acceptodds
Under review as a conference paper at ICLR 2027

EquiSolv: O(3)-Equivariant Neural Networks for Predicting Protein Solvation Energy Fields

Abstract

Drug design is an arduous process that depends on the discovery of functional protein-molecule interactions, whose energetic favourability depends on displacing solvent from the binding pocket. In practice, Grid Inhomogeneous Solvation Theory (GIST) is a commonly used method that provides accurate, computationally costly solvation thermodynamics through molecular dynamics (MD) simulations, rendering it infeasible for large-scale drug screenings. We propose an -equivariant graph neural network architecture, named EquiSolv, built on Euclidean neural network (e3nn) primitives that predicts GIST-derived scaled enthalpy (eGIST) fields. The field is defined at any coordinate and is evaluated on the ŠMD grid. On peptides, EquiSolv fuses ESM-C embeddings with geometric graph features; on DUD-E proteins EquiSolv omits ESM-C, which leaves MSE essentially unchanged in ablations. We predict per-atom parametric functions constructed as linear combinations, with learnable coefficients, of products of radial basis functions and spherical harmonics. We bound the approximation error of the population-optimal network, assuming a smooth interpolant. The bound does not cover finite-sample error. Of 1100 DBAASP peptide systems we train on 880 and test on 110 peptides; of the DUD-E proteins we train on 69 and test on 10 proteins. On the held-out peptides, we achieve an RMSE of kcal/mol/ų.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
O(3)-Equivariant Learning, Graph Representation Learning, Grid Inhomogeneous Solvation Theory (GIST), GIST-derived scaled enthalpy, Drug Discovery

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_rLTgOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes EquiSolv, an O(3)-equivariant graph neural network for predicting continuous molecular solvation energy fields derived from Grid Inhomogeneous Solvation Theory (GIST). Unlike voxel-based methods that predict values on a fixed three-dimensional grid, EquiSolv predicts atom-specific coefficients of a radial-basis-function and spherical-harmonic expansion and aggregates the resulting local expansions to reconstruct the field at arbitrary spatial coordinates. The model combines molecular graph features, optional ESM-C protein sequence embeddings, and equivariant message passing using Gated Convolution and SE(3)-Transformer layers. The paper also presents an asymptotic approximation-error analysis under smoothness assumptions, without analyzing finite-sample statistical error. Experiments on 1,100 peptide systems from DBAASP and 94 protein targets from DUD-E evaluate field prediction, compare EquiSolv with a 3D U-Net and DeepGIST, and assess the recovery of strongly bound water sites. The results suggest that EquiSolv can accurately predict GIST-derived energy fields with substantially lower inference cost than generating the corresponding labels through molecular dynamics simulations.

Critical strengths
  1. The paper addresses the high computational cost of obtaining solvation thermodynamics through explicit-solvent molecular dynamics and GIST simulations. Directly predicting solvation energy fields from molecular structures is a promising approach for large-scale screening.
  2. The use of O(3)-equivariant layers is appropriate for molecular geometry, while atom-centred expansions and spherical harmonics enable continuous field reconstruction. The global positional-awareness term based on displacement from the molecular centre of mass is also a specific architectural design choice that complements local message passing.
  3. The paper reports inference time and memory consumption, describes the simulation and training setups, and provides extensive appendices covering features, hyperparameters, ablations, evaluation metrics, and baseline comparisons. Releasing the generated GIST data and implementation would further support reproducibility and future comparisons.
Critical weaknesses
  1. The paper compares EquiSolv with a 3D U-Net and DeepGIST, but DeepGIST predicts a target that includes an entropy term and is not retrained specifically for the eGIST target used by EquiSolv. Therefore, the substantial RMSE difference against DeepGIST does not provide a controlled comparison of model architectures. Although the 3D U-Net is a more directly comparable baseline, further comparisons using consistent targets, data splits, preprocessing, and hyperparameter selection would better isolate the benefits of the proposed approach.
  2. A central advantage of EquiSolv is its ability to evaluate the field at arbitrary coordinates rather than only at fixed grid points. However, the main regression results are evaluated on a 1.0 Å grid. Although the paper investigates different grid-coarsening settings, direct evaluation at off-grid coordinates or at resolutions different from the training grid would provide stronger evidence for the practical benefits of continuous reconstruction over voxel-based prediction.
  3. The training objective assigns greater weight to voxels with larger log-space target magnitudes to emphasize non-bulk regions. Although the paper includes ablations of loss weighting and related hyperparameters, further analysis of how these choices affect prediction errors across different molecular systems would help establish the robustness of the selected configuration.
  4. The paper reports aggregate regression metrics and evaluates non-bulk regions separately, but these results do not fully reveal where prediction errors occur within molecular structures. An analysis of errors around different residue types, charged groups, or regions with complex solvent interactions would provide a clearer understanding of the model's strengths and limitations.
Questions for authors

The questions and concerns I intended to raise have already been outlined in the Weaknesses section above.

3 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer rLTg

We thank the reviewer for their detailed and positive comments. We now address the weaknesses point-by-point.

W1: DeepGIST did not expose training code nor did they publish data, so it was not possible to get a same training comparison with it. The 3D U-Net was trained and scored on the same train/eval/test split, following the same preprocessing.

W2: Off-grid ground-truth is not available, since solvers are run on grids. Though the proposition of training on coarser and testing finer grid resolutions is possible (at 0.5 distance increments, as this is the radius of our truth grid), we have now run inference using our model trained on 1.0 coarseness against 0.5 and 2.0 coarseness grids and find: Predictably 1.0 (0.176 ±\pm 0.026) and 2.0 (0.139 ±\pm 0.028) are close, as there are no extra points added but rather a mean is drawn to create the coarsening. While 0.5 (0.279 ±\pm 0.048) shows a slightly poorer RMSE as we already discovered through ablations. We include this table in Appendix Q, and upload a revised PDF.

W3: If we correctly understood what you mean by molecular systems, this would require retraining the loss ablation on additional datasets which is beyond what we can provide for this revision. The weighted loss was already removed on both systems we study. Validation MSE gets worse on DBAASP peptides and on DUD-E proteins (0.0101 to 0.0144 on DUD-E, with the rest of the configuration held fixed). Within DBAASP, the selected model also keeps its top-1000 F1 lead in every size tercile (Appendix J), so the result is not confined to one peptide size.

W4: These are insightful and inspiring suggestions about refining our model’s settings and potentially improving its performance on specific tasks, especially for future downstream applications, these directions will definitely help tailor the model to better solve more concrete problems. We already have aggregate regression metrics in Table 1 and the all-region rows of Table 2. Pooled RMSE, 99th-percentile absolute error, and max absolute error are in Table 13 in Appendix N. Non-bulk regions separately are in Table 2, rows under “non-bulk region …“. The native grid version is in Table 6 in Appendix H. Per-residue errors on these same voxels are in Table 16 in Appendix O. The coarser split, ionisable residues {R, K, H, D, E} against neutral residues, is Table 15 in Appendix N. Errors for voxels within 4 Å of one residue type, and of two or more residue types, are in Table 17 in Appendix P.

The authors have addressed some of my concerns, but I have a few remaining questions about their response. It would be helpful to clarify whether the 3D U-Net was tuned using the same validation set as EquiSolv. Comparing EquiSolv with another baseline trained on the same eGIST target and data split would help assess the benefits of the proposed architecture. The values of 0.831 and 0.766 in the main text do not appear to match the RMSE values of 0.522 and 0.269 in Table 17.

Thank you for your response.

To answer your questions:

  1. The 3D U-Net was indeed tuned with the same validation set as Equisolve, we used the same training/validation/test split.
  2. Thank you for pointing our the error, these were the wrong numbers in the text. We correct them and upload a revision.
Reviewer_RehDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper uses an equivariant neural network architecture to tackle the problem of predicting the energy values of water molecules near a binding pocket, an important problem in drug design that is currently tacked via expensive MD simulations. The method shows strong empirical performance on peptides and proteins. Crucially, it is the first equivariant NN to tackle this problem; the previous SOTA, DeepGIST, does not use an equivariant design.

Critical strengths

The paper is the first to directly model the desired energy density field at any coordinate. They are able to do so via thoughtful architectural choices, including a weighted aggregation scheme and incorporating each atom's direction relative to the protein's center of mass. The paper shows improved empirical performance as a result of these changes on peptide and protein tasks compared to the previous SOTA, DeepGIST, a non-equivariant model that does not directly predict the energy field.

Critical weaknesses

I'm unclear what the paper's novelty is; the architecture used seems straightforward as far as equivariant neural networks go. The paper has components - the weighted aggregation scheme or incorporating each atom's direction relative to the protein's center - that appear to be novel. However, these are more limited contributions. I am not sure if applying a relatively straightforward architecture to this particular problem constitutes sufficient novelty for a top-3 ML conf paper.

Questions for authors

Can you please clarify the paper's algorithmic/technical novelty? To me, it seems like a relatively straightforward application of egnns to this problem.

In order to recommend acceptance, it would help me to have a clear story for conceptual novelty or a new takeaway.

2 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer RehD

We thank the reviewer for raising this concern and for approving of our aggregation scheme and the global-positional awareness as strengths. We would like to clarify that the objective of this work is not to architecturally improve existing models. Instead, we consider our main contributions to be the dataset with our performance as a competitive baseline and proposing an efficient way of representing and reconstructing a continuous scalar field in this particular biomolecular application.

We would like to point out that the target solvent energy field, being a molecular property, is determined collectively by all atoms of the molecule. In our proposed representation methodology, applying the atom-centred decomposition to a collective/global solvent property involves a nontrivial modelling assumption, and we justified this design in our theoretical analysis (Lemma C.1 and the proof of the error behaviour), in addition to the strong empirical results.

Reviewer_RehDOpenReview ↗

Reviewer Response

Thank you for your comments. I feel that predicting the atom-centered coefficients directly is somewhat standard in the literature. While I agree that you achieve strong empirical performance on this particular task, I don't feel there is sufficient technical novelty. As such, I retain my current score.

Reviewer_dU3DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a method to approximate the spatial distribution of solvent energy surrounding a molecule, eg a protein. The model takes as input the known atomic structure of the molecule, together with geometric and physicochemical features and, optionally, embeddings from the pretrained protein language model ESM-C. Then an O(3) equivariant backbone predicts per atom coefficient of radial basis x spherical harmonics expansions. The obtained fields can be queried and averaged aggregated across nearby atoms to obtain the final field prediction, which is then evaluated against the ground truth molecular simulations. The authors created their own datasets for empirical evaluation, consisting of 1,100 peptide systems from DBAASP and 94 protein systems from DUD-E, and compared their method against several existing baselines.

Critical strengths
  1. The paper is generally well-written and readable, with a clear motivation and problem setup. The writing is natural and easy to follow in most sections, except for the architecture description.

  2. The authors promise to supply their own generated datasets for use in later works.

  3. The paper includes a theoretical error analysis, which is a valuable addition, although its scope is somewhat limited, particularly as it does not address finite-sample generalization.

  4. The proposed method offers some interpretability, as the predicted field can be decomposed into per-atom contributions through the learned coefficients. This allows for a more detailed analysis of how individual atoms contribute to the predicted field.

  5. The model predicts a continuous field that can be evaluated at arbitrary points in space rather than being restricted to a fixed grid.

Critical weaknesses
  1. The comparison against existing methods is limited. In particular, DeepGIST is not retrained for the same prediction target, making the reported RMSE comparison difficult to interpret. When evaluated at its original higher-resolution grid, DeepGIST achieves better Pearson correlation in the full region and nearly identical performance in the non-bulk region.

  2. The proposed backbone relies on relatively older architectures, such as Tensor Field Networks and SE(3)-Transformers. Comparisons with more recent equivariant architectures, eg as MACE or EquiformerV2, would be valuable, given the otherwise limited architectural novelty.

  3. The proposed modification of the message-passing update using center-of-mass positional information is not compared directly against conventional equivariant self-interaction. Consequently, it remains unclear how much this particular modification contributes to the performance.

  4. The model is trained to predict only one energy-density target, namely the GIST-derived scaled enthalpy. However, GIST provides several additional thermodynamic quantities, including separate energy contributions and entropy-related information. It would be interesting to investigate whether the proposed architecture generalizes to these quantities.

  5. The generalizability of the proposed method is not sufficiently established. In particular, the protein dataset consists of only 94 proteins, with just 10 used for testing.

  6. Needs extensive preparation from multiple sources, which constitutes a large chunk of its runtime.

Additional comments:

  • Overall, I find the paper potentially valuable for a rather niche proportion of the machine learning community. However, there are some concerns which make it not acceptable in its current form. The first and most important is the lack of comparable baselines (old and not directly comparable). The components used in the architecture are not novel, and quite old, so the novelty of the paper is more the problem formulation and the reuse of existing components in this exact setup.
  • Presentation: The paper is generally well-(and seemingly human)-written. The motivation and problem setup is introduced properly in detail. However, the section introducing the architecture has significant room for improvement. In a lot of cases, (eg s(.) in line 210 , VDW in line 183) the notation and abbreviations are not introduced properly before usage. This makes it harder to follow for a general audience.
  • Informativeness: The application of equivariant geometric deep learning to continuous solvent-energy field prediction is interesting, and the paper provides useful insight into this particular molecular modeling problem. The atom-centered reconstruction of continuous fields is also a potentially useful design choice. However, much of the underlying architecture builds on previously established methods, limiting the amount of fundamentally new methodological insight.
  • Soundness: The overall approach is technically reasonable, and the experiments indicate that the proposed method can approximate GIST-derived energy fields effectively. Nevertheless, there are important limitations in the empirical evaluation, particularly concerning baseline comparability, generalization, and the attribution of performance improvements to individual architectural components. I consider these concerns potentially addressable through additional experiments, although they currently limit the strength of the paper's conclusions.
Questions for authors

See Weaknesses 1-4.

1 comment
AuthorsOpenReview ↗

Rebuttal to Reviewer dU3D

We thank the reviewer for the careful reading and constructive suggestions. We respond to the identified weaknesses point-by-point below.

W1: To our knowledge, DeepGIST is the most appropriate and recent baseline we could find, yet we do not have access to its training code. We would also like to stress that we are working on releasing our datasets and preparing the release of our model upon publication, so as to contribute a first step towards comparable models and making this an established machine learning task.

W2: Thank you for the suggestion, we have run ablations using the MACE layers, these performed at best 2% worse compared to 1GC+3TR. Please see the updated table of ablations in Appendix G. Our primary contribution lies in the proposed framework for predicting spatially resolved solvent energy distributions around biomolecules, rather than in the architecture of the underlying equivariant backbone. The choice of TFN-type architectures is also motivated by theoretical considerations. To our knowledge, there is no known theoretical result about the universality of the suggested newer architectures. On the other hand, Dym and Maron (2021) established a universality result for rotation-equivariant point-cloud networks based on TFN-type architectures, providing a well-established theoretical starting point for our analysis of the approximation properties and asymptotic behaviour of the proposed framework.

W3: This modification was initially introduced during preliminary experiments on a toy dataset, where it yielded encouraging results, and was subsequently retained in the model used for our main experiments. Though we do not claim this modification being an improvement of the architecture, we agree that a direct comparison with conventional equivariant self-interaction would help isolate the contribution of the COM-relative update. We are still queueing on our server to run this ablation, and we will add the results in the CR version of our manuscript.

W4: Given the necessary data, our framework is able to learn any scalar field with EquiSolv’s performance depending on the regularity of the target field and the statistics of the training data. That is the scope of the contribution we present, and we test it with enthalpy-only MD values (motivated by Balius et al.’s findings [1] on different GIST calculation methods). While ablating an enthalpy+entropy modus would be chemically interesting, it would come at a significant computational cost, requiring us to run both the required MD simulations as well as retrain EquiSolv. .

W5: Generating a larger protein MD GIST dataset is currently outside of our rather reasonable academic computing budget.

W6: We agree with this and argue that our model is still overall very cheap and significantly faster than MD work. It should be clearly stated that the model loading step of 6.7s needs to be performed only once when processing multiple proteins, and that the rest of the combined steps take ~2.94 seconds. The evaluation step (0.34s) depends on the number of points queried, we evaluated on the full GIST grid but a pocket for example would take less time. The post-processing step (0.63s) can be skipped if all that is desired is an output GIST field.

We additionally respond to the reviewer’s additional comments:

C1: On our baselines: DeepGist was published in 2026 and it was the most similar work available; so we did our best to include recent and appropriate baselines.

C2: On notation issues: We fixed these details, and now upload an updated PDF with all notation defined.

[1] Trent E. Balius, Marcus Fischer, Reed M. Stein, Thomas B. Adler, Crystal N. Nguyen, Anthony Cruz, Michael K. Gilson, Tom Kurtzman, and Brian K. Shoichet. Testing inhomogeneous solvation theory in structure-based ligand discovery. Proceedings of the National Academy of Sciences, 114 (33):E6839–E6846, 2017

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.