acceptodds
Under review as a conference paper at ICLR 2027

Complete Donor-Set Guidance for Frozen Diffusion Docking Models

Abstract

Coordination information can guide a pretrained docking model, but its value depends on how specifically donor atoms are identified and how that information is used. We study a common Complete-Identity Contract Guidance (CICG) interface that applies a donor set during the final two denoising steps without retraining the generator or its confidence model. Historical Contract Transfer (HCT) supplies reference-mapped atom identities; trajectory-specific nearest-donor policies instead use historical donor counts, with or without elemental quotas, to select atoms from intermediate poses. No reference ligand coordinates are used as pose targets. We evaluate these policies in a pooled follow-up analysis of 86 assessable queries in 33 structural groups. Across two seeds, historical guidance raises joint pose, coordination, and validity success from 21.51% to 36.63%: +15.12 percentage points (95% group-bootstrap interval, 8.00-24.63). Composition-matched nearest-donor guidance reaches 33.14%, leaving the historical advantage unresolved at +3.49 points [-0.42, 8.87]. A retrospective audit using native donor labels finds that all 57 successes of the composition-matched policy occur on selected trajectories whose assigned sets match the native donor set, while 50 other exact assignments fail. Under the same donor-aware selector, a post-hoc comparison finds an additional 7.56-point gain [1.11, 14.04] from coordinate guidance, accompanied by lower ligand accuracy. These results support the usefulness of several donor-information policies within the evaluated scope, rather than a general advantage of fixed historical membership.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular docking, diffusion models, metalloproteins, protein–ligand interactions, metal coordination, test-time conditioning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Eax3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how donor-atom information can guide pretrained diffusion docking models without retraining. The authors introduce CICG, which applies metal–donor distance corrections during the last two denoising steps of SurfDock. They compare historical donor identities with simpler nearest-donor selection strategies, as well as donor-aware reranking. The experiments show improved coordination-aware docking success, although the benefit of transferring exact historical donor identities over simpler alternatives remains uncertain.

Critical strengths
  1. Incorporating metal coordination constraints into a frozen diffusion model is a useful idea. I like that the method does not require retraining and can operate directly on translation, rotation, and torsional coordinates. The comparisons are also reasonably well controlled, with matched sampling conditions and the same correction procedure across different donor-selection policies.

  2. Historical guidance improves joint Top-1 success from 21.51% to 36.63% (Table 1), which is a meaningful improvement. I also appreciate the comparisons with nearest-donor selection and donor-aware reranking. The authors are fairly transparent about the limitations, including incorrect donor assignments, reduced ligand accuracy, and increased clashes. These analyses make the results more informative.

Critical weaknesses
  1. My main concern is whether the complete historical donor set is actually necessary. Composition-matched Nearest-K already achieves 33.14% joint success, compared with 36.63% for historical guidance. The difference is only 3.49 percentage points, with a confidence interval that includes zero (Table 2). Even the count-only strategy recovers much of the improvement. Donor-aware reranking also improves success without modifying coordinates, suggesting that part of the gain comes from better pose selection. These are useful findings, but they make the specific contribution of complete historical donor identities less convincing. The paper demonstrates that donor information helps, but not clearly that transferring exact donor membership is substantially better than simpler alternatives. I would also like to see a comparison with a simple post-docking coordinate correction using the same donor information. This would help establish whether applying the correction during diffusion offers a real advantage.

  2. The improvement in joint success is encouraging, but Table 5 shows some concerning trade-offs. Historical guidance increases the candidate clash rate from 16.83% to 27.22%, while geometry validity decreases from 83.17% to 72.78%. Ligand-only Top-1 success also drops from 106 to 101 out of 172. This is important because correct metal coordination alone does not guarantee a physically reasonable binding pose. The current correction mainly enforces metal–donor distances without explicitly considering receptor clashes, coordination angles, or ligand strain during optimization. The applicability of historical transfer is another limitation. Only 62 of the original 304 queries receive historical donor assignments, and the method requires highly similar receptor sites with available reference complexes. I am therefore not sure how much the reported improvement would help in practical docking campaigns involving new targets or ligands without suitable historical references.

Questions for authors
  1. Could the authors compare CICG with a simpler post-docking correction using the same donor sets? Does intervention during diffusion provide an advantage over correcting completed poses?

  2. Could the authors incorporate basic clash or ligand-strain penalties into the correction and evaluate whether the coordination improvement can be maintained without reducing overall pose quality?

  3. Given the limited historical coverage, what is the expected improvement across the full original query set when queries without reliable donor assignments fall back to standard docking? How well would this approach generalize to less familiar protein families?

6 comments
AuthorsOpenReview ↗

Coordinate correction versus selection

Thank you for asking us to separate donor information, candidate generation, and final selection. We ran a bank-by-selector comparison that addresses this directly, using the same 86 assessable queries, two seeds and 33 structural groups.

What does changing the coordinate bank contribute? (W1)

Coordinate bankFrozen-confidence selectionIdentical donor-aware selector
No correction (C0)37/17251/172
Historical donor correction (HCT)63/17264/172

With the selector held fixed, CICG-corrected coordinates increase selected joint success from 51/172 to 64/172: +7.56 percentage points (pp), paired 95% CI [1.11, 14.04]. There are 21 improved and eight worsened query–seed banks. This is our most direct evidence that the method contributes useful candidate geometry beyond donor-aware ranking. Selection of unmodified poses helps, but it does not recover all the successes obtained from the corrected bank. (see §5.4, Figure 2, Table 3.)

The same comparison also shows how the gain arises. Banks containing at least one joint-successful pose increase from 91 to 96, while cases where such a pose is available but not selected decrease from 40 to 32. These are related descriptive changes rather than separately identified causal contributions. Under the same selector, ligand-only success falls from 107/172 to 99/172 and selected fatal clashes increase from 12 to 21. A different radial/clash selector gives 48/172 for C0 and 38/172 for HCT, confirming that the bank effect depends on the selection rule. (Appendix B.9, Table 27; G.1, Table 37.)

How specific must the donor information be?

SurfDock donor inputJoint Top-1
None37/172
Historical donor count56/172
Count + element composition57/172
Historical atom membership63/172

The HCT−composition increment is +3.49 pp [−0.42, 8.93] on SurfDock, +0.58 [−2.03, 3.52] on DiffDock-L, and +3.20 [−0.85, 9.15] across four SurfDock seeds. The extra gain from exact historical identity is therefore uncertain. The supported result is broader and useful: the same inference-time operator improves frozen docking with several donor-information policies, and its coordinate-bank contribution is measurable even when ranking is controlled. Historical references offer one practical source of a complete donor set; the interface can also work from donor count and composition when those quantities are available. (see §5.2, Table 2; Appendix G.7.)

Is correction during diffusion necessary? (Q1)

HCT policy on the 40-pose bankJoint Top-1
Last two sampling updates63/172
Two correction calls after sampling, then reselection65/172
Correct only C0's preselected Top-141/172

Interleaved−terminal is −1.16 pp [−3.19, 1.15] using the same donor sets and maximum correction budget. Correcting all 40 rather than only the original Top-1 adds +13.95 pp [6.35, 22.88], with 29 gains and five losses. Thus the correction is effective both inside and after sampling under these conditions, while full-bank processing and reselection matter substantially. (see §5.6; Appendix G.2, Table 38; G.9, Table 49.)

AuthorsOpenReview ↗

Physical reliability and applicability

We also tested where donor-conditioned correction helps and where its present objective remains insufficient. The correction reduces metal–donor distance residuals, but its radial objective does not directly penalize receptor overlap or ligand strain. The physical and coverage studies make this distinction visible.

Physical-quality evidence (W2, Q2)

Original 40-pose-bank measureC0HCT
All-pose fatal clashes16.83%27.22%
All-pose protein-overlap compliance71.66%45.84%
Confidence-selected ligand success106/172101/172
Joint ∩ initial nine-check battery33/17254/172

An overlap-acceptance gate increases HCT all-pose overlap compliance to 74.04%, but brings joint Top-1 back to 37/172 (from 63/172). This negative control identifies a real trade-off: simply rejecting overlapping correction steps can remove successful coordination fixes. (Appendix G.12.)

We then applied the same output-only 1.2-Å clash filter to C0 and HCT:

Selected measure / 172C0 before → afterHCT before → after
Joint Top-137 → 3863 → 64
Fatal clashes11 → 422 → 4
Nine-check pass, Q148 → 150127 → 135
Joint ∩ Q35 → 3658 → 58

Under the pinned PoseBusters 0.6.3 environment, filtered HCT retains a +12.79-pp [5.48, 22.22] advantage over filtered C0 for joint∩Q; the HCT-versus-C0 Q-only result remains lower by −8.72 pp [−14.02, −4.12]. The filter changes only which saved pose is returned, and joint success already excludes fatal clashes. Its contribution is to reduce avoidable selected collisions without disturbing the corrected candidate bank. The initial checker run (33/54 joint∩Q) and the pinned run (35/58 before filtering) use different environments, which we report separately. Applying the filter without retuning to seeds 44/45 likewise reduces HCT selected clashes from 21 to 3. (see §5.5; Appendix G.3, Table 39; G.13, Table 54.)

Full-roster accounting and unseen families (W2, Q3)

MeasureResult
Historical contract coverage in original roster62/304 (20.39%)
Covered 62 queries, C0 vs HCT, two seeds23/124 → 38/124
Whole-roster effect with identical C0 fallback+15/608 = +2.47 pp
Separate family/site-separated cohort: transferable contracts0/26

The +2.47-pp figure is accounting under the declared unchanged-C0 fallback, not a new 304-query docking run. The separately frozen 26-query cohort lacked suitable transferable historical references, so it tests the availability of HCT input rather than HCT accuracy on new families. When no eligible reference exists, the policy abstains and retains C0. The 86-query efficacy estimates apply to the HCT-covered, assessable group; we also report the 100-query failure-counted sensitivity for the reporting-stage restriction. (see §§4 and 6; Appendix A.2, B.1, E, G.12.)

Taken together, the experiments establish a specific role for CICG: given a supported donor assignment, it adds useful corrected poses to a frozen docker's candidate bank, with a benefit beyond donor-aware selection alone. Improving reference coverage and enforcing coordination, receptor compatibility and strain together are the next steps. We would welcome any further concern about the evidence or population definitions presented here.

Manuscript revision. We have incorporated the additional experiments and clarifications into the revised manuscript, while retaining the core method, evaluation cohort, and joint-success criterion. Key updates are highlighted in blue, with detailed results and protocols in Appendix G.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer Eax3,

Thank you again for your helpful comments. We would like to briefly follow up on our responses addressing the bank-by-selector controls, post-docking correction, physical-quality trade-offs, and transfer coverage.

When you have a chance, we would appreciate your feedback on whether any concerns remain about these analyses or their interpretation. We would be glad to provide further clarification or supporting details and respond promptly to any follow-up questions.

Thank you for taking the time to consider our responses.

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

This paper proposes Complete-Identity Contract Guidance (CICG), an inference-time framework for incorporating metal-coordination information into pretrained diffusion docking models without retraining. Historical Contract Transfer (HCT) transfers donor atom identities from structurally compatible complexes, while trajectory-specific nearest-donor policies select atoms using historical donor counts and elemental composition. The framework applies geometric corrections during the final denoising steps and evaluates the effects on coordination-aware docking success, ligand accuracy, and geometric validity.

Critical strengths

The paper addresses a meaningful limitation of general-purpose docking models: accurate ligand placement does not necessarily imply correct metal coordination. The proposed inference-time guidance is relatively simple and does not require retraining the underlying diffusion model, making it potentially practical.

The authors conduct controlled comparisons between historical donor transfer and trajectory-specific nearest-donor policies. The results suggest that donor information can improve coordination-aware docking success over an unguided baseline. The paper also examines failure cases and distinguishes donor assignment accuracy from geometric realization.

The evaluation considers multiple aspects of docking quality, including ligand accuracy, coordination recovery, and receptor clashes. This provides a more informative assessment than relying on pose success alone.

Critical weaknesses

The evidence for the advantage of historical donor identities remains limited. Although historical guidance improves joint Top-1 success over the unguided baseline, its advantage over composition-matched nearest-donor selection is not statistically established. The results therefore support the usefulness of donor guidance more strongly than the claimed value of transferring exact donor identities.

The evaluation has limited scope. The pooled analysis includes only 86 assessable queries and two random seeds, while historical transfer covers 62 of the initial 304 queries. The additional cohort broadens the evaluation, but does not establish generalization across a wide range of metalloproteins. Moreover, the main analyses include post-hoc comparisons and a reporting-stage restriction of the evaluation population, which complicate the interpretation of the overall evidence.

The guidance mechanism introduces a notable trade-off between coordination success and physical validity. The all-pose clash rate increases from 16.83% without guidance to 27.22% with historical guidance, while overall geometric validity decreases. Since the correction objective focuses on radial metal–donor distances without explicitly penalizing receptor clashes, the practical reliability of the generated poses remains a concern. Comparisons with established metal-specific docking methods are also missing.

Questions for authors

Can the authors better establish the advantage of historical donor identities over composition-matched nearest-donor selection? The current confidence interval includes zero, making the central claim about the benefit of complete donor-set transfer inconclusive.

Can the authors evaluate CICG on a broader and more diverse set of metalloproteins, with additional random seeds and an independent pretrained docking backbone? The current evidence does not fully establish generalizability beyond the evaluated setting.

Can the authors reduce the increased clash rate and compare against existing metal-specific docking methods under matched evaluation protocols? Such experiments would help determine whether the improvement in joint success translates into a practically reliable docking method.

5 comments
AuthorsOpenReview ↗

Donor information and the evaluation population

Thank you for pointing out that we must separate the value of donor guidance from the additional value of exact historical atom identities. We added the corresponding controls and have made this distinction central to our interpretation.

What information produces the gain? (W1, Q1)

Donor conditionSurfDockDiffDock-L
No guidance37/17243/172
Historical donor count only56/172—
Historical count + element composition57/17259/172
Historical donor identities (HCT)63/17260/172

On SurfDock, HCT improves joint Top-1 over C0 by +15.12 pp [8.00, 24.63]; on DiffDock-L, the gain is +9.88 pp [3.09, 17.16]. Importantly, we obtain substantial improvements without retraining either generator or confidence model. Both models can use donor information supplied at inference time through the same kinematic correction interface.

You are right that composition matching recovers much of this benefit. HCT−composition is +3.49 pp [−0.42, 8.93] on SurfDock and +0.58 pp [−2.03, 3.52] on DiffDock-L. Across four SurfDock seeds it is +3.20 pp [−0.85, 9.15]. These comparisons do not establish a reliable additional gain from historical atom identity. They do, however, identify an important practical feature: the interface works with different levels of donor information, including count/composition when those quantities are known but full atom membership is uncertain. Historical transfer provides one concrete way to supply a complete set, rather than the only possible input policy. (see §5.2, Table 2; Appendix G.7, Tables 47–48.)

We also tried a single composition-matched set fixed across trajectories, chosen from seed-42 sample 0 before correction. It yields 35/172, versus 57/172 for trajectory-local selection. Excluding that defining sample from the final ranking leaves the joint counts unchanged. This rules out a simple self-inclusion explanation for the reported numbers, although set-selection quality and trajectory adaptation remain coupled. (Appendices G.8 and G.14, Tables 48 and 55.)

Robustness and evaluation scope (W2, Q2)

SurfDock seedsC0HCTHCT−C0 [95% CI], pp
42/4337/17263/172+15.12 [8.00, 24.63]
44/4528/17255/172+15.70 [6.67, 25.71]
All four65/344118/344+15.41 [8.05, 24.37]

The 86-query analysis contains 33 structural groups and was restricted to HCT-covered, assessable inputs. The 100→86 assessability restriction was adopted during reporting after seeing earlier results. Counting all 14 unsupported inputs as failures in the full-100 sensitivity still gives HCT−C0 +13.00 pp [6.70, 21.25]. These are repeated tests on the same biological population, and we present the population restriction explicitly. (see §4; Appendix B.1.)

We separately froze a family/site-separated 26-query cohort to test applicability. The existing transfer rules find zero eligible contracts (0/26), primarily because suitable historical family/site references are absent. This tells us where the current reference-based policy cannot yet operate; it cannot estimate the accuracy of HCT correction on unseen families. The next test is reference-supported analogue queries selected independently, with a frozen transfer/fallback rule. (Appendix G.12.)

AuthorsOpenReview ↗

Physical quality and metal-aware baselines

Your concern about physical reliability led us to evaluate both the returned pose and the full candidate bank, with the same rules applied to C0 and HCT. This distinction matters because local metal correction can improve coordination while introducing clashes elsewhere in the ligand–protein complex.

Physical checks and the tested safeguards (W3, Q3)

In the original 40-pose banks, HCT increases joint Top-1 from 37/172 to 63/172, but all-pose protein-overlap compliance falls from 71.66% to 45.84%. An overlap-acceptance gate raises HCT overlap compliance to 74.04%, at the cost of reducing joint success to 37/172. We report this unsuccessful safeguard because it shows why receptor compatibility must be treated jointly with donor geometry.

We next applied an identical, output-only 1.2-Å clash filter to C0 and HCT. It demotes poses with severe receptor contacts while preserving the original confidence order otherwise; poses are left unchanged.

Selected outcome / 172C0 before → afterHCT before → after
Joint Top-137 → 3863 → 64
Fatal receptor clashes11 → 422 → 4
Nine-check conjunction, Q148 → 150127 → 135
Joint ∩ Q35 → 3658 → 58

After filtering, HCT exceeds filtered C0 by +15.12 pp [7.39, 25.00] on joint success and +12.79 pp [5.48, 22.22] on joint∩Q. This is useful as an inexpensive output safeguard: selected fatal clashes decrease substantially while the joint gain remains. The filter does not improve the generated bank's geometry, and HCT remains below C0 on Q alone (135 versus 150/172). The initial physical-check run gave 33/54 joint∩Q successes for C0/HCT; the later pinned PoseBusters 0.6.3 run gives 35/58 before filtering. That difference comes from the checker environment, not from the filter. (see §5.5; Appendix G.3, Table 39; G.12–G.13, Table 54.)

Without retuning, the same rule on seeds 44/45 reduces HCT selected clashes from 21 to 3; joint changes 55→56/172 and joint∩Q stays 53/172. This supports the filter's selected-clash effect under additional sampling of the same cohort. The next correction design needs to consider donor geometry, protein overlap and ligand strain together.

Specialized docking controls

90-s warm-query endpointSurfDock + HCTMetalloDock, native 10 Å
Joint Top-165/17272/172
Joint Top-40102/17275/172
Joint ∩ Q63/1728/172

MetalloDock−HCT Top-1 is +4.07 pp [−19.59, 20.27]; HCT−MetalloDock Top-40 is +15.70 pp [−3.66, 39.06]. Neither comparison resolves a general ranking. HCT−MetalloDock joint∩Q is +31.98 pp [19.34, 47.33] for the tested raw-output paths, with optional MetalloDock refinement disabled. MetalloDock has higher all-candidate protein-overlap compliance (94.58% versus 45.74%), while its raw poses fail more ligand-internal checks. The models retain different receptor preparations and ranking policies. Our 8-Å MetalloDock input test is reported separately as a sensitivity experiment. (see Table 4; Appendix G.11, Tables 52–53; G.15.)

On the matched seed-42 template comparison, HCT achieves 31/86 against TEMPL-200's 17/86; the Zn-only AutoDock4Zn-potential/Vina run yields 0/44 Top-1 and 22/44 Top-40 (HCT: 22/44 and 32/44). Search budgets and method-specific inputs are reported alongside each comparison (Appendix G.10, Tables 50–51).

These results make the contribution and its physical boundary more precise: donor-conditioned correction improves selected joint docking on two frozen generators, while receptor compatibility needs additional optimization. Please let us know if any of these comparisons leave a specific concern unaddressed.

Manuscript revision. We have incorporated the additional experiments and clarifications into the revised manuscript, while retaining the core method, evaluation cohort, and joint-success criterion. Key updates are highlighted in blue, with detailed results and protocols in Appendix G.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer aRUT,

Thank you again for your thoughtful review. We would like to follow up on our responses concerning donor-information controls, robustness across seeds and backbones, applicability, and physical-quality comparisons under the same filtering rules.

We would appreciate the opportunity to clarify any remaining questions about the evaluation conditions or interpretation of the results. When time permits, could you please take a look at our responses and let us know if further explanation would be helpful? We remain available to respond promptly.

Thank you again for your time and consideration.

Reviewer_EzwTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Complete-Identity Contract Guidance (CICG), an inference-time method for incorporating ligand donor-set information into a frozen diffusion docking model. The method applies a donor-dependent coordinate correction during the final two denoising steps of SurfDock, using translation, rotation, and torsional updates to reduce metal–donor distance residuals. The paper compares historical donor transfer, nearest-donor selection, composition-matched nearest-donor selection, and a native-donor diagnostic oracle.

Critical strengths
  1. The paper addresses an important problem in metalloprotein docking: a pose may fit the binding pocket while assigning the wrong ligand atoms to metal coordination.

  2. The donor-set construction policies are clearly defined, and the paper distinguishes historical transfer, trajectory-specific nearest-donor selection, and native-donor diagnostic guidance.

Critical weaknesses

The main concern is that the methodological novelty and experimental evidence do not sufficiently support the paper’s central claims.

  1. the core CICG operation appears to be a local coordinate correction applied near the end of the docking trajectory. In principle, a similar donor-distance correction could be applied after the docking process has finished. Both procedures would modify an already generated pose without retraining the model. Without this comparison, it is unclear whether the timing of the correction provides a meaningful algorithmic advantage or whether the method is essentially a post-docking refinement procedure inserted before the final denoising steps.

  2. the “no retraining” advantage is not sufficiently distinctive. Many docking post-processing and refinement methods can be applied without retraining the generator. The paper needs to show that applying the correction during sampling produces better results than applying the same donor constraint after sampling, under matched candidate, optimization, and computational budgets.

  3. The choice of SurfDock as the main backbone is insufficiently justified. The manuscript does not establish that SurfDock is a current state-of-the-art model for metalloprotein docking, nor does it provide a direct comparison with stronger metal-aware docking methods, template-based methods, or other recent diffusion docking backbones. The native-donor experiment with DiffDock-L does not resolve this issue because it does not establish historical donor transfer on that backbone.

Questions for authors
  1. What are the results when the same donor-distance correction is applied after all docking steps have finished? Please compare final-two-step correction, post-docking correction, and no guidance under identical candidate and optimization budgets.

  2. If post-docking correction is applied to all 40 candidates followed by confidence reranking, does it obtain similar performance? If only the Top-1 pose is refined, how much performance is lost because candidate selection occurs before refinement?

  3. What is the specific advantage of performing the correction during the last two denoising steps rather than after sampling? Does the correction alter the subsequent diffusion score, pose diversity, or confidence ranking in a measurable way?

  4. Why is SurfDock an appropriate primary backbone? Please compare against stronger recent diffusion docking models, metal-aware docking methods, and constraint or template-based baselines.

5 comments
AuthorsOpenReview ↗

Correction timing and what CICG adds

Thank you for suggesting a direct post-docking control. It helped us clarify the main contribution: CICG lets us impose a supplied ligand donor set on a pretrained docking model through translation, rotation, and torsion updates, without retraining the generator or copying a reference ligand pose. We tested whether this correction has to occur during diffusion and whether correcting only the initially preferred pose is sufficient.

Timing and candidate correction (W1–W2, Q1–Q2)

SurfDock settingJoint Top-1
No correction (C0)37/172
CICG at the last two sampling steps63/172
CICG after sampling, all 40 poses65/172
After sampling, only the preselected Top-141/172

The two 40-pose correction conditions use the same historical donor assignments and two correction calls, each capped at five relinearizations. Terminal minus interleaved correction gives +1.16 percentage points (pp); the paired 95% CI for interleaved minus terminal is [−3.19, 1.15] pp. The planned +3-pp interleaving advantage was not observed. Importantly, the same donor-conditioned correction remains effective when applied after sampling. This gives the method a useful implementation option across frozen models, even though timing-specific superiority is unsupported.

Correcting all 40 poses and reselecting gives 65/172, compared with 41/172 when only C0's original confidence Top-1 is corrected: +13.95 pp [6.35, 22.88], with 29 improvements and five regressions. This is an actual 172-bank experiment, independently checked against the saved sample IDs. It demonstrates the value of correcting a candidate bank rather than committing to its original Top-1. The contrast includes additional corrections and reselection; both settings first generate 40 poses. (see §5.6; Appendix G.2, Table 38; G.9, Table 49.)

Mechanistic evidence (Q3)

Output comparison: interleaved vs terminalMeasured result
Selected sample changes31/172 banks
Median confidence-rank Spearman correlation0.9594
Median corresponding-pose RMSD0.208 Å
Change in median within-bank pairwise RMSD+0.0364 Å [0.0226, 0.0484]

These output diagnostics show that placement of the correction can affect both coordinates and the selected pose. They do not establish a beneficial change to the diffusion score field; intermediate score tensors were not saved. We have kept that distinction explicit. (Appendix G.14, Table 55.)

We also tested whether our coordinate correction contributes beyond applying a donor-aware ranker to unmodified poses: with the same selector, joint success increases from 51/172 on C0 poses to 64/172 on CICG-corrected poses (+7.56 pp [1.11, 14.04]). This is the strongest controlled evidence that CICG adds useful candidate coordinates beyond donor-informed selection alone (revised §5.4, Table 3).

AuthorsOpenReview ↗

Transfer across backbones and docking baselines

To test whether the method depends on SurfDock, we ran actual HCT-guided correction on independently pretrained DiffDock-L. The two backbones share the same 86 biological queries but have different pretrained weights and sampling implementations.

Frozen backboneC0HCT + CICGPaired gain, pp [95% CI]
SurfDock37/17263/172+15.12 [8.00, 24.63]
DiffDock-L43/17260/172+9.88 [3.09, 17.16]

This gives us evidence for a portable inference-time correction interface on two pretrained generators, with no weight updates in either case. The experiments use the same cohort, so the second backbone tests model dependence rather than biological out-of-family generalization. (see Table 2; Appendix G.4, Table 45.)

We also ran comparisons using historical templates and Zn-specific docking:

Common evaluation populationCICG/HCTComparator
86 queries, seed 42, joint Top-131/86TEMPL-200: 17/86
22 Zn queries × 2 seeds, joint Top-122/44AutoDock4Zn potential + Vina: 0/44
Same Zn subset, joint Top-4032/44AutoDock4Zn potential + Vina: 22/44

HCT exceeds the evaluated TEMPL-200 configuration by +16.28 pp [1.28, 30.43]. TEMPL uses up to 200 conformers, whereas CICG uses 40 diffusion candidates; this is a comparison of the specified pipelines with different search procedures. AutoDock4Zn was evaluated with Vina 1.2.7 search and its own scoring, including the two preparation-compatibility failures in the denominator. Its Top-40/Top-1 gap highlights ranking sensitivity in this configuration. (Appendix G.10, Tables 50–51.)

MetalloDock with its native receptor preparation (W3, Q4)

We repeated MetalloDock using its original 10-Å receptor crop, on all 86 queries and both seeds. This matters: its earlier shared-8-Å input substantially changed its receptor graph and performance, so we report that setting as an input-sensitivity test, alongside the native configuration.

Joint endpointSurfDock HCTMetalloDock 10 Å
One generated pose, Top-115/17268/172
40 generated poses, Top-163/17268/172
90-s warm-query, Top-165/17272/172
90-s warm-query, Top-40102/17275/172
90-s warm-query, Joint ∩ nine physical checks63/1728/172

MetalloDock has a clear advantage in the one-pose operating comparison (+30.81 pp [12.20, 44.58]); the 40-pose Top-1 and 90-s Top-1 differences have intervals spanning zero. At 90 s, the HCT−MetalloDock Top-40 difference is +15.70 pp [−3.66, 39.06]. The joint-and-nine-check contrast favors HCT by +31.98 pp [19.34, 47.33], but here MetalloDock's optional refinement was disabled. Receptor preparation, model architecture and information access also differ. These results characterize the tested operating paths rather than establish an across-the-board model ranking. (see Table 4; Appendix G.11, Tables 52–53; G.15.)

Our case for CICG is its complementary capability: a general-purpose frozen docker can accept externally supplied metal donor information and improve joint docking success, with a measurable coordinate-bank contribution beyond selection alone. We would welcome any remaining questions about these comparisons or the mechanism.

Manuscript revision. We have incorporated the additional experiments and clarifications into the revised manuscript, while retaining the core method, evaluation cohort, and joint-success criterion. Key updates are highlighted in blue, with detailed results and protocols in Appendix G.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer EzwT,

Thank you again for your time and constructive comments. We would like to briefly follow up on our responses regarding correction timing, candidate correction and reselection, and model-native baseline comparisons.

When you have an opportunity, we would be grateful if you could review our responses and let us know whether any aspect of these comparisons requires further clarification. We would be happy to address any follow-up questions promptly and provide additional supporting details.

Thank you for your continued 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.