acceptodds
Under review as a conference paper at ICLR 2027

TSea: A Large-Scale Dataset of Synthetic Ternary Complexes for Molecular Glue Modeling

13 views

Abstract

Molecular glues (MGs) are an emerging therapeutic modality that induces or stabilizes protein–protein interactions, with demonstrated clinical impact in oncology. However, the scarcity of experimentally resolved MG ternary structures severely limits data-driven molecular glue modeling. Here, we introduce **TSea**, a large-scale dataset comprising over 100,000 synthetic ternary-complex structures, more than 450 times the size of MG-PDB. Starting from experimentally determined protein–ligand complexes, we construct synthetic ternary complexes through structurally constrained protein-chain partitioning and contact-based screening against MG-PDB reference distributions. Comparative analyses show that TSea captures selected structural and interaction characteristics of MG-PDB. We further develop **TSeaDiff**, a diffusion-based molecular generator trained on TSea. Among the evaluated models, TSeaDiff achieves the closest agreement with MG-PDB in protein–ligand interaction profiles. These results support TSea as a scalable structural resource for data-driven molecular glue modeling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular glues, Ternary complexes, Synthetic structural dataset, Structure-based drug design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WmvdOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper targets the scarcity of experimentally resolved molecular-glue (MG) ternary complexes. It begins with binary protein–ligand structures, partitions a single protein chain into two components under secondary-structure and fragment-length constraints, and retains synthetic ternary complexes whose contact statistics match distributions derived from the MG-PDB reference set. The resulting resource contains more than 100 000 structures. The authors compare ligand properties, interaction profiles, and buried-surface-area statistics against experimental complexes, then train the diffusion generator TSeaDiff on the synthetic data. On held-out, locally perturbed synthetic pockets, TSeaDiff yields protein–ligand interaction (PLI) profiles closer to the MG-PDB reference than the evaluated baselines.

Critical strengths
  1. The motivation is well founded. Experimentally resolved MG ternary structures remain limited (MG-PDB contains 221 complexes), and a substantially larger structural resource can support pre-training and hypothesis generation for ligand-mediated protein–protein interfaces.
  2. The construction pipeline is described with sufficient detail for understanding. Source filtering (BioLiP2 small-molecule subset, ligand size 10–74 heavy atoms, contact-coverage threshold, exclusion of short contacts suggestive of clashes or covalent linkage), protein-chain partitioning constraints, and MG-PDB-referenced contact selection are all specified.
  3. Evaluation examines multiple structural descriptors rather than size alone, and the authors explicitly note that the proxy metrics do not establish molecular-glue activity in biological systems and that the partitioned protein components are not guaranteed to fold independently.
Critical weaknesses
  1. The principal similarity analysis is partly circular. Synthetic structures are retained precisely because their contact patterns match the MG-PDB reference distribution; the same reference is then used to quantify how closely the synthetic set resembles experimental complexes. Agreement is therefore expected by construction, even if the partitioned fragments do not recapitulate the biology of genuine MG assembly. A held-out portion of the experimental reference would substantially strengthen the claim.
  2. Structural validation of the synthetic complexes remains limited. Partitioning one continuous protein chain into two fragments can generate an interface between sequence-contiguous regions that do not exist as separate chains in nature. The paper itself states that the resulting components are not guaranteed to fold independently. Independent checks of fragment foldability, interface stability, steric clashes at the newly created interface, and post-construction relaxation would be required before the structures can be treated as realistic ternary hypotheses; if a substantial fraction fail such filters, the usable size of the resource should be reported.
  3. The generator is assessed only on synthetic pockets drawn from a held-out partition of TSea and on proxy metrics (clash rate, valid yield, Vina scores, PLI-profile distance, HDR99 inclusion). Because the training distribution was deliberately constructed to imitate the reference, a favorable PLI-profile distance may partly reflect the selection rule rather than genuine generalization. Evaluation on experimentally resolved MG ternary complexes, ideally with baselines trained under matched data budgets, is needed to establish practical utility.
Questions for authors
  1. Can the MG-PDB reference distribution used for contact-based selection be fitted on one subset of the experimental complexes and the similarity statistics evaluated on a disjoint held-out subset?
  2. Which independent structural or energetic checks demonstrate that the partitioned fragments and the interfaces between them are physically plausible (e.g., predicted folding propensity, interface energy after relaxation, clash statistics)?
  3. How does TSeaDiff perform when conditioned on experimental MG ternary pockets, relative to baselines trained on comparable quantities of data?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Wmvd Part 1/2

Response to Weakness 1/ Question 1

We thank the reviewer for highlighting this limitation. Given the scarcity of experimentally determined MG ternary structures, our original construction uses the full MG-PDB reference to make the most of the available structural information. To address this concern, we reconstruct the synthetic dataset with disjoint experimental sets for reference fitting and evaluation, and reassess its similarity to experimental complexes. We also examine whether the reconstructed dataset retains its utility for downstream generation.

First, we separate the experimental complexes used for reference fitting from those used for evaluation. Specifically, we use MGBench, the predefined 88-complex subset of MG-PDB, as the held-out evaluation set. Excluding these 88 complexes from the 221-complex MG-PDB reference leaves 133 complexes for fitting the contact distribution. Feature standardization, KDE bandwidth selection, and HDR estimation are all based exclusively on these remaining 133 complexes. Using this refitted reference, we then select synthetic complexes while keeping the other construction criteria unchanged, obtaining 102,538 synthetic complexes.

We evaluate the resulting collection using the PLI/PPI profile definitions and fixed-pose Vina scoring protocol described in the manuscript. For each collection, we recompute PLI and PPI distances against MGBench and also report distances to the full MG-PDB reference to connect this analysis with the original results. As shown in the table below, the reconstructed collection achieves PLI/PPI distances to MGBench of 0.938/1.064, comparable to those of the original TSea collection (0.949/1.080) and lower than those of TernaryDB (1.454/1.370). Its Vina summaries likewise remain close to those of the original collection. These results indicate that the closer interaction-profile agreement with experimental MG complexes observed in our original analysis is maintained when the evaluation complexes are excluded from reference fitting.

DatasetFull-pocket VinaWeaker-side VinaStronger-side VinaPLI dist. (MG-PDB) ↓PPI dist. (MG-PDB) ↓PLI dist. (MGBench) ↓PPI dist. (MGBench) ↓
TSea (refitted reference)−6.62±2.30-6.62\pm2.30−1.88±1.02-1.88\pm1.02−4.75±1.84-4.75\pm1.840.6550.8800.9381.064
Original TSea HDR99−6.70±2.29-6.70\pm2.29−1.90±1.03-1.90\pm1.03−4.80±1.84-4.80\pm1.840.6630.8970.9491.080
TernaryDB−4.56±3.20-4.56\pm3.20−0.97±1.19-0.97\pm1.19−3.59±2.78-3.59\pm2.781.1341.1791.4541.370
MGBench−8.46±2.31-8.46\pm2.31−2.05±1.43-2.05\pm1.43−6.41±1.62-6.41\pm1.620.1590.1890.0000.000
MG-PDB−7.91±2.23-7.91\pm2.23−1.85±1.44-1.85\pm1.44−6.06±1.79-6.06\pm1.790.0000.0000.2150.292

Second, we assess whether the reconstructed dataset retains its utility for downstream generation. We train a new TSeaDiff model from scratch using the reconstructed dataset and perform 20 generation attempts for each of the 88 MGBench complexes excluded from reference fitting. For all models, we recalculate generation PLI distances using MGBench as the reference.

As shown in the table below, the retrained model achieves a PLI distance of 0.198, comparable to that of the original TSeaDiff (0.236) and lower than those of the other evaluated methods (0.326–0.460). Its weak-side Vina score and Vina centrality also remain comparable to those of the original TSeaDiff. These results support the retention of MG-relevant interaction characteristics in molecules generated using the reconstructed training data.

ModelFull-pocket VinaWeaker-side VinaPLI dist. (MGBench) ↓
TSeaDiff (refitted-reference data)−8.25±1.61-8.25\pm1.61−2.85±0.96-2.85\pm0.960.198
TSeaDiff (original)−8.47±1.63-8.47\pm1.63−2.77±0.80-2.77\pm0.800.236
PocketXMol−9.63±1.87-9.63\pm1.87−3.29±0.98-3.29\pm0.980.362
TargetDiff−7.20±2.17-7.20\pm2.17−2.38±1.14-2.38\pm1.140.326
DiffSBDD−5.97±2.85-5.97\pm2.85−1.73±1.46-1.73\pm1.460.369
ODesign A/B−8.32±2.36-8.32\pm2.36−2.37±1.14-2.37\pm1.140.460
ODesign merged−8.73±2.47-8.73\pm2.47−2.58±1.18-2.58\pm1.180.446

Together, these experiments show that TSea’s interaction-profile agreement is maintained against held-out experimental MG complexes that do not contribute to fitting the contact-selection reference. The downstream generation results further support the ability of the reconstructed dataset to capture MG-relevant interaction features that a generative model can learn.

AuthorsOpenReview ↗

Response to Reviewer Wmvd Part 2/2

Response to Weakness 2/ Question 2

Thank you for highlighting this. We perform independent fragment prediction, binary and ternary complex prediction, and post-construction relaxation.

Fragment structure recovery. We use ESMFold2 to predict each fragment independently. Among 212,175 fragments, 135,792 (64.00%) achieve TM-score ≥0.7\ge 0.7 against their corresponding reference conformations. For 42,363 candidates, both fragments meet this criterion when predicted separately, representing 39.80% of the full 106,436-candidate resource.

Complex and interface recovery. We additionally predict two-protein AB complexes and ABL ternary complexes. Among 89,813 systems evaluated under all three prediction conditions, the proportion of fragment chains achieving TM-score ≥0.7\ge 0.7 increase from 66.15% in separate monomer predictions to 79.76% in binary predictions and 83.86% in ternary predictions. Across 89,804 systems with interface measurements available for both binary and ternary predictions, median A–B interface C$\alpha$ RMSDs are 1.464 A˚\text{Å} and 1.146 A˚\text{Å}, respectively.

Relaxation and energetic assessment. We apply the same ligand-present relaxation protocol to TSea and MG-PDB and score A–B interfaces before and after relaxation, obtaining paired results for 64,565/106,436 and 208/221 complexes, respectively. Values below are medians, with scores shown before →\to after relaxation and chain B RMSD measured after chain A alignment. Among successfully evaluated complexes, median structural adjustments are comparable between TSea and the MG-PDB reference.

MetricTSeaMG-PDB
Interface score (REU/100 A˚2\text{Å}^{2})3.47→−3.503.47\to-3.504.78→−2.634.78\to-2.63
Chain B RMSD (A˚\text{Å})1.521.52
A–B contact retention (%)58.4551.15

We do not perform a separate clash check during construction because partitioning the experimental complexes preserves the coordinates of all retained atoms and therefore introduces no new steric clashes at the A–B interface.

Together, these analyses provide complementary support for fragment conformational recovery, recovery of the proposed interface geometry, and local energetic compatibility.

Response to Weakness 3/ Question 3.

Thank you for raising this point. To assess performance beyond synthetic pockets, we evaluate TSeaDiff and the baselines on 88 experimentally resolved MG ternary complexes from MGBench, which are not included in TSeaDiff’s training data. For each model, we perform 20 generation attempts per structure and evaluate the generated molecules using the same protocol in the manuscript.

As shown in tables below, TSeaDiff achieves the lowest PLI-profile distance (0.155), the lowest clash rate (0.42%), and the highest HDR99 inclusion rate (99.43%) among the evaluated models. These results support the generalization of TSeaDiff from synthetic training pockets to experimentally resolved MG systems.

Regarding training-data scale, the TargetDiff and DiffSBDD baselines each use 100,000 protein–ligand training pairs, and their training datasets are comparable in size to that of TSeaDiff. Compared with these two baselines, TSeaDiff shows better performance on the MG-related metrics highlighted above, while maintaining a valid yield comparable to TargetDiff and higher than DiffSBDD.

For general molecular quality, TSeaDiff shows less favorable QED and synthetic accessibility than TargetDiff and DiffSBDD, while maintaining comparable molecular diversity and uniqueness. This may reflect the emphasis of TSeaDiff’s training data on ternary interaction compatibility, which does not necessarily translate into better general drug-likeness or synthetic accessibility. Improving these properties while preserving favorable ternary-pocket interactions is therefore an important direction for the future work.

Overall, the MGBench evaluation shows that TSeaDiff maintains favorable performance on MG-related computational metrics in experimentally resolved ternary pockets, relative to baselines trained on comparable amounts of data.

ModelClash (%) ↓Valid (%) ↑Full-pocket VinaWeaker-side VinaPLI dist. ↓HDR99 (%) ↑QED ↑SA ↓Diversity ↑Uniqueness (%) ↑
TSeaDiff0.4280.0−8.47±1.63-8.47\pm1.63−2.77±0.80-2.77\pm0.800.15599.430.322±0.1940.322\pm0.1945.86±0.835.86\pm0.830.611±0.0860.611\pm0.086100.00
PocketXMol0.4588.5−9.63±1.87-9.63\pm1.87−3.29±0.98-3.29\pm0.980.22298.330.426±0.2050.426\pm0.2053.35±1.303.35\pm1.300.610±0.1040.610\pm0.10499.60
TargetDiff5.0280.6−7.20±2.17-7.20\pm2.17−2.38±1.14-2.38\pm1.140.17987.390.449±0.2180.449\pm0.2185.29±1.075.29\pm1.070.601±0.1100.601\pm0.110100.00
DiffSBDD11.3759.4−5.97±2.85-5.97\pm2.85−1.73±1.46-1.73\pm1.460.21690.050.423±0.2080.423\pm0.2085.05±1.055.05\pm1.050.612±0.1220.612\pm0.122100.00
ODesign A/B8.2990.5−8.32±2.36-8.32\pm2.36−2.37±1.14-2.37\pm1.140.34477.280.454±0.2350.454\pm0.2355.19±1.205.19\pm1.200.574±0.1110.574\pm0.11199.94
ODesign merged8.6390.3−8.73±2.47-8.73\pm2.47−2.58±1.18-2.58\pm1.180.33582.500.462±0.2370.462\pm0.2375.20±1.265.20\pm1.260.561±0.1120.561\pm0.11299.94
AuthorsOpenReview ↗

Looking Forward to Your Feedback

Dear Reviewer Wmvd,

We have revised the relevant sections of the manuscript and posted point-by-point responses to your comments. We would appreciate your feedback on the revisions and our responses, and would be happy to clarify any remaining questions.

Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_6zn7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work addresses data scarcity in machine learning for molecular glue (MG) structural modeling. The authors propose TSea, a new pipeline to generate pseudo-ternary complexes from existing binary protein–ligand structures. The pipeline cuts a single protein chain at loop regions to create two protein fragments that jointly bind one ligand. All candidate cuts are filtered by strict structural rules. The final dataset contains 106,436 valid structures (HDR99) and a filtered subset of 88,277 structures (HDR95).

The authors benchmark TSea against TernaryDB and MG-PDB using Vina scoring, PLIP interaction analysis and buried surface area (BSA). They further build TSeaDiff, a diffusion model designed for pairwise-to-ternary complex generation. Evaluated on perturbed held-out pockets, TSeaDiff achieves the lowest standardized PLI profile distance among compared methods.

Critical strengths
  1. The pipeline implementation is rigorous and fully reproducible. All filtering thresholds are clearly defined in Table 2. Table 3 presents the complete data filtering process with exact numbers at each step. Algorithm 1 provides full procedural details.

  2. Statistical screening is correctly implemented. The 2D KDE fitting, leave-one-out bandwidth optimization, HDR threshold construction and probability correction are all technically sound.

  3. The authors clearly state all limitations of their synthetic dataset. They explicitly clarify that HDR filtering only enforces geometric similarity and does not guarantee ternary stability or glue activity. They also note the absence of structural relaxation, side-chain rebuilding and redocking, and report systematic biases in protein–protein burial and evaluation missing values. This transparent discussion is a major strength.

Critical weaknesses
  1. Missing related work. The core idea of splitting single protein chains to generate synthetic complex training data has been widely explored in protein modeling, but relevant literature is not cited. Missing works include Teddymer/Proteina-Complexa, SeedProteo and previous pseudo-complex construction methods. These studies adopt highly similar validation strategies and provide clear theoretical assumptions. In addition, several closely related molecular glue generation methods are omitted, including TriGlue, Pep2Mol and GlueFinder. The paper fails to compare and discuss why artificial chain splitting is better than native interfacial pocket mining. This weakens the novelty statement.

  2. The TSea pipeline cuts a single folded protein chain into two fragments. The resulting A–B interface is essentially a cross-section of the protein hydrophobic core, not a natural protein–protein interaction interface. Additionally, will the core property of molecular glues, be absent in TSea synthetic structures? Real molecular glues stabilize ternary complexes via glue-induced new interactions that do not exist in binary states. In TSea structures, two protein fragments already interact before cutting, so no glue-driven cooperative stabilization exists. I want to understand the rationale behind how the dataset was constructed.

  3. All evaluations confound dataset effect and model architecture effect. The authors train their new TSeaDiff model on TSea data and compare it with external baseline models without fair control. Existing ablations only test network design choices, not dataset quality. Further ablation, e.g., comparison of the same model trained on TSea, raw unfiltered data and TernaryDB, is required.

Questions for authors

Current results only prove that TSea preserves local ligand pocket features, but whether it distorts natural MG interfacial properties and cooperative mechanisms still needs further exploration.

If the authors can address the above weaknesses, I would be willing to raise my score.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 6zn7 Part 1/2

Response to Weakness 1

We thank the reviewer for highlighting these studies. We acknowledge the precedents of chain partitioning in Teddymer/Proteina-Complexa, SeedProteo, and earlier pseudo-complex methods. Here, TSea focuses specifically on constructing ligand-centered ternary environments with MG-relevant structural characteristics, combining structurally constrained partitioning, bilateral ligand engagement, and experimentally informed contact screening. We will expand the related-work discussion to clarify these connections and distinctions.

At the model level, TriGlue jointly generates ligands and protein assembly poses, while Pep2Mol targets competitive PPI inhibitors. TSeaDiff adopts a pairwise-to-ternary architecture inspired by biological processes, explicitly integrating protein–protein and ligand–protein interactions. Its design and evaluation emphasize MG-like interaction patterns, particularly ligand engagement with both protein partners and agreement with experimental MG interaction profiles, rather than relying solely on conventional binding and drug-likeness metrics.

GlueFinder primarily addresses candidate MG discovery through interface-pocket identification, template matching, and virtual screening. TSea addresses the construction of a large-scale structural data resource. Accordingly, TernaryDB, which mines ligand-containing native interfaces, is the principal dataset baseline compared with TSea in our manuscript.

Our construction strategy offers two advantages over this native interfacial pocket-mining baseline. First, TSea has a larger scale by drawing on more source structures and producing more usable ternary complexes. Mining at interface requires interfacial ligand in a multi-chain source structure, whereas chain-split strategy accesses the substantially larger collection of binary protein–ligand structures. PLINDER reports that only 25% of its holo protein–ligand systems involve multiple interacting protein chains [1]^{[1]}. Applying MG-derived contact screening further reduces the number of qualifying native-mined complexes: only 9,191 of the 22,760 TernaryDB entries (40.4%) analyzed in our study fall within the MG-derived HDR95 contact region, compared with 88,277 structures in the corresponding TSea subset.

Second, TSea more closely matches experimental MG complexes on key interaction and local-pocket metrics, including PLI profiles and local BSA, as demonstrated in the manuscript. Thus, for constructing MG-oriented structural resources, our constrained partitioning and curation strategy improves on the evaluated native interfacial pocket mining baseline in both dataset scale and agreement with the experimental MG characteristics assessed in our study.

Response to Weakness 2

We thank the reviewer for raising this question. TSea’s central objective is to approximate the structural and interaction distributions of experimentally resolved MG complexes. We clarify its construction rationale in three points.

First, intramolecular contacts among secondary-structure elements and other segments share the same basic noncovalent interaction principles as intermolecular PPI contacts, although their composition and interaction preferences differ$^{[2]}$. These differences motivate our comparison against experimental MG structures. Although hydrophobic contacts are enriched in TSea, hydrogen bonds, salt bridges, and aromatic interactions also contribute to its interfaces. Across the evaluated PPI and PLI profiles, TSea exhibits smaller overall distances to MG-PDB than TernaryDB.

Second, MGs can stabilize pre-existing protein interactions. The approved drugs lenalidomide and pomalidomide enhance basal CRBN–CK1α and CRBN–IKZF1 interactions, respectively [3]^{[3]}, while fusicoccin stabilizes an existing 14-3-3–target-peptide interaction [4]^{[4]}. Moreover, protein-chain fragmentation has an wet-lab experimental precedent: the complementary E. coli DHFR fragments already associate without ligand, yet adding methotrexate or natural ligands further restores native-like secondary and tertiary structures [5]^{[5]}. Thus, existing contacts and a shared single-chain origin are compatible with additional ligand-dependent stabilization.

Third, contacts within the intact holo-state protein do not imply that the separated fragments would form the same stable interface without the ligand. In response to requests from other reviewers, we performed additional folding predictions for the same fragment pairs with and without the ligand. These predictions showed a lower median PPI-interface RMSD relative to the reference structure when the ligand was included. Through these screening and evaluation steps, we constructed TSea to preserve, as far as possible, structural features potentially relevant to the core function of MGs and to help address the severe scarcity of structural data for MG modeling.

AuthorsOpenReview ↗

Response to Reviewer 6zn7 Part 2/2

Response to Weakness 3

We thank the reviewer for this suggestion. TSea’s construction preserves the source atomic coordinates while reassigning chain IDs to define two protein components. Models whose input representations ignore this chain partition cannot directly exploit the additional information it provides. We therefore use TSeaDiff, which explicitly processes the two protein components, as a fixed architecture to assess the contribution of the training dataset.

Regarding unfiltered data, our existing construction ablation (Appendix D.5) shows that removing HDR screening leads to larger deviations from MG-PDB in interaction profiles and interfacial geometry. Given these findings and computational and time constraints, we prioritized the comparison with TernaryDB and did not additionally train TSeaDiff on the no-HDR dataset.

Specifically, we trained the same TSeaDiff architecture from scratch on TernaryDB. We then evaluated the resulting model using the same test pockets, sampling budgets, and evaluation protocol as in the manuscript. The results are reported below.

Molecular-glue-related evaluation.

Training datasetClash (%) ↓Valid (%) ↑Full-pocket VinaWeaker-side VinaPLI dist. ↓HDR99 (%) ↑
TSea3.1876.1−5.23−1.190.20595.53
TernaryDB2.5146.7−4.53−1.020.40492.08

Molecular-quality evaluation.

Training datasetClash (%) ↓Valid (%) ↑QED ↑SA ↓Diversity ↑Uniqueness (%) ↑
TSea14.6366.50.3135.020.68199.36
TernaryDB6.8844.70.2264.510.69798.48

With the architecture held fixed, training on TSea yields a higher valid yield, more favorable full-pocket and weaker-side Vina scores, closer PLI profile agreement with MG-PDB, and a higher HDR99 inclusion rate than training on TernaryDB. Although improvements are not uniform across all molecular-quality metrics, these results suggest that TSea offers an advantage over TernaryDB for learning MG-like interaction patterns under this evaluation setting.

Overall, TSea provides a substantially larger structural training resource with closer agreement to experimental MG interaction profiles than the native-interface-mining baseline evaluated here. Through structurally constrained protein-chain splitting, TSea constructs ternary environments in which a single ligand interacts with both protein components. We trained TSeaDiff on TSea to use this bilateral interaction information to generate molecules whose interaction types and distributions resemble those of experimental MGs. These results demonstrate TSea’s utility for MG-oriented ligand generation, and we hope it will also provide a useful data resource for other areas of MG research.

References

[1]. Durairaj J et al. (2024). bioRxiv, 2024.07.17.603955 (preprint).

[2]. Cohen M et al. (2008). Proteins 72(2), 741–753.

[3]. Cao S et al. (2022). Nat. Commun. 13, 815.

[4]. Würtele M et al. (2003). EMBO J. 22(5), 987–994.

[5]. Smith VF, Matthews CR. (2001). Protein Sci. 10(1), 116–128.

Thanks for the rebuttal. My concerns have been largely addressed. Exploring how to obtain more and better data in data-scarce setting is a meaningful direction. Therefore, I decided to increase my score to 3.

AuthorsOpenReview ↗

Thank You for Your Feedback

Dear Reviewer,

Thank you for reviewing our responses and reconsidering your assessment. We appreciate your recognition of the value of expanding structural data resources for molecular glue modeling. Your constructive feedback has helped us clarify TSea’s contribution and strengthen its evaluation.

Best regards,

The Authors

Reviewer_ARprOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the scarcity of experimentally resolved molecular glue ternary complexes, of which MG-PDB contains 221, by constructing synthetic ones from binary protein-ligand structures.

The first part characterizes MG-PDB to define screening criteria. Two contact descriptors are defined at a 4.0 A heavy-atom cutoff, the mean ligand-protein contact-pair count over the two sides and the mean number of distinct contacting protein atoms across the protein-protein interface. These are standardized and fitted with a bivariate Gaussian kernel density estimate whose bandwidth is chosen by leave-one-out likelihood, from which 95 and 99 percent highest-density regions are derived numerically. Ligand-level criteria, size between 10 and 75 heavy atoms, contact coverage above 0.4 and contact centrality above 0.3, come from the same analysis.

The second part applies these to BioLiP2. Single-chain complexes with a non-nuisance ligand are partitioned by excising two adjacent loop residues, requiring both fragments to have at least 20 residues, loop secondary structure at the excised residues, no cross-fragment beta ladder, and at least 6.0 A from the cut to the ligand; surviving partitions must satisfy the centrality criterion and lie inside the 99 percent region. The pipeline runs from 501,072 source records and 79.9 million candidate anchors to 106,436 structures from 70,979 source interactions, with 88,277 in the nested 95 percent subset. Evaluation compares TSea against TernaryDB and MG-PDB on scale and ligand coverage, fixed-pose Vina scoring, PLIP interaction profiles and buried surface area, plus a no-screening construction ablation.

The third part trains TSeaDiff, a pairwise-to-ternary equivariant diffusion model with a protein-protein expert and a parameter-shared binding expert applied separately to the ligand with each protein, fused by symmetric residual averaging and refined by a common equivariant block. It is compared against PocketXMol, TargetDiff, DiffSBDD and two ODesign input configurations on 100 locally perturbed held-out TSea pockets.

Critical strengths
  • Every threshold is frozen with its inequality direction stated as part of the data definition (L975-995).
  • The two-panel waterfall gives counts at both source-record and partition level (L997-1018), and the appendix reports the standardization constants, bandwidth selection and deduplication rules (L634-699, L868-878).
  • An efficient all-cut contact evaluation scheme is documented, making the pipeline auditable at scale (L886-898).
  • The no-screening ablation moves descriptors that were not used to define the screen, from 2.174 to 0.898 profile distance and 4368 to 2604 square angstroms of interfacial burial (L1176-1186).
  • That a two-descriptor contact criterion constrains quantities it does not directly see is a real, non-obvious result, and the part that most supports the pipeline design.
Critical weaknesses
  • Excising two loop residues from a single folded chain (L241-246) produces a domain interior, not an interface induced between partners that otherwise associate weakly (L028-032).
  • The paper's own numbers confirm this, with 2632 against 1640 square angstroms of interfacial burial and 18.3 against 7.3 hydrophobic events, attributed to exactly that cause (L305-312).
  • Independent folding is listed as a limitation (L461-467), but it is the premise rather than a caveat, and cooperativity is unlearnable from this data by construction.
  • MG-PDB protein-ligand contacts are hydrophobic-dominated at 7.05 against 2.83 while TSea is hydrogen-bond-dominated at 3.66 against 4.18, yet the aggregate distance still ranks TSea as closer.
  • PocketXMol, which is not trained on TSea, wins on clash rate, valid yield, both Vina scores, region inclusion, QED and synthetic accessibility; TSeaDiff leads only on the resemblance metric (L417-426).
  • No fine-tuning control separates the dataset's contribution from the architecture's, so the utility claim has no direct evidence.
  • Nothing is evaluated on real molecular glue systems despite the MGBench motivation (L036-039), and the prediction use case is deferred to future work (L1465-1468).
Questions for authors
  • Does training TargetDiff or DiffSBDD on TSea improve those models over their released checkpoints on the same 100 test pockets? Without this, the generation table cannot separate the dataset's contribution from the architecture's (L417-426).
  • What happens when TSeaDiff and the baselines generate into the 221 MG-PDB pockets or the MGBench subset? That would test transfer rather than in-domain reconstruction, which is what the motivation at L036-039 calls for.
  • Given the hydrophobic versus hydrogen-bond inversion, can you report per-interaction-type distances alongside the aggregate and a distributional distance rather than a standardized distance between dataset means (L1122-1126), which also does not account for the difference in sampling error between 221 and 106,436 structures?
  • The 106,436 structures derive from 70,979 source interactions and all preserve the source ligand pose (L997-1018, L868-878). How many distinct source PDB entries are involved, and what is the effective number of independent pockets after clustering receptors at 40 percent identity?
  • Fixed-pose Vina scores a pose docked against the intact chain, not the partitioned system (L1020-1025). Does redocking into the two-fragment system change the weaker-side distribution, and if so what does the agreement with MG-PDB weaker-side scores mean?
  • Did you check what fraction of partitions yield fragments that remain folded, for example by predicting each fragment separately and comparing to its conformation in the complex? That would make the limitation at L461-467 quantitative.
  • Test pockets are selected as the energy-qualified repacked conformation with the largest ligand-proximity-weighted side-chain displacement (L1404-1457). Do the comparative conclusions hold on unperturbed pockets and on randomly selected rather than maximum-displacement conformations?
4 comments
AuthorsOpenReview ↗

Response to Reviewer ARpr Part 1/3

Response to Weakness 1 / Weakness 2

We thank the reviewer for raising this concern. We acknowledge that chain partitioning brings part of the hydrophobic protein interior to the newly defined interfacial surfaces.

TSea is constructed through structurally constrained protein-chain partitioning and contact-based screening against MG-PDB reference distributions. Nevertheless, the protein–protein and ligand–protein contact metrics of all passed synthetic ternary complexes fall within the MG-PDB-derived HDR99 region defined in Section 3. This threshold constrains the accepted contact region without requiring the distribution of these metrics in TSea to exactly match the experimental MG-PDB reference.

In the construction ablation using the same subset of 10,000 source PDB entries, HDR99 screening reduced the mean protein–protein BSA from 4,368 to 2,604 A˚2\text{Å}^2 (Table 8). Comparative analyses also showed smaller PPI and PLI profile distances to MG-PDB than TernaryDB (Table 6), together with closer agreement in BSA at ligand-centered pocket scales (Table 7). These results show that TSea captures selected structural and interaction characteristics of real MG complexes and support it as a scalable structural resource for data-driven MG modeling.

Response to Weakness 3 / Question 6

We thank the reviewer for this valuable suggestion. Following the reviewer’s recommendation, we independently predicted the structure of each protein fragment using ESMFold2 and compared the predicted structures with their corresponding conformations in the original complexes.

Of the 212,175 protein fragments evaluated, 135,792 (64.00%) achieved a TM-score ≥ 0.7, indicating substantial structural similarity to their original complex-bound conformations.

These results suggest that approximately two-thirds of the evaluated fragments can recover their original conformations through independent structure prediction, providing quantitative evidence of their computational structural recoverability.

Response to Weakness 4 / Question 3

Thank you for this helpful suggestion. We have extended our analysis to compare the interaction distributions of TSea and TernaryDB with those of MG-PDB using Wasserstein-1 distance:

Interaction type, metricTSea (HDR99)TernaryDB
Hydrogen bonds, W1W_1 ↓0.08210.1488
Hydrophobic contacts, W1W_1 ↓0.08270.1042
Salt bridges, W1W_1 ↓0.01090.0284
Aggregate distributional distance, mean W1W_1 ↓0.03760.0576

To assess sampling uncertainty associated with the different dataset sizes, we repeated the comparison on 2,000 resampled versions of the datasets. Structures sharing the same source PDB were sampled together, and MG-PDB was also resampled each time to include uncertainty in the smaller reference set. The difference in aggregate distance (TSea − TernaryDB) was −0.0200 (95% CI: −0.0220 to −0.0182), which supports the conclusion that TSea is closer to MG-PDB overall than TernaryDB.

Although the hydrogen-bond/hydrophobic balance differs from MG-PDB, TSea still has smaller distributional distances than TernaryDB for hydrogen bonds, hydrophobic contacts, and salt bridges, which together account for 95.3% of the PLI events in MG-PDB. These results support closer overall agreement with MG-PDB.

Response to Weakness 5

We acknowledge PocketXMol’s stronger performance on several general molecular-generation metrics. However, its reported data resources include approximately 12 million small molecules, roughly 800 times the 14,893 unique ligand in the training set of TSeaDiff. Despite these different training resources, TSeaDiff remains competitive on other metrics, ranking second in clash rate (3.18% versus 1.95%) and HDR99 inclusion rate (95.53% versus 96.26%). It also achieves the highest mean ligand diversity on the separate 10-pocket evaluation panel (Table 10).

More importantly, our primary objective is to learn MG-like interfacial interactions. TSeaDiff achieves the lowest MG-PDB-standardized PLI distance among all evaluated models (0.205 versus 0.377 for PocketXMol), indicating closer agreement with the interaction-type profiles of experimentally resolved MG structures. This provides task-relevant structural evidence supporting TSea’s utility for MG modeling, even though further improvements in general molecular quality remain necessary.

AuthorsOpenReview ↗

Response to Reviewer ARpr Part 2/3

Response to Weakness 6 / Question 1

We thank the reviewer for highlighting the need to distinguish the contributions of training data and model architecture. TSea preserves the original atomic coordinates while assigning distinct chain IDs to the two protein components. Models whose input representations ignore this partition, including TargetDiff and DiffSBDD, treat the components as a combined protein environment and cannot directly exploit the additional partition information TSea provided. We therefore use TSeaDiff, which explicitly models the ligand’s interactions with both protein components, to verify TSea’s utility.

To provide additional evidence, we trained TSeaDiff from scratch on TernaryDB and compared it with the TSea-trained model using the same test pockets, sampling budgets, reconstruction procedure, and evaluation metrics as described in the manuscript.

Training datasetClash (%) ↓Valid (%) ↑Full-pocket VinaWeaker-side VinaPLI dist. ↓HDR99 (%) ↑
TSea3.1876.1−5.23−1.190.20595.53
TernaryDB2.5146.7−4.53−1.020.40492.08

The TSea-trained model achieves a higher valid yield, closer PLI profile agreement with MG-PDB, more favorable full-pocket and weaker-side Vina scores, and higher HDR99 inclusion.

We also assessed the architectural design through ablations (Appendix F.6, Table 11), with the full model and both variants trained on the same TSea training set and evaluated under the same protocol. The full model achieves higher valid yield and lower PLI profile distance than the common-only and joint-expert variants. Since all three models receive both protein components, these results support the pairwise-to-ternary design. Together with the additional training-dataset comparison, these findings support TSea’s practical utility as a training resource for generating ligands with MG-like interaction profiles.

Response to Weakness 7 / Question 2

We thank the reviewer for highlighting the need for evaluation on real MG systems. We conducted an additional generation experiment on 88 MGBench (subset of MG-PDB) structures that were not used for training. Each model made 20 generation attempts per pocket, and the outputs were assessed following the evaluation procedures described in the manuscript.

The results show that TSeaDiff achieves the lowest clash rate and PLI profile distance, together with the highest HDR99 inclusion rate, producing 1,408 valid molecules (80.0% yield). These results support its applicability on real MG pockets. PocketXMol achieves better Vina scores and a higher valid yield, while TSeaDiff’s QED and SA scores indicate room for improvement.

ModelClash (%) ↓Valid (%) ↑Full-pocket VinaWeaker-side VinaPLI dist. ↓HDR99 (%) ↑
TSeaDiff0.4280.0−8.47−2.770.15599.43
PocketXMol0.4588.5−9.63−3.290.22298.33
TargetDiff5.0280.6−7.20−2.380.17987.39
DiffSBDD11.3759.4−5.97−1.730.21690.05
ODesign A/B8.2990.5−8.32−2.370.34477.28
ODesign merged8.6390.3−8.73−2.580.33582.50
ModelQED ↑SA ↓Diversity ↑Uniqueness (%) ↑
TSeaDiff0.3225.860.611100.00
PocketXMol0.4263.350.61099.60
TargetDiff0.4495.290.601100.00
DiffSBDD0.4235.050.612100.00
ODesign A/B0.4545.190.57499.94
ODesign merged0.4625.200.56199.94

Response to Question 4

We thank the reviewer for this question. The 106,436 TSea structures derive from 70,979 source interactions spanning 32,498 distinct PDB entries. Clustering the original receptor sequences at 40% sequence identity yields 4,835 clusters, while applying the same procedure to TernaryDB yields 2,192 clusters from 4,386 source PDB entries. Detailed results are reported below:

Counting unitTSeaTernaryDBTSea / TernaryDB (fold)
Structure records106,43622,7604.68
Distinct source PDB entries32,4984,3867.41
Distinct source receptor chains58,52718,5443.16
Exact unique observed receptor sequences29,8177,2624.11
Receptor clusters at 40% sequence identity4,8352,1922.21
AuthorsOpenReview ↗

Response to Reviewer ARpr Part 3/3

Response to Question 5

We thank the reviewer for raising this point. The original analysis scores the ligand pose from the source experimental structure against the partitioned A+B receptor and each partner separately. To directly assess the effect of pose search in the two-fragment system, we additionally evaluate local ligand optimization (Vina M) and redocking (Vina D). On matched cases, both yield more favorable mean weaker-side scores than fixed-pose scoring (Vina S) in TSea and MG-PDB:

Mean weaker-side scoreVina SVina MVina D
TSea−1.89−2.06−2.08
MG-PDB−1.46−1.63−1.88

Using all available native-pose MG-PDB scores as a fixed reference, the between-dataset Wasserstein-1 distance changes little after TSea redocking, from 0.378 to 0.403 kcal/mol. Thus, redocking shifts the weaker-side scores toward more favorable values, while their overall distribution remains similarly close to the native-pose MG-PDB reference. The weaker-side score distributions of TSea and MG-PDB remain similarly close after redocking.

Response to Question 7

Thank you for raising this point. We originally introduced pocket perturbations to support a fair comparison with the baselines. For models that do not distinguish protein chain IDs, an unperturbed pocket can be geometrically equivalent to the corresponding pocket in the source binary protein–ligand complex. Side-chain perturbations were therefore introduced to reduce potential bias from directly reusing ligand-bound pocket geometries that may have been used for baseline training.

To assess whether our comparative conclusions hold under alternative pocket conditions, we performed two additional generation experiments on the same 100 test pockets used in the manuscript: one using unperturbed pockets and the other using randomly selected energy-qualified repacked conformations. We evaluated TSeaDiff and the baselines using the same analysis pipeline and evaluation criteria as in the original setting.

As shown in tables below, TSeaDiff consistently achieves the lowest PLI-profile distance and maintains lower clash rates and higher HDR99 inclusion than most baselines. The main comparative patterns reported in the manuscript remain consistent on both unperturbed and randomly selected repacked pockets.

MG-related properties on unperturbed pockets.

ModelClash (%) ↓Valid (%) ↑Full-pocket VinaWeaker-side VinaPLI dist. ↓HDR99 (%) ↑
TSeaDiff0.1282.2−7.27±2.16-7.27\pm2.16−1.96±1.16-1.96\pm1.160.27194.77
PocketXMol1.8087.4−8.04±2.33-8.04\pm2.33−2.14±1.23-2.14\pm1.230.41494.97
TargetDiff3.8382.9−6.84±2.17-6.84\pm2.17−1.78±1.18-1.78\pm1.180.41189.63
DiffSBDD15.4161.5−6.01±2.40-6.01\pm2.40−1.59±1.30-1.59\pm1.300.43089.59
ODesign A/B10.2288.7−6.21±2.54-6.21\pm2.54−1.58±1.34-1.58\pm1.340.48190.76
ODesign merged7.1991.7−6.45±2.52-6.45\pm2.52−1.68±1.29-1.68\pm1.290.46391.38

Molecular quality on unperturbed pockets.

ModelClash (%) ↓Valid (%) ↑QED ↑SA ↓Diversity ↑Uniqueness (%) ↑
TSeaDiff0.1375.10.350±0.1700.350\pm0.1705.18±1.205.18\pm1.200.649±0.1070.649\pm0.107100.00
PocketXMol0.0091.00.420±0.1800.420\pm0.1803.22±1.083.22\pm1.080.624±0.1180.624\pm0.11895.84
TargetDiff4.8384.80.471±0.1880.471\pm0.1885.03±1.095.03\pm1.090.632±0.1390.632\pm0.139100.00
DiffSBDD19.7660.90.463±0.1890.463\pm0.1894.64±1.124.64\pm1.120.657±0.1370.657\pm0.137100.00
ODesign A/B5.0793.60.415±0.1980.415\pm0.1984.48±1.504.48\pm1.500.620±0.1250.620\pm0.12592.75
ODesign merged3.2395.80.420±0.2050.420\pm0.2054.58±1.384.58\pm1.380.643±0.1470.643\pm0.14796.99

MG-related properties on randomly selected pockets.

ModelClash (%) ↓Valid (%) ↑Full-pocket VinaWeaker-side VinaPLI dist. ↓HDR99 (%) ↑
TSeaDiff3.0878.6−5.86±2.62-5.86\pm2.62−1.44±1.48-1.44\pm1.480.33193.00
PocketXMol1.5986.9−7.28±2.37-7.28\pm2.37−1.93±1.21-1.93\pm1.210.43393.44
TargetDiff7.2777.8−6.02±2.61-6.02\pm2.61−1.45±1.48-1.45\pm1.480.42787.92
DiffSBDD21.8556.5−5.17±2.72-5.17\pm2.72−1.23±1.55-1.23\pm1.550.40984.78
ODesign A/B12.6186.6−5.35±2.75-5.35\pm2.75−1.15±1.56-1.15\pm1.560.46591.34
ODesign merged10.5187.7−5.60±2.58-5.60\pm2.58−1.32±1.35-1.32\pm1.350.45991.22

Molecular quality on randomly selected pockets.

ModelClash (%) ↓Valid (%) ↑QED ↑SA ↓Diversity ↑Uniqueness (%) ↑
TSeaDiff12.5866.70.311±0.1830.311\pm0.1835.15±1.185.15\pm1.180.666±0.1150.666\pm0.115100.00
PocketXMol10.4777.00.416±0.1830.416\pm0.1833.37±1.043.37\pm1.040.646±0.1180.646\pm0.11899.39
TargetDiff17.1867.00.438±0.1980.438\pm0.1985.12±1.005.12\pm1.000.648±0.1290.648\pm0.129100.00
DiffSBDD44.3838.60.424±0.1950.424\pm0.1954.70±1.154.70\pm1.150.690±0.1350.690\pm0.135100.00
ODesign A/B21.4576.90.375±0.1810.375\pm0.1814.52±1.604.52\pm1.600.600±0.1160.600\pm0.11690.07
ODesign merged18.1580.70.395±0.1890.395\pm0.1894.42±1.514.42\pm1.510.645±0.1260.645\pm0.12695.12
AuthorsOpenReview ↗

Looking Forward to Your Feedback

Dear Reviewer ARpr,

We have revised the relevant sections of the manuscript and posted point-by-point responses to your comments. We would appreciate your feedback on the revisions and our responses, and would be happy to clarify any remaining questions.

Thank you again for your time and consideration.

Best regards,

The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response and Summary of Revisions

General Response and Summary of Revisions

Dear Reviewers and Area Chair,

Thank you for your careful reading and constructive feedback. We have revised the manuscript and responded to each review in its own thread. Section and table references below refer to the revised PDF.

Main Revisions

  • Evaluation on experimental MG pockets

    • Generation on 88 MGBench complexes, with 20 attempts per pocket. (Sec. 5.2, Table 2)
    • Molecular-quality results and discussion of the remaining trade-offs. (Appendix F.5, Table 12)
  • Training-data comparison

    • TSeaDiff trained from scratch on TernaryDB and evaluated under the same held-out-pocket protocol as the TSea-trained model. Results cover both MG-related properties and molecular quality. (Sec. 5.2; Appendix F.7, Table 14)
  • Related work

    • Added HumanPPI, Teddymer/Proteína-Complexa, and SeedProteo, and clarified connections to TriGlue, Pep2Mol, and GlueFinder. (Sec. 2)

Additional Analyses in the Reviewer Responses

  • Structural recovery: Independent fragment prediction, binary and ternary complex prediction, and relaxation with interface scoring. (ARpr, Part 1/3; Wmvd, Part 2/2)
  • Contact-reference fitting: Dataset reconstruction and model retraining after excluding MGBench from reference fitting. (Wmvd, Part 1/2)
  • Distributional agreement: Per-interaction-type Wasserstein-1 distances and uncertainty estimates from 2,000 resampling iterations grouped by source PDB. (ARpr, Part 1/3)
  • Dataset coverage: Source PDB, receptor-chain, unique-sequence, and 40%-identity cluster counts. (ARpr, Part 2/3)
  • Evaluation robustness: Generation on unperturbed and randomly repacked pockets, together with local ligand optimization and redocking comparisons. (ARpr, Part 3/3)

The revised manuscript and individual responses provide the detailed settings and results. We appreciate the reviewers’ suggestions, which helped strengthen the evaluation and clarify the scope of TSea.

Best regards,

The Authors

AuthorsOpenReview ↗

Follow-up on Our Response and Revisions

Dear Reviewer,

Thank you for your detailed and thoughtful review. We have now posted a full response to your comments. In particular, we would be grateful if you could kindly reconsider your evaluation of our manuscript based on our analysis and explanations.

If there are any remaining questions or concerns, we would be glad to provide further clarification. We sincerely hope that our revisions have addressed your concerns and that the improvements we made will be favorably considered during your re-evaluation.

Thank you again for your time and consideration.

Best regards,

The Authors

AuthorsOpenReview ↗

Rebuttal Acknowledgment

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair for their leadership in overseeing the review process, and all reviewers for their careful reading and constructive feedback.

We appreciate the reviewers’ recognition of several strengths of our work:

  • Reviewer ARpr highlighted the explicit construction criteria and detailed data accounting, describing the pipeline as “auditable at scale” and the screening ablation as “a real, non-obvious result.”
  • Reviewer 6zn7 described the implementation as “rigorous and fully reproducible,” the statistical screening procedures as “technically sound,” and the transparent discussion of limitations as “a major strength.”
  • Reviewer Wmvd recognized that the motivation is “well founded” and the pipeline is described with “sufficient detail for understanding,” while noting the evaluation across multiple structural descriptors and the explicit acknowledgment of biological limitations.

Our work introduces TSea, a dataset of over 100,000 synthetic ternary-complex structures constructed through structurally constrained chain partitioning and contact-based screening. It provides a scalable resource for learning MG-like interfacial interactions. We further demonstrate its utility through TSeaDiff, which achieves the closest protein–ligand interaction-profile agreement with the experimental reference among the evaluated models on both synthetic and experimentally resolved MG pockets.

During the discussion, we conducted extensive additional experiments and analyses, provided point-by-point responses to all reviews, and incorporated key results and clarifications into the revised manuscript.

We are pleased that Reviewer 6zn7 stated that their concerns had been “largely addressed” and raised their score to 3 following our responses on related work, the construction rationale, and the dataset’s contribution. We particularly appreciate their recognition of the value of obtaining more and better data in this data-scarce setting.

We regret that we have not had further discussion with Reviewers ARpr and Wmvd. We remain confident that our additional evidence and detailed explanations substantively address their concerns about structural plausibility, evaluation validity, and dataset utility. We respectfully ask that these responses and the revised manuscript be considered in the final assessment.

Thank you again for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.