acceptodds
Under review as a conference paper at ICLR 2027

From Bond Angles to Molecular Curvature: A Geometry-Aware Graph Transformer for Ground-State Conformation Prediction

Abstract

Accurate prediction of molecular ground-state conformations is essential for understanding structure-property relationships and guiding rational molecular design. Despite progress in deep learning, most current methods lack explicit modeling of key geometric priors, particularly atomic bond angles and their derived higher-order curvature features. This often results in conformations with excessive local strain and suboptimal non-bonded interactions that deviate from the true energy minimum. To address this limitation, we propose GeoGT, a Graph Transformer-based model that integrates two geometric constraints into its core design. The first constraint quantifies local geometric non-ideality using an atomic bond angle variance metric, guiding the model away from high-strain structures. The second is a molecular curvature measure derived from angular defect, which extends Gaussian curvature to discrete molecular graphs to capture global spatial bending and packing, thereby improving long-range non-bonded interactions. GeoGT achieves the best distance-based performance across all evaluated datasets and splits, with particularly pronounced gains on QM9. These results highlight the critical role of explicit geometric regularization in improving high-fidelity ground-state conformation prediction. The code and data are available on https://anonymous.4open.science/r/GeoGT-D826.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Ground-state conformation prediction, Geometry-aware graph transformer, Bond angle variance, Molecular curvature

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_oc9kOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on Molecular Ground-State Conformation Prediction, aiming to predict the ground-state three-dimensional conformation of a molecule from its two-dimensional molecular graph topology. The core idea is to enhance the model’s ability to capture molecular 3D structures by incorporating additional geometric inductive biases into the attention mechanism of a Graph Transformer. Specifically, the authors propose integrating two types of geometric information—Bond Angle Variance and Molecular Curvature—into the attention mechanism. The main technical contribution lies in the design and implementation of mechanisms that effectively incorporate these two geometric features into attention computation. Experimental results demonstrate that the proposed method achieves significant performance improvements on both the QM9 and Molecule3D datasets.

Critical strengths
  1. The main experimental results demonstrate significant performance improvements.
  2. The ablation studies and efficiency evaluations further validate the effectiveness and computational efficiency of the proposed method.
  3. Incorporating Bond Angle Variance and Molecular Curvature as geometric constraints into molecular geometry modeling offers a certain degree of novelty and research value. Their potential benefits may extend beyond the current task, suggesting broader applicability to other 3D molecular modeling tasks.
Critical weaknesses
  1. Although incorporating the two geometric constraints improves the model’s performance to some extent, the overall novelty and academic contribution of this work remain relatively limited. From a methodological perspective, the proposed approach primarily extends the existing GTMGC framework by introducing two additional geometric constraints, without offering substantial methodological innovation.
  2. The discussions of Bond Angle Variance and Molecular Curvature in the fourth- and fifth-to-last paragraphs of the Introduction lack sufficient support from the literature. The authors should cite relevant studies on molecular force fields, clarify the physical definition of angular strain, and provide a theoretical justification for using bond angle variance as a proxy for molecular strain. If such theoretical or empirical support is unavailable, the authors should explicitly acknowledge the heuristic nature of this design choice or provide experimental evidence demonstrating its correlation with actual bond-angle strain energy.
  3. In Equation (2), the authors use bond angle variance to quantify the extent to which local molecular geometry deviates from an ideal configuration. However, this metric measures the dispersion of bond angles around their mean rather than their deviations from equilibrium bond angles. For example, the two sets of bond angles, {109.5∘,109.5∘,109.5∘}\{109.5^\circ, 109.5^\circ, 109.5^\circ\} and {100∘,100∘,100∘}\{100^\circ, 100^\circ, 100^\circ\}, both have zero variance, even though they may correspond to different degrees of angular strain.
Questions for authors
  1. Why did the authors choose not to incorporate these geometric constraints directly into the loss function as additional optimization objectives? Would such an approach be feasible, and have the authors explored this possibility?
  2. Regarding the optional iterative decoding mechanism, why does the model’s performance deteriorate when the number of decoding iterations exceeds two? Could the authors provide any plausible explanations or insights into this phenomenon?
  3. Have the authors experimented with using multiple decoders with independent parameters, rather than sharing the same decoder parameters across iterations, to implement iterative refinement?

Post-Rebuttal Update: I appreciate the authors’ detailed rebuttal, which largely addresses my concerns regarding Bond Angle Variance. The additional experiments and explanations also provide useful answers to my questions. Although my concern about limited methodological novelty remains, I have decided to raise my rating from 2 to 3.

6 comments

w1:

Thank you for your comments. GeoGT introduces methodological innovations in both geometric representation construction and the mechanisms through which geometry guides conformation prediction. Its central contribution is a unified modeling approach that connects local angular organization with global conformational representation by jointly designing geometric descriptors, positional encoding, and attention computation.

Existing approaches incorporate structural information through attention biases based on adjacency and pairwise distances, or modify molecular connectivity through force-based graph rewiring, as in ReBIND. GeoGT advances a different modeling direction: angular geometry explicitly conditions both the positional representations of atoms and their information exchange. Our methodological innovations are detailed below.

(1) An explicit computational mechanism connecting angular distributions to geometric attention. We transform bond angle variance into learnable atomic geometric embeddings and combine them with geometric positional representations to compute interaction weights between atoms. These weights enter the geometric residual attention term, allowing the angular organization of atomic neighborhoods to directly influence attention scores and information aggregation. This design extends the geometric conditioning of atom interactions from pairwise connectivity and distance to the organization of angles involving neighboring atom triplets. The resulting mechanism establishes a concrete connection between local angular statistics and the computation of interatomic interactions during ground state conformation prediction.

(2) A curvature descriptor constructed from angular geometry and local spatial scales. To construct a curvature descriptor directly on molecular graphs, we combine an angular-defect-inspired quantity with normalization by a local area scale:

mci=4π−∑θ∈Θi2π(1−cos⁡θ2)4πRi2+ϵ.mc_i = \frac{4\pi - \sum_{\theta\in\Theta_i} 2\pi\left(1-\cos\frac{\theta}{2}\right)}{4\pi R_i^2+\epsilon}.

Here, 2π(1−cos⁡(θ/2))2\pi(1-\cos(\theta/2)) is the solid angle of a spherical cap with angular radius θ/2\theta/2. Aggregating these angular terms and incorporating the local spatial scale RiR_i provides a geometrically motivated construction that jointly describes angular organization and spatial extent without requiring an explicit molecular surface mesh. The descriptor is invariant to translations and rotations because its constituent bond angles and distances are invariant to these transformations. It also preserves information complementary to bond angle variance: identical variances can produce different curvature descriptors because the latter depend on actual angle values and local scales. These analytical properties underpin the construction of the geometric representation used by GeoGT.

(3) A coupled mechanism for global positional modeling and local geometric interaction. We further design how this atomic curvature descriptor participates in global conformational modeling. Curvature-dependent weights are integrated with molecular connectivity, intermediate distances, and graph positional information to construct geometric positional representations. These representations participate in attention computation and are jointly used with variance embeddings in geometric residual attention. The curvature branch therefore conditions the positional representation across the molecular graph, while the variance branch introduces local angular organization into geometric interaction weights. Their coupling enables local and global geometric information to jointly guide the decoder.

These designs establish a unified computational pathway from bond angle distributions to curvature construction, positional representation, and attention-based information aggregation. This coordinated design of geometric representations and their integration mechanisms constitutes the core methodological innovation of GeoGT for ground state conformation prediction.

w2:

Thank you for your suggestion. We have revised the two relevant paragraphs in the Introduction, added references on molecular force fields and discrete geometry, and clarified the distinction between physical energy terms and the geometric descriptors used in our method. The revisions are detailed below.

(1) Added references and clarified the physical definition of angular strain. We have cited UFF[1] and MMFF94[2] to explain that angular strain in molecular force fields is associated with deviations of actual bond angles from reference angles determined by the chemical environment. The corresponding energetic penalties depend on parameters and force constants for angle bending. These references provide the physical background for the relationship between angular deformation and energy.

(2) Clarified the geometric role and heuristic nature of bond angle variance. We have replaced “physically grounded measure of angular strain” with “a geometric descriptor of local angular nonuniformity” and explicitly stated that this descriptor serves as a heuristic geometric inductive bias for encoding local angular distributions. We have also added a citation to [3] to provide background on the use of angular distortion statistics in coordination geometry analysis.

(3) Added the geometric motivation for molecular curvature and clarified its interpretation. We have cited the work of [4] to explain the use of angle deficits for estimating discrete Gaussian curvature on triangulated surfaces. Our molecular curvature descriptor is now explicitly described as a geometric quantity inspired by angular defect, computed from the intermediate conformation and incorporated into positional encoding for global conformational modeling.

[1] A. K. Rappé, C. J. Casewit, K. S. Colwell, W. A. Goddard III, and W. M. Skiff. UFF, a Full Periodic Table Force Field for Molecular Mechanics and Molecular Dynamics Simulations. Journal of the American Chemical Society, 114(25):10024–10035, 1992. DOI: https://doi.org/10.1021/ja00051a040.

[2] T. A. Halgren. Merck Molecular Force Field. III. Molecular Geometries and Vibrational Frequencies for MMFF94. Journal of Computational Chemistry, 17(5–6):553–586, 1996. DOI: https://doi.org/10.1002/(SICI)1096-987X(199604)17:5/6<553::AID-JCC3>3.0.CO;2-T.

[3] K. Robinson, G. V. Gibbs, and P. H. Ribbe. Quadratic Elongation: A Quantitative Measure of Distortion in Coordination Polyhedra. Science, 172(3983):567–570, 1971. DOI: https://doi.org/10.1126/science.172.3983.567.

[4] M. Meyer, M. Desbrun, P. Schröder, and A. H. Barr. Discrete Differential-Geometry Operators for Triangulated 2-Manifolds. In H.-C. Hege and K. Polthier (Eds.), Visualization and Mathematics III, pp. 35–57. Springer, Berlin, Heidelberg, 2003. DOI: https://doi.org/10.1007/978-3-662-05105-4_2.

w3:

Thank you for highlighting this important distinction.

GeoGT combines bond angle variance and a curvature descriptor to capture complementary aspects of molecular geometry. Bond angle variance characterizes differences among angles within an atomic neighborhood, while the curvature descriptor directly incorporates actual bond angle values and local spatial scales, providing angular information that variance does not retain. For the two sets of angles provided by the reviewer, when the local spatial scale is held fixed, their different angle values yield different curvature descriptors, even though both sets have zero variance. Thus, identical bond angle variances do not imply identical geometric information available to GeoGT.

Within this framework, bond angle variance contributes to attention computation through learned atomic geometric embeddings, while the curvature descriptor participates in global conformational modeling through positional encoding, linking local angular organization to the representation of the overall molecular structure. Together with atomic attributes, molecular connectivity, and distance information, these descriptors support ground state conformation prediction. The model learns to use these features through supervision from interatomic distances in ground truth conformations and does not directly equate zero variance with an ideal geometry or a minimum energy state. Accordingly, we interpret these two descriptors as complementary geometric inductive biases rather than direct estimates of angular strain energy.

p1:

Thank you for your suggestion. Incorporating these geometric descriptors into the loss function is feasible. To evaluate this possibility directly, we conducted an additional experiment on QM9 by adding auxiliary losses for both bond angle variance and curvature to the original training objective:

L=Ldist+λvarLvar+λcurvLcurv.\mathcal{L} = \mathcal{L}_{\mathrm{dist}} + \lambda_{\mathrm{var}}\mathcal{L}_{\mathrm{var}} + \lambda_{\mathrm{curv}}\mathcal{L}_{\mathrm{curv}}.

Here, Ldist\mathcal{L}_{\mathrm{dist}} is the original distance supervision applied to the intermediate and final conformations, while Lvar\mathcal{L}_{\mathrm{var}} and Lcurv\mathcal{L}_{\mathrm{curv}} denote the auxiliary supervision terms for bond angle variance and curvature, respectively. The coefficients λvar\lambda_{\mathrm{var}} and λcurv\lambda_{\mathrm{curv}} control their contributions. The results are reported in Table R.4.

Table R.4: Evaluation of auxiliary geometric losses on QM9. The auxiliary-loss variant adds both bond angle variance and curvature losses to GeoGT. Lower values are better; the best results are highlighted in bold.

ValidationTest
VariantD-MAE↓D-RMSE↓K-RMSD↓D-MAE↓D-RMSE↓K-RMSD↓
GeoGT + auxiliary geometric losses0.1720.3500.2750.1740.3530.277
GeoGT0.1250.2890.2010.1300.2980.210

As shown in Table R.4, adding the two auxiliary losses increases all three prediction errors on both the validation and test sets. On the test set, D-MAE increases from 0.130 to 0.17357, D-RMSE from 0.298 to 0.35293, and K-RMSD from 0.210 to 0.27696. Thus, the additional geometric supervision does not improve conformation prediction under the evaluated configuration.

Geometric descriptors can be useful for representation learning without requiring their errors to be optimized as separate objectives. Bond angle variance and curvature summarize particular aspects of molecular geometry. Adding losses on these descriptors changes the relative emphasis placed on different geometric errors during training. Reducing descriptor discrepancies does not necessarily yield a corresponding reduction in the distance and coordinate errors of the complete conformation. The distinction is therefore between using geometric information to guide prediction and assigning it additional weight in the optimization objective.

GeoGT incorporates these descriptors into geometric embeddings, positional encoding, and attention computation, allowing them to guide representation construction and information aggregation. The original distance objective supervises both the intermediate and final conformations, directing learning toward the target molecular geometry. The additional experiment supports retaining this combination of internal geometric modeling and distance supervision for the current model configuration.

p2:

Thank you for your question. Table 10 of the submitted manuscript shows that, within the evaluated settings, a moderate increase in the number of decoding passes improves prediction accuracy, with two passes achieving the best results. An additional pass allows the model to recompute geometric features from the previously predicted conformation and further refine local structures and the overall spatial arrangement using the updated geometric reference.

However, additional iterations also lengthen the sequence of updates, increasing the effective computational depth of the unrolled decoding process. Each iteration involves recomputing geometric features, positional encodings, and attention scores. As the number of iterations increases, residual errors in the preceding conformation may repeatedly influence subsequent predictions through these geometric features. When the gains from further refinement diminish, continued updates may introduce unnecessary adjustments to structures that are already reasonably accurate, potentially increasing prediction errors. We consider this a plausible explanation for the performance degradation observed with excessive iterations. Moreover, this mechanism performs learned conformational updates without imposing a constraint that prediction errors must decrease at every iteration. Greater computational depth therefore does not guarantee continued improvements in accuracy.

These results indicate that the number of iterations should balance prediction gains against computational cost. Two decoding passes can further improve accuracy, whereas a single pass provides higher computational efficiency. GeoGT therefore uses a single decoding pass by default and retains additional iterations as an optional refinement strategy, with the number of passes selected according to validation performance and the available computational budget.

Reviewer_Vy67OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose GeoGT, a method where a encoder-decoder based transformer networks predicts 3D position of the atoms DFT level of theory from its 2D graph (SMILES). The encoder output derives a per-atom bond-angle variance embeddings, which feed geometric residual attention in the decoder. The authors evaluate their method on QM9 and Molecule3D datasets and demonstrates superior performance compared to baselines.

Critical strengths
  1. In Table 8 the bond angle pas rate of proposed method, GeoGT is the highest on every split.
  2. It's especially a convincing where authors demonstrated the effectiveness of GeoGT on downstream tasks on QM9. This demonstrates strong evidence of authors effectiveness.
Critical weaknesses
  1. Authors didn't compare their work with MPSU-Mamba (arXiv 2511.09564), DiffFormer, CFD. The MPSU-Mamba which reports lower MAE on most of the metrics compared to authors work.
  2. Tables 8-9 doesn't isolate bond angle variance or curvature, the Appendix F.6 without angle removes the modulation and all of the BGeoB^{Geo}. Therefore, only removing the variance embedding, only mc$_i$, or only Eq. 4, plus a shuffled control are missing.
Questions for authors
  1. Was Table 1 on QM9 generated with one decoder pass?
  2. With angle on, removing T leaved Table 8 test K-RMSD mostly identical, while D-MAE nearly doubles. As K-RMSD covers heavy atoms only, does T mainly correct hydrogen placement?
  3. Per W1, MPSU-Mamba (arXiv 2511.09564, Communications Chemistry, 6 July 2026) also utilizes 2D-input also reports similar datasets/metrics as in Table 1. Can authors add it to comparison, if not provide rational behind it.
  4. Appendix E.4 does not say whether the baseline rows of Table 1 were rerun or taken from the GTMGC and ReBIND papers.
4 comments

w1 and p3:

Thank you for suggesting these additional baselines. We have expanded the QM9 comparison to include DiffFormer and an adapted version of MPSU-Mamba.

We also checked the code link provided for CFD, but the linked repository did not contain implementation code at the time of our inspection. We explain the additional results and the relevant experimental settings below.

(1) Additional baseline comparisons.

The additional results are reported in Table R.2. GeoGT achieves the lowest errors across all six validation and test metrics. Compared with the adapted MPSU-Mamba, GeoGT reduces test D-MAE, D-RMSE, and K-RMSD by 26.7%, 13.4%, and 31.4%, respectively.

Table R.2: Additional baseline comparisons on QM9. Lower values are better; the best results are highlighted in bold.

ValidationTest
MethodD-MAE↓D-RMSE↓K-RMSD↓D-MAE↓D-RMSE↓K-RMSD↓
MPSU-Mamba0.176770.344080.304500.177060.344490.30524
DiffFormer0.285840.478460.414680.287280.480810.41896
GeoGT0.124690.288970.200770.129770.298420.20954

(2) Experimental settings of MPSU-Mamba.

We inspected the official implementation of MPSU-Mamba and found that the reference conformation is used to determine the atom order for subsequent decoding. In models/modules/task_head.py, ConformerPredictionHead.forward() calls align_conformer_hat_to_conformer() to align the intermediate predicted coordinates to the ground-truth conformation. In models/mpsu/modeling_mpsu.py, MPSUMamba.forward() then passes the aligned coordinates to sort_by_coordinates(), which generates an atom ordering for each of the xx, yy, and zz axes. The decoder uses these orderings in subsequent computation. The ground-truth conformation can therefore influence the atom orderings used by the decoder. When only a two-dimensional molecular graph is available as input, the target conformation is unknown. The reference information required to construct these orderings through alignment to the target is therefore unavailable in this setting.

To remove the dependence of atom ordering on the target conformation, we retained MPSU-Mamba's sorting procedure and applied it directly to the intermediate predicted coordinates before alignment to the reference conformation. We retrained and evaluated this adapted version of MPSU-Mamba using the same QM9 data split, tokenizer, model selection procedure, and evaluation protocol as GeoGT.

(3) Code availability of CFD.

At the time of our inspection, the repository linked by CFD did not contain implementation code. We were therefore unable to reproduce CFD under our evaluation protocol. Accordingly, CFD is not included in the numerical comparison.

w2:

Thank you for pointing out that the original ablations did not sufficiently separate the contributions of the individual components and lacked shuffled controls. We agree that the original “w/o Angle” setting jointly removes the spectral modulation in Eq. (4) and the entire geometric residual attention term BGeoB^{\mathrm{Geo}}, making it difficult to distinguish their respective effects.

To address this concern, we have expanded the QM9 ablation study to separately remove: (i) only the bond-angle variance embedding eiae_i^a; (ii) only the curvature descriptor mcimc_i; and (iii) only the spectral modulation in Eq. (4). Each variant retains the remaining components of GeoGT. We also report ablations of BGeoB^{\mathrm{Geo}}, TT, and their combination.

In addition, we have included the following two shuffled controls:

Shuffled variance embedding. We randomly permute the variance embeddings eiae_i^a among atoms within each molecule. The correct curvature information, spectral modulation, and complete attention structure are retained.

Shuffled curvature. We randomly permute the curvature values mcimc_i among atoms within each molecule and then recompute the curvature-related branch using these permuted values. The correct variance embeddings, spectral modulation, interatomic distances, and adjacency information are retained.

These controls preserve the network architecture and parameter count, as well as the collection of variance embeddings or curvature values within each molecule, while disrupting their correspondence with individual atoms. The complete results are reported in Table R.3.

Table R.3: Fine-grained ablation and shuffled control results on QM9. Lower values are better; the best results are highlighted in bold.

ValidationTest
VariantD-MAE↓D-RMSE↓K-RMSD↓D-MAE↓D-RMSE↓K-RMSD↓
w/o spectral modulation0.1360.3000.2260.1430.3120.239
w/o variance embedding eiae_i^a0.1590.3220.2630.1640.3300.272
w/o curvature mcimc_i0.1960.3620.3130.2030.3710.322
w/o BGeoB^{\mathrm{Geo}}0.1800.3400.2700.1860.3480.280
w/o TT0.2350.4450.2190.2430.4590.228
w/o BGeoB^{\mathrm{Geo}} and TT0.2870.4960.3820.2910.4940.390
Shuffled variance embedding eiae_i^a0.1760.3410.2860.1840.3510.298
Shuffled curvature mcimc_i0.2220.3920.3550.2280.4000.363
Full GeoGT0.1250.2890.2010.1300.2980.210

As shown in Table R.3, individually removing any of the three components increases all three prediction errors on both the validation and test sets. Specifically, removing only the variance embedding, only curvature, or only the spectral modulation increases test D-MAE from 0.130 to 0.164, 0.203, and 0.143, respectively, with corresponding increases in D-RMSE and K-RMSD.

The shuffled controls also degrade performance across all three metrics on both splits. Shuffling the variance embeddings increases test D-MAE, D-RMSE, and K-RMSD from 0.130, 0.298, and 0.210 to 0.184, 0.351, and 0.298, respectively. Shuffling curvature increases these errors to 0.228, 0.400, and 0.363. Together with the component removal experiments, these results support the importance of correctly associating geometric information with individual atoms: prediction errors increase when these associations are disrupted, even though the corresponding network components are retained.

p1:

Thank you for the question. Yes, the QM9 results in Table 1 of the submitted manuscript were obtained using a single decoding pass (T=1T=1), without additional iterative refinement. As specified in Section 3.4, this is the default setting of GeoGT.

p2:

Thank you for this observation. The positional encoding TT incorporates geometric information into the representations of all atoms and is not specifically designed to correct hydrogen positions.

D-MAE measures pairwise-distance errors involving all atoms, including heavy–heavy, heavy–hydrogen, and hydrogen–hydrogen pairs, whereas K-RMSD measures coordinate deviations of heavy atoms after alignment. These metrics therefore differ in both their atom coverage and the geometric quantities measured. In particular, a nearly unchanged heavy-atom K-RMSD does not imply that heavy–heavy distance errors are unchanged.

These results demonstrate a clear reduction in the average interatomic-distance error evaluated over all atoms. Improvements in hydrogen-related geometry are a plausible contributor, but these aggregate metrics alone do not establish that the improvement comes primarily from hydrogen placement.

p4:

Thank you for pointing out this omission. The baseline results in the original Table 1 were taken from the GTMGC and ReBIND papers and were not obtained by rerunning these methods.

Based on ours clarifications on the baselines and additional results I'm raising my score to 3.

We sincerely thank the reviewer for the constructive feedback and for raising the score to 3 upon reviewing our additional results. Your insightful suggestions were instrumental in expanding our baseline comparisons, strengthening our ablation study through finer-grained experiments and shuffled controls, and clarifying several key aspects of the manuscript. Thank you again for your time and effort dedicated to evaluating our work.

Reviewer_zb5gOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoGT is a graph transformer that predicts a single ground-state conformer from a 2D graph, similar to ReBIND and GTMGC. An encoder first produces an intermediate conformation CimC_{im}. Then, featurization is applied by extracting curvature and feeding it into the decoders' positional encoding and attention bias. The loss is simple MAE loss upon both the intermediate and final conformations.

Critical strengths

The architecture itself is sensible (though identical in terms of ReBIND): a predict → extract coordinate-derived geometry → re-decode recycling loop at O(N2d)O(N^2 d) cost, with T=2T=2 passes giving further gains and the fastest inference among baselines. Second, the evaluation is broad for this benchmark line, covering three splits, PoseBusters validity, size-binned error, downstream property prediction, and ablations.

Critical weaknesses

The main question I have is on whether the curvature itself is adding useful information. The "w/o T" experiment, as up to my understanding, is removing both the intermediate distance and curvature together, thus the performance drop cannot be solely attributed to curvature. It would be helpful if the authors provide ablations regarding this aspect. Also, the Eq.7 seems a bit inconsistent in terms of dimensionality. Could you clarify this as well?

Questions for authors

Mentioned above.

3 comments

w1:

Thank you for highlighting the need to isolate the contribution of curvature and for pointing out the ambiguity in Eq. (7).

(1) Fine-grained ablations to isolate the contribution of curvature.

We agree that the original “w/o TT” ablation evaluates the positional encoding branch as a whole: it removes both curvature and intermediate interatomic distance information, and therefore cannot attribute the performance difference solely to curvature. To address this concern, we have expanded the ablation study to separately examine the spectral modulation in Eq. (4), the bond-angle variance embedding, the curvature descriptor mcimc_i, and the geometric residual attention term BGeoB^{\mathrm{Geo}}. The results on QM9 are reported in Table R.1.

Table R.1: Fine-grained ablation results on QM9. Lower values are better; the best results are highlighted in bold.

ValidationTest
VariantD-MAE↓D-RMSE↓K-RMSD↓D-MAE↓D-RMSE↓K-RMSD↓
w/o spectral modulation0.1360.3000.2260.1430.3120.239
w/o variance embedding eiae_i^a0.1590.3220.2630.1640.3300.272
w/o curvature mcimc_i0.1960.3620.3130.2030.3710.322
w/o BGeoB^{\mathrm{Geo}}0.1800.3400.2700.1860.3480.280
w/o TT0.2350.4450.2190.2430.4590.228
w/o BGeoB^{\mathrm{Geo}} and TT0.2870.4960.3820.2910.4940.390
Full GeoGT0.1250.2890.2010.1300.2980.210

Removing curvature increases test D-MAE from 0.130 to 0.203, D-RMSE from 0.298 to 0.371, and K-RMSD from 0.210 to 0.322, corresponding to relative error increases of 56.2%, 24.5%, and 53.3%, respectively. Consistent degradation is also observed on the validation set.

(2) Clarification of the construction of PiP_i in Eq. (7).

We have clarified the notation for the first part of Eq. (7):

P_i &= \exp\\!\left(-\beta \left|mc_i - mc_{\min}\right|\right)e_i^{\top} + A_i + D_i, b = \sum_{i=1}^{N} mc_i \tilde{u}_i, [U, \Sigma, V] = \operatorname{SVD}(b), T_i = P_i + U. \end{aligned}$$ Here, $e_i \in \mathbb{R}^{N}$ denotes the $i$-th standard basis column vector, while $A_i$ and $D_i$ denote the corresponding adjacency and intermediate-distance row vectors. The exponential term is a scalar, and $e_i^\top$, $A_i$, and $D_i$ all have shape $1 \times N$. Their addition therefore yields $P_i \in \mathbb{R}^{1 \times N}$, **matching the dimensionality of $U$, so that the addition $T_i = P_i + U$ is dimensionally consistent**.

I thank the authors for their rebuttal. My main concern has been resolved and thus will raise my score leaning towards acceptance.

We sincerely thank the reviewer for the constructive feedback and for taking the time to review our rebuttal. We are very encouraged to hear that our clarifications regarding the curvature ablation and Eq. (7) have resolved your main concerns, and we greatly appreciate your decision to raise the score. Thank you again for your valuable time and support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.