acceptodds
Under review as a conference paper at ICLR 2027

ALPHASURF: ON-THE-FLY SURFACE COMPUTATIONS FOR PROTEIN REPRESENTATION LEARNING

Abstract

Several protein surfaces have been proposed, notably for visualization purposes. Machine learning methods have incorporated these surfaces as protein representations, resulting in strong performance, at the cost of heavy computations. In this paper, we show that this burden can be avoided by introducing AlphaSurf, a coarse meshing method tailored for learning on proteins that leverages -complexes. We characterize the surfaces theoretically, including their topological relation to the SAS surface, and evaluate them experimentally on downstream tasks. Our method runs on-the-fly during training while maintaining performance. This enables use cases previously out of reach for surface-based methods: large-scale training as well as coordinate-noise data augmentation and training on molecular dynamics frames which improves performance. The code is available here https://anonymous.4open.science/r/AlphaSurf-C5A1/README.md

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
proteins, representation learning, surface

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_td4SOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose a faster way to calculate protein surface meshes with AlphaSurf, discarding the slow MSMS triangulated molecular surface generator.

Critical strengths

AlphaSurf is a protein surface mesh generation method that is substantially faster than MSMS and EDTSurf. Fig 4 shows that AlphaSurf is orders of magnitude faster than MSMS, EDTSurf, and NanoShaper. In addition, AlphaSurf is much more usable at 1% compared to the aforementioned methods. It appears that AlphaSurf retains the performance of other methods while being much faster. The priors of AlphaSurf are strong - one does not need to render the dense surface mesh if not using the mesh for visualization, thus yielding a speedup if only using it for feature propagation.

Critical weaknesses

"I am not fully convinced by the analyses in Fig 6. Aside from MaSIF-Ligand and AF2 hetero, it appears that AlphaSurf is within error or worse compared to methods that fall under the non shaded areas. Do error bars denote SEM, SD, or confidence intervals?

Looking at supplemental tables S1 and S2, MSMS+simplification is reasonably fast. The main text highlights 10x improvement of AlphaSurf on PINDER, but the supplemental table shows a much smaller speedup. EDTSurf at gs=0.4 is also quite fast and has comparable PINDER processing times to AlphaSurf, can the authors elaborate on what gs a layperson should look at to determine speedup?

The paper uses a lot of technical jargon, making reading difficult, especially if one is not well-versed in the field of surface mesh calculation. I am not an expert in this area, and understanding the latter parts of the introduction and the method itself is quite difficult."

Questions for authors

Suggestion: the x axis labels in Figure 4 are difficult to read and are messy, perhaps cleaning up the formatting and making it clear which colors correspond to MSMS, EDTSurf, NanoShaper would be helpful in understanding this figure.

3 comments
AuthorsOpenReview ↗

Answer to reviewer 3 (1/2)

Thank you for your detailed comments. We would like to draw your attention to the general answer, in which we resume the additional results obtained for the rebuttal. Concerning your points, we corrected our PINDER runs, which now show clear gains over EDTSurf on the Apo-Hetero and AF2-Hetero splits, added a comparison with EDTSurf on molecular dynamics (MISATO), revised Figure 4, and reorganized the first section to make the paper easier to follow. In addition, we provide a point-by-point answer to your comments below.

W1.

W1. I am not fully convinced by the analyses in Fig 6. Aside from MaSIF-Ligand and AF2 hetero, it appears that AlphaSurf is within error or worse compared to methods that fall under the non shaded areas. Do error bars denote SEM, SD, or confidence intervals?

A-W1. First, the error bars denote standard deviation over 3 seeds.

The non-shaded area corresponds to EDTSurf, used with much smaller grid sizes than its default. We agree that it represents a reasonable baseline but would like to highlight that our work is the first to assess its performance for learning applications. The revised version adds three results where AlphaSurf is clearly better than EDTSurf at grid scale 0.4:

  1. With the singular-edge fix (Appendix E), AlphaSurf usability reaches 100% (EDTSurf fails on some proteins).
  2. We found an error in our SLURM PINDER launch script that didn't enable the tufting to be applied for this specific dataset, and we are very sorry about this. We have updated Figures 6-7 with the new results, and we now see that AlphaSurf also clearly outperforms EDTSurf on Apo-Hetero and AF2-Hetero. The Appendix considering Tufting (Appendix I.7 Mesh Ablation) has also been updated, showing that tufting matters also for PINDER-Pair.
PINDER-Pair AUROC (%)HoloApoAF2
No tufting92.12 ± 0.1886.40 ± 0.1385.46 ± 0.22
Tufting92.36 ± 0.0386.67 ± 0.0585.96 ± 0.32
Tufting + linking edges (updated algorithm)92.35 ± 0.0386.72 ± 0.2285.89 ± 0.43
SurfaceApo HeteroAF2 HomoAF2 Hetero
EDTSurf (gs = 0.4)84.29 ± 0.4886.22 ± 0.1278.18 ± 0.35
AlphaSurf85.79 ± 0.3286.60 ± 0.3881.36 ± 0.72

PINDER-Pair AUROC (%), mean ± std. Full results in Figure 6.

  1. On MISATO with random-frame training, EDTSurf meshes lower performance by 6.9 F1, 3.8 AUROC and 9.0 AUPRC points relative to AlphaSurf. This could stem from EDTSurf oversmoothing small pockets. Moreover, EDTSurf benefits less from dynamic augmentions.

Holo-conformation training

SurfaceF1AUROCAUPRC
EDTSurf (gs = 0.4)42.9 ± 0.682.3 ± 0.840.3 ± 0.9
AlphaSurf48.9 ± 0.585.7 ± 0.347.7 ± 0.8

Random-frame training

SurfaceF1AUROCAUPRC
EDTSurf (gs = 0.4)43.4 ± 0.483.0 ± 0.441.4 ± 0.6
AlphaSurf50.3 ± 0.286.8 ± 0.150.4 ± 0.5

Binding-site prediction on MISATO (%, mean ± std), evaluated on holo conformations. Bold: best surface in each training setting.

AuthorsOpenReview ↗

Answer to reviewer 3 (2/2)

W2.

W2. Looking at supplemental tables S1 and S2, MSMS+simplification is reasonably fast. The main text highlights 10x improvement of AlphaSurf on PINDER, but the supplemental table shows a much smaller speedup. EDTSurf at gs=0.4 is also quite fast and has comparable PINDER processing times to AlphaSurf, can the authors elaborate on what gs a layperson should look at to determine speedup?

A-W2. Apologies for this confusion: as explained in our general comments and in the revised manuscript, one needs to consider separately the preprocessing time and the learning time. Our 10x improvement is for the former, but it still translates into a 2-3× speedup of the whole learning pipeline versus MSMS. EDTSurf's cost depends on its grid scale: at gs = 0.4, where its meshes match ours in size, it is comparably fast, but this grid scale must be tuned dataset-wise in order to be on the fly, whereas ours doesn't necessitate any tuning. At this matched cost, AlphaSurf is more robust (0% vs close to 1% failures) and more accurate on Apo-Hetero, AF2-Hetero and MISATO (see W1).

W3.

W3. The paper uses a lot of technical jargon, making reading difficult, especially if one is not well-versed in the field of surface mesh calculation. I am not an expert in this area, and understanding the latter parts of the introduction and the method itself is quite difficult.

A-W3. We acknowledge that the alpha-complex is a non-trivial construction from computational geometry, which does not belong to the usual background of ML practitioners, and we rely on precise mathematical terms to describe molecular surfaces and the associated objects unambiguously. We did our best to provide a gentle and illustrated introduction to it in the main text, and the revised version reorganizes the first section to start with learning rather than surface representations, which should make it easier to follow.

Q1.

Q1. Suggestion: the x axis labels in Figure 4 are difficult to read and are messy, perhaps cleaning up the formatting and making it clear which colors correspond to MSMS, EDTSurf, NanoShaper would be helpful in understanding this figure.

A-Q1. We acknowledge that this figure wasn't as clear as it could be; we provide a revised version (for MaSIF-Ligand also) which makes the numbers and the legend more understandable.

We have tried to address all of the comments and questions by the reviewer, and thank the reviewer again for helping us improve our submission. We invite the reviewer to also consider the general feedback we provided in our general comment to all reviewers. We will make sure to address all comments or concerns in the potential final version of our paper.

AuthorsOpenReview ↗

Follow up to reviewer 3

We thank you again for your review and your remarks, which we believe have made the paper better.

As the discussion period closes soon, we would be grateful to know whether our answers above address your questions. We believe they do, and we would be happy to clarify any remaining points.

Reviewer_JbAZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a new representation for proteins for machine learning tasks. The surface is related to alpha-shapes from Edelsbrunner and Muecke and it is easy to calculate. Authors demonstrate that the performance is comparable to other more computationally costly surface representations.

Critical strengths

The research on efficient representations for molecular interaction predictions is valid especially for dynamic scenarios. The new representation is easy to compute as compared to other methods.

Critical weaknesses

This paper presents the alpha-shapes with all the topological theory which effectively fills up the entire paper and the supplementary material. While all this theory might be nice and elegant, I do not think it is necessary at all. Because authors use only alpha=0 everywhere (except of the augmentation) and all what the surface does is to connect all the most outer atom centers into a mesh in case their SAS spheres intersect. I would argue that this is not really a protein surface, it is a protein sub-surface. Therefore authors need to artificially inflate it so that the binding experiment delivers plausible results. Another problem is that the surface sometimes in thin area collapses into a single triangle in the double sided collapse. Then the fix is to offset it, so that it is no longer a valid alpha-surface, to prevent the diffusion to go from one side of the surface to another. Therefore I see the contribution as a nice engineering work to develop a simple surrogate of a real molecular surface such that several computations can be made faster. It is a nice workshop paper, which it was already at ICLR 2026. But this whole alpha-shape theory just obfuscates the simplicity for no apparent reason. Authors do not provide any evidence that this whole alpha-shape framework presentation is insightful in the context of this paper at all. To make this work publishable at ICLR it would need to be entirely rewritten.

Another weakness is that the "preprocessing" substantially penalizes the overall performance and the technique's computational efficiency is overstated.

Questions for authors

I have concerns about what the fins and "no tufting" reveals. The overall performance without tufting experiences a negligible drop. If that is so, why using Laplace-Bertrami operator that should operate on 2-manifolds? Does this perhaps mean that we can use a 3D spatial proximity graph and run a standard graph neural network directly on those 3D coordinates and forget about the surface concept entirely?

6 comments
AuthorsOpenReview ↗

Answer to reviewer 2 (1/3)

Thank you for your detailed comments. We would like to draw your attention to the general answer, in which we resume the additional results obtained for the rebuttal. Concerning your points, we added a control replacing the surface encoder with a GVP-GNN on the same vertices, corrected our PINDER tufting ablation, quantified the effect of tufting on non-manifold defects, reported new results on molecular dynamics (MISATO), and reorganized the first section of the paper. In addition, we provide a point-by-point answer to your comments below.

W1.

W1. This paper presents the alpha-shapes with all the topological theory which effectively fills up the entire paper and the supplementary material. While all this theory might be nice and elegant, I do not think it is necessary at all. Because authors use only alpha=0 everywhere (except of the augmentation) and all what the surface does is to connect all the most outer atom centers into a mesh in case their SAS spheres intersect.

A1. We may not have made this point clear enough: using alpha-shapes with alpha=0 is the optimal way to connect atoms in a way compatible with classical molecular surfaces. We have clarified this point and made explicit the role of alpha-shapes in the Abstract and contributions.

AuthorsOpenReview ↗

Answer to reviewer 2 (2/3)

W2 / Q1.

W2. I would argue that this is not really a protein surface, it is a protein sub-surface. Therefore authors need to artificially inflate it so that the binding experiment delivers plausible results. Another problem is that the surface sometimes in thin area collapses into a single triangle in the double sided collapse. Then the fix is to offset it, so that it is no longer a valid alpha-surface, to prevent the diffusion to go from one side of the surface to another. Therefore I see the contribution as a nice engineering work to develop a simple surrogate of a real molecular surface such that several computations can be made faster. It is a nice workshop paper, which it was already at ICLR 2026. But this whole alpha-shape theory just obfuscates the simplicity for no apparent reason. Authors do not provide any evidence that this whole alpha-shape framework presentation is insightful in the context of this paper at all. To make this work publishable at ICLR it would need to be entirely rewritten.

Q1. I have concerns about what the fins and "no tufting" reveals. The overall performance without tufting experiences a negligible drop. If that is so, why using Laplace-Bertrami operator that should operate on 2-manifolds? Does this perhaps mean that we can use a 3D spatial proximity graph and run a standard graph neural network directly on those 3D coordinates and forget about the surface concept entirely?

A-W2-Q1. We agree that the presentation of the theory may have obscured the main message, and we have reorganized the first section to start with learning rather than surface representations. That said, we would like to clarify a few points:

  • Sub-surface and offset. We define a simplicial complex, the boundary of the alpha-complex, rather than a surface. Theorem 1 proves that its offset is isotopic to the molecular surface: our complex is the definition domain of the molecular surface.
  • Usefulness of the theory. The exhaustive classification of defects (Appendix C) directly led to tufting (defect 5), which yields consistent gains on PINDER-Pair and MISATO (Appendix I.7), and to the linking-edge fix (defect 2, Appendix E), which brings usability to 100%. To our knowledge, no other algorithm offers such a characterization.
  • Laplace-Beltrami operator. Tufting fixes 95% of singular faces and reduces finned edges (shared by three or more triangles) from 5.4% to 0.17% of all edges on Pinder, which justifies using the Laplace-Beltrami operator on our complex.
  • Tufting. A bug in our SLURM launch script prevented tufting from being applied on PINDER only, please excuse us for that. With the corrected PINDER runs (Figures 6-7 and Appendix I.7 updated), tufting improves all three splits. This reinforces our tufting results, whose effect is even clearer in the more detailed MISATO ablation, which was not affected by the bug:
PINDER-Pair AUROC (%)HoloApoAF2
No tufting92.12±0.1886.40±0.1385.46±0.22
Tufting92.36±0.0386.67±0.0585.96±0.32
Tufting + linking edges92.35±0.0386.72±0.2285.89±0.43
MISATO (%)TuftingF1AUROCAUPRC
Holo-conf.No47.8±0.784.8±0.446.2±0.2
Yes48.9±0.585.7±0.347.7±0.8
Random-frameNo48.9±0.486.2±0.548.0±0.6
Yes50.3±0.286.8±0.150.4±0.5

Encoder ablations

  1. GNN on 3D coordinates. AtomSurf already includes a GVP-GNN for graph–surface communication, so the question is whether the surface encoder adds information. We replaced DiffusionNet with a GVP-GNN over a 16-NN graph on the same AlphaSurf vertices, inspiring ourself from S3F that uses a GVP-GNN on a point of cloud to encode a protein' surface. We use S3F's edge features (a 16-dimensional RBF of the edge length together with the edge vector) and removing the geometric surface inputs (curvature, HKS), any other parameter/input/hyperparameter is the same as when using DiffusionNet.

  2. dMaSIF on its own point clouds. We replaced the surface branch with dMaSIF, following the dMaSIF repository setup. On PINDER-Pair, the inefficient forward+backward pass of KeOps [1] (>300 h per seed) required restricting each point to its 32 nearest neighbours, the largest tractable value. Even with this restriction the runs remain too compute heavy, so the PINDER numbers are reported for a single seed. Performance drops on every benchmark: |Surface branch|MaSIF-Lig. (%)|Holo|Apo|AF2| |-|-|-|-|-| |AlphaSurf + DiffusionNet|84.1±2.6|92.35±0.03|86.72±0.22|85.89±0.43| |GVP-GNN (16-NN)|76.3±2.9|90.78±0.09|84.86±0.18|83.45±0.17| |dMaSIF (32-NN)|81.16±2.69|90.44|84.14|82.82|

The GVP-GNN control has 10% more parameters than DiffusionNet and dMaSIF 1.0% more in total parameters, so the drops are not a capacity effect. This shows that using a mesh based encoder with Laplace Beltrami operators is still state of the art for protein multimodal learning.

[1] FlashSinkhorn: IO-Aware Entropic Optimal Transport on GPU

AuthorsOpenReview ↗

Answer to reviewer 2 (3/3)

W3.

W3. Another weakness is that the "preprocessing" substantially penalizes the overall performance and the technique's computational efficiency is overstated.

A-W3. Our preprocessing actually improves the overall performance. Surface construction was the bottleneck: AlphaSurf makes it 10× faster, giving a 2-3× end-to-end speedup (Figure 6) that permits computing surfaces on the fly, which matters for the dynamic scenarios the reviewer mentioned. Beyond speed, not all fast representations are equal: on MISATO with random-frame training, AlphaSurf clearly outperforms the similarly fast EDTSurf (gs = 0.4). This could stem from EDTSurf oversmoothing small pockets. Moreover, EDTSurf benefits less from dynamic augmentions.

Holo-conformation training

SurfaceF1AUROCAUPRC
EDTSurf (gs = 0.4)42.9 ± 0.682.3 ± 0.840.3 ± 0.9
AlphaSurf48.9 ± 0.585.7 ± 0.347.7 ± 0.8

Random-frame training

SurfaceF1AUROCAUPRC
EDTSurf (gs = 0.4)43.4 ± 0.483.0 ± 0.441.4 ± 0.6
AlphaSurf50.3 ± 0.286.8 ± 0.150.4 ± 0.5

Binding-site prediction on MISATO (%, mean ± std), evaluated on holo conformations. Bold: best surface in each training setting.

We have tried to address all of the comments and questions by the reviewer, and thank the reviewer again for helping us improve our submission. We invite the reviewer to also consider the general feedback we provided in our general comment to all reviewers. We will make sure to address all comments or concerns in the potential final version of our paper.

Reviewer_nhL2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose to use the boundary of alpha complex of atoms to build the protein surface mesh. The authors build upon AtomSurf model and applied the proposed mesh idea and show that AlphaSurf is competitive with the proposed meshes. The authors showed three use cases: a smaller point cloud for S3F, coordinate-noise augmentation on PINDER-Pair, and training on molecular dynamics frames of MISATO.

Critical strengths
  1. The idea is simple and, as far as I know, new for this field: use the alpha complex boundary itself as the learning mesh. It removes the external surface program and the resolution parameter, the vertices map one to one to atoms, and the mesh can be rebuilt every time the atoms move.
  2. The mesh does what the paper says. Usable meshes are close to 100%, while NanoShaper at the same size gives less than 70%.
  3. The comparison in Fig. 6 is well designed: same model, seven mesh settings, three seeds, and one common test set. The paper also reports results that go against it, for example that low-resolution EDTSurf matches its size and speed.
  4. The heat diffusion on AlphaSurf is closer to the full-resolution reference than on EDTSurf, its equal-cost competitor.
  5. Some side results are solid: training on molecular dynamics frames helps on MISATO (Table 2), and the alpha-complex point cloud works in S3F with about 7 times fewer points (Table 1).
Critical weaknesses
  1. The MSMS baseline in Fig. 6 is much worse than its published result. The paper says it follows AtomSurf and that "Input features and hyperparameters are retained without modification, except for the surface mesh used". So the MSMS plus coarsening should reproduce AtomSurf. In Fig. 6 this experiment matches on Holo but is about 5 to 15 points lower on Apo and about 15 points lower on AF2. Appendix G lists several changes but does not say which of them differ from AtomSurf.

This matters because Fig. 6 is the only experiment that compares AlphaSurf with other meshes in the same model. Against the authors' own baseline, AlphaSurf is up to about 15 points better on Apo and AF2. Against the published AtomSurf numbers, it is about 1 to 2 points worse on all three splits. Both readings are "competitive", but a reader cannot tell which one is true. An ablation that finds the cause, or both model run in the published AtomSurf setup, would settle it.

  1. The gain is modest once the claims are stated carefully. Total preprocessing is 1.9x faster on PINDER and 3.1x on MaSIF-Ligand, not the tenfold the abstract suggests, because computing the spectral operators takes about three quarters of the time and is not changed. EDTSurf at grid scale 0.4 costs the same on PINDER, can also run on the fly, and is within about one AUROC point of AlphaSurf in five of the six PINDER-Pair panels of Fig. 6. The use-case gains are 0.5 to 0.9 AUROC points and 1.4 F1 points. I still think the contribution is useful, but at this size it cannot carry the open problem in weakness 1.

  2. Theorem 1 is presented as proved, but there is no 3D proof. Appendix B proves only a 2D version of the first step, gives a sketch of the second step. I am not saying the theorem is wrong. But it should be presented as a 2D result plus a conjecture, or the full 3D proof should be given. No experiment depends on it, so this does not change my rating on its own.

  3. It is not clear whether the gain from noise augmentation actually comes from the surface. In Fig. 7, noise is added to both the graph and the mesh, while the reported AUROC comes only from the graph head. As the paper itself notes, adding coordinate noise to graph models is already common. Without an experiment that adds noise only to the graph while keeping the surface clean, it is hard to tell whether the improvement actually depends on the on-the-fly surfaces. That distinction matters because the abstract presents this as one of the benefits of generating surfaces on the fly.

Questions for authors
  1. Can you give the detailed settings and setup for AtomSurf and explain the difference between the results?

  2. How were the values in fig 6 computed, is there any task where AlphaSurf is clearly better than EDTSurf at grid scale 0.4?

  3. Can you give the full proof for 3D including the assumptions?

  4. In Fig. 7 you add noise to the atom coordinates and rebuild the surface from the noisy atoms. Could you add one run where only the graph gets the noise and the surface is built from the clean atoms? Also, your graph encoder adds 0.5 Å noise: is that noise on or off in the Fig. 7 baseline?

5 comments
AuthorsOpenReview ↗

Answer to reviewer 1 (1/3)

Thank you for your detailed comments. We would like to draw your attention to the general answer, in which we resume the additional results obtained for the rebuttal. Concerning your points, we identified the cause of the MSMS gap and retrained it with random rotations, corrected our PINDER runs (where tufting had not been applied), added the full 3D proof of Theorem 1, and ran the graph-only noising control you suggested. In addition, we provide a point-by-point answer to your comments below.

W1 / Q1.

W1. The MSMS baseline in Fig. 6 is much worse than its published result. The paper says it follows AtomSurf and that "Input features and hyperparameters are retained without modification, except for the surface mesh used". So the MSMS plus coarsening should reproduce AtomSurf. In Fig. 6 this experiment matches on Holo but is about 5 to 15 points lower on Apo and about 15 points lower on AF2. Appendix G lists several changes but does not say which of them differ from AtomSurf.

This matters because Fig. 6 is the only experiment that compares AlphaSurf with other meshes in the same model. Against the authors' own baseline, AlphaSurf is up to about 15 points better on Apo and AF2. Against the published AtomSurf numbers, it is about 1 to 2 points worse on all three splits. Both readings are "competitive", but a reader cannot tell which one is true. An ablation that finds the cause, or both model run in the published AtomSurf setup, would settle it.

Q1. Can you give the detailed settings and setup for AtomSurf and explain the difference between the results?

A-Q1. The AtomSurf repository does not include the training setup for PINDER, so we contacted the authors and used their setup in the results shown in the paper; we use the hyperparameters highlighted in Appendix H Hyperparameters. However, we identified the cause of the Apo and AF2 gap: the AtomSurf surface encoder receives normals as input, so it is not rotation-invariant. With MSMS meshes, the training loss sometimes drops sharply at a given step: the model learns a shortcut that aligns the normals of the two partners, which overfits to Holo and fails on Apo and AF2. This collapse does not occur in every run, and AtomSurf reports a single seed, which likely explains why it is absent from the published results. Moreover, even for the other benchmarks where the code is available, rotations were not used. We argue that this happens to MSMS and not to EDTSurf or AlphaSurf because the MSMS surface, even coarsened, has fine-grained geometric details that permit this shortcut.

Following DiffusionNet, we now apply independent random rotations to each protein during training for the MSMS retraining; AlphaSurf was not affected by this shortcut. MSMS is now mostly closing the gap with the official AtomSurf numbers which provides only one seed (AtomSurf-noESM 93.0 / 87.6 / 87.0 AUROC on Holo / Apo / AF2). Figure 6 has been updated with those new results. MSMS still obtains the strongest results on Holodimers, though closer to AlphaSurf. For the APO and AF2 conformations tests, both methods perform comparably on Homodimers and Alphasurf still obtains the strongest results on Heterodimers.

Homodimers

PINDER-Pair AUROC (%)HoloApoAF2
MSMS no rotations94.84 ± 0.1173.17 ± 0.7371.38 ± 5.03
MSMS + random rotations93.49 ± 0.0786.99 ± 0.1486.74 ± 0.07
AlphaSurf92.85 ± 0.0586.83 ± 0.2186.60 ± 0.38

Heterodimers

PINDER-Pair AUROC (%)HoloApoAF2
MSMS no rotations92.21 ± 0.1182.25 ± 2.1571.57 ± 1.68
MSMS + random rotations90.20 ± 0.0985.23 ± 0.3880.24 ± 0.35
AlphaSurf89.53 ± 0.1285.79 ± 0.3281.36 ± 0.72
AuthorsOpenReview ↗

Answer to reviewer 1 (2/3)

W2 / Q2.

W2. The gain is modest once the claims are stated carefully. Total preprocessing is 1.9x faster on PINDER and 3.1x on MaSIF-Ligand, not the tenfold the abstract suggests, because computing the spectral operators takes about three quarters of the time and is not changed. EDTSurf at grid scale 0.4 costs the same on PINDER, can also run on the fly, and is within about one AUROC point of AlphaSurf in five of the six PINDER-Pair panels of Fig. 6. The use-case gains are 0.5 to 0.9 AUROC points and 1.4 F1 points. I still think the contribution is useful, but at this size it cannot carry the open problem in weakness 1.

Q2. How were the values in fig 6 computed, is there any task where AlphaSurf is clearly better than EDTSurf at grid scale 0.4?

A-Q2. For Figure 6, the protocol is highlighted in Appendix H Hyperparameters.

Coarsened MSMS is dominated by surface construction, not by the operators. AlphaSurf removes this step, which moves the bottleneck to the spectral computations. With 20 data-loading workers, the end-to-end speedup is 2-3× (Figure 6). This speedup is what permits computing the surfaces on the fly during training.

The revised version adds three results where AlphaSurf is clearly better than EDTSurf at grid scale 0.4:

  1. With the singular-edge fix (Appendix E), AlphaSurf usability reaches 100% (EDTSurf fails on some proteins).
  2. We found an error in our SLURM PINDER launch script that didn't enable the tufting to be applied for this specific dataset, please excuse us for that. We have updated Figures 6 and 7 with the new results, and we now see that AlphaSurf also clearly outperforms EDTSurf on Apo-Hetero and AF2-Hetero. The Appendix considering Tufting (Appendix I.7 Mesh Ablation) has also been updated, showing that tufting matters also for PINDER-Pair.
PINDER-Pair AUROC (%)HoloApoAF2
No tufting92.12 ± 0.1886.40 ± 0.1385.46 ± 0.22
Tufting92.36 ± 0.0386.67 ± 0.0585.96 ± 0.32
Tufting + linking edges (updated algorithm)92.35 ± 0.0386.72 ± 0.2285.89 ± 0.43
SurfaceApo HeteroAF2 HomoAF2 Hetero
EDTSurf (gs = 0.4)84.29 ± 0.4886.22 ± 0.1278.18 ± 0.35
AlphaSurf85.79 ± 0.3286.60 ± 0.3881.36 ± 0.72

PINDER-Pair AUROC (%), mean ± std. Full results in Figure 6.

  1. On MISATO with random-frame training, EDTSurf meshes lower performance by 6.9 F1, 3.8 AUROC and 9.0 AUPRC points relative to AlphaSurf. This could stem from EDTSurf oversmoothing small pockets. Moreover, EDTSurf benefits less from dynamic augmentions.

Holo-conformation training

SurfaceF1AUROCAUPRC
EDTSurf (gs = 0.4)42.9 ± 0.682.3 ± 0.840.3 ± 0.9
AlphaSurf48.9 ± 0.585.7 ± 0.347.7 ± 0.8

Random-frame training

SurfaceF1AUROCAUPRC
EDTSurf (gs = 0.4)43.4 ± 0.483.0 ± 0.441.4 ± 0.6
AlphaSurf50.3 ± 0.286.8 ± 0.150.4 ± 0.5

Binding-site prediction on MISATO (%, mean ± std), evaluated on holo conformations. Bold: best surface in each training setting.

AuthorsOpenReview ↗

Answer to reviewer 1 (3/3)

W3 / Q3.

W3. Theorem 1 is presented as proved, but there is no 3D proof. Appendix B proves only a 2D version of the first step, gives a sketch of the second step. I am not saying the theorem is wrong. But it should be presented as a 2D result plus a conjecture, or the full 3D proof should be given. No experiment depends on it, so this does not change my rating on its own.

Q3. Can you give the full proof for 3D including the assumptions?

A-Q3. The appendix now provides the full 3D proof (Appendix B Isotopy).

W4 / Q4.

W4. It is not clear whether the gain from noise augmentation actually comes from the surface. In Fig. 7, noise is added to both the graph and the mesh, while the reported AUROC comes only from the graph head. As the paper itself notes, adding coordinate noise to graph models is already common. Without an experiment that adds noise only to the graph while keeping the surface clean, it is hard to tell whether the improvement actually depends on the on-the-fly surfaces. That distinction matters because the abstract presents this as one of the benefits of generating surfaces on the fly.

Q4. In Fig. 7 you add noise to the atom coordinates and rebuild the surface from the noisy atoms. Could you add one run where only the graph gets the noise and the surface is built from the clean atoms? Also, your graph encoder adds 0.5 Å noise: is that noise on or off in the Fig. 7 baseline?

A-Q4. Our mention that ProNet used noise was in reference to the practice of graph augmentation, hence the noise was off in our baselines. We added the graph-only control at each noising scale (Figure 7 updated). The resulting performances are slightly worse than the joint noising of the graph and surface, we believe that overall the joint noising permitted by the AlphaSurf method overperforms the graph as it holds for each modality the best overall AUC, especially on APO and AF2 which are important problems.

Holo (no noise: 92.35 ± 0.03)

σGraph onlyJoint
0.192.40 ± 0.0992.32 ± 0.09 ↓
0.392.51 ± 0.1492.62 ± 0.03
0.592.32 ± 0.15 ↓92.42 ± 0.12
0.7591.95 ± 0.04 ↓91.94 ± 0.22 ↓
1.091.59 ± 0.28 ↓90.95 ± 0.02 ↓

Apo (no noise: 86.72 ± 0.22)

σGraph onlyJoint
0.186.68 ± 0.16 ↓86.64 ± 0.19 ↓
0.387.02 ± 0.0187.06 ± 0.24
0.586.99 ± 0.2587.02 ± 0.15
0.7586.76 ± 0.0587.15 ± 0.09
1.086.56 ± 0.11 ↓86.47 ± 0.25 ↓

AF2 (no noise: 85.89 ± 0.43)

σGraph onlyJoint
0.185.92 ± 0.1886.05 ± 0.27
0.386.03 ± 0.3986.40 ± 0.31
0.585.84 ± 0.18 ↓86.18 ± 0.34
0.7585.33 ± 0.17 ↓86.00 ± 0.15
1.085.00 ± 0.08 ↓85.31 ± 0.16 ↓

PINDER-Pair AUROC (%), mean ± std. Bold: better of graph-only and joint at each noise level. ↓: below the no-noise baseline.

We have tried to address all of the comments and questions by the reviewer, and thank the reviewer again for helping us improve our submission. We invite the reviewer to also consider the general feedback we provided in our general comment to all reviewers. We will make sure to address all comments or concerns in the potential final version of our paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General answer

We thank reviewers for their constructive comments and feedback. We acknowledge the fact that we did not put the emphasis on the key novelties of our approach, which calls for the following two general comments.

  1. Learning on protein surfaces and the alpha-shape construction

Learning on protein surfaces is a powerful paradigm operating in two steps: producing a geometric representation and using it to learn a model for a given task. While previous work has mostly focused on triangle meshes for the former, we use alpha-shapes to produce a simpler geometric model in two steps.

First, we show that the boundary of the alpha-shape, which is the simplest geometric complex coding the molecular surface, has an offset which is isotopic to this molecular surface.

Second, in search for a representation striking a balance between geometric accuracy and computational complexity, we derive a simplicial complex termed the 'tufted complex' from the boundary of the alpha-shape; the tufted complex has linear size (whence optimal) and exhibits learning performances on par with SOTA methods.

We also updated our surface creation algorithm to obtain 100% validity by fixing our only failure cases highlighted in the paper, more details can be seen in Appendix E Singular Edges.

  1. Significance of our contribution

Prior to our work, the bottleneck for learning on surfaces was the computation of the geometric representation. The 10x factor claimed in the main text relates to the creation of this representation, the gain for the entire pipeline being 2-3x (Figure 6). This end-to-end speedup is what permits computing the surfaces on the fly during training.

Since, as explained above, our encoding has linear size and is the simplest possible (constant omitted), our work implies that cutting down the preprocessing time will now be very challenging. Further work may focus on alternative post-processing / learning representations, spectral operators being of special interest since they currently define the bottleneck.

Summarizing, we believe that four facts warrant the adoption of our work:

  1. our geometric encoding is optimal complexity-wise when surfaces are used,
  2. it is parameter free and is 100% robust,
  3. it achieves learning performances on par with SOTA methods.
  4. opens the use of surface based method for large scale and dynamic data.

To clarify our contribution, we have re-organized the first section, to start with learning rather than surface representations and stress the points just discussed. Beyond this reorganization, the revised version includes the following changes and new results, each detailed in our answers to the reviewers:

  • Singular-edge fix, giving 100% validity of our representation (Appendix E).
  • Corrected PINDER runs, where a launch-script bug had disabled tufting (Figure 6 and Appendix I.7 updated).
  • MSMS and NanoShaper baseline retrained with random rotations, mostly closing the Apo/AF2 gap with AtomSurf.
  • Full 3D proof of the isotopy theorem (Appendix B).
  • Graph-only noising control (Figure 7).
  • Control replacing the surface encoder with a GVP-GNN on the same vertices or using the dMaSIF pipeline.
  • MISATO random-frame training, where AlphaSurf clearly outperforms EDTSurf.
  • Clarified speedups: 10× for building the representation, 2-3× end-to-end (Figure 6).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.