acceptodds
Under review as a conference paper at ICLR 2027

Interaction-Aligned Surface Representations for Generalizable Blind Docking

Abstract

Recent advances in machine learning have substantially improved protein–ligand docking, yet blind docking still generalizes poorly to unfamiliar receptors and binding environments. We argue that part of this difficulty arises from a mismatch between structural representation and interaction-level invariance: global structural novelty need not imply interaction novelty, as substantial receptor variation can leave the geometry and chemistry of relevant binding interfaces largely unchanged. Under limited supervision, however, Euclidean geometric encoders must infer such invariances implicitly and may entangle transferable interaction patterns with receptor-specific structural variation. We introduce SADock, a geometric diffusion framework that addresses this mismatch by explicitly representing both proteins and ligands through molecular surfaces, the natural boundaries on which molecular recognition occurs. Surface geometry and physicochemical fields expose interaction-facing patterns more directly, while a hierarchical heterogeneous graph integrates surface representations with residue- and atom-level context across multiple spatial scales for receptor-wide search and fine-grained pose modeling. We further establish conditions under which boundary-preserving structural changes retain binding-basin preferences, providing a theoretical characterization of why surface-aligned representations can remain transferable under structural variation. Across multiple benchmarks, SADock achieves higher docking success rates and lower pose RMSD. On challenging DockGen, our method improves Top-1/Top-5 docking success rates from 22.8%/29.6% for the strongest evaluated baseline to 27.0%/38.6%, despite using less data augmentation and without inference tuning tricks. Further analysis reveals substantial gains in pocket localization, a key bottleneck in generalizable blind docking, supporting surface-aware interaction modeling as a principled inductive bias under structural distribution shift.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Blind Docking, Surface, Diffusion Model, Out-of-Distribution Generalization

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

This paper introduces SADock, a surface-aware diffusion framework for blind protein-ligand docking. Motivated by the observation that globally distinct receptors share similar local interaction geometry and chemistry. Extending DiffDock, SADock combines surface representations with residue- and atom-level features via hierarchical message passing and time-conditioned modulation. The paper reports strong docking performance on PDBBind and DockGen, with further analysis pointing to improved pocket localization as the key driver of these gains.

Critical strengths

This paper tackles a key bottleneck with a clear and well-reasoned hypothesis. The observation that global receptor dissimilarity does not imply a novel interaction interface provides a solid motivation for surface-aware blind docking.

The empirical claims are supported by a strong set of complementary experiments.

Critical weaknesses

The experiments do not yet fully isolate the proposed representation-level explanation. Introducing the surface branch also brings in extra geometric and physicochemical descriptors, more computational pathways, and potentially higher model capacity.

While the ablation study confirms that the surface-augmented system is useful, it doesn't quite distinguish whether the gains come from organizing information on molecular surfaces or simply from feeding the model richer inputs.

While the theoretical analysis aligns with this interpretation, it relies on energy-perturbation and model-sensitivity assumptions that haven't quite been verified for the evaluated receptors.

A few reporting and implementation details need clarification to make the core results fully interpretable. For instance, Figure 3 is described as using the five highest-confidence candidates, meaning its RMSD success rate at 2 Å should match the corresponding Top-5 result in Table 1 under the same configuration. However, the plotted value appears noticeably lower, so these numbers should be reconciled.

The construction and updating of ligand surface geometry and descriptors as torsions change could use a clearer operational description, especially since this mechanism sits right at the heart of the proposed representation.

Questions for authors

Could you reconcile the data between Figure 3 and Table 1? Given that Figure 3 uses the top-five confidence candidates, its RMSD success rate at 2 Å should match the corresponding Top-5 figures in Table 1. Clarifying whether the exact same checkpoints, candidate budgets, and aggregation rules were used would clear up any apparent discrepancies.

How are ligand surface points, normals, geodesic neighborhoods, and physicochemical descriptors handled when torsional changes happen during training and inference?

Are the Euclidean-only and surface-augmented models truly comparable in terms of capacity and underlying chemical information?

Section 3.1 refers to results on 428 PoseBusters complexes, but those numbers didn't seem to show up in the current manuscript. Could you point us to where they are or correct the statement?

2 comments
AuthorsOpenReview ↗

Response to Reviewer zqJG: W1/Q3 & W2 — Controlled Model Capacity, Surface Geometry, and Hierarchical Modeling

We thank the reviewer for recognizing that SADock “tackles a key bottleneck with a clear and well-reasoned hypothesis” and that our empirical claims are supported by “a strong set of complementary experiments.” We address your concerns point by point below.

W1 & Q3: Representation-level contribution and controlled model capacity.

Thanks for raising this important concern. We would like to clarify that molecular surfaces are derived from the same molecular coordinates and chemical identities used by Euclidean representations, without introducing additional external information. By explicitly organizing geometry and physicochemical properties around the molecular boundary, surface representations provide an interaction-aligned inductive bias that facilitates transferable pocket localization under structural distribution shift.

Regarding model capacity, we deliberately match SADock/SADock-L (~20M/30M parameters) with DiffDock/DiffDock-L, respectively. More importantly, our Euclidean-only ablation is retrained with the same atom/residue inputs, training data, optimization budget, and parameter count (~20M) as SADock, rather than using the released DiffDock checkpoint.

Under these controlled settings, introducing surface representations improves DockGen Top-5 success from 15.3% to 28.6% and oracle success from 16.4% to 36.5%. Further component ablations confirm the contributions of geodesic encoding and surface–Euclidean interactions. These results support the benefits of surface-aligned representation learning beyond increased model capacity or additional molecular information.

W2: Contribution of surface geometry and hierarchical modeling.

To distinguish the benefits of surface representation learning from simply introducing richer descriptors, we conducted finer-grained component ablations, using the same pretrained DiffDock confidence model for consistent ranking.

(1) Hierarchical surface modeling. While retaining the same surface descriptors, removing F2C/C2F reduces DockGen Oracle success from 36.5% to 32.3%, demonstrating the benefit of hierarchical surface communication.

(2) Intrinsic surface geometry. Replacing geodesic distances with Euclidean distances further reduces Oracle success to 30.7, highlighting the importance of modeling intrinsic surface geometry beyond conventional Euclidean proximity.

These results demonstrate that SADock's improvements depend not only on richer surface descriptors, but also on how interaction-relevant information is geometrically organized and propagated. We have incorporated these ablations into the revised manuscript.

Table R17: Component ablation results.

ModelPDBBind OraclePDBBind Top1PDBBind Top5DockGen OracleDockGen Top1DockGen Top5
SADock61.744.656.736.515.328.6
w/o Surface (Euclidean-only)46.035.840.816.48.515.3
w/o Geodesic Distance60.641.954.330.714.324.9
w/o F2C/C2F59.542.253.432.318.524.3
w/o Ligand Surface56.837.549.031.814.823.3
w/o Hetero-MP49.335.544.917.59.013.8
AuthorsOpenReview ↗

Response to Reviewer zqJG: W3, W4/Q1, W5/Q2 & Q4 — Theoretical Assumptions, Reporting Consistency, Ligand Surfaces, and PoseBusters

W3: Empirical validation of theoretical assumptions.

Thanks for this constructive comment. We agree that our theoretical analysis establishes sufficient conditions for transferability. To further substantiate the underlying assumptions, we conducted two additional analyses.

(1) Physical-level assessment (Eq. 7). We performed 50 geometric perturbations on each of 20 DockGen complexes, obtaining 989 valid comparisons of Lennard–Jones and Coulomb interaction energies under fixed ligand configurations. We calibrated per-complex empirical bounds with at least 90% coverage:

δUQ≤KℓδS+εℓ,\delta U_{\mathcal Q}\leq K_\ell\delta_S+\varepsilon_\ell,

yielding Kℓ∈[0.1007,5.1053]K_\ell\in[0.1007,5.1053] and εℓ∈[0,0.03223]\varepsilon_\ell\in[0,0.03223] kcal/mol. These results quantitatively characterize the relationship between surface discrepancy and interaction-energy variations, providing an initial empirical assessment of Eq. 7. See implementation details in Reply W2-1 to Reviewer W4gW.

(2) Model-level stability (Eq. 8). On 300 PDBBind complexes, we quantified the Euclidean residual sensitivity under controlled perturbations while keeping surface inputs fixed. At t=1, perturbations of 0.01/0.1 Å produce only 0.55%/1.84% relative score variations, with sensitivity increasing as denoising progresses (Pearson r=0.91 with (1-t)). This supports the intended reduced Euclidean sensitivity during early pocket localization.

Together, these results provide empirical evidence for the physical and model-level stability assumptions underlying our theoretical formulation, while the theoretical guarantees remain conditional on the stated assumptions. We have incorporated these analyses and clarified their scope in the revised manuscript.

W4 & Q1: Consistency between Figure 3 and Table 1.

Thanks for pointing out this inconsistency. Table 1 used the EMA checkpoint (SADock-400K-EMA), whereas Figure 3 mistakenly used its non-EMA counterpart (SADock-400K). We have corrected Figure 3 to ensure consistency with Table 1.

W5 & Q2: Dynamic ligand surface construction during denoising.

During training and inference, the ligand surface is reconstructed using NanoShaper after each conformational update, before the next score evaluation. Surface points, normals, areas, geodesic neighborhoods, and atom–surface connections are updated accordingly. Atomic chemical attributes remain unchanged and are incorporated into surface representations through Hetero-MP; no separate surface hydrophobicity or electrostatic descriptors are used. Ligand surface generation and graph construction take only ~20 ms per update, and CPU preprocessing can overlap with GPU inference of other batches, limiting the additional overhead.

Q4: Missing PoseBusters results.

Thanks for pointing out this omission. We have supplemented the evaluation on all 428 PoseBusters complexes, using 40 candidates per target.

Table R18: PoseBusters results.

MethodOracleTop1Top5
DiffDock51.4035.5144.63
DiffDock-L61.9250.9356.07
SADock64.4942.5256.07
SADock-L75.4753.5066.59

Notably, SADock-L outperforms DiffDock-L across all three metrics, improving Oracle, Top-1, and Top-5 success by 13.55%, 2.57%, and 10.52% (absolute), respectively. We have added these results to the revised manuscript.

Reviewer_8TQXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SADock is a geometric diffusion model for blind protein–ligand docking. It represents both protein and ligand with molecular surfaces and uses a dual-branch network (SANet) that combines surface features with euclidean geometry. The main claim is that surface-aligned representations help under structural distribution shift. On DockGen, SADock-L improves top-1 / top-5 success over DiffDock-L.

Critical strengths

The evaluation focuses on a hard blind-docking setting and reports solid improvements over DiffDock / DiffDock-L.

Critical weaknesses
  1. The overall framework largely follows DiffDock’s setup and uses surface information in a manner similar to SurfDock. SANet is a solid engineering improvement, but the contribution remains incremental.
  2. The main diffusion baselines are DiffDock and DiffDock-L. There is no comparison against current co-folding models (AF3, Boltz, Chai-1, etc.), which have become strong baselines and often outperform earlier diffusion docking methods.
  3. The authors claim that “We additionally report results on PoseBusters,” but no such results appear in the paper.
  4. The main tables focus on RMSD; PB-Valid would better show the physical plausibility of the predicted poses.
  5. SADock is designed and evaluated only for blind docking. Most real-world applications provide pocket information, and recent strong methods focus on the pocket-conditioned setting (e.g., DiffDock-Pocket, SurfDock, Interformer, Uni-Mol Docking v2). The paper does not show whether the model works, or can be adapted, in that more common regime.
Questions for authors
  1. Can the authors provide comparisons against current co-folding models (AF3, Boltz-2, Chai-1)?
  2. Can SADock be evaluated, or easily adapted, in the pocket-conditioned setting? If so, please compare it with DiffDock-Pocket, SurfDock, and Interformer.
  3. Could the authors report success rates with PB-Valid?
  4. Why are several entries in Table 1 missing, when similar metrics were reported normally in the DiffDock papers?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 8TQX: W1 — Novelty beyond DiffDock and SurfDock

We thank the reviewer for recognizing our evaluation in a “hard blind-docking setting” and the “solid improvements over DiffDock / DiffDock-L.” We appreciate your constructive suggestions on broader comparisons and practical applicability, and address your concerns point by point below.

W1: Novelty beyond DiffDock and SurfDock.

We thank the reviewer for this comment and would like to clarify our core contributions, particularly their distinctions from DiffDock and SurfDock.

(1) Beyond DiffDock: Interaction-aligned representations for OOD generalization. Our central insight is that global structural novelty does not necessarily imply interaction novelty, as structurally distinct receptors can preserve similar ligand-facing interaction environments. Under limited supervision, Euclidean-based models may overfit to receptor-specific structural patterns, limiting pocket localization on unfamiliar receptors. This motivates our surface-aligned representation and dedicated surface–Euclidean interaction framework for generalizable blind docking. Our contribution is therefore orthogonal to DiffDock's diffusion formulation, which we deliberately retain to isolate the benefits of interaction-aligned modeling.

(2) Beyond SurfDock: From pocket-conditioned docking to receptor-wide localization. While SurfDock incorporates surface information for pose generation within predefined pockets, SADock extends surface modeling to joint pocket discovery and pose generation, directly addressing the critical localization bottleneck in blind docking.

To evaluate this distinction, we construct a blind-docking baseline using P2Rank's top four predicted pockets, running SurfDock with 10 poses per pocket (40 candidates per target), jointly ranked by its native MDN scores (SurfScore). The results (Top-1 success rate (%) / median RMSD (Å)) are:

Table R11: Blind-docking results.

MethodPDBBindPoseBustersDockGen
P2Rank+SurfDock6.9 / 16.76.5 / 12.73.7 / 18.4
SADock44.6 / 2.842.5 / 2.815.3 / 5.2
SADock-L47.9 / 2.253.5 / 1.827.0 / 3.2

To further investigate the gap, we compare pocket localization success (centroid distance < 4 Å):

Table R12: Pocket localization success rates (%).

MethodPDBBindPoseBustersDockGen
P2Rank:Top4 Oracle58.461.452.4
SADock-L:Top184.682.767.2
SADock-L:Top591.289.776.2
SADock-L:Oracle (N=40)95.695.385.7

Notably, on DockGen, SADock-L achieves 67.2% Top-1 localization success, exceeding P2Rank's 52.4% even with oracle selection among four pockets. These results highlight the limitations of two-stage pipelines, where pocket prediction errors and potential mismatches between predicted and training pockets can constrain downstream pose generation.

In contrast, SADock integrates interaction-aligned surface representations directly into receptor-wide score modeling, enabling pocket localization and pose generation to be optimized jointly. This distinction, rather than simply introducing surface features into docking, constitutes our primary methodological contribution.

AuthorsOpenReview ↗

Response to Reviewer 8TQX: W2/Q1, W3 & W4/Q3 — Co-folding Comparisons, PoseBusters, and Physical Validity

W2 & Q1: Comparison with co-folding models.

We thank the reviewer for this suggestion. We have extended our evaluation to include AlphaFold 3 (AF3), with and without MSA, on PDBBind, PoseBusters, and DockGen, using 40 candidates per target.

Table R13: Comparison with co-folding models. Top-1 / Top-5 success rates (RMSD < 2 Å, %) and inference time.

MethodPDBBindPoseBustersDockGenInference time
AF361.4 / 66.955.1 / 60.355.0 / 61.4274.9s
AF3 w/o MSA3.3 / 3.31.9 / 2.63.2 / 3.7210.5s
SADock-L47.9 / 60.653.5 / 66.627.0 / 40.732.3s

(1) Performance and efficiency. Full AF3 achieves stronger performance on PDBBind and DockGen, while SADock-L remains competitive on PoseBusters, achieving a higher Top-5 success rate (66.6% vs. 60.3%). Notably, SADock-L is 8.5× faster than AF3. Without MSA inputs, AF3 exhibits substantial performance degradation, highlighting its reliance on evolutionary information. In contrast, SADock-L achieves competitive docking performance without explicit MSA inputs, at substantially lower computational cost.

(2) Training-data overlap. We also note that AF3 was trained on PDB structures released before September 30, 2021 [1]. Since DockGen was not constructed using this cutoff, a substantial fraction of its complexes may overlap with AF3's training data, as also noted by [2]. Such potential overlap should be considered when interpreting AF3's DockGen performance as evidence of generalization to unseen interactions. These results highlight the complementary strengths of large-scale co-folding models and SADock's efficient, interaction-aligned blind docking framework.

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

[2] Assessing the potential of deep learning for protein--ligand docking. Nature Machine Intelligence 2026

W3: Additional PoseBusters results.

Thanks for pointing out this omission. We have completed the evaluation on all 428 PoseBusters complexes, using 40 candidates per target for all methods.

Table R14: PoseBusters results.

MethodTop1(%)Top5(%)
DiffDock35.544.6
DiffDock-L50.956.1
SADock42.556.1
SADock-L53.566.6

SADock and SADock-L consistently outperform their DiffDock counterparts. Notably, SADock-L improves Top-1/Top-5 success by 2.6%/10.5% (absolute) over DiffDock-L, confirming its effectiveness on the independent PoseBusters benchmark. We have added these results to the revised manuscript.

W4 & Q3: Physical validity of predicted poses.

Thanks for this suggestion. We further evaluate PB-valid-joint success rates, requiring both RMSD < 2 Å and passing PoseBusters physical-validity checks, across PDBBind, PoseBusters, and DockGen.

Table R15: Top-1 / Top-5 PB-valid-joint success rates.

MethodPDBBindPoseBustersDockGen
DiffDock21.5 / 28.915.7 / 25.01.1 / 1.6
DiffDock-L21.2 / 30.625.2 / 36.23.2 / 5.3
SADock23.4 / 36.922.0 / 38.15.8 / 11.1
SADock-L32.5 / 47.429.0 / 48.43.7 / 9.0

SADock and SADock-L consistently outperform their DiffDock counterparts across all three benchmarks under the additional physical-validity constraints. Notably, SADock-L achieves 32.5%/47.4% on PDBBind and 29.0%/48.4% on PoseBusters, demonstrating that its improvements extend beyond RMSD accuracy to the recovery of physically valid binding poses. We have added these results to the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 8TQX: W5/Q2 & Q4 — Pocket-Conditioned Docking and Missing Table Entries

W5 & Q2: Extension to pocket-conditioned docking.

Although SADock primarily targets generalizable blind docking, its surface-aware architecture can be naturally adapted to pocket-conditioned docking. To verify this, we retrained SADock-L following SurfDock's pocket-conditioned setting, with two efficiency-oriented differences: (i) using 10× sparser surface meshes and (ii) adopting SurfDock's pretrained SurfScore for pose ranking instead of training a dedicated confidence model.

For a consistent comparison, we re-evaluated SurfDock using its released checkpoint, with 40 candidates per complex for both methods.

Table R16: Pocket-conditioned docking results. Oracle / Top-1 / Top-5 success rates (RMSD < 2 Å, %).

MethodPDBBindPoseBustersDockGenInference time(s)
SurfDock-Rerun78.2 / 66.4 / 72.492.9 / 82.0 / 89.779.8 / 67.2 / 77.2110.7
SADock-L+SurfScore77.7 / 64.2 / 73.392.3 / 77.80 / 85.381.5 / 63.5 / 76.232.3

SADock-L achieves comparable oracle performance across all three benchmarks, even slightly outperforming SurfDock on DockGen (81.5% vs. 79.8%). Its Top-1/Top-5 performance is generally somewhat lower, potentially reflecting the sparser surface representation and the use of a confidence model not specifically trained on SADock poses. Notably, SADock-L requires only 29.2% of SurfDock's inference time (32.3 vs. 110.7 s).

These results demonstrate that SADock readily extends to pocket-conditioned docking with competitive accuracy and substantially lower computational cost, while its primary contribution remains improving receptor-wide localization and pose generation under structural distribution shift. We have included these additional results in the revised manuscript.

Q4: Missing entries in Table 1.

Thanks for pointing out this omission. After cross-checking the original DiffDock papers, we confirmed that Top-5 success rates and median RMSD results for SMINA and GNINA on PDBBind were inadvertently omitted. We have corrected Table 1 and included the missing results in the revised manuscript.

Reviewer_UH2eOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SADock augments DiffDock-style blind docking with explicit molecular surfaces for both receptor and ligand. SANet combines geodesic, hierarchical, heterogeneous, and intermolecular message passing with time-conditioned FiLM. The paper motivates transfer through boundary-stability arguments and evaluates receptor-wide docking on PDBBind and DockGen, with analyses focused on pocket localization under structural distribution shift.

Critical strengths
  1. The method targets an important failure mode: locating a binding region on an unfamiliar receptor. It uses receptor-wide surface information rather than assuming a known pocket. The reported gains are substantial for the smaller model on DockGen: top-1 success rises from 5.3% to 15.3%, and top-5 from 11.1% to 28.6%. The larger model also improves from 22.8%/29.6% to 27.0%/38.6%. Improvements in centroid localization and oracle candidate recovery provide useful evidence about where the gain arises.

  2. The direct DiffDock comparisons are stronger than a comparison of unrelated published numbers alone. The authors regenerate predictions from released DiffDock checkpoints, use matched candidate counts within the relevant model pairs, and retain failed predictions in the denominator. Surface and FiLM ablations are directionally consistent with the proposed mechanism. The improvement is obtained without the larger baseline's vdM augmentation and inference tuning, which makes the practical result interesting.

Critical weaknesses
  1. The theory is a conditional stability argument, not a demonstration that the learned representation achieves the required invariance. Equation 7 assumes that surface discrepancy controls pose-dependent interaction-energy perturbations, with a residual term; Equation 8 additionally leaves Euclidean-branch sensitivity uncontrolled. Appendix B gives useful sufficient conditions, but does not measure the relevant constants or residuals for the evaluated receptors. Lower domain separability and feature overlap are consistent with transfer but do not establish that useful interaction information, rather than general representational detail, has been preserved. The empirical gains are more convincing than the claim of a demonstrated mechanistic explanation.

  2. Robustness and practical validity are not yet characterized well enough for the strongest generalization claim. The large-model top-1 improvement is 4.2 percentage points on only 189 DockGen cases, without reported paired uncertainty or training-seed variability. Low RMSD alone also does not establish valid ligand geometry or absence of receptor clashes. Section 3.1 mentions a 428-complex PoseBusters evaluation, but I could not locate its results in the supplied 20-page manuscript. A common physical-validity analysis and timing that includes surface construction would materially strengthen the practical case.

Questions for authors
  1. Please provide paired confidence intervals or success/failure transitions for the DockGen comparisons, particularly SADock-L versus DiffDock-L, and clarify the source of the small DiffDock differences between Tables 1 and 2.

  2. Where are the stated PoseBusters results? What fraction of poses is both below 2 Å RMSD and physically valid, and what is the end-to-end cost including receptor preprocessing and ligand-surface updates during denoising?

  3. Can the existing analysis isolate the dual-surface/geodesic contribution from a simpler surface-feature addition, and quantify score sensitivity under controlled boundary-preserving perturbations? If not, please present the theory as sufficient-condition motivation rather than verified explanation of the observed OOD gains.

3 comments
AuthorsOpenReview ↗

Response to Reviewer UH2e: W1 — Empirical Validation of Interaction-Aligned Transferability

We thank the reviewer for recognizing that SADock targets an “important failure mode,” highlighting the “useful evidence” from pocket localization and oracle recovery, and appreciating our rigorous DiffDock comparisons and practical gains. We address your concerns point by point below.

W1: Empirical validation of interaction-aligned transferability.

Thanks for this constructive assessment. We agree that our theoretical analysis establishes sufficient conditions for transferability, and have conducted three additional analyses to empirically substantiate the underlying assumptions and their relevance to SADock.

(1) Physical-level assessment of Eq. 7. We performed 50 geometric perturbations on each of 20 DockGen complexes, obtaining 989 valid comparisons of Lennard–Jones and Coulomb interaction energies under fixed ligand configurations. We calibrated per-complex empirical bounds with at least 90% coverage:

δUQ≤KℓδS+εℓ,\delta U_{\mathcal Q}\leq K_\ell\delta_S+\varepsilon_\ell,

yielding Kℓ∈[0.1007,5.1053]K_\ell\in[0.1007,5.1053] and εℓ∈[0,0.03223]\varepsilon_\ell\in[0,0.03223] kcal/mol. These results quantitatively characterize the relationship between surface discrepancy and interaction-energy variations, with small fitted residuals under the examined conditions. While not establishing a uniform bound, they provide an initial empirical assessment of the boundary-dependent assumption in Eq. 7. See implementation details in Reply W2-1 to Reviewer W4gW.

(2) Model-level validation of Eq. 8. We further quantified the Euclidean residual sensitivity on 300 PDBBind complexes. Under 0.01/0.1 Å perturbations, relative score variations remain small at early denoising (t=1: 0.55%/1.84%) and increase as denoising progresses (Pearson r=0.91 with (1-t)). Moreover, the correlation between Euclidean feature magnitude and (1-t) increases from r=0.39 before FiLM to r=0.99 after FiLM, supporting the intended time-adaptive modulation of Euclidean information. See implementation details in Reply W2-2 to Reviewer W4gW.

(3) Representation informativeness and transferability. We clarify that the surface and Euclidean features in Fig. 1 are dimension-matched and standardized, controlling for differences in feature scale and variance. More importantly, removing surface representations under matched training conditions reduces DockGen oracle success from 36.51% to 16.40%, demonstrating their substantial task-relevant contribution. We further find that surface discrepancy exhibits a stronger Pearson correlation with DockGen oracle RMSD than global TM-score (∣r∣=0.402|r|=0.402 vs. 0.1310.131), linking interaction-boundary similarity to OOD docking performance. See implementation details in Reply W2-3 to Reviewer W4gW.

Together, these results provide complementary physical-, model-, and representation-level evidence supporting our interaction-aligned transferability hypothesis. We have incorporated these analyses into the revised manuscript and clarified the sufficient-condition nature of our theoretical results.

AuthorsOpenReview ↗

Response to Reviewer UH2e: W2 & Q2 — Training Stability, Physical Validity, and Computational Efficiency

W2 & Q2: Training stability, physical validity, and computational efficiency.

Thanks for these constructive suggestions; ‡^\ddagger denotes results ranked using SAConf. We have supplemented our evaluation with multi-seed experiments, physical-validity assessments across three benchmarks, and end-to-end computational cost analysis. For consistency with the original submission, all SADock results reported here use the released DiffDock confidence model for pose ranking.

(1) Training stability. We trained SADock-L with two additional random seeds and evaluated all three runs using the 400K-step EMA checkpoints and 40 candidates per target on DockGen. The Top-1 success rates are 26.98%, 28.57%, and 26.98%, yielding a mean of 27.51% ± 0.92%. The small variation across independent training runs supports the stability of SADock-L's OOD performance.

(2) PoseBusters and physical validity. We have completed the evaluation on all 428 PoseBusters complexes and applied consistent physical-validity checks to PDBBind and DockGen. The table reports Top-1 / PB-valid-joint Top-1 success rates, N=40:

Table R9: Top-1 / PB-valid-joint Top-1 success rates (%), N=40.

MethodPDBBindPoseBustersDockGen
DiffDock38.8 / 21.535.5 / 15.75.3 / 1.1
DiffDock-L45.2 / 21.250.9 / 25.225.4 / 3.2
SADock44.6 / 23.442.5 / 22.015.3 / 5.8
SADock-L47.9 / 32.553.5 / 29.027.0 / 3.7

On PoseBusters, SADock/SADock-L improve Top-1 success over DiffDock/DiffDock-L by 7.0%/2.6% (absolute). Their PB-valid-joint success rates also consistently outperform the corresponding baselines across all three benchmarks, with average absolute improvements of 4.3%/5.2%, respectively.

Additionally, as detailed in reply W1-1 & Q2/Q5 to Reviewer W4gW, we trained an SANet-based confidence model (SAConf) on SADock-generated poses, yielding further gains. On DockGen, SAConf improves SADock-L's Top-1 success from 27.0% to 33.9%‡^\ddagger and PB-valid-joint success from 3.7% to 19.0%‡^\ddagger.

(3) End-to-end computational cost. We benchmarked inference using one NVIDIA A100 (80 GB), generating 40 candidates per complex. Average wall-clock time is measured over the complete PDBBind test set, including ESM feature generation, receptor preprocessing, ligand surface updates during denoising, docking, and confidence ranking. Peak GPU memory is measured independently at batch size 1, corresponding to one complex and one conformer. Receptor surface generation and geodesic graph construction take 0.785 s and 2.330 s, respectively, and can be reused across diffusion steps and ligand queries. Online ligand surface generation and graph construction take 9.07 ms and 10.65 ms per update, respectively. All these costs are included in the end-to-end measurements below.

Table R10: Inference time and peak GPU memory.

MethodInference time(s/complex)Peak GPU Memory(GiB)
DiffDock22.45.99
DiffDock-L55.610.12
SADock30.81.98
SADock-L32.32.20
AuthorsOpenReview ↗

Response to Reviewer UH2e: Q1 & Q3 — Paired Oracle Recovery, Ablation Protocol, and Model Stability

Q1: Paired success analysis and clarification of ablation baselines.

We provide the following clarifications and additional analysis.

(1) Paired success/failure transitions. Using oracle best-of-40 success (RMSD < 2 Å), SADock-L recovers 43 complexes on which DiffDock-L fails, while losing 16, yielding a net gain of 27 cases among all 189 DockGen complexes. These results provide further evidence of SADock's improved recovery of challenging targets.

(2) DiffDock differences between Tables 1 and 2. We should have clarified this distinction. Table 1 evaluates the publicly released DiffDock checkpoints, whereas the Euclidean-only baseline in Table 2 was retrained from scratch using DiffDock's official architecture (~20M parameters), with the same training data and budget as SADock. The small numerical differences therefore arise from independently trained models rather than inconsistent evaluation protocols. We have clarified this in the revised manuscript.

Q3: Architectural contributions and model-level stability.

Thanks for this constructive suggestion. We have conducted additional analyses to clarify the contributions of our surface modeling and empirically quantify the model sensitivity motivated by Eq. 8.

1) Fine-grained architectural ablations. We further isolate the contributions of geodesic encoding, ligand surface representations, and heterogeneous message passing under matched training settings and comparable model capacities. On DockGen, removing geodesic distance encoding or ligand surface representations reduces Top-5 success from 28.6% to 24.9% and 23.3%, respectively, while removing Hetero-MP further reduces it to 13.8%. These results demonstrate that SADock's gains arise not merely from adding surface features, but also from explicitly modeling intrinsic surface geometry and surface–Euclidean interactions.

  1. Model-level sensitivity. On 300 PDBBind complexes, we measured score variations under controlled Euclidean perturbations while keeping surface inputs fixed. At t=1, perturbations of 0.01/0.1 Å produce only 0.55%/1.84% relative score changes, with sensitivity increasing as denoising progresses (Pearson r=0.91 with (1-t)). This supports reduced Euclidean sensitivity during pocket localization, while retaining structural sensitivity for pose refinement. See implementation details in Reply W2-2 to Reviewer W4gW.

3) Theoretical scope. We agree that our theory establishes sufficient conditions rather than a universal guarantee of learned invariance. These additional experiments provide empirical support for the proposed mechanism, and we have clarified this distinction in the revised manuscript.

Reviewer_W4gWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the generalization challenge of learning-based blind protein-ligand docking when encountering structurally novel receptors. The authors argue that global structural novelty does not imply interaction novelty, and that Euclidean atom/residue encoders entangle transferable interface patterns with receptor-specific structural variation. They propose SADock, which retains DiffDock's pose-diffusion framework (translation, rotation, torsion diffusion) but replaces the score network with SANet, a dual-branch architecture that introduces molecular surface point cloud representations for both proteins and ligands. SANet comprises four key components: (i) geodesic message passing on surfaces using both Euclidean and mesh geodesic distances, (ii) fine-to-coarse / coarse-to-fine hierarchical communication pathways for receptor-wide surface information propagation, (iii) heterogeneous message passing coupling surface points with residues/atoms, and (iv) an equivariance-preserving FiLM module that modulates Euclidean features with diffusion-time-conditioned surface context.

The paper provides theoretical analysis (Proposition 1): if pose-dependent interaction energy perturbations are bounded by the surface discrepancy δ_S, then free-energy basin preferences are preserved. Experimentally, SADock/SADock-L (~20M/30M parameters, fewer training steps, no vdM augmentation or inference tuning) outperforms DiffDock/DiffDock-L on the PDBBind 2019 time split, with more pronounced improvements on DockGen-full (Top-1: 22.8%→27.0%, Top-5: 29.6%→38.6%). Further analysis demonstrates improved pocket centroid localization and performance on structurally unfamiliar targets. Ablation confirms that both the surface branch and FiLM module contribute to the gains.

Critical strengths
  1. Substantial OOD performance gains on DockGen with analysis linking improvements to the pocket localization mechanism.

DockGen is the benchmark most relevant to the paper's core claims. The improvements over DiffDock-L are considerable (Top-1: 22.8%→27.0%, Top-5: 29.6%→38.6%, median RMSD: 4.3→3.4 Å), achieved with fewer training steps (400K vs. 1250K), no vdM augmentation, and no inference-time tuning. More importantly, the paper goes beyond reporting metric numbers. The RMSD and centroid-distance success-rate curves (Fig. 3), TM-score-stratified oracle success-set overlap analysis (Fig. 4a), and the case study (Fig. 4b) collectively demonstrate that the gains arise primarily from improved full-receptor pocket discovery on structurally unfamiliar targets, which is precisely the mechanism the paper hypothesizes. This complete evidence chain from claim to analysis is the paper's most valuable aspect.

  1. A clearly designed, well-motivated architectural innovation that fills a genuine gap in blind docking.

Existing surface-based docking methods such as SurfDock require a predefined pocket, and P2Rank+docking pipelines decouple pocket prediction from pose generation. SADock is among the first methods to maintain full-receptor and ligand surface representations throughout the blind diffusion docking process. The design choices are sensible and well-described: fine-scale geodesic encoding before pooling, explicit Euclidean-geodesic distance differentials Δ_{ij}, area-weighted F2C/C2F, and equivariance-preserving FiLM (Appendix A). The ablation (Table 2) shows the surface branch approximately doubles DockGen Top-1 over the Euclidean-only model. As a drop-in score-network replacement, SANet could be adopted by other diffusion/flow docking frameworks.

Critical weaknesses
  1. Experimental evidence is insufficient to support the generalization claims, and some comparisons may be confounded.

Missing benchmarks and baselines. PoseBusters (428 complexes) is listed in Sec. 3.1 but no results appear in the paper. PoseBusters physical-validity metrics are also absent. The most relevant surface-based method, SurfDock (e.g., combined with P2Rank for blind docking), is not compared. Recent blind/flexible docking methods cited in the introduction, including DeltaDock, FABind variants, and DynamicBind, are not evaluated.

Confounds in ablation and ranking. It is unclear whether the "Euclidean-only" baseline in Table 2 is a retrained model with matched training budget and parameter count, or the public DiffDock checkpoint (whose PDBBind Top-1 of 38.8% happens to match). All SADock poses are ranked using DiffDock's confidence model, which was trained on DiffDock's pose distribution. The large Top-5 vs. Top-1 gap (e.g., DockGen: SADock 15.3%/28.6%) suggests ranking is a performance bottleneck, yet the paper does not discuss this issue.

No computational cost analysis. NanoShaper mesh generation, full-receptor Dijkstra geodesic computation, and fine-resolution surface message passing add substantial preprocessing and inference overhead. Runtime and GPU memory comparisons with DiffDock/DiffDock-L are absent, which is critical given the paper's implicit claims about practical screening applications.

These issues are consequential because the paper's primary contribution is experimental. They determine whether the DockGen gains are a robust, practically meaningful improvement or an artifact of the evaluation setup.

  1. The theoretical analysis does not substantively demonstrate the superiority of surface representations.

Proposition 1 is a standard corollary of placing a uniform bound on pose-dependent energy perturbations: a uniform bound b on U_{X'} − U_X implies basin free-energy oscillations ≤ 2b. All substantive content resides in the assumption of Eq. 7 (|U_{X'} − U_X − c| ≤ K_ℓ δ_S + ε_ℓ), which is not verified. The paper does not estimate K_ℓ or ε_ℓ for actual DockGen/PDBBind receptor pairs, and Appendix B.3 derives the bound only under strong idealizations (common rigid dielectric boundary, fixed-parameter LJ potentials, atoms at fixed distances).

The model-level bound (Eq. 8) is a triangle inequality applicable to any network and any input decomposition. The residual ε_{E,θ} is uncontrolled, so the bound cannot explain why surface representations, as opposed to any learned feature, should be more transferable.

No experiment bridges theory and data. For instance, showing that δ_S (Eq. 22) between DockGen test pockets and their nearest training pockets is a better predictor of SADock success than global TM-score would have made Proposition 1 a testable claim.

The domain classifier analysis (Fig. 1b/c) compares feature separability from two different models with potentially different dimensionalities and variances. Lower separability may merely reflect less informative or lower-variance features rather than better transferability.

This matters because the paper positions its theoretical analysis as a "principled" explanation and a core contribution. As written, it provides qualitative intuition rather than quantitative evidence.

Questions for authors
  1. Holo vs. apo structures. Are all receptor surfaces computed from holo crystal structures? Can you provide SADock vs. DiffDock-L comparisons on apo or predicted receptor structures (e.g., ESMFold/AF2 versions of PDBBind test and DockGen)? This is the single most important question for the generalization claim.

  2. PoseBusters results. Sec. 3.1 claims results on PoseBusters (428 complexes) but I could not locate them. Please provide Top-1/Top-5 success rates and PB-valid rates. Also report pose physical validity on PDBBind/DockGen.

  3. Ablation protocol. Is the "Euclidean-only" model in Table 2 a retrained model with the same training steps, data, and approximately matched parameter count as SADock, or the public DiffDock checkpoint? Please add a parameter-matched Euclidean-only baseline and finer-grained component ablations (geodesic distance / Δ_{ij}, F2C/C2F, ligand surface, Hetero-MP ablated separately).

  4. Confidence model. Are SADock-L poses ranked with DiffDock or DiffDock-L's confidence model? Was this confidence model retrained on SADock poses? The large Top-5 vs. Top-1 gap suggests ranking is a bottleneck. What are SADock's oracle (best-of-N) success rates vs. DiffDock-L?

  5. Additional baselines. How does SADock compare with SurfDock under a blind docking setup (e.g., SurfDock + P2Rank pocket)?

  6. Computational cost. What are the preprocessing (surface mesh + geodesic computation) and inference wall-clock times, and per-complex GPU memory, compared with DiffDock-L?

  7. Empirical validation of theory. Can you empirically estimate δ_S (Eq. 22) between DockGen test pockets and nearest training pockets, and show whether it predicts SADock success better than global TM-score?

  8. Domain classifier controls (Fig. 1). Are the feature dimensionalities and variances of the domain classifier inputs comparable? How do you rule out that surface features are simply less informative?

10 comments
AuthorsOpenReview ↗

Response to Reviewer W4gW: W1-1 & Q2/Q5 — Benchmarks, Baselines, and Physical Validity

We thank the reviewer for recognizing SADock’s “substantial OOD performance gains” and “well-motivated architectural innovation,” and for highlighting the “complete evidence chain from claim to analysis.” We address your concerns point by point below.

W1-1 & Q2/Q5:

Thanks for these comprehensive suggestions; ‡^\ddagger denotes results ranked using SAConf. We have expanded our evaluation to include PoseBusters, additional competitive baselines, and PB-valid metrics. We also trained a SANet-based confidence model (SAConf) on SADock-generated poses for more effective pose ranking and use it for all SADock results reported below. All sampling-based methods use 40 conformations per target for consistent comparison. The updated results are summarized below.

Tabel R1: Top1 success rate (%) / Median RMSD (Å)

MethodPDBBindPoseBustersDockGen
FABind+41.6 / 2.821.3 / 4.53.2 / 7.4
DiffDock38.8 / 3.335.5 / 3.35.3 / 7.8
DiffDock-L45.2 / 2.550.9 / 1.925.4 / 3.7
DynamicBind29.2 / 3.626.9 / 4.22.6 / 7.1
P2Rank+SurfDock6.9 / 16.76.5 / 12.73.7 / 18.4
SADock$^\ddagger$46.8 / 2.244.4 / 2.421.2 / 5.3
SADock-L$^\ddagger$48.8 / 2.157.5 / 1.633.9 / 3.2
  1. PoseBusters results. SADock-L achieves a 57.5% Top-1 success rate on PoseBusters, outperforming DiffDock-L (50.9%) by 6.6 percentage points, while reducing median RMSD from 1.9 Å to 1.6 Å. This improvement is consistent with its performance on PDBBind and DockGen.

  2. Additional baselines. We have included DynamicBind and P2Rank + SurfDock, while FABind+ was already evaluated in the original submission. DeltaDock was not included because evaluation-ready code and model checkpoints were unavailable. SADock-L consistently outperforms these additional baselines, achieving particularly substantial gains on DockGen.

  3. Physical validity. We further report PB-valid-joint Top-1/Top-5 success rates, requiring predicted poses to satisfy both RMSD < 2 Å and PoseBusters physical-validity checks. As shown in Table R2, SADock-L achieves the best results across all benchmarks. Notably, on DockGen, SADock and SADock-L achieve 8.5-fold (9.0% vs. 1.1%) and 6.0-fold (19.0% vs. 3.2%) higher PB-valid success rates than DiffDock and DiffDock-L, respectively. These results demonstrate that SADock improves not only docking accuracy but also the recovery of physically valid binding poses.

Table R2: success rate / PB-valid-joint success rate (%)

MethodPDBBind Top1PDBBind Top5PoseBusters Top1PoseBusters Top5DockGen Top1DockGen Top5
FABind+41.6 / 8.047.4 / 19.321.3 / 6.827.1 / 12.13.2 / 0.04.8 / 0.0
DiffDock38.8 / 21.546.0 / 28.935.5 / 15.744.6 / 25.05.3 / 1.111.1 / 1.6
DiffDock-L45.2 / 21.252.9 / 30.650.9 / 25.256.1 / 36.225.4 / 3.232.3 / 5.3
DynamicBind29.2 / 13.235.8 / 18.226.9 / 11.234.1 / 15.22.6 / 0.04.8 / 0.5
P2Rank+SurfDock6.9 / 3.310.2 / 4.76.5 / 3.79.8 / 5.83.7 / 0.55.3 / 0.5
SADock$^\ddagger$46.8 / 29.555.6 / 38.644.4 / 24.558.2 / 39.321.2 / 9.031.2 / 19.6
SADock-L$^\ddagger$48.8 / 33.661.2 / 49.357.5 / 37.668.7 / 50.533.9 / 19.043.9 / 25.9
  1. Comparison with SurfDock. To establish a competitive blind-docking baseline, we use P2Rank's top four predicted pockets as inputs to SurfDock, generating 10 poses per pocket (40 candidates per target). Candidates from all pockets are jointly ranked using SurfDock’s native MDN scores. Nevertheless, P2Rank + SurfDock achieves only 3.7% Top-1 success on DockGen, compared with 33.9% for SADock-L.

To investigate this gap, we further compare pocket localization success rates (centroid distance < 4 Å):

MethodPDBBindPoseBustersDockGen
P2Rank:Top4 Oracle58.461.452.4
SADock-L$^\ddagger$:Oracle (N=40)95.695.385.7
SADock-L$^\ddagger$:Top184.884.370.4
SADock-L$^\ddagger$:Top590.989.776.7

Notably, on DockGen, SADock-L achieves 70.4% Top-1 localization success, substantially exceeding P2Rank's 52.4% even with oracle selection among four pockets. This highlights a key limitation of two-stage pipelines: upstream pocket localization errors constrain downstream docking, while mismatches between predicted pockets and SurfDock's pocket-conditioned training inputs may further affect pose generation. In contrast, SADock jointly optimizes receptor-wide pocket localization and pose generation through interaction-aligned surface representations, avoiding this separation and achieving substantially stronger OOD generalization.

AuthorsOpenReview ↗

Response to Reviewer W4gW: W1-2 & Q4 — Ablation Protocol and Confidence Ranking

W1-2 & Q4: Ablation protocol and confidence ranking.

We thank the reviewer for highlighting these concerns; ‡^\ddagger denotes results ranked using SAConf. We clarify the ablation and ranking protocols and provide additional oracle evaluations.

  1. Ablation protocol. We should have clarified this distinction. While Table 1 evaluates DiffDock using its released checkpoints, the Euclidean-only baseline in Table 2 was retrained from scratch with matched training data, budget, and parameter count as SADock, ensuring a controlled comparison.

  2. Confidence ranking. We agree that confidence ranking is important for docking performance. To isolate the contribution of our surface-aware score model, we originally used DiffDock's pretrained confidence model without retraining on SADock poses for a controlled comparison. To address this concern, we further trained an SANet-based confidence model (SAConf) on SADock-generated poses and reranked the same candidates. This consistently improves Top-1 success rates, particularly on DockGen, where SADock/SADock-L improve from 15.3%/27.0% to 21.2%/33.9%. Similar gains are observed on PDBBind and PoseBusters, demonstrating the benefit of surface-aware confidence estimation. For consistency, we report SAConf-ranked SADock results throughout our responses.

Table R3: Top1/Top5 Success Rate, % (40 samples for all methods)

ModelConfidencePDBBindPoseBustersDockGen
SADockDiffDockConf44.6 / 56.742.5 / 56.115.3 / 28.6
SADock$^\ddagger$SAConf46.8 / 55.644.4 / 58.221.2 / 31.2
SADock-LDiffDockConf47.9 / 60.653.5 / 66.627.0 / 40.7
SADock-L$^\ddagger$SAConf48.8 / 61.257.5 / 68.733.9 / 43.9
  1. Oracle success rates. To further disentangle pose generation from ranking, we report oracle best-of-40 success rates across three benchmarks:
MethodPDBBindPoseBustersDockGen
FABind+50.134.18.5
DiffDock50.751.413.8
DiffDock-L59.261.938.6
DynamicBind42.741.810.6
P2Rank+SurfDock13.514.57.4
SADock61.764.536.5
SADock-L68.075.555.0

SADock and SADock-L consistently outperform their DiffDock counterparts across all benchmarks. Notably, SADock-L improves DockGen oracle success from 38.6% to 55.0%, showing that its advantage extends beyond confidence ranking to pose generation.

AuthorsOpenReview ↗

Response to Reviewer W4gW: W1-3 & Q6 — Computational Cost and Inference Efficiency

W1-3 & Q6: Computational cost and inference efficiency.

We should have clarified the computational overhead and efficiency of SADock; ‡^\ddagger denotes results ranked using SAConf.

  1. Computational cost. Receptor surface generation and geodesic graph construction take 0.785 s and 2.330 s per receptor, respectively, and are performed only once, with the results reusable across diffusion steps and ligand queries. Online ligand surface generation and graph construction take only 9.07 ms and 10.65 ms per update, respectively. Moreover, our hierarchical F2C/C2F design reduces surface interaction costs by performing expensive cross-representation and intermolecular message passing on coarse surface representations. All preprocessing overhead is included in the end-to-end inference times reported below.
Surface generation(ms)Geodesic surface graphing(ms)
Receptor785.372,330.46
Ligand9.0710.65
  1. Inference efficiency. We benchmark all methods using a single NVIDIA A100 (80 GB; 8 vCPUs), with 40 candidates per target for sampling-based methods. Average wall-clock time is measured over the complete PDBBind test set, including ESM feature generation, receptor preprocessing, and docking. Peak GPU memory is measured per model forward pass with batch size 1 (1 target with 1 conformer). Notably, SADock-L reduces inference time by 42% (32.3 vs. 55.6 s) and peak GPU memory by 78% (2.20 vs. 10.12 GiB) compared with DiffDock-L. Although SADock incurs additional runtime relative to the smaller DiffDock, it substantially reduces memory usage. These results demonstrate that hierarchical surface modeling enables competitive computational efficiency, supporting SADock's practicality for large-scale virtual screening.
MethodEnd-to-end Inference Time(s/complex)Peak GPU Memory(GiB)
DiffDock22.45.99
DiffDock-L55.610.12
FABind+29.12.23
P2Rank+SurfDock113.22.05
DynamicBind94.810.57
SADock$^\ddagger$30.81.98
SADock-L$^\ddagger$32.32.20

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.