acceptodds
Under review as a conference paper at ICLR 2027

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

13 views

Abstract

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://anonymous.4open.science/r/molecular-glue-design-806B.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Glue, Ternary Complex Design, Targeted Protein Degradation, Geometric Deep Learning, SE(3)-Equivariant

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_P5e4OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary
  • Research problem and motivation. This work investigates de novo molecular glue design, a challenging problem in computational drug discovery that requires jointly generating a small molecule and determining the relative arrangement of two proteins to form a plausible ternary complex. Unlike conventional structure-based drug design, the protein–protein interface is unknown beforehand, making molecular generation and protein docking inherently interdependent.

  • Methodological innovation. The authors propose TriGlue, a biology-inspired generative framework that decomposes ternary complex generation into two stages. First, an SE(3)-equivariant interface estimation module predicts the potential protein–protein interface from unbound monomer structures. Then, an interface-conditioned ternary flow matching network jointly generates the molecular glue, predicts its atomic and bond attributes, and determines the target protein’s rigid-body transformation.

  • Experimental results. Experiments on TernaryDB demonstrate promising performance in molecular generation and ternary complex reconstruction. TriGlue achieves favorable molecular quality metrics compared with existing molecular generation baselines and improves protein docking performance over methods such as DeepTernary, EquiDock, and DiffDock-PP. The authors also demonstrate the benefits of explicit interface conditioning and few-shot adaptation across different molecular glue families.

Critical strengths
  • Molecular glue design is challenging because it requires identifying a suitable protein–protein interface while simultaneously designing a small molecule that facilitates ternary complex formation. By formulating this task as joint molecular generation and ternary complex reconstruction, the work addresses a meaningful gap in conventional structure-based drug design.

  • TriGlue combines SE(3)-equivariant interface estimation with interface-conditioned flow matching to jointly generate molecular glues and predict protein docking poses, with experiments demonstrating improvements over selected baselines.

Critical weaknesses
  • The lack of comparisons with all-atom approaches for docking and ternary complex generation. TriGlue uses a residue-level interface representation to guide protein docking and molecular glue generation, but does not compare its interface estimation module with existing interface prediction methods such as CPORT, or its docking strategy with all-atom approaches such as HADDOCK and RosettaDock. Furthermore, comparisons with all-atom complex prediction models, such as AlphaFold 3 and Boltz-1, could help assess whether explicitly modeling atomic-level interactions improves ternary complex assembly and clarify the advantages of TriGlue's approach.
  • Limited exploration of alternative interface hypotheses. TriGlue relies on a single predicted protein–protein interface to condition ternary complex generation, potentially limiting its ability to capture alternative binding modes. Generating and evaluating multiple plausible interface hypotheses could improve the diversity and success rate of ternary complex design, particularly when the interface is uncertain.
Questions for authors
  • Could the authors explore generating multiple plausible protein–protein interface hypotheses rather than relying on a single predicted interface? This could allow TriGlue to capture alternative binding modes and improve ternary complex generation, particularly when the interface is uncertain.
  • Regarding all-atom modeling: Could the authors compare TriGlue with at least one interface prediction and docking methods, such as CPORT, HADDOCK, and RosettaDock, as well as all-atom complex prediction models such as AlphaFold 3 and Boltz-1? Such comparisons could clarify the advantages of the residue-level interface representation and assess whether explicitly modeling atomic-level interactions improves ternary complex generation.
2 comments
AuthorsOpenReview ↗

Response to Reviewer P5e4

We thank the reviewer for recognizing the value of our work and for suggesting these additional comparisons.

W1 & Q2: “Please compare TriGlue with representative interface-prediction, docking, and all-atom complex-prediction methods to clarify the advantages of its residue-level interface representation.”

A: We added dMaSIF [1] with all-atom inputs and AlphaFold3 [2] as complementary benchmarks for interface estimation and ternary-complex prediction, respectively.

(1) Interface estimation. We compared TriGlue with dMaSIF using the same interface-evaluation metrics. As shown in the following Table, TriGlue achieves lower errors across all four metrics, supporting the effectiveness of its interface representation relative to these alternatives, including the baseline using all-atom inputs.

MethodJSD ↓W2 ↓Center-L2 ↓LogDet ↓
dMaSIF (All-Atom)0.69 ± 0.0158.21 ± 40.1257.20 ± 40.289.85 ± 7.62
TriGlue0.47 ± 0.1512.65 ± 7.8910.99 ± 7.473.83 ± 5.01

(2) All-atom structure prediction. We evaluated AlphaFold3 on the test set using both protein sequences and reference ligand SMILES, we measured ligand localization and geometric agreement with the reference. We observe that TriGlue yields lower ligand localization and geometric matching errors, with Hungarian-RMSD of 1.96 Å versus 6.18 Å for AlphaFold3. These comparisons complement the existing docking baselines and directly assess interface estimation and ligand placement within the assembled complex.

MethodCenter-L2 (Å) ↓Center < 2 Å ↑Hungarian-RMSD (Å) ↓
AlphaFold3 (Supplied-Ligand)5.84 ± 23.3085.2%6.18 ± 23.25
TriGlue0.12 ± 0.05100.0%1.96 ± 0.41

W2 & Q1: “Reliance on a single predicted interface may limit exploration of alternative binding modes. Could the authors generate and evaluate multiple plausible interface hypotheses, particularly when the interface is uncertain?”

A: We thank the reviewer for this valuable suggestion. In future versions, we plan to sample multiple spatially distinct interface hypotheses, generate ternary complexes conditioned on each, and rank the candidates by interface compatibility and steric clashes. This would broaden exploration of alternative binding modes when the interface is uncertain.

The new comparisons with all-atom dMaSIF (Appendices H.2, H.5, and I.6) and AlphaFold3 (Appendix I.4) have been added in blue to the revised Appendix. We hope these additions address your points.

[1] Fast End-to-End Learning on Protein Surfaces. CVPR, 2021.

[2] Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 2024.

With the discussion deadline approaching, we would like to kindly check whether our rebuttal and additional evaluations (including the all-atom dMaSIF interface comparison, AlphaFold3 complex benchmark, and multi-interface discussion) satisfactorily resolve the points in your review. Please let us know if any further feedback or clarification would be helpful. An overview of all current rebuttal updates is also available in our General Response at the top of the thread.

This year's ICLR process has been particularly challenging for reviewers and authors. We are truly grateful for your time, constructive feedback, and dedicated service to the community.

Reviewer_HBVtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TriGlue, a generative framework for jointly designing a molecular glue and assembling a ternary complex from two input protein structures. The method first predicts a pair of protein interfaces using attention-weighted virtual points, supervised through their means and covariances. An interface-conditioned flow network then jointly models the target protein's rigid-body pose, ligand coordinates, and ligand atom and bond types through a shared ternary EGNN. Predicted interface geometry and a Kabsch-derived pose prior condition the rigid transformation flow. The paper also introduces TriConf, an interface-stability score associated with docking quality.

The evaluation uses TernaryDB and held-out molecular-glue complexes, comparing molecular generation, protein docking, and interface estimation separately against relevant baselines. TriGlue reports strong molecular-generation results and improved docking performance, including an acceptable-docking rate of 47.347.3\\% versus 31.531.5\\% for DeepTernary. Additional experiments investigate interface conditioning, few-shot adaptation, and robustness to predicted receptor structures.

Critical strengths
  1. The framework provides a coherent approach to jointly modeling ligand generation and protein assembly. The shared ternary representation allows ligand geometry and chemistry to inform protein docking while the evolving protein arrangement conditions ligand generation. The interface estimator provides a compact geometric condition without requiring a predefined pocket. This is a useful contribution beyond evaluating molecule generation or protein docking in isolation.

  2. The reported experiments provide encouraging evidence across the framework's main components. TriGlue achieves the best or tied-best results in most columns of Table 1 and improves the reported docking success rate over the evaluated baselines. Table 2 also supports the quality of its interface estimates. The matched-noise conditioning interventions show that the trained model makes meaningful use of interface information, particularly the virtual-interface geometry. Together, these results support the practical promise of the proposed framework, subject to the protocol clarifications below.

Critical weaknesses
  1. The unspecified inference-time ligand size prevents a confident assessment of the claimed generation setting, and ligand RMSD requires a separate clarification. Problem 1 and Algorithm 1 describe generation conditioned on two proteins, but the paper does not explain how the number of ligand atoms is determined at inference. This is consequential: if the reference ligand's atom count is supplied, the experiments evaluate generation with an additional size condition rather than the stated protein-only setting. Separately, Equation (90) assumes corresponding atoms between generated and reference ligands. De novo generation can change atom identities and molecular connectivity, and equal atom counts alone do not establish a chemically meaningful correspondence. Without an explicit matching and alignment protocol, it is difficult to interpret the ligand RMSD comparison in Table 1. These unresolved protocol questions are the main reason for my current recommendation; they do not demonstrate a fundamental flaw in the architecture.

  2. The baseline protocols and conditioning ablations do not fully establish the source of the reported advantages. Appendix H.2 identifies the baseline methods but does not sufficiently specify the pocket construction and protein inputs used for molecular generation. Appendix H.4 denotes the Vina protein environment by PP without clearly identifying whether it contains one or both proteins and whether their arrangement is predicted or taken from the reference complex. The conversion of PeSTo and ScanNet predictions into partner-specific interface ellipsoids also needs clarification. In addition, Appendix I.1 states that all conditioning ablations use the same full-model checkpoint and mask inputs during inference. This is informative about the trained model's reliance on those inputs, but it does not fully isolate their benefit relative to models trained without them, because masking may introduce a distribution shift. The Kabsch-prior intervention alone is not statistically significant (p=0.172p=0.172), so its independent benefit should be described cautiously.

Questions for authors
  1. Ligand atom count — clarification and requested evaluation: How is the ligand atom count determined at inference in Algorithm 1? Does any evaluation use the reference ligand's atom count? If not, please specify the size-selection procedure and confirm that the reported results use it. If so, please explicitly state this additional conditioning information and report an evaluation in which ligand size is determined without access to the reference ligand, applying a comparable size-conditioning protocol to the molecular-generation baselines where appropriate. This clarification and, if needed, the additional results would materially affect my recommendation.

  2. Ligand RMSD — clarification: How are atoms matched between generated and reference ligands in Equation (90), particularly when their element identities, connectivity, or atom counts differ? Please specify the coordinate-alignment procedure, treatment of molecular symmetries, and handling of unmatched or excluded samples. Does the reported quantity measure chemically matched ligand pose recovery, or a different form of coordinate correspondence?

  3. Baseline inputs and evaluation — clarification: For AR, TargetDiff, and PocketXMol, how are pockets defined, and do the inputs contain one protein or both proteins? What exactly is the protein environment PP used for Vina evaluation, and is it constructed from reference or predicted protein poses? Please also specify the candidate counts and selection rules for the compared methods. For Table 2, how are PeSTo and ScanNet residue predictions converted into ellipsoids corresponding to the particular protein partner being evaluated? These details are needed to interpret the comparisons; the concern is not a presumption that the comparisons are unfair.

  4. Interface-conditioning ablations — requested control or narrower interpretation: Can the authors provide at least one model retrained without interface conditioning, using otherwise matched training and evaluation settings? This would help distinguish the benefit of the conditioning design from sensitivity to removing inputs from a model trained to use them. If this experiment is unavailable, please distinguish the current inference-time masking study from a retrained architectural ablation and qualify the claim about the independent contribution of the Kabsch pose prior given p=0.172p=0.172.

7 comments
AuthorsOpenReview ↗

Response to Reviewer HBVt [1/3]

Thank you for your careful review, which has greatly helped improve the quality of our paper. Below we address your questions.

W1 & Q1 & Q2: “Please clarify how ligand atom count is determined at inference and, if the reference count is used, provide an evaluation with reference-independent size selection. Specify the atom-matching and coordinate-alignment protocols for ligand RMSD, including molecular symmetries and unmatched samples, and clarify whether it measures chemically meaningful pose recovery.”

A: Thank you for highlighting these points. Below, we address these one by one.

(1) Ligand atom count (Q1). Table 1 follows the exact same atom-count protocol (NM=NrefN_{\mathcal{M}} = N_{\mathrm{ref}} per test complex) used by non-autoregressive 3D baselines (TargetDiff with sample_num_atoms=ref [1] and PocketXMol [2]) to ensure a strictly fair comparison. Importantly, TriGlue requires no retraining to accept arbitrary NMN_{\mathcal{M}} at inference, sampling NM∼ptrain(NM)N_{\mathcal{M}} \sim p_{\mathrm{train}}(N_{\mathcal{M}}) when no reference ligand exists:

MethodAtom-Count Rule (n = 92)Per-Target Count Examples (Range: 10–69)Protein InputGenerated from Noise
TargetDiff [1] (ref)N_M = N_ref5FQD (CRBN): 19; 6SJ7 (GSPT1): 24; 2O98 (14-3-3): 48; 1FAP (FKBP–FRB): 65Bound holo pocketAtom types, 3D coords (post-hoc bonds)
PocketXMol [2] (ref)N_M = N_refIdentical per-target N_ref across all 92 test complexes (5FQD: 19; 6SJ7: 24; 2O98: 48; 1FAP: 65)Bound holo pocketAtom types, bonds, 3D coords
TriGlue (Ours)N_M = N_refIdentical per-target N_ref across all 92 test complexes (5FQD: 19; 6SJ7: 24; 2O98: 48; 1FAP: 65)Unassembled P1, P2Atom types, bonds, 3D coords + Target pose (R, t)

(2) Ligand RMSD (Q2). Equation (90) (RMSD) pairs generated and reference atoms by index (NM=NrefN_{\mathcal{M}} = N_{\mathrm{ref}}), providing an upper bound on 3D pose error within the complex. To further address your concerns, we additionally provide Hungarian-RMSD and Center-L2 in the same unaligned P1P_1 frame to measure permutation-invariant pose agreement and size-independent binding-site localization as follows:

Metric (Å)↓TriGlue
RMSD4.95 ± 1.15
Hungarian-RMSD1.97 ± 0.41
Center-L20.13 ± 0.05

(3) Chemical geometry (Q2). To address the concern that RMSD does not account for differing element identities or connectivity in molecules, we additionally evaluate Bond-Length JSD, the Jensen–Shannon divergence between generated and reference crystal bond-length distributions across covalent bond types. The results indicate that TriGlue outperforms TargetDiff on four of six bond types and matches the remaining two, confirming accurate 3D covalent geometry alongside binding-site placement:

MethodC=CC:CC=NC:NC–OC=O
TargetDiff0.37630.29460.77610.38700.36610.4842
TriGlue0.32350.29620.62670.37250.36710.3089

[1] 3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction, ICLR, 2023.

[2] Unified modeling of 3D molecular generation via atomic interactions with PocketXMol, Cell, 2026.

AuthorsOpenReview ↗

Response to Reviewer HBVt [2/3]

W2 & Q3: “Clarify baseline pocket construction and protein inputs, the composition of the Vina protein environment and whether its poses are reference or predicted, and how PeSTo/ScanNet predictions are converted into partner-specific interface ellipsoids.”

A: Thank you for requesting these clarifications regarding our baseline and evaluation protocols. All compared methods strictly follow their official implementations, and in fact operate under more favorable conditions than TriGlue because they receive known pocket or ligand inputs. Below, we clarify these details point by point.

(1) Baseline inputs and pocket construction. Baseline methods (including AR with a 23 Å cube, TargetDiff, and PocketXMol) extract single-protein pockets centered on the ground-truth reference ligand (oracle binding site), whereas TriGlue receives no ground-truth pocket coordinates and localizes the binding site de novo via its self-predicted interface. All metrics are aggregated over all valid generated molecules under each method's official sampling setup without post-hoc selection:

MethodPocket Location & InputCandidatesRMSD (Å) ↓Vina Score ↓QED ↑
TargetDiffGround-truth pocket (P1, 10 Å COM cutoff)506.27 ± 1.99-2.110.47
PocketXMolGround-truth pocket (P1, 15 Å residue cutoff)1006.74 ± 0.91-6.170.47
TriGlue (Ours)Self-predicted pocket (unassembled P1, P2; no GT pocket)1004.95 ± 1.15-6.570.72

(2) Vina protein environment. All methods are evaluated against the same experimental receptor P1P_1, using identical receptor preparation and Vina scoring settings. The second protein P2P_2 is excluded for every method, including TriGlue. This ensures consistent scoring conditions across methods.

(3) PeSTo and ScanNet. To ensure a strictly fair comparison, all methods receive the exact same unassembled monomers (P1,P2P_1, P_2), and predicted interface residues from PeSTo and ScanNet (official weights, standard probability ≥0.5\ge 0.5 or top-1 fallback) are mapped to Cα\text{C}\alpha ellipsoids via the exact same mean-and-covariance formula used for TriGlue and ground-truth interfaces (Section 3.2, Eq. (8); Appendix C, Eqs. (21)–(22)). This is fair because evaluating each protein in its own monomer frame decouples site prediction from rigid-body docking (so single-protein baselines are never penalized for lacking pose prediction) while comparing all methods in an identical geometric metric space (Appendix H.5).

(4) Supplied-ligand AlphaFold3 baseline. For a strictly fair comparison, we supply AlphaFold3 with both protein sequences and the ground-truth ligand SMILES, and Kabsch-align its predicted P1P_1 onto TriGlue's exact receptor frame [3] (Appendix I.4; consistent across alternative alignment frames). Even de novo without reference SMILES, TriGlue substantially outperforms this supplied-ligand baseline across all evaluated metrics, raising localization within 2 Å from 85.2% to 100.0% (Center-L2: 0.12 Å vs. 5.84 Å) and lowering mean Hungarian-RMSD from 6.18 Å to 1.96 Å.

MethodCenter-L2 (Å) ↓Center < 2 Å ↑Hungarian-RMSD (Å) ↓
AlphaFold3 (Supplied-Ligand)5.84 ± 23.3085.2%6.18 ± 23.25
TriGlue (De Novo)0.12 ± 0.05100.0%1.96 ± 0.41

[3] Accurate structure prediction of biomolecular interactions with AlphaFold 3, Nature, 2024.

AuthorsOpenReview ↗

Response to Reviewer HBVt [3/3]

W2 & Q4: “Inference-time masking may induce distribution shift and does not isolate the benefit of interface conditioning. Please provide a model retrained without conditioning under matched settings.”

A: To ensure a strictly fair architectural ablation free of inference-time masking shift, we trained an end-to-end interface-free counterpart (TriGlue w/o Interface) from scratch (Appendix I.1). Specifically, we remove only Stage 1 (the Interface Estimation Module) and all interface-derived conditions, including both the interface points (IP) and the Kabsch pose prior (R,t)(\mathbf{R},\mathbf{t}), from ΦRT\Phi_{\mathrm{RT}}. Meanwhile, we keep the 4-layer unified ternary EGNN trunk, cross-entity message passing among P1P_1, P2P_2, and the ligand, the local force-torque head, flow losses Lflow\mathcal{L}_{\mathrm{flow}}, optimizer, and 200-epoch budget, allowing the flow network to learn (R,t)(\mathbf{R},\mathbf{t}) and ligand generation directly from the three components.

Combining this retrained ablation with our component-level ablations (Fig. 4b, Appendix I.1) yields two clear conclusions: (i) while retraining recovers the distribution-shift drop of naive inference-time masking (DockQ 0.097 to 0.211), TriGlue still consistently outperforms the retrained interface-free model in Best DockQ (0.231 vs. 0.211), trajectory-wide Mean DockQ (0.100 vs. 0.077, +29.9%, Wilcoxon p=1.3×10−4p=1.3\times 10^{-4}), and Ligand Center-L2 (0.113 Å vs. 0.142 Å); and (ii) fine-grained ablations confirm that explicit Interface Points (w/o IP: 0.169) provide the primary geometric guidance alongside the auxiliary Kabsch pose prior (w/o Pose Prior: 0.204).

VariantAblation SettingBest DockQ ↑Mean DockQ ↑Ligand Center-L2 (Å) ↓
TriGlue w/o Pose PriorComponent ablation (mask Kabsch prior)0.204N/AN/A
TriGlue w/o IPComponent ablation (mask interface points)0.169N/AN/A
TriGlue w/o All InterfaceComponent ablation (mask both at inference)0.097N/AN/A
TriGlue w/o Interface (Retrained)Trained from scratch (200 ep, no Stage 1)0.211 ± 0.1770.077 ± 0.0610.142 ± 0.071
TriGlue (Ours)Matched full model (200 ep / Fig. 4b)0.231 ± 0.2080.100 ± 0.1020.113 ± 0.072

All corresponding updates appear in the revised draft, including NMN_{\mathcal{M}} in Algorithm 1 (Appendix F), the ligand pose and Bond-Length JSD analyses (Appendices H.4 and I.5), the baseline and evaluation protocols (Appendices H.2, H.4, H.5, and I.4), and the retrained ablation (Appendix I.1). We hope these clarifications and experiments fully answer your questions.

Reviewer_NBHZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TriGlue for jointly generating molecular glues and assembling the corresponding ternary complexes. Given two protein structures, it first predicts their interaction regions using an interface estimation module with an ellipsoidal representation. It then uses an interface-conditioned flow matching model to jointly generate the glue’s coordinates and chemical attributes, together with the rotation and translation of the target protein. Both proteins’ internal conformations remain fixed during generation.

Critical strengths
  • The joint problem formulation is useful. Generating ligand chemistry while assembling two proteins addresses a meaningful limitation of methods that assume either a predefined binding pocket or a known ligand. The proposed integration therefore goes beyond simply applying an existing docking model to another dataset.

  • The paper provides relevant geometric controls. The interface-conditioning ablations, family adaptation experiments, and receptor-only AlphaFold3 evaluation help establish that the interface representation contributes to the reported structural performance.

Critical weaknesses

The ternary-complex reconstruction baselines appear incomplete. Although no prior method may jointly perform de novo molecular-glue generation and complex assembly, several contemporary cofolding models—including AlphaFold3, Boltz-1, Chai-1, Protenix, and RoseTTAFold All-Atom—have been directly evaluated for molecular-glue ternary structure prediction (Liao et al., 2025). Representative cofolding baselines should therefore be included under a clearly defined supplied-ligand protocol, or their omission should be justified. DeepTernary is a relevant ternary baseline, but comparison with it and binary protein–protein docking methods leaves the strength of the reconstruction component relative to contemporary cofolding models unclear.

Although the paper motivates TriGlue through glue-induced surface remodeling and allosteric conformational rearrangement, the proposed model keeps both protein conformations fixed and models assembly solely through a rigid-body transformation of the target. Consequently, the model selects contact regions on fixed protein surfaces rather than modeling the protein conformational changes described in the motivation. The paper should clarify this distinction and discuss the resulting limitations, particularly because the dataset components are extracted from resolved ternary complexes and may already adopt bound-compatible conformations. The receptor-only AlphaFold3 experiment is useful, but retains the experimental target structure and does not address cases where assembly requires internal conformational changes.

Questions for authors

Why were recent cofolding models not included in the ternary-reconstruction evaluation?

Could the authors include a representative comparison under a clearly defined supplied-ligand setting, or explain why such a comparison would not be appropriate?

2 comments
AuthorsOpenReview ↗

Response to Reviewer NBHZ

Thanks for your thoughtful feedback. Below we address your concerns.

W1 & Q1 & Q2: “The ternary-complex reconstruction baselines appear incomplete.”

A: We thank the reviewer for this suggestion. Our initial baseline selection reflected two considerations:

(i) Cofolding models such as AlphaFold3 are trained on extensive PDB data, posing a risk of training/test overlap with our PDB-derived benchmark [1].

(ii) Our design task focuses on molecular-glue generation and pose prediction given target and receptor structures, whereas cofolding jointly predicts protein conformations and complex assembly. These considerations motivated our emphasis on docking and ternary-assembly baselines.

To further address your concern, we used AlphaFold3 to predict ternary structures on the test set, supplying both protein sequences and reference ligand SMILES. After receptor alignment, we evaluated ligand placement using centroid distance (Center-L2), localization success (Center < 2 Å), and Hungarian-RMSD.

Accurate molecular-glue positioning is critical for establishing the contacts that stabilize ternary complexes[2]. As shown in the Table below, AlphaFold3 leaves a subset of supplied ligands mislocalized, whereas TriGlue achieves 100.0% localization success, compared with 85.2% for AlphaFold3. TriGlue also yields a lower Hungarian-RMSD (1.96 Å versus 6.18 Å), indicating more accurate ligand placement and geometric agreement with the reference.

MethodCenter-L2 (Å)↓Center < 2 Å↑Hungarian-RMSD (Å)↓
AlphaFold3 (Supplied-Ligand)5.84 ± 23.3085.2%6.18 ± 23.25
TriGlue (De Novo)0.12 ± 0.05100.0%1.96 ± 0.41

W2: “Although motivated by glue-induced surface remodeling and allosteric conformational rearrangement, TriGlue keeps both protein conformations fixed and models assembly solely through a rigid-body transformation of the target.”

A: We appreciate the reviewer for raising this concern. We kindly remind that the submitted manuscript includes an experiment using an AlphaFold3-predicted apo receptor as input, while retaining the experimental target structure (We respectfully refer the Reviewer to Lines 410–415).

To further address this concern, we applied small, spatially correlated Gaussian perturbations at and around the interfaces of both proteins across the dataset. We then rebuilt the training dataset and retrained and evaluated TriGlue to assess robustness to approximate an unbound-input setting [3]. Under the same sampling budget, we compared the original and perturbation-trained models on bound and dual-perturbed inputs. Ligand RMSD was computed using Hungarian atom assignment to remove dependence on atom ordering.

ModelReceptor / Target InputDockQ ↑DockQ > 0.23Ligand RMSD (Å) ↓QED ↑
TriGlue (Original)Bound / Bound0.313 ± 0.24549.45%1.97 ± 0.410.566
TriGlue (Original, Zero-Shot)Perturbed / Perturbed0.305 ± 0.20654.95%1.98 ± 0.420.571
TriGlue (Perturbed-Trained)Perturbed / Perturbed0.328 ± 0.22156.04%1.92 ± 0.480.583

TriGlue maintains comparable performance when both protein inputs are perturbed, even without retraining, with similar DockQ and ligand RMSD. Retraining on perturbed inputs further improves all reported metrics, achieving a DockQ of 0.328 and a docking success rate of 56.04%. These results support robustness to local conformational deviations in both proteins.

We have incorporated the AlphaFold3 benchmark (Appendices H.2 and I.4) and the dual-protein interfacial perturbation experiments (Appendix I.3) in the revised manuscript, and we hope these results resolve your concerns.

[1] Benchmarking Cofolding Methods for Molecular Glue Ternary Structure Prediction. 2025.

[2] Structural basis of lenalidomide-induced CK1α degradation by the CRL4^CRBN ubiquitin ligase, Nature, 2016.

[3] Robust deep learning–based protein sequence design using ProteinMPNN. Science, 2022.

As the discussion period is coming to a close, we would like to kindly follow up to see if our response and new experiments (specifically the supplied-ligand AlphaFold3 benchmark and dual-protein interfacial perturbation retraining) have addressed your concerns, or if there are any remaining questions we can clarify in time. For your convenience, we have also summarized our latest rebuttal progress and manuscript revisions in the updated General Response.

We fully understand that this ICLR cycle has been a demanding challenge for both reviewers and authors, and we sincerely appreciate the time and effort you have dedicated to evaluating our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Summary of Rebuttal Response

Dear Reviewers and Area Chairs,

We sincerely thank Reviewers P5e4, HBVt, and NBHZ and the Area Chair for your careful evaluation and constructive feedback. This year's ICLR cycle has been particularly demanding for both reviewers and authors, and we deeply appreciate your time, rigor, and dedicated service to our research community.

Significance, Novelty & Strengths Recognized by Reviewers

Molecular glues offer a transformative therapeutic paradigm for modulating historically "undruggable" disease targets by inducing cooperative protein-protein interactions, yet rational in silico design remains largely unexplored. We are encouraged that all Reviewers recognized the value and contributions of our work:

  • Addressing an open gap via joint formulation (Sections 2 & 3). Unlike conventional methods that require either a predefined binding pocket or a known ligand, TriGlue jointly generates novel molecular glues and assembles unbound proteins into coherent ternary complexes (highlighted by all Reviewers: P5e4, HBVt, NBHZ).
  • Biology-inspired two-stage generative design (Sections 3.2 & 3.3). TriGlue couples SE(3)-equivariant ellipsoid interface estimation with a shared ternary flow network that jointly evolves ligand chemistry, 3D coordinates, and target pose, while naturally providing the GT-free confidence score TriConf (Reviewers P5e4, HBVt, NBHZ).
  • Strong empirical performance without privileged pockets (Section 4). Even without ground-truth pocket coordinates, TriGlue outperforms oracle-pocket molecule generators, docking baselines, and interface predictors, with effective few-shot transfer across diverse glue families (Reviewers P5e4, HBVt, NBHZ).

Summary of New Experiments & Manuscript Revisions

Guided by the Reviewers' suggestions, we have conducted new experiments, clarified evaluation protocols, and uploaded a revised manuscript (updates in blue) alongside detailed point-by-point responses:

  1. All-atom and cofolding baselines (Reviewers NBHZ, HBVt, P5e4; Appendices H.2, H.5, I.4, I.6). We added all-atom dMaSIF under a fair monomer-frame ellipsoid protocol for interface estimation and evaluated AlphaFold3 under a supplied-ligand protocol, confirming superior de novo binding-site localization and pose agreement by TriGlue.

  2. Unbound-input robustness via dual-protein perturbation retraining (Reviewer NBHZ; Appendix I.3). We constructed a dataset with smooth interfacial perturbations on both receptor and target, then retrained and evaluated TriGlue from scratch to verify robustness beyond idealized bound conformations.

  3. Retrained end-to-end interface-free ablation (Reviewer HBVt; Appendix I.1). We trained TriGlue w/o Interface from scratch under strictly matched settings and unified it with component-level ablations to confirm the architectural benefit of interface conditioning free of inference-time masking shift.

  4. Explicit generation and evaluation protocols (Reviewer HBVt; Appendices F, H.2, H.4, H.5, I.5). We specified the inference-time atom-count condition NMN_{\mathcal{M}} in Algorithm 1, clarified fair baseline inputs, candidate counts, single-receptor Vina scoring, and pose-decoupled ellipsoid evaluation, and added Hungarian-RMSD, Center-L2, and Bond-Length JSD analyses.

  5. Reference-free generation without reference atom counts (Reviewer HBVt follow-up; Table 4, Appendices F, H.2, I.7). Following Reviewer HBVt's follow-up, we evaluated TriGlue under reference-free size sampling (NM∼ptrain(NM)N_{\mathcal{M}} \sim p_{\mathrm{train}}(N_{\mathcal{M}})) against TargetDiff (prior) and PocketXMol (pocket_atoms_based) under their official default settings, confirming sub-Ångström localization (0.42 Å Center-L2), stronger binding affinity (-10.01 Vina Dock, 0.95 Aff), competitive QED (0.79) and SA (0.88), and size-distribution insights without access to NrefN_{\mathrm{ref}}.

Summary of Discussion Status

  • Reviewer HBVt (Score raised: 2 →\rightarrow 3). Following our multi-round discussion and new experiments, Reviewer HBVt raised their score from 2 to 3, and we have resolved all of their concerns and follow-up questions.
  • Reviewers NBHZ and P5e4. We have not yet received follow-up comments from Reviewer NBHZ or Reviewer P5e4 (initial rating 3). We would like to highlight that all concerns raised in their initial reviews have been addressed through dedicated experiments. As the discussion period draws to a close, we kindly invite Reviewers NBHZ and P5e4 to review our responses and let us know if any additional feedback or clarification would be helpful.

We hope these additions and clarifications fully address all concerns. We thank the reviewers and the Area Chair again for their consideration and remain available to provide any further clarification.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.