acceptodds
Under review as a conference paper at ICLR 2027

GuideDock: Geometry-Guided Diffusion for Rigid Protein-Protein Docking

Abstract

Modeling 3D protein-protein interactions is central to structural biology and the design of drugs and therapeutics. Recent deep learning methods for rigid binder pose prediction can be more efficient than classical docking software and expensive co-folding models. However, these methods typically operate directly in pose space with limited explicit priors over interface geometry. As a result, they must jointly learn which binder-receptor regions are compatible and which rigid transformations realize that compatibility, often leading to unreliable sampling and suboptimal accuracy. In this work, we introduce GuideDock, a framework that guides diffusion-based pose generation with distance maps recovered by latent flow matching. GuideDock treats recovered distance maps as soft geometric constraints that steer the diffusion process toward physically plausible rigid poses. Across DIPS-Het and Docking Benchmark 5.5, GuideDock outperforms competing deep learning-based rigid docking baselines, improving mean DockQ by 30% and reducing mean complex RMSD by up to 15% with only a single generated pose, while running up to 300 times faster than the best co-folding models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
rigid docking, protein-protein interactions, generative models

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_M5CYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes GuideDock, a geometry-guided diffusion framework for rigid protein-protein docking. It uses latent flow matching to generate inter-protein distance maps, which subsequently guide SE(3) pose diffusion through Feynman–Kac steering and gradient-based rigid-body updates. Experiments on DIPS-Het and DB5.5 demonstrate improvements over existing MSA-free neural docking methods.

Critical strengths
  • Well-motivated formulation: Using predicted distance maps as soft geometric constraints is a reasonable approach to improving the reliability of rigid-body pose generation.

  • Interesting integration of geometric guidance: Combining distance-based potentials with FK steering and force-based rigid-body updates provides a structured approach to guiding diffusion in SE(3).

  • Promising empirical improvements: GuideDock achieves better DockQ and RMSD than several MSA-free neural docking baselines, with relatively low reported inference latency.

Critical weaknesses
  • Limited methodological novelty and insufficient justification for architectural complexity: The framework primarily integrates existing techniques, including latent flow matching, distance map, SE(3) diffusion, FK steering, and torque-based updates. While assembling these tools for rigid docking is sensible, the paper reads largely as an intricate combination of existing techniques without fundamental methodological novelty in diffusion theory or structural geometry. In addition, Table 6 compares the generative distance-map model with a simple MLP distance regressor and shows only a modest absolute DockQ improvement of about +0.03 on both DIPS-Het (0.24 → 0.27) and DB5.5 (0.06 → 0.09). This is a comparison of the distance-map generation module, not of the full GuideDock pipeline against a regressor, but it still raises a legitimate question: the paper needs to justify more convincingly why the complex latent flow matching / VAE pipeline is necessary relative to a stronger deterministic distance predictor.

  • The absolute docking accuracy remains low, limiting the demonstrated practical utility: On DB5.5, GuideDock achieves a mean DockQ of only 0.09 and a median of 0.05, both below the conventional acceptable-quality threshold of 0.23. Although the relative improvements over several MSA-free neural baselines are encouraging, their practical significance remains unclear without reporting the proportion of successfully docked complexes under standard CAPRI/DockQ thresholds. Evaluation on genuinely unbound protein conformations would further strengthen the applicability claims, since the current experiments use fixed conformations and do not test robustness to conformational change.

  • The inference-time efficiency claims lack sufficient computational accounting: Despite emphasizing single-pose prediction, GuideDock requires 100 latent flow matching steps, 50 pose diffusion steps, and four particles for FK steering. The reported runtime of approximately 7.5 seconds per complex lacks a component-wise breakdown, and it is unclear whether feature extraction and geometric preprocessing are included. More transparent profiling and comparisons under matched computational budgets are needed. Moreover, the runtime comparisons appear to use numbers taken from other papers, which may have been measured under different hardware, batch sizes, and preprocessing assumptions. Since high-throughput inference is a central motivation, the paper should provide transparent profiling and comparisons under matched computational budgets.

  • Missing baseline DFMDock (Chu L S, Sarma S, Xu D, et al. Unified sampling and ranking for protein docking with DFMDock[J]. Proteins: Structure, Function, and Bioinformatics, 2026.).

Questions for authors

See weakness

4 comments

We thank the reviewer for the detailed feedback and questions and highlighting our integration of geometric guidance and the empirical improvements of our work. Please see our point-by-point response to each weakness below.

Weakness 1: Novelty, stronger deterministic baseline, and architecture justification.

Our contribution is a way to use predicted inter-protein geometry to guide rigid docking. The distance map supplies a global geometric target, the pose model searches over rigid placements of the supplied monomers, FK steering allocates particles toward promising poses, and the rigid-body updates refine their translation and rotation. Treating the map as a soft constraint allows the pose model to use informative distances even when the full predicted map cannot be realized exactly in three dimensions.

The generative map model is motivated by the ambiguity of inter-protein geometry. A deterministic predictor trained with squared error can average distinct plausible distance maps, whereas a generative predictor can represent their distribution. The benefit of sampling multiple distance maps is exhibited in Figure 7 of the paper, and this capability is not available to deterministic distance predictors. To test a stronger geometric predictor than the MLP baseline supplied, we trained a deterministic model with four invariant point attention layers, joint attention between the partners, and a direct distance-prediction head, all using the same sequence embeddings. Below are the results formatted as mean / median / std.

DatasetModelDockQIRMSD (Å)CRMSD (Å)
DIPS-HetOurs0.27 / 0.16 / 0.288.95 / 6.68 / 7.718.17 / 5.94 / 7.29
DIPS-HetPredictor0.25 / 0.14 / 0.259.34 / 7.30 / 7.738.43 / 6.17 / 7.22
DB5.5, 25Ours0.09 / 0.05 / 0.1011.40 / 10.74 / 5.1511.46 / 11.52 / 4.98
DB5.5, 25Predictor0.07 / 0.04 / 0.0811.90 / 11.48 / 5.4011.82 / 13.07 / 5.07

GuideDock's manuscript results have higher mean DockQ and lower mean IRMSD and CRMSD than the added IPA baseline on both datasets. These results support the generative map predictor against both simple and geometry-aware deterministic alternatives. Separately, holding the backbone and embeddings fixed, our guidance ablation increases DIPS-Het mean DockQ from 0.20 to 0.26, demonstrating the contribution of guidance during pose sampling. We emphasize that the main contribution of the work is not the specific architecture used to predict distance maps but the approach of using distance maps to guide pose diffusion.

Weakness 2: CAPRI categories and conformational change.

Thank you for this suggestion. The quality distribution makes the practical gain clearer than mean DockQ alone. Below we reproduce the CAPRI-style categories from our earlier evaluation; entries are percentages of predictions.

DatasetMethodIncorrectAcceptableMediumHigh
DIPS-HetEBMDock69.5714.0115.940.48
DIPS-HetGuideDock61.179.2224.764.85
DB5.5EBMDock100.000.000.000.00
DB5.5GuideDock88.0012.000.000.00

GuideDock increases the fraction of medium- or high-quality DIPS-Het poses from 16.42% to 29.61%. On DB5.5, the earlier evaluation yields acceptable poses in 12% of predictions, compared with none for EBMDock. These results show that the improvement includes a greater yield of useful poses, while the DB5.5 distribution identifies accurate interface recovery as a priority.

Regarding conformational change, the current evaluation uses bound monomer conformations and measures rigid docking. To accommodate substantial conformational changes, a future extension could update internal coordinates alongside rigid-body motion, with the distance map guiding both interface placement and conformational adjustment. For experimental results, please see our response to Reviewer EAbw under Weakness 7 where we report performance on rigid vs non-rigid complexes in DB5.5.We see a clear degradation in performance for the flexible complexes over the rigid-body complexes across the board given the scope of the method, although the gap in performance far from catastrophic. This is expected given that apo and holo structures often share high backbone similarity and often undergo only minor induced fit changes [1, 2].

[1] Clark et al. Inherent versus induced protein flexibility: Comparisons within and between apo and holo structures. PLoS Computational Biology 15 (2019).

[2] Ruvinsky et al. Side-Chain Conformational Changes upon Protein–Protein Association. Journal of Molecular Biology 408 (2011).

Weakness 3: Runtime breakdown and profiling.

GuideDock returns one pose after 100 latent-flow steps, 50 pose-diffusion steps, and four-particle FK steering. We agree that these internal computations should be explicit when interpreting its runtime. To support this, we profiled the components of GuideDock more closely on RTX Pro 6000 GPUs since 3090 GPUs were unavailable in the given time. The following timings are means of three repeated measurements per complex, averaged over the ten complexes.

GuideDock componentMean time (s)
Cached feature collation and transfer0.0011
Map conditioning, generation, and decoding0.7691
Pose initialization0.0005
Pose sampling, selection, and CPU transfer1.8951
Independently measured total2.6668

The profile includes map generation, geometric guidance, and all four particles. Static sequence-embedding generation and model loading are excluded. Pose sampling accounts for approximately 71% of the measured time and map generation for 29%, identifying where the computational cost lies.

We also measured the baselines on the same GPU and complexes, using four internal particles or candidates per method and returning one selected pose.

MethodMean inference time (s)
Direct pose diffusion1.83
EBMDock0.52
DFMDock1.53
GuideDock2.67

Guidance adds 0.840 s over the same pose model without guidance in this profile. These measurements make the accuracy–runtime tradeoff explicit: geometric guidance adds computational cost while retaining seconds-scale inference. For matched sampling budgets, please see our response to Reviewer EAbw under Weakness 3.

Weakness 4: DFMDock baseline results.

Thank you for identifying this baseline. We evaluated the released DFMDock model checkpoint on 100 DIPS-Het and 25 DB5.5 complexes, selecting its prediction by the lowest energy score. We use 40 denoising steps per their documented code config and 4 candidates per-target to ensure an at-parity sampling budget.

DatasetMethodDockQIRMSD (Å)CRMSD (Å)
DIPS-HetDFMDock0.08 / 0.03 / 0.1315.52 / 14.04 / 9.0213.89 / 12.95 / 8.68
DIPS-HetGuideDock0.27 / 0.16 / 0.288.95 / 6.68 / 7.718.17 / 5.94 / 7.29
DB5.5DFMDock0.09 / 0.05 / 0.1211.93 / 10.51 / 5.8012.25 / 12.61 / 5.68
DB5.5GuideDock0.09 / 0.05 / 0.1011.40 / 10.74 / 5.1511.46 / 11.52 / 4.98

On DIPS-Het, GuideDock's manuscript result achieves more than three times DFMDock's mean DockQ, alongside lower mean IRMSD and CRMSD. On DB5.5, the two methods have similar mean DockQ at the reported precision, while GuideDock has lower mean IRMSD and CRMSD. Thus, the added baseline supports a substantial docking-quality advantage on DIPS-Het and improved pose accuracy on both datasets. Our response to reviewer wJ6i, Question 1, additionally reports DFMDock in the expanded DIPS-Het evaluation.

Thank you for your clarifications. However, I remain somewhat conservative regarding the paper's novelty and overall contribution to the field. I will make my final recommendation after taking the other reviewers' comments and discussions into consideration.

Thank you for engaging during the discussion period. We respect your decision, and we are happy to answer any outstanding questions or concerns in the time remaining.

Reviewer_wJ6iOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

GuideDock is a framework that uses distance maps to guide protein-protein interaction pose generation. The method combines latent flow-matching model and pose diffusion model, while combining distance maps as soft constraints. The tool is validated with DIPS-Het and DB5.5, demonstates meaninful RMSD improvement comparing to other MSA-free rigid and substantially much faster than co-folding methods.

Critical strengths
  • the paper provides valuable insight of protein-protein interaction generation: using distance to guide the diffusion model. The distance maps is an interpretable representation to approach the pose generation problem.
  • the result section clearly discusses the advantages of this method: highest mean DockQ among MSA-free rigid methods; while states that the efficiency is the focus when comparing to co-folding methods.
Critical weaknesses
  • the number of test examples is relatively small. The experiment only uses 100 samples from DIPS-Het and 25 from DB5.5
Questions for authors
  • How are the 100 test examples selected from the 1153 samples? What is the permutation result for this test dataset?
  • Which baseline and averaging procedure produce the reported 30% DockQ improvement and 15% CRMSD reduction?
  • Why is the DB5.5 difficult for MSA-free deep docking methods? Could the authors clarify which types of complexes these methods handle well, and what are the factors associated with their failures? Do you think there is a way to solve it?
3 comments

Thank you for the valuable feedback and acknowledging the insightfulness of our contribution to protein-protein interaction prediction. Please see our point-by-point reply to each question below.

Weakness 1 + Question 1: Test set selection and test set size concern.

We first clarify the test set size and selection process and apologize for the lack of clarity. The reported “100 examples from 1,153” does not describe the recovered evaluation data. The subset follows the 91 PDB identifiers listed in Appendix F.2 of EBMDock. We match these identifiers against the DIPS-Het test list and retain the corresponding chain-pair examples. This produces 216 examples, of which 206 examples have usable inputs. We will issue a correction in the revised manuscript describing this selection process, the example counts, and the corresponding PDB and chain identifiers.

To address the concern about test-set size directly, we additionally evaluated all 1,103 usable complexes from 633 PDB structures in the full DIPS-Het test list. Results are formatted as mean / median /std.

MethodDockQIRMSD (Å)CRMSD (Å)
EBMDock0.26 / 0.15 / 0.258.32 / 6.33 / 9.498.79 / 6.26 / 10.03
DiffDock-PP0.20 / 0.03 / 0.301.36 × 10⁶ / 13.91 / 4.50 × 10⁷1.37 × 10⁶ / 11.94 / 4.54 × 10⁷
DFMDock0.10 / 0.03 / 0.1615.49 / 14.69 / 9.7314.03 / 12.94 / 9.22
GuideDock0.35 / 0.35 / 0.287.31 / 3.56 / 12.146.49 / 3.16 / 8.20

We see that GuideDock improves mean DockQ by 33.8% and reduces mean CRMSD by 26.2% relative to EBMDock, the strongest baseline in this expanded evaluation. It reaches DockQ ≥ 0.23 on 628/1,103 complexes (56.9%), compared with 444/1,103 (40.3%) for EBMDock. The advantage therefore extends beyond the original subset to the full usable test set.

To further address the concern regarding the small test set size, we also include results on 50 dimers from PPIRef-50K [1], retaining structures deposited after September 30, 2021 and excluding pairs with greater than 80% sequence similarity between binding partners. This provides an additional test beyond the DIPS-Het and DB5.5 benchmarks.

MethodDockQIRMSD (Å)CRMSD (Å)
Boltz-1x0.27 / 0.18 / 0.255.46 / 2.97 / 5.246.75 / 3.21 / 6.62
Chai-10.26 / 0.16 / 0.235.40 / 3.25 / 5.216.53 / 3.73 / 6.68
DiffDock-PP0.04 / 0.02 / 0.0620.21 / 13.19 / 32.9213.25 / 13.22 / 8.68
EBMDock0.08 / 0.06 / 0.0810.82 / 9.87 / 5.1911.68 / 10.77 / 5.49
GuideDock0.12 / 0.06 / 0.129.61 / 9.03 / 4.688.05 / 7.47 / 5.60

On these additional dimers, GuideDock increases mean DockQ from 0.08 for EBMDock to 0.12 and reduces mean CRMSD from 11.68 Å to 8.05 Å. It has the highest mean DockQ and lowest mean RMSDs among the evaluated neural rigid-docking methods, supporting the benefit of geometric guidance on a separate dataset.

Question 2: Matching headline numbers with results.

Thank you for pointing out this important clarity issue. Please see our reply to Reviewer EAbw under Weakness 2.

Question 3: Difficulty of DB5.5 and failure analysis.

Please see our reply to Reviewer EAbw under Weakness 1. We hypothesize that DB5.5 is harder because its smaller binding interfaces provide fewer geometric constraints for locating and orienting the binding partners. Consistent with this hypothesis, we observe a smaller median interface fraction on DB5.5 than DIPS-Het (8.13% versus 11.74%), fewer native contacts (48 versus 73.5), and an association between more accurate predicted distance maps and better docking performance. These findings motivate improving interface localization by emphasizing interface contacts during map prediction and adapting guidance strength to predicted map confidence.

[1] Bushuiev et al. Learning to design protein-protein interactions with enhanced generalization. ICLR 2024.

Thank you for the author's response and the experiments. The author basically answered my questions, but the IRMSD/CRMSD value of first table of DiffDock-PP, which is 1.36×10⁶, seems to be an abnormal value. I'm concerned that this might affect the comparability of the mean values. Furthermore, I believe that compared to existing similar methods, there have been certain improvements. However, there is still room for improvement in the performance of some tests. The impact on the development of the field has not been fully discussed in more extensive cases and in more real-world applications. I believe my rating has given a positive assessment of this article and I have decided to keep it unchanged.

Thank you for the reply. Regarding the abnormally large DiffDock-PP RMSD values, we diagnosed the issue and found that the abnormalities are caused by certain poses with >1000Å RMSD, i.e. completely degenerate poses. To remedy this, we filtered out all cases where DiffDock-PP generated a pose that situated the binder and receptor centroids more than 1000Å apart. Only 13 complexes are affected. The corrected results are below:

MethodDockQIRMSD (Å)CRMSD (Å)
EBMDock0.26 / 0.15 / 0.258.32 / 6.33 / 9.498.79 / 6.26 / 10.03
DiffDock-PP0.20 / 0.03 / 0.3116.44 / 13.64 / 20.9914.02 / 11.72 / 17.66
DFMDock0.10 / 0.03 / 0.1615.49 / 14.69 / 9.7314.03 / 12.94 / 9.22
GuideDock0.35 / 0.35 / 0.287.31 / 3.56 / 12.146.49 / 3.16 / 8.20

We agree that our method has room for improvement in benchmarks, and we hope to test it in more expansive practical settings. We thank the reviewer for their positive review and for engaging in the rebuttal discussion.

Reviewer_EAbwOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses rigid protein-protein docking, where binder and receptor conformations are given and only the relative pose must be inferred. The authors argue that existing MSA-free methods operate directly in SE(3) pose space with few explicit geometric priors, so the model must simultaneously learn which regions are compatible and which rigid transformation realizes that compatibility. GuideDock splits the problem into two stages: a latent flow matching model that generates an inter-residue distance map between binder and receptor, and a pose diffusion model whose reverse trajectory is steered by a differentiable potential measuring agreement between the sampled pose and the predicted distance map. Steering combines Feynman-Kac particle resampling with a gradient update expressed in the Lie algebra of SO(3). Experiments on DIPS-Het and DB5.5 report RMSD, DockQ, and runtime against classical docking software, co-folding models, and MSA-free rigid docking baselines.

Critical strengths

Using distance maps as an intermediate representation is a well-motivated idea. They are invariant to global rotation and translation, richer than contact maps, and easier to learn than full coordinates, which makes them a reasonable level of abstraction for constraining pose space.

Expressing the guidance gradient as a rigid-body update on the rotation manifold is a more careful treatment than naive Euclidean updates on a rotation parameterization, and the derivation is presented clearly.

The empirical presentation is unusually candid. Co-folding models are explicitly framed as high-cost references rather than direct competitors, possible training-set leakage for AlphaFold-Multimer is acknowledged, and an oracle experiment with ground-truth distance maps is included, which openly exposes how much headroom the method still has.

The observation that averaging multiple sampled distance maps hurts performance while using them as separate seeds helps is an interesting finding in its own right, and the explanation offered is consistent with the multi-modal nature of docking.

Critical weaknesses

My main concern is attribution of the gains. The ablation suggests that on one of the two datasets the backbone diffusion model without any guidance already performs comparably to or better than the baselines, so the contribution of the geometric guidance that the paper is named after appears modest there, while being substantial on the other dataset. Q1: Can the authors separate the contribution of the stronger backbone (architecture and sequence embeddings) from the contribution of distance-map guidance, and explain why the benefit differs so much between the two datasets?

Some headline numbers in the abstract are hard to match to the tables, since the relative improvements seem to depend on which baseline and which dataset is used as the reference. Q2: Could each reported percentage and speedup be tied explicitly to a specific baseline and dataset?

The claim of using only a single generated pose sits somewhat uneasily with the use of multiple particles in the steering procedure, followed by selecting the lowest-energy one. Q3: How does the comparison look under a matched effective sampling budget?

Absolute performance remains low on both benchmarks, with mean DockQ values that would mostly not reach acceptable quality by CAPRI conventions. Q4: Could the authors report the distribution across CAPRI quality categories rather than only mean DockQ, so readers can judge practical usefulness?

The strongest classical baseline appears to substantially outperform the proposed method on accuracy, and the paper responds mainly on runtime grounds. Q5: In the intended high-throughput screening setting, is the current recall quality sufficient to serve as a coarse stage ahead of a slower but more accurate method?

A few technical details could be clarified. The reliability weight derived from the MDS reconstruction error does not obviously stay within a sensible range, and its geometric justification is not fully argued. Q6: Is this weight clipped, and is there an ablation with the weight disabled?

Finally, results appear to come from a single training run and a single selected checkpoint, with no seed variance or significance testing, and one of the two test sets is quite small. Q7: Could multi-seed results and paired confidence intervals be provided, particularly for the smaller benchmark? There is also no analysis of why performance degrades on the harder benchmark, which would help readers understand whether the rigid-body assumption itself is the limiting factor.

On presentation, one paragraph in the method section is repeated almost verbatim, and several figure cross-references do not match the figure numbering.

Questions for authors

see weakness

4 comments

Thank you for the thoughtful and comprehensive review and for acknowledging our paper's motivation and experimental strengths. We answer each weakness point-by-point below.

Weakness 1: Contribution of guidance and backbone + performance gap between datasets.

Distance-map guidance contributes accuracy beyond the backbone and sequence embeddings. See our response to Weakness 1 + Question 1 from 2WJs holding the architecture, checkpoint, and embeddings fixed. Guidance increases mean DockQ from 0.20 to 0.26 on DIPS-Het and from 0.05 to 0.09 on DB5.5. This directly attributes the improvement to guided inference.

Regarding the contribution of the residue sequence embeddings, we trained EBMDock with the same E1-600M embeddings and evaluated GuideDock with ESM2-650M embeddings. Results are formatted as mean / median /std. The ESM2-650M results are directly from Appendix H.2.

DatasetMethodDockQIRMSD (Å)CRMSD (Å)
DIPS-HetEBMDock, E1-600M0.20 / 0.12 / 0.189.61 / 7.26 / 7.949.52 / 7.27 / 7.98
DIPS-HetGuideDock, ESM2-650M0.24 / 0.13 / 0.259.08 / 6.78 / 7.308.34 / 6.92 / 6.60
DIPS-HetGuideDock, E1-600M (ours)0.27 / 0.16 / 0.288.95 / 6.68 / 7.718.17 / 5.94 / 7.29
DB5.5EBMDock, E1-600M0.06 / 0.04 / 0.0711.08 / 10.16 / 3.4515.25 / 14.39 / 5.01
DB5.5GuideDock, ESM2-650M0.08 / 0.04 / 0.1112.12 / 12.00 / 4.3712.77 / 12.72 / 5.55
DB5.5GuideDock, E1-600M (ours)0.09 / 0.05 / 0.1011.40 / 10.74 / 5.1511.46 / 11.52 / 4.98

GuideDock retains higher DockQ and lower CRMSD with either encoder than EBMDock with E1. Together with the controlled guidance comparison, this shows that the gains cannot be explained by the sequence encoder alone.

Regarding the performance gap between datasets, guidance is in principle most useful when the predicted map identifies the interface accurately enough to improve its registration. To understand this difference, we examined native interface geometry, defining a contact as an inter-partner C$\alpha$ pair within 8 Å and interface percentage as the percentage of all residues across both binding partners involved in a contact. This yields the below table:

DatasetMedian interface percentageMedian contact pairs
DB5.58.13%48
DIPS-Het11.74%73.5

We see that DB5.5 has generally smaller interfaces than DIPS-Het with fewer residues participating in an interface and each interface being composed of fewer contacts. Furthermore, on DIPS-Het, the interface percentage is positively associated with GuideDock DockQ (Spearman ρ=0.24\rho = 0.24). Together, these results suggest that the smaller interfaces in DB5.5 are harder to locate and provide fewer native contacts to constrain the docking orientation, making interface localization a plausible contributor to the performance gap.

We also examined the relationship between distance-map accuracy and the benefit of guidance and break down distnace map MSE versus mean DockQ. Map MSE measures error against native inter-protein distances.

DIPS-Het

MSE percentileMap MSE range (Ų)Mean DockQ
0–25th1.16–12.610.657
25–50th12.63–29.870.287
50–75th30.04–123.880.116
75–90th124.86–255.630.041
90–100th257.58–1857.910.014

DB5.5

MSE percentileMap MSE range (Ų)Mean DockQ
0–25th20.51–112.930.1362
25–50th117.45–155.410.0637
50–75th161.39–279.430.0430
75–90th311.32–384.350.0426
90–100th498.97–1174.450.0602

As expected, the lowest-MSE tiers have the highest mean DockQ scores, linking more accurate maps with better absolute pose quality. Together with DB5.5's smaller interfaces, this supports examining whether the map accurately locates the interacting residues. Guidance therefore provides the strongest improvement when the map supplies an accurate geometric signal. Together with the smaller interfaces observed on DB5.5, this suggests that the quality of the predicted interface geometry is an important factor in how much guidance helps across datasets.

Weakness 2: Matching headline numbers with results.

Thank you for pointing out this important clarity issue. We derive the 30% DockQ improvement and 15% CRMSD reduction numbers from averaging the relative improvement across the two datasets, using the strongest MSA-free neural rigid-docking baseline for each metric on each dataset. See the below table for an in-depth breakdown.

MetricDatasetBest baselineBaseline meanGuideDock meanRelative improvement
DockQDIPS-HetDiffDock-PP0.240.2712.50%
DockQDB5.5EBMDock / ElliDock0.060.0950.00%
CRMSD (Å)DIPS-HetEBMDock9.098.1710.12%
CRMSD (Å)DB5.5ElliDock14.4111.4620.47%

Averaging the two dataset-wise improvements gives 31.25% for DockQ and 15.30% for CRMSD, which round to the reported 30% and 15%. We will make this averaging explicit in the manuscript and describe the CRMSD reduction as an average rather than “up to 15%.”

Regarding runtime numbers, the co-folding runtime comparison refers to AlphaFold-Multimer. Its reported times are 1,940 s versus 7.5 s for GuideDock on DIPS-Het, and 2,540 s versus 7.4 s on DB5.5, giving reported runtime ratios of 258.7× and 343.2×, respectively. We will update the manuscript to identify the comparator and datasets explicitly.

Weakness 3: Matching sampling budgets.

By “single pose,” we mean one returned prediction after four-particle FK steering. We will make this sampling budget explicit in the manuscript. Regarding a matched sampling budget, we note that DiffDock-PP and EBMDock already use multiple samples in their procedure, e.g. EBMDock uses 5 samples during denoising and selects the lowest-energy final pose for evaluation, so in in this sense, our methods are already at near-sampling budget parity. For full parity, we conduct experiments where each baseline has four candidates instead of 5 to match GuideDock's four internal particles. See the table below for the results.

DatasetMethodDockQIRMSD (Å)CRMSD (Å)
DIPS-HetEBMDock0.21 / 0.11 / 0.229.79 / 8.30 / 7.689.50 / 7.99 / 7.38
DIPS-HetDiffDock-PP0.24 / 0.03 / 0.3314.77 / 13.71 / 12.4712.40 / 10.69 / 11.01
DIPS-HetGuideDock0.27 / 0.16 / 0.288.95 / 6.68 / 7.718.17 / 5.94 / 7.29
DB5.5EBMDock0.06 / 0.03 / 0.0714.38 / 13.49 / 5.9315.58 / 14.15 / 6.25
DB5.5DiffDock-PP0.04 / 0.02 / 0.0516.44 / 14.60 / 6.0516.17 / 16.38 / 6.12
DB5.5GuideDock0.09 / 0.05 / 0.1011.40 / 10.74 / 5.1511.46 / 11.52 / 4.98

GuideDock's advantage therefore remains at four internal poses per method.

Weakness 4: CAPRI quality categories.

Thank you for this suggestion. Below we report CAPRI-style quality distributions and show that GuideDock yields higher-quality poses over EBMDock, the strongest neural rigid-body baseline, across datasets.

DatasetMethodIncorrectAcceptableMediumHigh
DIPS-HetEBMDock69.5714.0115.940.48
DIPS-HetGuideDock61.179.2224.764.85
DB5.5EBMDock100.000.000.000.00
DB5.5GuideDock88.0012.000.000.00

We see that GuideDock shifts predictions toward higher-quality categories compared to EBMDock. On DIPS-Het, it increases the fraction of medium- or high-quality poses from 16.42% for EBMDock to 29.61%. On DB5.5, GuideDock produces acceptable poses in 12% of predictions, compared with none for EBMDock.

Weakness 5: High-throughput screening setting and tradeoff.

To substantiate our method's utility as a coarse-stage ahead of a slower, more accurate method, we evaluated a two-stage workflow in which GuideDock predicts poses for all candidate pairs for DIPS-Het, its highest-scoring predictions are accepted directly, and full-search HDOCK handles the remaining pairs. GuideDock scores are defined by agreement between its generated distance map and predicted pose. We also evaluated the same workflow with EBMDock as the first stage, sorting poses by lowest energy. HDOCK alone achieves 191/206 acceptable-or-better predictions (DockQ ≥ 0.23). See below for the two-stage results. We test performance when accepting the top-10%, 20%, and 50% of the first-stage candidates.

First-stage acceptanceGuideDock → HDOCK successesEBMDock → HDOCK successesGuideDock → HDOCK mean DockQEBMDock → HDOCK mean DockQGuideDock → HDOCK estimated time saved
10%191/206189/2060.87220.85906.65%
20%192/206183/2060.83460.798213.90%
50%169/206148/2060.64350.598037.67%

Time saved is relative to HDOCK. We use the RTX 3090 GuideDock sampling time of 5 seconds per case, add 5.25 seconds for confidence scoring, and add the measured HDOCK time for the remaining cases.

Accepting the top 20% of GuideDock predictions avoids 42 HDOCK calls, preserves 189/191 (98.95%) of HDOCK-successful cases, and recovers three additional cases. The resulting acceptable-pose yield is comparable to HDOCK alone, with an estimated 13.9% reduction in sequential inference time.

Compared to EBMDock at the same acceptance fraction, both workflows make 164 HDOCK calls and retain nearly identical HDOCK computation, 85.20% and 85.51% of the full workload, but GuideDock yields nine more successful final predictions. More aggressive acceptance produces a larger accuracy-cost tradeoff, as the 50% rows show. The results support selective use of GuideDock to reduce expensive docking calls while maintaining acceptable-pose yield under these benchmark conditions.

Weakness 6: Reliability weight clarification and ablation.

The weight uiju_{ij}, for binder residue ii and receptor residue jj, is not clipped. Its motivation is that the predicted interchain distances are coupled: they must fit a shared three-dimensional arrangement of the two proteins. We therefore use MDS to reconstruct the predicted interchain distances together with the supplied intrachain distances. This gives each constraint access to the geometry of the full distance map when determining its relative influence on the pose update.

Intuitively, a pair that receives full weight agrees with the reconstruction, and increasing relative disagreement lowers its weight. For example, if the predicted distance is 10 Å, a reconstruction of either 5 Å or 15 Å gives uij=0.5u_{ij}=0.5, attenuating the corresponding pair's contribution to the forces and torques used in the rigid-body update.

Because the weight is not clipped, it could become negative when the reconstructed distance exceeds 2x the predicted distance, and this is by design. In this case, a negative weight seeks to reverse a pair's guidance contribution, pushing the pair's residues outward rather than pulling them together to reverse the effect of a prediction that is likely incorrect. We use this weight as a geometric heuristic rather than a calibrated score. Improving it and finding optimal ways of tuning it is an outstanding research direction. We report the results of an ablation below.

DatasetUse uiju_{ij}?DockQIRMSD (Å)CRMSD (Å)
DIPS-HetYes0.27 / 0.16 / 0.288.95 / 6.68 / 7.718.17 / 5.94 / 7.29
DIPS-HetNo0.26 / 0.15 / 0.288.93 / 6.69 / 7.748.17 / 5.98 / 7.36
DB5.5Yes0.09 / 0.05 / 0.1011.40 / 10.74 / 5.1511.46 / 11.52 / 4.98
DB5.5No0.08 / 0.05 / 0.1011.59 / 10.79 / 4.9311.67 / 11.78 / 4.76

The ablation shows that weighting has a modest effect on performance, very mildly increasing mean DockQ on all datasets and mildly benefitting RMSD on DB5.5. The two versions perform similarly overall. We will revise the description around Equation 17 to clarify this role. We leave research for improved weighting schemes to future work.

Weakness 7: Seed variance, paired CIs, and explaining performance degradation.

Regarding test set size, please see our reply to Reviewer wJ6Ri under Weakness 1 where we report the full DIPS-Het test set performance as well as performance on new dimers from PPIRef-50k [1]. Regarding seed variance and CIs, we evaluated our method on 3 seeds and report the SDs and their 95% CIs below.

DB5.5

MetricMean ± SD95% CI
DockQ ↑0.08 ± 0.01[0.05, 0.10]
C-RMSD ↓ (Å)11.92 ± 0.54[10.58, 13.27]
I-RMSD ↓ (Å)12.02 ± 0.90[9.80, 14.25]
L-RMSD ↓ (Å)27.57 ± 0.95[25.21, 29.93]

DIPS-Het

MetricMean ± SD95% CI
DockQ ↑0.27 ± 0.00[0.26, 0.28]
C-RMSD ↓ (Å)8.13 ± 0.11[7.87, 8.40]
I-RMSD ↓ (Å)9.00 ± 0.12[8.72, 9.29]
L-RMSD ↓ (Å)20.00 ± 0.25[19.38, 20.61]

Regarding performance degradation, please see our previous response to Weakness 1 where we examine the interface identifiability of DB5.5 versus DIPS-Het. Regarding whether the rigid-body assumption is the culprit, we examine performance according to whether the complex is classified as rigid-body, defined as having a holo-apo IRMSD <= 1.5Å and fraction of absent native contacts <= 0.4 from the DB5.5 authors [2]. Below, we see a clear degradation in performance for the flexible complexes over the rigid-body complexes across the board. This is expected given the scope of the method, although the gap in performance far from catastrophic. This is expected given that apo and holo structures often share high backbone similarity and often undergo only minor induced fit changes [3, 4]. We therefore attribute the performance gap largely to the higher difficulty in residue distance prediction and harder interface identifiability.

GroupDockQIRMSDCRMSDDistance-map MSE
Rigid-body0.092 / 0.053 / 0.12111.58 / 10.10 / 5.7911.28 / 11.00 / 5.31203.64 / 148.97 / 159.60
Flexible0.077 / 0.042 / 0.05911.08 / 11.22 / 3.7411.78 / 13.35 / 4.32290.87 / 162.12 / 349.53

Weakness 8: Manuscript errors.

Thank you for identifying these presentation issues. We will remove the duplicated passage, add the missing overall captions for Figures 3 and 6, and correct the corresponding references.

[1] Bushuiev et al. Learning to design protein-protein interactions with enhanced generalization. ICLR 2024.

[2] Guest et al. An Expanded Benchmark for Antibody-Antigen Docking and Affinity Prediction Reveals Insights into Antibody Recognition Determinants. Structure 29 (2019).

[3] Clark et al. Inherent versus induced protein flexibility: Comparisons within and between apo and holo structures. PLoS Computational Biology 15 (2019).

[4] Ruvinsky et al. Side-Chain Conformational Changes upon Protein–Protein Association. Journal of Molecular Biology 408 (2011).

Reviewer_2WJsOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GuideDock addresses rigid protein-protein docking by generating inter-protein distance maps with a VAE and latent flow matching model, then using these maps to guide pose diffusion. The predicted distances act as soft geometric constraints. Experiments on DIPS-Het and DB5.5 show improvements over the reported MSA-free deep learning docking baselines, with competitive inference time.

Critical strengths
  1. I find the method technically solid. Using predicted distances as soft constraints makes sense when the distance maps contain errors, and the rigid-body updates preserve the input monomer structures. The formulation connects distance prediction and pose generation in a coherent way.

  2. The experiments cover docking accuracy, runtime, ablations, and sensitivity to sampling choices. Table 5 shows the benefit of distance guidance over the backbone model, while Table 6 supports the generative distance model over the tested MLP baseline. These results support the effectiveness of the approach in the evaluated setting.

Critical weaknesses
  1. My main reservation is the motivation for the full modeling pipeline. The ablations support the proposed approach, but the paper could better explain why the combination of VAE, latent flow matching, and pose diffusion is worth the added complexity. In particular, the discussion of MDS does not directly address fitting the predicted distances by optimizing only the rigid rotation and translation of two fixed monomers. A clearer account of the expected trade-offs would strengthen the motivation.

  2. The absolute DockQ gains over the strongest reported MSA-free baselines are modest, particularly on DB5.5. I would have liked more interpretation of where these gains come from. Figure 2 relates distance-map accuracy to pose quality, but leaves open which docking cases benefit most and what guidance changes in those cases. This limits the insight into when the proposed inductive bias is most helpful

Questions for authors

My questions concern the design rationale and interpretation of the existing results. I would welcome a discussion or an illustrative case study; additional benchmarking or scaling experiments are not necessary for my assessment.

  1. Could the authors explain what motivates the full generative pipeline over simpler alternatives, such as directly fitting the same predicted distances over rigid-body poses? In which settings do they expect the additional stages to offer the greatest benefit, and how might this depend on the capacity of the pose model?

  2. Could the authors discuss an example from the existing results that shows what distance guidance improves geometrically? An example where guidance offers little benefit or leads to a worse prediction would also help explain the method’s useful scope.

3 comments

We thank the reviewer for their feedback and noting that our work is technically solid. We give a point-by-point response to each weakness and question below.

Weakness 1 + Question 1: Motivation for the full pipeline.

The motivation for the full pipeline is that directly fitting predicted distances over rigid-poses often fails to give valid 3D poses. Since the distances are predicted by an unconstrained neural network, important physical constraints like the triangle inequality are not guaranteed, meaning the MDS solution may yield an invalid Euclidean embedding. See Appendix A for an example. In our experiments, direct rigid-fitting failed for all complexes since the resulting distance matrices failed to be PSD with rank-3, which is required for an exact 3D Euclidean embedding. Therefore, to avoid this problem, our method incorporates the predicted distance maps as a soft constraint to guide diffusion in pose space instead. We see the greatest improvement over baselines when using this distance-based guidance, per our distance map guidance ablation below. Results are formatted as mean / median /std. "Direct pose diffusion" refers to our baseline backbone pose diffusion model without any geometric guidance.

DIPS-Het

AblationDockQIRMSDCRMSD
Direct pose diffusion0.20 / 0.12 / 0.209.70 / 7.90 / 7.768.73 / 6.38 / 7.55
GuideDock0.27 / 0.16 / 0.288.95 / 6.68 / 7.718.17 / 5.94 / 7.29

DB5.5

AblationDockQIRMSDCRMSD
Direct pose diffusion0.05 / 0.03 / 0.0714.08 / 13.27 / 5.7214.86 / 14.09 / 5.96
GuideDock0.09 / 0.05 / 0.1011.40 / 10.74 / 5.1511.46 / 11.52 / 4.98

Regarding settings that would benefit most from distance guidance, we expect distance guidance to be most useful when the predicted map captures the approximate interface but contains local errors that prevent all distances from being satisfied simultaneously. In this setting, the map can help identify promising regions of pose space, while the pose model combines that information with its learned prior over protein geometry. The benefit should depend on whether the map provides information beyond what the pose model can infer from the individual proteins.

Finally, regarding the impact of the capacity of the pose model, we hypothesize that guidance may be especially helpful when the pose model is low-capacity and has difficulty identifying a plausible interface. A higher-capacity model with a stronger learned pose prior may rely less on the map, but it may also be better able to use an informative map while resolving inconsistent constraints. The balance between model capacity and guidance quality is therefore an important question for future study. Our current experiments do not isolate the effect of pose-model capacity, but examining performance as each component scales is an outstanding research question.

Weakness 2 + Question 2: Geometric benefits and failure cases of guidance.

Since mean distance-map MSE averages over all inter-protein residue pairs, we hypothesize that accurate distances between many distant residues can obscure errors at the binding interface. To test this, we examine distances separately for the residue pairs that form the native interface (C$\alpha$ distance below 8 Å). The table below shows how accurately the final poses recover these native distances. Results are formatted as unguided → guided.

ComplexFinal-pose MSE w.r.t. native distances (Ų)Native contacts recoveredDockQ
5JDC_D_ABC9.38 → 0.3698/221 → 210/2210.471 → 0.939
7SZ0_C_ABD5.81 → 72.2631/58 → 23/580.506 → 0.255

In 5JDC, guidance corrects the distances between the interacting residues and recovers most of the native interface. The binder center barely changes (0.84 to 0.83 Å error), while orientation error decreases from 20.9° to 2.6°. Thus the improvement comes from aligning the correct residue pairs across the interface. On the other hand, in 7SZ0, the generated map has an overall MSE of 6.06 Ų against native distances, but its MSE on native interface pairs is 39.95 Ų. Those interface pairs account for just 58 of 59,831 distances, so their errors have little influence on the overall mean. Although guidance brings the binder center closer to its native position, the final pose has much larger errors in the distances between the residues that should interact, alongside worse orientation and lower DockQ.

In sum, we find that guidance is most useful when it helps recover the distances between the residue pairs that define the binding interface. Low mean distance-map MSE does not by itself establish that the interface-relevant distances are accurate or that guidance will improve the pose. Using a classifier to first identify residues that are most likely to be at the interface is a viable remedy for this problem for future research.

Thank you for your response, which has addressed my concerns. However, I believe my current score already reflects a fairly positive assessment of the paper.

Thank you for engaging during the discussion period and for the positive assessment of our work. We are happy to answer any outstanding questions or concerns in the time remaining.

Comments on the paper

2 comments

Thank you to the reviewers for your careful reading and constructive feedback. We have revised the manuscript to address several presentation and clarity concerns, with changes highlighted in blue. In particular:

  • Headline results and runtime (Reviewers EAbw and wJ6i): We clarified that the reported 30% DockQ improvement and 15% C-RMSD reduction are averages across DIPS-Het and DB5.5 relative to the best reported MSA-free neural rigid-docking baseline for each metric and dataset. We also identified AlphaFold-Multimer as the comparator for the co-folding speedup and added the dataset-specific runtime ratios.
  • Sampling budget (Reviewers EAbw and M5CY): We clarified that GuideDock returns one final pose after four-particle Feynman–Kac steering.
  • DIPS-Het test selection (Reviewer wJ6i): We corrected the description of the recovered evaluation data. The selection contains 216 chain-pair examples before filtering, of which 206 are usable. We added the corresponding PDB and chain identifiers to the appendix.
  • Reliability weight (Reviewer EAbw): We lightly revised the discussion around Equation 17 to explain how the MDS reconstruction provides geometric context for reweighting residue-pair contributions to the rigid-body update.
  • Presentation and figure references (Reviewer EAbw): We removed the duplicated method passage and added the missing overall captions for Figures 3 and 6, correcting the associated panel references.

In our responses, we also provided additional analyses addressing the pipeline motivation and geometric examples requested by Reviewer 2WJs, and the deterministic predictor, quality categories, runtime breakdown, and DFMDock comparison requested by Reviewer M5CY.

We appreciate the reviewers’ feedback helping us clarify the paper and examine the method’s strengths and limitations. We hope these revisions and our responses address their concerns, and we welcome any further questions and score revisions in light of our rebuttal.

AuthorsOpenReview ↗

Follow-up on Rebuttal Responses

Dear Reviewers,

Thank you again for your time and valuable feedback on our work. We would love to hear your thoughts on our rebuttal responses, which we believe address the central concerns brought up during the review process.

We greatly appreciate any further discussion and would be happy to clarify any unresolved questions. If our responses have addressed your concerns, we would also be grateful if you would consider increasing your scores accordingly.

Thank you again for your valuable time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.