groups. The framework may broaden molecular-glue design and reduce early drug-discovery burden by enabling computational ternary-complex exploration.

  • H Group Analysis: TriGlue samples molecular glues for each test-set group to evaluate generated ternary-complex structural quality.The evaluation combines sampled glues with the sequences of two proteins before AlphaFold3 prediction.
  • H Group Analysis: AlphaFold3 predicts complex structures from each sampled molecular glue and the two protein sequences.The resulting predictions are used to assess ternary-complex structure quality.
  • H Group Analysis: The evaluation computes pTM and ipTM confidence scores for the predicted ternary complexes.These confidence measures provide a structure-prediction perspective on generated complexes.
  • H Group Analysis: Structurally reasonable ternary complexes are predicted for most groups.This result provides additional evidence for the quality of TriGlue’s samples.
  • I Broader Impacts: Molecular glue design could expand the druggable proteome and accelerate therapies for diseases difficult to treat with conventional small molecules.The broader-impact claim follows from the modality’s therapeutic potential described in the paper.
  • I Broader Impacts: Computational exploration of ternary complex formation may reduce experimental trial-and-error, shorten early-stage discovery timelines, and lower research costs.These effects could ultimately improve access to therapies.
Loading 2607.22143v2…