acceptodds
Under review as a conference paper at ICLR 2027

From Surfaces to Volumes: Registered Geometry for Protein Representation Learning

Abstract

Existing protein geometry models typically represent molecular surfaces using local geometric features such as sampled points, normals, and curvature. While effective for capturing exposed molecular shape, these representations do not explicitly model the volumetric organization beneath the surface or provide a consistent coordinate system for residue-wise volumetric structure. We introduce Protein-TetSphere, a registered residue-wise volumetric representation for proteins. Each protein chain is tetrahedralized to obtain local volumetric regions associated with individual residues, which are then registered to a shared fixed-topology tetrahedral reference and represented in a common Laplacian basis. This registration establishes consistent volumetric coordinates across residues, enabling local three-dimensional deformation to be integrated with surface and chemical information in a multimodal protein representation. We evaluate Protein-TetSphere on ligand-binding pocket classification, protein–protein interface prediction, and de novo protein binder design. Across the three tasks, Protein-TetSphere improves ligand-binding pocket balanced accuracy from to , Pinder-Pair/Site AUROC from to , and binder-design success from to on the BoltzGen Challenge Set and from to at the ProtDBench backbone level. These results show that registered volumetric geometry provides complementary spatial information beyond molecular surfaces across protein recognition, interaction, and design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein representation learning, volumetric geometry, molecular geometry, geometric deep learning, protein design, protein-protein interface, ligand-binding pocket

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

This work introduces Protein-TetSphere, a registered residue-wise volumetric representation for proteins. Specifically, it registers residue-wise protein volumes to a shared tetrahedral reference and encodes them in a common Laplacian basis, which can be viewed as a geometric representation learning strategy. The authors evaluate the proposed representation on three tasks, including ligand-binding pocket classification, protein–protein interface prediction, and de novo protein binder design, and report improvements across all three tasks. However, the baseline comparisons are limited. In particular, for ligand-binding pocket classification and protein–protein interface prediction, AtomSurf is the only main baseline, which provides relatively weak evidence for the overall competitiveness of the proposed method.

Critical strengths
  1. Protein-TetSphere explicitly models the three-dimensional volumetric organization within and around protein residues. The motivation is clear, and the idea is interesting.
  2. The method is evaluated on three protein representation-related tasks, providing relatively broad task coverage.
Critical weaknesses
  1. The number of baselines is limited. For pocket classification and PPI prediction, the comparisons mainly include AtomSurf, Ours w/o TetSphere, and Ours, although the related work discusses several relevant geometric models, such as MaSIF, dMaSIF, HMR, ScanNet, and PeSTo. The matched comparisons help demonstrate the additional value of TetSphere, but they do not fully establish the competitiveness of Protein-TetSphere against a broader range of existing methods.
  2. The paper presents Protein-TetSphere as a general representation learning approach, but separate pretrained models are trained for different downstream tasks, rather than learning a single protein representation that can be directly transferred across tasks.
  3. The construction of Protein-TetSphere appears to involve relatively high computational cost. The authors could provide further analysis and clarification on its computational efficiency and scalability.
Questions for authors
  1. The authors should clarify the rationale behind the baseline selection and more clearly explain the key methodological innovation that distinguishes Protein-TetSphere from the original TetSphere representation.
  2. The computational cost and time complexity should be reported more explicitly. An effective method should not introduce substantial computational overhead for only limited performance gains, especially for proteins, where 3D structural information is already computationally expensive to process.
  3. Protein binder design is one of the key evaluation tasks, but the paper does not provide a clear case study demonstrating that the proposed strategy produces better binders than the baseline methods.
2 comments

W1 / Q1 — Baseline selection and relationship to TetSphere Splatting

Our method improves all three prediction metrics over both AtomSurf and the matched control without volumetric input, with the released backbone held fixed (Appendix C; Tables 1–2, mean ± SD over five seeds). These results show that registered volumetric geometry provides complementary information. We deliberately select strong backbones: AtomSurf reports state-of-the-art MaSIF-ligand balanced accuracy of 0.88 (0.84 without ESM; Mallet et al., 2025), and BoltzGen offers an open all-atom generator with experimentally validated binders in its original study (Stark et al., 2025). Our adaptation of TetSphere Splatting establishes shared residue-wise coordinates: one fixed-topology template preserves vertex correspondence; bidirectional point-to-point fitting to local surfaces and point-to-triangle fitting to chain-derived residue regions replace rendering losses; two-stage optimization adds per-update Gauss–Seidel volume repair. A common Laplacian basis then provides comparable deformation descriptors for learning (Sections 3.1–3.2, Appendix A.3).

Task and metricAtomSurfOurs w/o TetSphereOurs
MaSIF-ligand balanced accuracy0.795 ± 0.0050.807 ± 0.0030.826 ± 0.004
Pinder-Pair AUROC0.914 ± 0.0020.919 ± 0.0020.932 ± 0.001
Pinder-Site AUROC0.852 ± 0.0020.855 ± 0.0010.866 ± 0.001

W2 — Task-specific pretraining and transfer

We demonstrate the same representation's utility in pocket classification, interface prediction, and binder design. Task-specific pretraining is a deliberate safeguard against leakage. For the two prediction tasks, we pretrain only on their training splits and select checkpoints on validation data, excluding test data from both. Pocket pretraining masks ligand identity and disables ligand-related targets; interface pretraining masks cross-chain geometry to prevent direct label leakage. These strict controls address the risk identified by Hermann et al., 2024. Protocols appear in Sections 2 and 4.1 and Appendices B–D.

W3 / Q2 — Construction cost and scalability

  • No added geometry computation at generation. Binder generation follows BoltzGen without volumetric fitting or the geometry teacher; geometric supervision is incorporated during training (Section 3.4).
  • Efficient, reusable preprocessing. Geometry is precomputed once and reused for the same structures. Mean preprocessing worker elapsed time is 28.98 s per chain; fitting averages only 3.65 GPU-s per chain, amortized over batches. Profiling used a Xeon Platinum 8481C, 32 CPU workers, and one H100 80 GB (Appendix H).
  • High reliability. Overall preprocessing success is 99.98% across the three tasks, with segment retries, local-surface fallback, and residue-map and positive-volume checks.

Table 8: Mean per-chain time and CPU use. GPU time is amortized over fitting batches.

Measure and unitMean
Preprocessing worker elapsed (s)28.98
Preprocessing (CPU core-s)39.91
Fitting batch amortization (GPU-s)3.65

Table 9: Medians by residue count. A separate timing set supplies estimates with GPU batch time allocated by residue count.

Residues per chainPreprocessing (CPU core-s)Fitting estimate (GPU-s)
<25624.542.21
256–51156.075.01
512–102378.975.16
≥1024167.247.52

Q3 — Binder case study

Under the same selection rule, our selected PDL1 binder scores better than +Cont. on all three AF2 interface and geometry metrics (Figure 14, Appendix I.4). In each arm we select the highest interface pTM among ProtDBench AF2-IG Easy passers across evaluated lengths, breaking ties by lower interface PAE, candidate identifier, then sequence index. Both complexes use the same target alignment, view, and scale; structures and metrics come from ProtDBench AF2.

Selected PDL1 candidate+Cont.Ours
Interface pTM ↑0.860.87
Interface PAE (Å) ↓3.843.63
Bound–unbound RMSD (Å) ↓0.320.16

Our method also achieves higher backbone success when all evaluated PDL1 lengths are pooled (Table 10):

PDL1 endpoint+Cont. (%)Ours (%)Difference [95% CI] (pp)
Backbone success40.2143.96+3.75 [−1.88, 9.17]

We plan to contact wet-lab collaborators to validate binding and measure the affinities of the designed candidates.

The authors claim substantial improvements across three related tasks. However, for the first two tasks, AtomSurf is the only backbone baseline considered. This makes it difficult to determine whether the improvements reflect the effectiveness of the proposed method or simply address limitations specific to AtomSurf. The authors' response has not fully resolved this concern. Therefore, I will maintain my original rating.

Reviewer_Qp1zOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, authors propose Protein-TetSphere, which is local MSMS surface and a chain level fTetWild mesh are fitted by fixed-topology tetrahedral template. The authors evaluated their approach on pocket classification, interface prediction and binder design. They show that their approach provides gain on the evaluated tasks in this work.

Critical strengths
  1. Carrying the representation into generation is a crucial result since binder geometry does ot exists before the design. The challenge set success improves from 14.95% to 19.90%, which strengthens the paper.
  2. The authors evaluated the binder design in four arms. Separating the continued training from alignment makes the improvements interpretable.
Critical weaknesses
  1. The sole representation level control deletes the volumentric input rather than contrasting shared mode indexing with an unregistered or shuffled one. Figure 3 only shows truncated-basis reconstruction error. A shuffled-index control and an M sweep would be needed to verify.
  2. Table 1 reports AtomSuff at 0.795 and full model at 0.826, while AtomSurf's own ESM-free ablation (arXiv:2309.16519) provides 0.84 on seemingly same split and the gap is not marginal (0.045).
Questions for authors
  1. Please see W2, what is the reason. behind this gap?
  2. Can conduct a paired per target test over the ten targets, or repeated fine-tuning runs per arm? (Figure 6)
  3. Can authors compare with an off-the-shelf sequence or structure encoder used as alignment teacher at the same weight? Only alternative teacher on Table 3 is the input ablated twin of the same pretrained model.
2 comments

W1 — Shuffled-index control and basis-size sensitivity

Our method outperforms shuffled indices on all six metrics below, supporting the importance of shared mode correspondence in the registered volumetric representation. Shuffling spectral indices during feature export preserves the volumetric input while disrupting its coefficient-to-mode correspondence. ProtDBench backbone success and Challenge Set final success fall by 10.38 and 6.15 percentage points (pp). Reducing M from 64 to 16 in separately pretrained models with otherwise identical settings lowers these endpoints by 0.42 and 1.30 pp. Tet Simple replaces Laplacian features with four scalar descriptors (Appendix G). Tables 4–5 are reproduced below; prediction scores are mean ± SD, and binder rates use bootstrap SD. Appendix F reports the complete binder screening results.

Task and metricShuffled indicesTet SimpleM = 16Full (M = 64)
MaSIF-ligand balanced accuracy0.774 ± 0.0060.811 ± 0.0040.817 ± 0.0030.826 ± 0.004
Pinder-Pair AUROC0.876 ± 0.0030.917 ± 0.0020.923 ± 0.0010.932 ± 0.001
Pinder-Site AUROC0.824 ± 0.0020.853 ± 0.0010.859 ± 0.0010.866 ± 0.001
ProtDBench backbone success (%)21.81 ± 0.4829.62 ± 0.5631.77 ± 0.5532.19 ± 0.57
ProtDBench sequence success (%)9.94 ± 0.2814.40 ± 0.3514.08 ± 0.3314.67 ± 0.34
BoltzGen Challenge Set final success (%)13.75 ± 0.7615.75 ± 0.7818.60 ± 0.8319.90 ± 0.84

W2 / Q1 — AtomSurf reproduction gap

The cited 0.84 and our baseline use the same standard MaSIF-ligand split but different implementations. The AtomSurf paper describes six width-128 DiffusionNet–ProNet–GVP blocks; the public backbone uses three ProNet/GVP blocks, and we found no public six-block configuration. Its Adam configuration specifies β₂ = 0.99, while the optimizer constructor uses the default 0.999. We reconstructed the six-block model, matched its reported parameter count, and trained it five times with the same hyperparameters; loss diverged at epoch 7 in every run. We therefore use the released three-block backbone for all MaSIF-ligand variants, with 200 epochs and explicit β₂ = 0.99 (Appendix C). Our method improves performance under this identical backbone and downstream protocol.

ConfigurationBalanced accuracy
AtomSurf paper, ESM-free0.84
Released AtomSurf backbone0.795 ± 0.005
Same backbone + our method0.826 ± 0.004

Q2 — Per-target comparison and repeated fine-tuning

Ours improves success on 9/10 ProtDBench targets and 7/10 Challenge Set targets. Rates are %, and differences are Ours minus +Cont. in pp with pointwise 95% bootstrap CIs (Appendices E and I). Δ uses unrounded rates. We will add fine-tuning repeats with independent training seeds in the final version.

ProtDBench: backbone success (Table 10).

TargetOurs+Cont.Δ [95% CI]
BHRF159.3853.47+5.903 [−1.39, 13.54]
H116.3628.49−12.132 [−16.18, −8.09]
IL17A3.451.97+1.480 [−0.33, 3.29]
IL7RA38.3329.79+8.542 [2.92, 13.55]
IR51.8451.47+0.368 [−4.78, 5.70]
PDL143.9640.21+3.750 [−1.88, 9.17]
SC2RBD9.729.03+0.694 [−4.17, 5.21]
TNFa6.253.33+2.917 [0.21, 5.42]
TrkA64.7957.92+6.875 [1.04, 12.50]
VEGFA35.8624.51+11.349 [6.41, 16.12]
Overall32.1929.29+2.896 [1.38, 4.35]

BoltzGen Challenge Set: final success (Table 12).

TargetOurs+Cont.Δ [95% CI]
1G135.506.50−1.000 [−5.50, 3.50]
1JQD13.003.00+10.000 [4.99, 15.50]
1NB048.0027.00+21.000 [12.00, 30.50]
2A1X35.0022.50+12.500 [3.50, 21.00]
2PNY17.509.00+8.500 [2.00, 14.50]
3APU18.0017.00+1.000 [−6.50, 8.01]
3CH49.007.00+2.000 [−3.00, 7.00]
3QKG22.5036.00−13.500 [−22.50, −4.50]
6M1U19.5013.00+6.500 [−0.50, 13.50]
7AAH11.0011.50−0.500 [−6.50, 5.50]
Overall19.9015.25+4.650 [2.40, 6.85]

Q3 — Off-the-shelf alignment teacher

The matched control already isolates the contribution of registered volumetric geometry: only the teacher's volumetric input changes, with teacher architecture, pretraining data, alignment objective and weight, and training schedule fixed (Table 3). Varying M, shuffling indices, and using Tet Simple provide complementary evidence (Tables 4–5). An external teacher would also change the architecture, pretraining data, or objective; equal alignment weights alone would not isolate this contribution.

The authors response attempts to resolve the issues, but main concerns remain regarding to limited comparison and marginal improvement. Therefore, I'm keeping my score.

Reviewer_12buOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the limitation that commonly used protein representations—such as surface features and graph representations—capture outer shapes well but fail to represent how the space beneath the surface is distributed among individual amino acids, they propose Protein-TetSphere. The core idea is to place a uniform tetrahedral framework over every amino acid and approximate its occupied 3D space by adjusting only the vertex positions. Because all amino acids share the identical framework, the structure enables comparative learning. They pre-train this volume-related feature by combining it with existing features and apply it to three downstream tasks, resulting in performance improvements across all of them.

Critical strengths
  • Unlike existing models that mainly use protein surfaces or relationships between atoms, this paper proposes a new way to represent the 3D volume of each amino acid in a consistent format. This approach allows the model to learn additional structural information that existing models may miss.

  • The paper uses well-designed comparison experiments to show that the performance improvements come from the new volumetric information, rather than simply increasing the model size or training longer. By comparing models without TetSphere and models with additional training, the authors demonstrate consistent improvements across three tasks: pocket classification, PPI prediction, and binder design.

  • The authors address the challenge that the 3D structure of a new binder is not available before generation by using volumetric information only during training. As a result, no additional computational cost is required during binder generation. The improvements in structural accuracy also suggest that this approach has potential for practical protein binder design.

Critical weaknesses
  • The main contribution of the proposed method is not fully validated. Previous studies have already shown that protein volume information is useful. The main novelty of this paper is representing different residue volumes in a shared format. However, the authors do not compare their method with simpler and less expensive features such as SASA, residue depth, Voronoi volume, or local packing density. Therefore, it is unclear whether the complex tetrahedral registration is really necessary. In addition, comparisons with recent models are mostly limited to AtomSurf and BoltzGen.

  • The generalizability of the model and its benefits for binder design are not fully demonstrated. The authors train separate models for the three tasks, so it is unclear whether a single pretrained model can be used across different tasks. Although backbone-level success improves on ProtDBench, sequence-level success is slightly lower than the continued-training baseline (14.67% vs. 14.80%). The model without TetSphere also performs better on some diversity metrics. These results raise questions about whether the proposed method consistently improves binder quality and diversity.

  • The computational cost and practical value of the method remain unclear. The improvements in binder design are based on computational screening results, without experimental validation of binding affinity or actual binding success. In addition, generating and fitting the volume representation for each residue may require significant computation, but the authors do not report preprocessing time or fitting failure rates. Therefore, it is difficult to judge whether the performance gains justify the additional complexity and computational cost.

Questions for authors

What are the pocket classification and PPI prediction results when simpler features, such as SASA, residue depth, Voronoi volume, or local atom density, are added to the same model under the same training conditions? The w/o TetSphere baseline only controls for model size. Without comparing these simpler features, it is difficult to tell whether the proposed registration method itself provides any additional benefit.

1 comment

W1 / Q1 — Simpler geometric features and scope of baselines

Our main contribution is a shared, registered residue-wise volumetric representation. We choose AtomSurf for its reported state-of-the-art performance on protein learning benchmarks and its joint surface–graph encoding (Mallet et al., 2025). BoltzGen is an experimentally validated all-atom binder generator with open training code and weights (Stark et al., 2025). Matched controls on these strong backbones test the representation's contribution to prediction and generation. Tet Simple replaces only Laplacian features with SASA, residue depth, Voronoi volume, and local atom density; all other model, training, and evaluation settings remain unchanged (Appendix G). Our method outperforms this control on all three tasks, including both binder benchmarks (Tables 4–5), supporting informative volumetric geometry beyond four scalars. Prediction scores are mean ± SD; binder rates use bootstrap SD.

Task and metricTet SimpleOur method (Full)
MaSIF-ligand balanced accuracy0.811 ± 0.0040.826 ± 0.004
Pinder-Pair AUROC0.917 ± 0.0020.932 ± 0.001
Pinder-Site AUROC0.853 ± 0.0010.866 ± 0.001
ProtDBench backbone success (%)29.62 ± 0.5632.19 ± 0.57
ProtDBench sequence success (%)14.40 ± 0.3514.67 ± 0.34
BoltzGen Challenge Set final success (%)15.75 ± 0.7819.90 ± 0.84

W2 — Generality across tasks, sequence-level success, and structural diversity

Using task-specific checkpoints lets us apply strict leakage controls within a shared representation framework. Pocket and interface pretraining use only their respective training splits; checkpoints are selected on validation data, with test data excluded from both steps. We further mask ligand identity and cross-chain geometry, respectively, to prevent direct label leakage. These safeguards address the pretraining leakage risk highlighted by Hermann et al., 2024.

Against +Cont. (Table 3, Appendix I.2). Our method improves backbone success and Challenge Set final success, with both 95% CIs for the differences above zero, while maintaining comparable sequence-level success in this evaluation. Differences below are Ours minus the control, in percentage points (pp), computed before rounding.

Metric+Cont. (%)Ours (%)Difference [95% CI] (pp)
ProtDBench backbone success29.2932.19+2.90 [1.38, 4.35]
ProtDBench sequence success14.8014.67−0.125 [−0.99, 0.79]
Challenge Set final success15.2519.90+4.65 [2.40, 6.85]

Against the no-TetSphere control (Table 3, Appendix I.3). Our method yields more successful designs on both benchmarks and more structurally distinct successful binders at TM 0.8. Cluster pass rates count successful clusters relative to the full generation budget; all-candidate diversity also counts failed designs (Appendix E). The TM 0.6 results reflect a tradeoff at the coarser clustering resolution.

MetricNo TetSphere (%)Ours (%)Difference (pp)
ProtDBench backbone success26.4632.19+5.73
Challenge Set final success16.9019.90+3.00
ProtDBench cluster pass, TM 0.614.0013.65−0.35
ProtDBench cluster pass, TM 0.822.4825.10+2.62
Challenge Set cluster pass, TM 0.614.7012.25−2.45
Challenge Set cluster pass, TM 0.815.9518.40+2.45

W3 — Computational cost, fitting reliability, and experimental validation

  • Efficient, reusable preprocessing. Geometry is precomputed once and reused. CPU preprocessing averages 28.98 s per chain; batched fitting takes only 3.65 GPU-s per chain on average (batch amortization). Tables 8–9 from Appendix H are reproduced below; profiling used a Xeon Platinum 8481C, 32 CPU workers, and one H100 80 GB.
  • No added geometry computation at generation. Binder generation follows BoltzGen without TetSphere fitting or the geometry teacher; the geometric information benefits the model through training.
  • High reliability. Preprocessing passed for 99.98% across the three tasks. Handling includes segment retries, local-surface fallback, residue-map and positive-volume checks.
  • Future experimental validation. We plan to contact wet-lab collaborators to validate binding and measure the affinities of the designed candidates.

Table 8: Mean per-chain time and CPU use.

Measure and unitMean
Preprocessing worker elapsed (s)28.98
Preprocessing (CPU core-s)39.91
Fitting batch amortization (GPU-s)3.65

Table 9: Medians by residue count. GPU estimates use batch times allocated by residue count.

Residues per chainPreprocessing (CPU core-s)Fitting estimate (GPU-s)
<25624.542.21
256–51156.075.01
512–102378.975.16
≥1024167.247.52

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all three reviewers for their detailed feedback. We summarize our main contributions below.

  • [Motivation] Proteins with similar surfaces can differ in the three-dimensional shape and packing of their residues. We study whether making those differences accessible to models helps explain molecular recognition and interactions.
  • [Method] Our method registers each residue-associated volume to a shared tetrahedral template. The shared topology and vertex ordering let us encode the fitted deformations in a common Laplacian basis and combine the resulting descriptors with existing surface and chemical features.
  • [Experiments] The representation improves ligand-binding pocket classification and protein interface prediction, and increases backbone-level success in computational binder design.

During the rebuttal period, we made the following additions and revisions in response to the reviewers' comments, marked in red in the revised manuscript.

  • Added shuffled spectral-index and M = 16 controls against the full M = 64 model to evaluate shared spectral correspondence and spectral resolution on all three downstream tasks (Section 4.5, Tables 4–5; Appendix F, Table 7).
  • Added the Tet Simple control, replacing Laplacian features with SASA, residue depth, Voronoi volume, and local atom density while keeping all other settings unchanged, to compare simple scalars with the registered volumetric representation (Section 4.5, Tables 4–5; Appendix G, Table 7).
  • Clarified the matched AtomSurf and BoltzGen controls and how they distinguish volumetric information from additional model capacity and training (Sections 4.1–4.4, Tables 1–3).
  • Clarified the released three-block AtomSurf backbone used in our experiments, its difference from the paper's six-block architecture, and the training and evaluation settings (Appendix C).
  • Further explained the adaptation of TetSphere Splatting to residue-wise registration and the encoding of fitted deformations in a shared Laplacian basis (Sections 3.1–3.2, Appendix A.3).
  • Clarified task-specific pretraining within the shared representation framework, including data splits and input restrictions to prevent leakage (Sections 2 and 4.1, Appendices B–E).
  • Added an analysis of CPU preprocessing and GPU fitting time by chain length, together with the overall preprocessing success rate of 99.98% across the three tasks, a documented failure case, and failure-handling procedures (Appendix H, Tables 8–9).
  • Expanded binder evaluation with separate backbone- and sequence-level success definitions, per-target differences and 95% confidence intervals, structural diversity analysis of successful designs, and a matched PDL1 case study (Section 4.4; Appendix I, Figures 12 and 14, Tables 10–12).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.