acceptodds
Under review as a conference paper at ICLR 2027

: Global Stereochemical Fields for Chiral Graph Transformers

Abstract

Enantiomers share atoms, bonds, and pairwise distances yet can behave differently in chiral environments, so molecular encoders must respect atom relabelings and proper rotations without becoming blind to reflection. We introduce GSF-, a graph transformer in which stereogenic units modulate all pairwise interactions rather than single out one atom as special. Each central or axial stereogenic unit creates a reflection-even phase field over all atoms, a handedness pseudoscalar sets the direction of a relative rotation on latent query–key blocks, giving a Chiral-RoPE that reflection inverts rather than leaves fixed. A projection separates mirror-even ECD peak counts and positions from mirror-odd peak signs. We prove the operator's even–odd decomposition and its annotation-inversion, permutation, and unit-order identities under explicit canonical-role conditions; property tests and a coordinate-reflection audit verify the laws end to end. GSF- leads every central-ECD output and improves axial Rotation and Symbol by and over the strongest baseline. Equal-budget controls attribute the Rotation advantage to global signed support rather than parameter or edge count; the projection yields exact enantiomer-pair consistency at a small raw-accuracy cost under complete supervision and becomes predictive when mirror supervision is scarce.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
graph transformer, chirality, ECD spectra prediction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UiYzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces GSF-χ, a graph transformer that explicitly models chirality by making pairwise interactions sensitive to molecular handedness. Its Chiral-RoPE mechanism reverses under reflection, while a symmetry-based readout separates mirror-even spectral features from mirror-dependent ECD peak signs. The authors validate these properties through mathematical identities, 85 property tests, and coordinate-reflection audits, then evaluate performance on ChIRo R/S and Ranking, CMCDS ECD, and ACMP axial chirality benchmarks, including comparisons with equal-budget controls and label-efficiency studies.

Critical strengths
  • The symmetry analysis is careful and well-scoped
  • Parity is treated as an output-dependent requirement, not a single property of the model.
Critical weaknesses
  • The paper is a little hard to read and understand; many things are not explained properly in the main text. More background on RoPE would be helpful.
  • The mechanism does nothing for the even targets it also claims to serve. The reflection-even control matches or beats the full model on Position and Number, so the contribution is confined to sign prediction
  • The paper solves a representation problem with a labeling solution. Chirality never arises from geometry inside the model. It is extracted upstream by a cheminformatics pipeline, reduced to a per-unit sign, and injected.
  • Scope is claimed much more broadly than the construction supports. The method covers central and axial units because both can be reduced to anchor sets with a sign. Planar and helical chirality, conformational chirality, and chirality that emerges from assembly rather than from a localized unit have no obvious encoding in this framework, yet the conclusion offers “a general representation that integrates chirality into graph transformers.”
  • Several design choices are justified by ablation rather than by the stated principle.
Questions for authors
  • If the same C₂ projection is wrapped around ChiDeK’s and ECDFormer’s Symbol heads, do they also achieve exact pair complement?
  • What is the computational complexity of the method compared to the baselines?
2 comments
AuthorsOpenReview ↗

Response to Reviewer UiYz (1/2)

1. Explanation and relation to RoPE. Ordinary RoPE rotates query and key channel pairs so that their dot product depends on a relative angle. Here, an atom–unit field supplies that relative angle, and the unit's handedness changes its direction. Reversing handedness therefore reverses the rotation rather than leaving the interaction unchanged. The revision defines every input and output at first use (Section 4), and shows the encoder and the output projection as separate stages in the redrawn Figure 1. We ran the standard two-channel construction (RoFormer) as a direct control with matched parameters, seeds and budgets: both rotary forms outperform a parameter-matched scalar handedness bias on Rotation (77.67% and 77.50% vs 73.17%; the three-channel block wins in all five seeds), while they perform on par with each other. The benefit therefore comes from handedness-signed relative rotation, which is the defining idea of Chiral-RoPE, rather than from the block dimension (details in our response to Reviewer ozQ3).

2. Even targets and the role of preprocessing. We agree that the ablations do not establish a universal improvement in Number or Position. An odd-sensitive encoder can yield an even output after projection, but this is a compatibility property, not a guarantee of better accuracy on even targets. The revision narrows the empirical claim accordingly. We also agree that chirality is extracted upstream. The intended question is how a supplied stereogenic signal should affect molecular interactions, not whether the network can discover that signal from reflection-invariant inputs. R/S accuracy alone does not establish the need for the full architecture.

3. Scope. The implementation and audits cover the stated central and axial units under explicit role/frame assumptions. They do not establish an encoding for planar, helical, conformational or assembly-level chirality. The revised conclusion is scoped to central and axial stereogenic units and describes broader chirality as future work.

4. Projection around existing baselines. Yes, at the output-function level the construction is model independent. For a baseline binary Symbol logit vector f(x)f(x) and a physical mirror operation MM, define Pf(x)=12[f(x)+Jf(Mx)]Pf(x) = \tfrac{1}{2}\big[f(x) + J f(Mx)\big], where JJ swaps the two classes. Since M2=IM^2 = I, Pf(Mx)=J Pf(x)Pf(Mx) = J\,Pf(x). The even heads can analogously average their original and mirrored logits. This enforces probability/logit consistency for any encoder when every mirror-sensitive input is transformed correctly. It is therefore not evidence that the GSF encoder is uniquely necessary. A zero score is an important exception for hard labels: both probabilities are 1/2 and a fixed argmax tie rule does not produce complementary labels. An encoder that is itself mirror blind can collapse to exactly this case.

We evaluated this wrapper on all five frozen matched checkpoints of each baseline, transforming coordinates, chiral frames and local tags as applicable. On the 120 held-out ACMP molecules, the mean valid-peak Symbol accuracies are:

Frozen encoderOriginalC2 projectedExact logit-swap residualProjected exact-tie fraction
ChiDeK65.07%63.67%00.067%
ECDFormer50.00%50.00%075.33%

These are CPU frozen-checkpoint evaluations, with the same evaluator before and after projection. We report valid-peak accuracy here to avoid the historical evaluators' different zero-peak slot bonuses; under a common seven-slot convention ChiDeK is 66.62%/65.29% and ECDFormer is 52.23%/52.23%. The physical-mirror hard-label flip rates after projection are 99.93% and 24.67%, respectively, with the remaining cases being exact ties. This demonstrates the algebraic guarantee while showing that it does not ensure predictive improvement or non-tied class decisions. The physical-mirror audit is distinct from comparing the independently supplied benchmark partners. Because a post-hoc wrapper does not measure adaptation during training, we also retrained each baseline with the projection inside its objective (original architecture, five matched seeds, 25 epochs, validation selection). Under the seven-slot convention, projected ChiDeK reaches 66.69% ± 2.30%, essentially unchanged from 66.62% without projection. Projected ECDFormer outputs tied Symbol decisions on every seed (50.00% true-slot accuracy; 52.83% seven-slot), so its mirror partners receive identical rather than complementary signs. Exact consistency is obtained in both cases; better prediction is not.

AuthorsOpenReview ↗

Response to Reviewer UiYz (2/2)

5. Complexity and measured cost. With LL layers, NN atoms, SS units and hidden width dd, the unit-wise rotary attention correction costs O(LSN2d)O(LSN^2d), in addition to the usual O(LN2d+LNd2)O(LN^2d + LNd^2) attention/projection terms and the atom–unit field networks. The direct implementation materializes per-unit pair tensors, so memory grows with SN2SN^2. ChiDeK's QQ anchor queries attend to NN atoms with an O(LQNd)O(LQNd) attention term; ECDFormer combines bond/angle message passing with ordinary dense transformer attention over padded atoms and peak queries. Thus neither baseline has the extra unit-wise SN2SN^2 interaction term. Their widths, query counts and head designs also differ. Parameter count alone does not establish lower computational cost.

The parity readout evaluates both handedness signs, adding approximately twice the encoder arithmetic; batching affects the observed latency. We measured inference on the same first 32 ACMP test molecules with preloaded inputs, FP32, four CPU threads, and an Intel Xeon Gold 6530. The values below are median forward times over three repetitions of five iterations after warm-up, using each model's native architecture and peak-slot count. They exclude preprocessing, transfer and metric computation.

ECD modelParametersCPU forward, batch 32
GSF-χ, three-channel rotation154,440413.46 ms
GSF-χ, two-channel rotation154,498453.69 ms
GSF-χ, signed attention bias154,348360.47 ms
ChiDeK8,028,283215.95 ms
ChiDeK + C2 projection8,028,283429.72 ms
ECDFormer795,43350.25 ms
ECDFormer + C2 projection795,433100.85 ms

GSF's smaller parameter count therefore does not translate into lower CPU latency in this batch setting. We repeated the measurement on one NVIDIA H100 80GB GPU (FP32, eval mode, inputs already on the device; median of five repetitions of 50 iterations after 10 warm-up iterations):

ECD modelGPU forward, batch 1GPU forward, batch 32Peak allocated memory, batch 32
GSF-χ, three-channel rotation4.82 ms5.59 ms763 MiB
GSF-χ, two-channel rotation4.42 ms5.20 ms763 MiB
GSF-χ, signed attention bias3.91 ms4.57 ms689 MiB
ChiDeK39.52 ms665.54 ms240 MiB
ChiDeK + C2 projection78.39 ms1306.79 ms241 MiB
ECDFormer5.95 ms9.20 ms98 MiB
ECDFormer + C2 projection11.67 ms17.95 ms99 MiB

On GPU, GSF-χ is the fastest of the three models at both batch sizes, but its per-unit pair tensors give the highest peak memory at batch 32 (about 3× ChiDeK and 8× ECDFormer), consistent with the SN2SN^2 term above. The output projection roughly doubles baseline latency because it requires a second, mirrored forward pass. These are inference measurements, not training times.

6. Design choices versus principle. The symmetry principle fixes what the construction must satisfy: reflection-even context, a single reflection-odd sign per unit, order-free aggregation over units, and a parity-split ECD readout. It does not determine sizes such as the rotary block dimension or the gate form, which we choose empirically and now test directly: Section 6.3 shows that the signed rotary form, rather than its block size, carries the Rotation gain.

Where to find the changes in the revised PDF (blue). Two-channel RoPE control: Section 6.3, Table 2; Appendix F.3. Projection around baselines: Appendix F.4, Table 29. Measured cost: Section 4.5; Appendix F.5, Table 30. Clearer method presentation: Section 4 now opens with a plain-language description of the model inputs and outputs (including what ECD measures), Figure 1 is redrawn from input to output, and every symbol is defined at first use. Scope: Conclusion.

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

The paper proposes GSF-χ, a molecular graph transformer in which stereogenic units generate reflection-even phase fields over atoms, while handedness determines the direction of relative rotations in latent query-key blocks. A projected ECD readout separates mirror-even peak counts and positions from mirror-odd peak signs. The paper provides conditional symmetry guarantees and evaluates central and axial chirality tasks.

Critical strengths

The separation between global chirality-dependent interactions and an explicitly parity-compatible ECD readout is a reasonable design. The reported improvements on axial Rotation and Symbol provide encouraging evidence that the approach is useful.

Critical weaknesses
  1. Some essential details of the method and training objective remain unclear. Appendix A.1 introduces a 26-dimensional atom–unit relation vector and a 342-dimensional unit descriptor, but only lists broad feature categories. Similarly, I could not find a definition of the structured-loss term in Eq. (20). The complete objective, its rationale, and its relationship to earlier ECD methods would benefit from further explanation.

  2. Several experimental descriptions would benefit from clarification. The meaning of ``deliberately a Symbol-focused case" in Figure 12 is unclear. In the upper row for pair 27, the model predicts a positive second peak near bin 11, whereas the reference sign appears negative. Separately, Section 6.3 and Appendix E.9 describe Rotation and Symbol as sharing one representation and checkpoint and differing only in projection, whereas Tables 3 and 9 report different training configurations and selected epochs for these tasks. Please clarify how these descriptions are consistent.

Questions for authors
  1. Please specify the input features, how the branch/side roles and geometric frames used to compute handedness are constructed, and which components are inherited from prior work.

  2. How is the structured-loss term in Eq. (20) defined? Which parts of the ECD objective are inherited and which are newly designed, and what motivates the additional terms and weights? Which choices are shared across models, and which are model-specific?

  3. Please verify the reference and predicted position–sign tuples for both configurations of pair 27 and clarify whether the apparent discrepancy arises from prediction, labeling, or plotting.

  4. Do the Rotation and Symbol results in Table 2 use the same encoder checkpoint?

  5. Please check the author order in the SPMS reference against the citation below. SPMS would also be more accurately described as a CNN descriptor-based approach rather than a message-passing model.

Li-Cheng Xu, Xin Li, Miao-Jiong Tang, Luo-Tian Yuan, Jia-Yu Zheng, Shuo-Qing Zhang, and Xin Hong. A molecular stereostructure descriptor based on spherical projection. Synlett, 32(18): 1837–1842, 2021. doi: 10.1055/s-0040-1705977.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Skqq (1/2)

1. Inputs, roles and provenance. Section 4 now opens with a "Model inputs" paragraph that states exactly what the model receives (per-atom 52-dimensional chemical descriptors without stereochemical tags, bonds, 3D coordinates, and the stereogenic units with handedness and confidence) and a "Model outputs" paragraph for each task, and it points to the new Appendix A.7 (Table 4), which lists every input coordinate. Each symbol is also defined at first use in Section 4.1, and the redrawn Figure 1 shows the pipeline from input to output. In short, each atom–unit relation has 26 entries: graph distances to the two anchors; Euclidean distances to the anchors; radial, signed longitudinal and absolute longitudinal distances; a graph-side indicator; two anchor indicators; two reflection-even shape quantities; anchor separation; three shortest-path bond statistics; nine branch/side-role indicators; and a central-unit indicator. Distances have the normalizations documented in the accompanying feature table. For a central unit, the two anchor distances coincide, axial-only entries are zero, and branch roles are assigned by ordered neighbor priorities.

The 342-dimensional unit descriptor contains six ordered 52-dimensional atom blocks and 30 scalar entries. The atom blocks encode the anchors and priority-ordered substituents, with zero blocks for absent roles. The scalars record unit type and shape, substituent presence, bond order, length, radial length and normalized priority, two within-side dot products, endpoint degrees and candidate counts. Appendix A.7 gives the exact offsets and zero conventions.

The graph, conformers, peak targets and stereogenic-unit preprocessing are inherited from the benchmark pipelines, including RDKit/ChiralFinder conventions. Our learned atom–unit phase, its use in global signed attention, and the parity-separated readout are the architectural additions. Handedness is supplied by preprocessing; it is not learned from an unsigned distance matrix. The central-ECD loader specifically obtains its sign from assigned R/S stereochemistry, whereas the R/S benchmark and axial pipeline have their own determinant/frame conventions. The revision describes these task-specific paths separately (Appendix A.7).

2. The undefined structured loss. This omission should be corrected explicitly. Let μN\mu_N and μk\mu_k be the posterior expected count and the posterior expected position of slot kk, with reference count NN and reference bins yky_k. In the implementation,

Lstructured=λNOE[(μN−N)2]+λNRE[∑cpN(c)(c−N)2]+λPOE[(μk−yk)2]+λPRE[∑bpk(b)(b−yk)2]+λME[ReLU(1+μk−μk+1)2]+0.5Ltransition.L_{\mathrm{structured}} = \lambda_{NO}\mathbb{E}\big[(\mu_N - N)^2\big] + \lambda_{NR}\mathbb{E}\Big[\sum_c p_N(c)(c - N)^2\Big] + \lambda_{PO}\mathbb{E}\big[(\mu_k - y_k)^2\big] + \lambda_{PR}\mathbb{E}\Big[\sum_b p_k(b)(b - y_k)^2\Big] + \lambda_{M}\mathbb{E}\big[\mathrm{ReLU}(1 + \mu_k - \mu_{k+1})^2\big] + 0.5L_{\mathrm{transition}}.

For an ordinal count head, μN=∑tσ(at)\mu_N=\sum_t \sigma(a_t) is the sum of threshold sigmoid probabilities, its base loss is threshold binary cross entropy, and the number-risk term is implemented as the expected-count squared error. Position and Symbol losses use the ground-truth valid-peak mask in these configurations. Ltransition\mathcal{L}_{\mathrm{transition}} is the cross entropy of adjacent sign-equality/flip labels and is zero for the reported parity and parity_split readouts, which return no transition logits. The complete loss also includes the base count and position losses, twice the Symbol cross entropy, and optional field regularizers. Appendix A.7 lists the coefficients for central and axial ECD and for each baseline.

Count/location/sign decomposition follows earlier peak-prediction pipelines; the additional ordinal/risk/ordering terms are choices of this implementation, not consequences of the symmetry theorem. The raw objectives and reductions of every comparator must be distinguished from “the same evaluator” or “the same epoch budget.” The revision states that budgets, labels and validation rules are shared across methods, whereas objectives follow each model's design.

AuthorsOpenReview ↗

Response to Reviewer Skqq (2/2)

3. Pair 27. The saved prediction tuples and the raw labels agree. With binary sign encoding 0/1, the records are:

ConfigurationReference (bin, sign)Count-decoded GSF prediction
mol_27(3,0), (7,1), (11,1), (16,1)(3,0), (11,1)
mol_-27(3,1), (7,0), (11,0), (16,0)(3,1), (11,0)

The predicted second slot has the sign of the second reference slot, whose position is bin 7, but places that prediction at bin 11. The model predicts only two of the four reference peaks. Moreover, the raw sampled amplitude at bin 11 is zero; its binary target is inherited from the dataset's sign/complement convention. Thus the “Symbol-focused” label referred to slot-sign accuracy, not agreement of the continuous curve or of all position–sign tuples. The revised caption (Figure 12) presents it as an example of correct slot signs but incorrect count and positions, gives the discrete tuples, and states the zero-amplitude convention.

4. Rotation and Symbol checkpoints. No: the headline Rotation and Symbol results do not come from one encoder checkpoint. Rotation is trained as a separate binary classification task; axial ECD jointly produces Number, Position and Symbol from one ECD checkpoint. The statements in Section 6.3 and Appendix E.9 conflated a common architecture with shared trained weights. The revision now distinguishes these two meanings (Sections 4.4 and 6.4; Appendix F.1).

5. SPMS. Thank you for the correction. The author order is Li-Cheng Xu, Xin Li, Miao-Jiong Tang, Luo-Tian Yuan, Jia-Yu Zheng, Shuo-Qing Zhang, and Xin Hong. The revision describes the cited method as a spherical-projection descriptor with a CNN predictor, rather than placing it among message-passing models. We checked this against the original paper.

Where to find the changes in the revised PDF (blue). Separate Rotation/ECD checkpoints: Sections 4.4 and 6.4; Appendix F.1. Pair 27: Figure 12 caption and Table 31. SPMS: Section 5.1 and the references. Indexed features, roles and complete loss: Appendix A.7 (Table 4); every input symbol is now defined at first use in Section 4, and Figure 1 is redrawn as an input-to-output framework.

Thanks for your response. Most of my original concerns have been addressed. I have one remaining question about the ECD training objective. With the expanded structured loss, Eq. (20) appears to combine a considerable number of loss terms. Could the authors discuss whether some of these terms overlap in their effects, which terms are active in the reported experiments, and whether jointly optimizing them introduces competing objectives or sensitivity to their relative weights? A brief discussion of the rationale for this combination would help to understand the ECD objective.

Overall, I am considering raising my score from 2 to 3.

AuthorsOpenReview ↗

Response to Reviewer Skqq

Thank you for raising this point and also thank you for considering raising the score. Both main ECD runs use three base losses with weights 1 for count, 1 for position, and 2 for sign. Each run adds three auxiliary terms.

In the main corrected central ECD run, the additions are count risk (0.25), position risk (0.125), and an ordering penalty (0.1) when adjacent predicted mean positions are less than one bin apart. The separate central rerun using the original appendix settings uses only count and position risk (0.15 each).

In matched axial ECD, the additions are squared error of the predicted mean count (1.0), squared error of the predicted mean position (0.1), and position risk (0.1). The base count loss uses binary cross-entropy over ordered thresholds. Matched ChiDeK uses the same auxiliary weights with categorical count cross-entropy. ECDFormer uses the base losses alone. The transition loss and both field regularizers are inactive in the two main GSF runs.

The two axial position terms do overlap. For a fixed target bin, expected squared distance to the target equals the squared error of the predicted mean plus the variance of the predicted bin distribution. Their combined contribution therefore gives weight 0.2 to squared mean error and 0.1 to variance. This lets the loss distinguish a near miss from a distant miss even when both predictions assign the same probability to the correct bin.

The weights are empirical. Because the heads share an encoder, their gradients can compete. For example, increasing the position-risk weight favors more concentrated bin predictions and can change the balance with sign accuracy. Our comparisons do not isolate sensitivity to each weight. The mirror identities follow from the architecture under the stated input assumptions and hold for any choice of these weights. We will clarify this in the objective discussion.

Reviewer_ozQ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper proposes GSF-χ, a graph transformer for predicting properties of chiral molecules from molecular graphs, 3D conformations, and preprocessed stereochemical information. Each stereogenic unit produces learned values over all atoms, and its handedness controls rotations of hidden query–key features to modify attention between atom pairs. For electronic circular dichroism (ECD) prediction, two evaluations with opposite handedness signs separate peak counts and positions, which should remain unchanged under reflection, from peak signs, which should reverse. Experiments cover R/S classification, enantiomer ranking, and central and axial chiroptical prediction.

Critical strengths
  1. A concrete way to incorporate chirality into molecular attention and prediction.

    The model allows each stereogenic unit to influence interactions throughout the molecule, with one interface for central and axial chirality. The output construction also accounts for the different reflection behavior of ECD peak counts, positions, and signs. The symmetry arguments are sound under their stated assumptions and explain how these design choices produce the intended transformations.

  2. Useful axial-chirality results and controlled comparisons.

    The paper reports improvements of 12.6 and 7.9 ppt over the matched ChiDeK retraining on optical-rotation sign and ECD peak-sign prediction, respectively. The equal-budget comparison is particularly informative: allowing signed corrections away from the anchor improves Rotation even when the number of corrected attention entries is fixed. This provides evidence about where chirality information should enter the model, although the same ordering does not hold for every output.

Critical weaknesses
  1. Writing and motivation need clearer explanation.

    The paper introduces uncommon jargon before explaining the basic idea and without appropriate definitions. E.g. the abstract uses “reflection-even phase field” and “global signed support” without explaining what these operations are/do. The task labels “Number,” “Position,” and “Symbol” are also used before being clearly defined.

    The motivation also skips a step. Distinguishing mirror images does not by itself require chirality to modulate every atom pair. The introduction calls the all-atom design a choice backed by the ablations in Section 6.1, but the Figure 2 caption begins with “Why chirality must enter all interactions.” What specific limitation of anchor-based methods does the global design address? The equal-budget comparison favors it for optical-rotation sign, though the paper notes that other outputs do not all follow that ordering. So how is the proposed architecture a necessity?

    Overall, the writing and motivation make it difficult to understand the paper’s contributions and the reasons for its design choices.

  2. The numerical results need a clearer correspondence to the original runs.

    Table 19 reports a paired Rotation improvement of 9.67 percentage points over retrained ChiDeK, whereas the means in Table 24 give 77.78 − 65.17 = 12.61. Under the stated use of the same complete run sets and test pairs, these values should agree. This limits confidence in the reported effect sizes and significance tests. Please explain the cause of the discrepancy.

  3. The benefit of the full architecture is not sufficiently isolated.

    The experiments do not establish whether the particular three-channel rotation is needed, compared with a simpler handedness-conditioned attention score or a two-dimensional rotary construction using the same inputs, gates, and budget. Similarly, applying the same even/odd output projection to existing baselines would help separate the benefit of the GSF encoder from that of the output constraint. The nearly saturated R/S task is less informative here because its input already contains a geometry-derived handedness sign. This is legitimate structural information, but the high score alone does not establish the need for the full model.

  4. The central-ECD implementation does not satisfy the premise of the mirror guarantee.

    The guarantee requires the input features other than the handedness signs to remain unchanged under reflection. However, atom_features(..., include_cip=False) still includes GetChiralTag() features, which enter the central-ECD atom and unit descriptors. Reflecting the coordinates and updating the corresponding stereochemical tags can therefore change the supposedly even inputs. The projection, which only reverses the handedness signs, no longer compares the actual physical mirror pair. This can break the claimed output identities even with fixed model weights. The central-ECD preprocessing violates an assumption required by the projection guarantee. The authors should correct the preprocessing and evaluate its impact on the reported results.

    Code locations in the anonymous repository: src/gsf_chi/features.py:36–56 retains the chiral-tag features; src/gsf_chi/data/central_ecd.py:90–101 builds the atom features and passes them to the unit descriptor, which concatenates the selected atom-feature blocks in src/gsf_chi/data/axial.py:389–402,442; src/gsf_chi/ecd.py:195–222 constructs the paired forwards by negating only chi and local_chi while retaining the other input tensors.

Questions for authors

See the above weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer ozQ3 (1/2)

1. Central-ECD preprocessing and the physical-mirror guarantee. You are correct. Disabling CIP labels did not disable RDKit's tetrahedral ordering tags. Those tags entered both the atom features and the unit descriptors, although the projection only negated chi and local_chi. Consequently, the submitted central preprocessing did not satisfy the even-input premise, and the existing annotation-inversion tests were insufficient to establish the corresponding physical-reflection guarantee.

We have corrected the featurizer so that both CIP and local chiral-tag channels are neutral in its even-input mode, retaining the original feature dimension. The central cache now records the feature version and rejects stale caches. A new regression test reflects the coordinates, changes the tetrahedral tags, rebuilds the complete input, and checks both the even features and the projected outputs with fixed weights. On 256 deterministically selected parent molecules, the old pipeline gives a maximum Symbol logit-swap residual of 0.7773 for one fixed random model; corrected even inputs match exactly and the corresponding float64 residual is below 1e−14. Float32 evaluation has a maximum residual of 2.58e−5 at the odd normalized head, so we distinguish the algebraic identity from bitwise equality. These checks do not measure prediction accuracy. We also distinguish exact reflected conformers from independently generated enantiomer conformers: the latter need not have identical geometric features, so exact prediction identities are not implied for arbitrary conformer pairs.

The algebraic projection remains valid under its stated premise, but this does not validate the submitted central accuracy numbers after the preprocessing correction. We measured that impact by retraining with the original and the corrected features, using identical splits, targets, seeds (0/1/2), budgets and validation selection, for both the appendix configuration and the configuration recovered from the run manifest (12 runs). Test results (mean ± s.d. over three seeds):

ConfigurationFeaturesSymbol ↑Number RMSE ↓Position RMSE ↓
Appendixoriginal (with chiral tags)52.19 ± 1.79%1.2012.288
Appendixcorrected52.43 ± 1.19%1.2172.281
Run manifestoriginal (with chiral tags)53.26 ± 0.48%1.1722.300
Run manifestcorrected52.64 ± 0.91%1.2032.305

The paired corrected-minus-original Symbol differences are +0.24 and −0.63 points, within seed-to-seed variation (per-seed differences range from −1.97 to +3.22 points); Number RMSE increases slightly (+0.017 and +0.031). Thus the leaked tags did not drive the central results. After correction, GSF-χ still leads ChiDeK (50.26%, 1.235, 2.623) and ECDFormer (50.42%, 1.316, 2.718) on all three central-ECD metrics, with every gap larger than the sum of the run standard deviations and two-sided Welch p < 0.05 in all six comparisons. The revised paper reports only the corrected-feature runs for central ECD (52.64% Symbol, 1.203 Number, 2.305 Position, seeds 0/1/2) and drops the separately selected seed set 41/168/552 (53.76%).

2. The 9.67 versus 12.61 percentage-point discrepancy. The discrepancy is an error in our correspondence between tables, not rounding or a consequence of bootstrapping. The five-run matched comparison uses seeds 101, 211, 307, 401 and 503. Its recorded Rotation means are 74.8333% for GSF-χ and 65.1667% for ChiDeK, giving 9.6667 percentage points, which agrees with the paired bootstrap. The 77.7778% value comes from a different three-run configuration screen. The experiment records explicitly identify that configuration as retained after inspecting test performance. It therefore should not have been presented as the same matched run set, or as a prespecified configuration.

The revision makes the five-run comparison primary (Table 1) and labels the three-run configuration as exploratory (Appendix F). This correction also affects Symbol: the matched means are 69.5541% versus 66.6242%, a 2.9299-point difference. The corresponding pair-bootstrap interval is [−4.46, +9.87] points, so we withdraw the claim that the matched Symbol improvement is statistically resolved. For Rotation, the matched interval is [+1.83, +17.83] points. We have rerun the checkpoint-level evaluation and 10,000 paired resamples, reproducing these values. Resampling keeps each of the 60 enantiomer pairs together and averages over the same five fixed checkpoints; these intervals characterize test-pair uncertainty conditional on those runs, not uncertainty over new training seeds. The run identifiers, selected epochs and checkpoint provenance are retained with the audit.

AuthorsOpenReview ↗

Response to Reviewer ozQ3 (2/2)

3. Writing and motivation. The abstract is rewritten in plain language without terms such as “phase field” or “signed support”; Section 4 now opens by stating what the model receives and predicts, including what an ECD spectrum is; Number, Position and Symbol are defined there before any use, and Section 5.1 defines every metric. Every symbol in Section 4 is defined at first use, and Figure 1 is redrawn from input to output.

4. Global support and simpler constructions. We agree that distinguishing enantiomers does not require all-pairs modulation. In an anchor-restricted layer, a direct signed attention correction is confined to rows or entries incident to a stereogenic anchor; remote interactions can receive its effects indirectly through later layers. Our global design permits a direct signed correction between two non-anchor atoms. That is an architectural choice with an empirical tradeoff, not a necessity theorem. The Figure 2 caption now reads “How a global field extends anchor-restricted attention,” and the empirical motivation is restricted to the targets and controls that support it.

We compared the original three-channel rotation with ordinary two-channel rotary blocks and a scalar handedness-conditioned attention bias. All three receive the same descriptors, field inputs and pair gates, and use the same five seeds (101/211/307/401/503), 25-epoch budgets and validation-selection rules for Rotation and axial ECD. Field widths of the simpler models are adjusted so that total parameters stay within 0.1% (154,440 / 154,498 / 154,348). All 30 runs were trained in the same environment (H100/H200); test accuracies (mean ± s.d. over five seeds) are:

ConstructionRotationSymbol (seven-slot)Number RMSEPosition RMSE
Three-channel rotation (original)77.67 ± 4.46%68.85 ± 4.03%1.0342.786
Two-channel rotary blocks77.50 ± 6.64%72.23 ± 2.14%1.0262.794
Signed attention bias73.17 ± 4.39%72.10 ± 1.38%1.0032.765

The results support signed rotary attention as the mechanism behind the Rotation gain. Chiral-RoPE beats the parameter-matched scalar handedness bias on Rotation in all five seeds (+4.50 points on average), so the improvement is not explained by simply injecting a signed scalar into attention. The standard two-channel rotary block performs on par with the three-channel one on Rotation (+0.17 points), showing that the benefit comes from signed relative rotation and is robust to the block dimension. On Symbol and the even outputs the three signed forms are not separated (paired p > 0.1; the simpler forms are numerically higher on Symbol). The revision therefore presents the three-channel block as one instance of signed rotary attention rather than a uniquely required design; the comparison is in Section 6.3 (Table 2) and Appendix F.3. The three-channel rerun here (77.67% Rotation, 68.85% Symbol) and the historical matched runs (74.83%, 69.55%) differ by run-to-run variation in a new environment; comparisons are made within each set.

We also applied the same output-level C2 projection to the five frozen matched checkpoints of both baselines, with their physical mirror inputs rebuilt. The projection enforces exact logit swapping independently of the encoder. Mean valid-peak Symbol accuracy changes from 65.07% to 63.67% for ChiDeK and stays at 50.00% for ECDFormer. The respective projected tie fractions are 0.067% and 75.33%, so logit consistency must not be equated with unconditional hard-label complement or improved prediction. These are frozen-checkpoint CPU evaluations. We also retrained each baseline with the projection inside its objective (original architecture, seeds 101/211/307/401/503, 25-epoch budget, validation selection; H100). Under the common seven-slot Symbol convention, projected ChiDeK reaches 66.69% ± 2.30% (mean ± s.d. over five seeds), versus 66.62% for the unprojected matched ChiDeK runs. Projected ECDFormer collapses to tied Symbol decisions on every seed: 50.00% true-slot accuracy (52.83% under the seven-slot convention), with mirror partners receiving identical rather than complementary signs. Training with the projection therefore does not make either baseline more accurate, and it does not repair a mirror-blind encoder.

Where to find the changes in the revised PDF (blue). Central-ECD feature correction: Appendix F.2, Table 27. Matched headline numbers: Table 1; Appendices F and F.1. Bootstrap: Appendix E.7, Table 20. Rotary-construction and projected-baseline controls: Section 6.3, Table 2; Appendices F.3–F.4. Revised Figure 2 caption: Section 4.

Thank you for the response and the additional experiments. The clearer method description, Fig 1 and corrected tables improve the presentation, so I raise Presentation from 1 to 2. Please note that not everything reported in the rebuttal is yet reflected in the abstract and the PDF. The response mostly answers my questions, but the answers also confirm my earlier concerns: after the corrections, the clear gain is mainly on axial Rotation. I see this as solid work, but not enough for me to recommend acceptance, so I keep my rating at 2.

AuthorsOpenReview ↗

Response to Reviewer ozQ3

Thank you for reading the revision and for pointing out the preprocessing error and the discrepancy between the tables. Your comments helped us correct both, and we're glad Figure 1 and the revised tables are clearer.

We agree that the clearest gain is on axial optical-rotation sign. In the matched five-seed comparison, the improvement over ChiDeK is 9.67 percentage points on Rotation. For ECD peak signs, the estimated gain is smaller and its paired confidence interval includes zero. The abstract's broad claim of state-of-the-art performance obscures that distinction. We will replace it with the task-specific finding and carry the same qualification through the results and conclusion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

General response

Thank you for the careful reviews. We agree that the submission needs a clearer account of the method, its assumptions, and the provenance of its numerical results. Our contribution is a way to distribute an available stereochemical signal across graph attention, together with an output construction that respects the different mirror transformations of spectral quantities. It is not a claim that chirality must affect every attention entry, that three-dimensional rotary blocks are uniquely necessary, or that the model discovers every form of chirality from unannotated coordinates.

We use Number for the number of retained ECD extrema, Position for their ordered indices in the 20-bin representation, and Symbol for the associated binary signs. The latter is a peak-slot metric and does not imply correct peak locations or complete spectral reconstruction. Rotation is the separate optical-rotation sign classification task.

Summary of the revision (changes marked in blue in the updated PDF). The main text remains within the 10-page limit.

  • Headline numbers. Table 1 now reports, for every method, the matched comparison with fixed run sets: seeds 101/211/307/401/503 for axial and seeds 0/1/2 for central ECD, with validation-only checkpoint selection. On axial chirality, GSF-χ improves Rotation by 9.7 points over ChiDeK (paired 95% CI [+1.8, +17.8], p = 0.015) and leads Number and Symbol. The earlier three-run configuration is relabelled as exploratory (Appendix F) and used only as the common base of the ablations.
  • Central ECD. The even-input leak is fixed (Appendix F.2, Table 27). With corrected inputs, GSF-χ still leads both baselines on all three central-ECD outputs (Welch p < 0.05 in all six comparisons).
  • New Section 6.3 and Table 2: is the signed rotary form needed? Under matched parameters and seeds, Chiral-RoPE beats a scalar handedness-conditioned attention bias on Rotation in all five seeds, and a standard two-channel rotary block performs on par. The gain therefore comes from handedness-signed relative rotation itself. Retraining ChiDeK and ECDFormer with the same C2 projection gives exact consistency but no accuracy gain. Full results are in Appendices F.3–F.4.
  • Cost (Appendix F.5). On an H100, GSF-χ is the fastest ECD model, running 119× faster than ChiDeK at batch size 32 with 52× fewer parameters, at higher peak memory.
  • Presentation. The abstract is rewritten in plain language. Section 4 now opens with a plain-language description of the model inputs (atom descriptors, bonds, 3D coordinates, stereogenic units with their handedness) and outputs (including what an ECD spectrum is and how it is summarized into peak Number, Position and Symbol). Figure 1 is redrawn as an input-to-output framework, and every symbol in Section 4 is defined at first use.
  • Corrections. The paired-bootstrap table (Table 20) is updated. The text now states that Rotation and ECD use separate checkpoints. The Pair 27 caption and the SPMS citation are corrected, and the stated scope is central and axial stereogenic units.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.