acceptodds
Under review as a conference paper at ICLR 2027

Phase Synchronization Dynamics for Molecular Property Prediction

Abstract

Accurate molecular property prediction is predominantly addressed through graph representation learning. However, conventional spatial message-passing neural networks encounter inherent structural bottlenecks when modeling long-range dependencies and complex global topologies. We propose MolSync (Molecular Synchronization Networks), which represents molecular graphs as networks of non-linear phase oscillators and maps structural priors into a dynamic phase space. The resulting synchronization process provides an intrinsic feature-binding mechanism that organizes local chemical environments and higher-order structural context through learned oscillator interactions. Theoretically, we show that continuous synchronization removes the hard finite-hop dependency constraint of fixed-depth message passing, heterogeneous intrinsic dynamics sustain non-trivial representation variation at phase-locked states, and spectral evolution can distinguish a class of 1-WL-indistinguishable graphs with different weighted Laplacian spectra. Extensive experiments show strong predictive performance, while energy-based attribution and oscillator analyses reveal chemically meaningful, task-dependent patterns in the learned dynamics.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Molecular Property Prediction, Phase Synchronization, Graph Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_GwWfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary
  • In the context of molecular GNNs, dense attention fails to provide physically meaningful attribution for feature binding (attention tends to pick up statistical artifacts). This paper aims to address this problem.
  • Proposal: MolSync, a molecular GNN based on nonlinear phase synchronization dynamics (a vector Kuramoto model on the sphere).
  • It learns molecular representations through feature binding (self-organization into functional substructures), and claims that continuous synchronization removes the multi-hop (finite-hop) constraint of fixed-depth MPNNs.
  • The core part of the model is imported from AKOrN (Miyato+, image domain) to molecular graphs. Architecture: multi-scale initialization via Geometric Scattering Transform (GST) -> driven Kuramoto updates (L blocks x T internal steps) -> phase-invariant readout. Energy-based attribution (EBA) is also provided as a sub-product.
  • Experimentally validated with the MoleculeNet test sets.
Critical strengths
  1. DeMol and Uni-Mol require 3D conformer information. Compared with these, MolSync achieves competitive scores on MoleculeNet using only 2D graph information.
  2. Introducing the Kuramoto model into molecular GNNs as an alternative inter-node interaction mechanism to attention is new to me and interesting.
Critical weaknesses
  1. Experiments are not strong.
  • The claim in Table 1 ("best on 6/8 datasets") sounds overstated to me. The margin over the strongest baseline DeMol is very thin. MolSync vs DeMol: BBBP +3.2, but Tox21 +0.5, SIDER +0.2, ClinTox +0.3, MUV +0.1, BACE +0.1. It seems these margins are not significant compared to the reported std (0.2–1.4) except for BBBP.
  • Ablations (Table 2) and qualitative (Figure 3) evaluation only on BBBP, where the effect is largest. Its generality is unclear since no validation on other datasets is provided.
  • Figure 3 lacks quantification. Without a comparison against a globally collapsed picture, or a quantitative metric, it is difficult for readers to tell whether the figure is consistent with Thm. 2. For example, do attention-based models exhibit a similar grouping structure or not?
  • In Table 2, the largest drops are w/o Kuramoto (73.8) and w/o GST (73.9), essentially tied. This means initialization (GST) matters a lot. However, the computational cost of GST (preprocessing) is not reported in the complexity analysis (Table 5).
  1. Rationale for the model selection, and induced theory
  • If my understanding is correct, the update rule is close to the original (scalar) Kuramoto model, and may be weaker than AKOrN (Miyato et al.) in terms of representation capacity. Is reproducing the Kuramoto form itself essential? Would an implementation faithful to AKOrN perform poorly?
  • Separation between global phase collapse and over-smoothing: Thm. 2 only shows that exact global collapse is impossible. Approximate homogenization of representations (over-smoothing in practice) is not ruled out, and how these are separated is unclear.
Questions for authors
  • Do you have any significance test results for the performance differences in Table 1?
  • What is the total number of trainable parameters? What is the computational cost of GST preprocessing?
  • Is the grouping structure in Figure 3 also observed for attention-based models? A relative comparison would be appreciated.
  • What is the essential implementation difference from AKOrN? How does performance compare against an implementation faithful to AKOrN?
  • What is the essential reason the dynamics must be Kuramoto-type?
  • Around L325: how do you distinguish the global phase collapse from the over-smoothing? How does "exact collapse does not occur" relate to avoiding over-smoothing?
11 comments
AuthorsOpenReview ↗

Rebuttal Part1

Response to Reviewer GwWf

We thank you for a detailed and technically precise review, and for recognising the synchronisation idea and the substructure connection as strengths. Four of your concerns led to changes that materially improved the paper: the ablation now covers all eight datasets, Figure 3 is quantified against an attention baseline, the computational cost of the GST component is measured and reported, and the discussion of Theorem 2 has been rewritten to make the role of the Dirichlet energy bound explicit. We have also added a finite-step separation result that speaks to Q5. We address every point below.



W1 & Q1 — Margins in Table 1, and significance testing

You are right that the margins were reported without a test. We have now run one on all eight benchmarks and we report it in full, including the dataset on which the test goes against us.

Test. For each benchmark we compare our five independent MolSync runs against the strongest published mean in Table 1 using Welch's unequal-variance tt-test computed from the reported summary statistics (mean, standard deviation, run count), two-sided, with Holm--Bonferroni correction across the eight datasets.

DatasetMolSyncStrongest baselineΔ\DeltaWelch ttHolm-adj. ppVerdict
BBBP78.3±1.278.3 \pm 1.2DeMol 75.1±0.675.1 \pm 0.6+3.2+3.25.335.330.013\mathbf{0.013}MolSync higher
Tox2181.4±0.381.4 \pm 0.3DeMol 80.9±0.480.9 \pm 0.4+0.5+0.52.242.240.350.35not separated
ToxCast67.8±0.267.8 \pm 0.2Uni-Mol 69.6±0.169.6 \pm 0.1−1.8-1.8−18.00-18.001.8×10−5\mathbf{1.8\times10^{-5}}baseline higher
SIDER68.6±0.268.6 \pm 0.2DeMol 68.4±0.368.4 \pm 0.3+0.2+0.21.241.241.001.00not separated
ClinTox92.9±0.292.9 \pm 0.2DeMol 92.6±0.792.6 \pm 0.7+0.3+0.30.920.921.001.00not separated
MUV82.6±0.482.6 \pm 0.4DeMol 82.5±1.482.5 \pm 1.4+0.1+0.10.150.151.001.00not separated
HIV80.9±1.280.9 \pm 1.2DeMol 81.2±0.981.2 \pm 0.9−0.3-0.3−0.45-0.451.001.00not separated
BACE89.1±1.289.1 \pm 1.2DeMol 89.0±1.189.0 \pm 1.1+0.1+0.10.140.141.001.00not separated

Outcome: one separation in our favour (BBBP), one against us (ToxCast), and six benchmarks on which MolSync and the strongest pre-trained 3D baseline are statistically indistinguishable. Note that this is the correct reading of HIV as well: the −0.3-0.3 deficit against DeMol is within run-to-run variation and is not a separation in either direction.

Robustness to the one unknown. The compilation does not state the run count behind the baseline rows. We therefore repeated every test under both nbase=3n_{\text{base}}=3 and nbase=5n_{\text{base}}=5. The verdicts are identical in all eight cases, and the two significant results move only in the third decimal (p=0.017p = 0.017 and 2.3×10−52.3\times10^{-5} at nbase=3n_{\text{base}}=3). The conclusion does not depend on that assumption.

What the test licenses. Because baseline per-seed values are not published, this is an unpaired comparison against summary statistics rather than a paired test on common splits, and it inherits whatever protocol differences exist between the cited studies and ours. We state this limitation with the numbers. Where both arms run inside our own pipeline on identical splits, paired testing is available and we use it: the GST-equipped GIN controls in W4 & Q2 and the faithful AKOrN comparison in W2.1 & Q4, each over five runs per arm.

How Table 1 is now described. The average column has been recomputed directly from the eight displayed entries for every row (DeMol 79.8879.88, MolSync 80.2080.20), and the surrounding text states the result the test supports: MolSync is statistically indistinguishable from large pre-trained 3D-conformer models on six of eight MoleculeNet classification benchmarks, higher on BBBP by 3.23.2 points, and lower on ToxCast, while being trained from scratch on each benchmark's training split from the 2D bond graph alone. We make no claim of a benchmark sweep, and the per-dataset verdicts above are reported as they came out.

The reason we still consider this worth reporting is the asymmetry in what each side consumes: GEM, Uni-Mol and DeMol are pre-trained on millions of unlabelled molecules and take 3D conformers as input. Reaching parity with them from 2D topology and no task-independent pre-training is the claim the architecture is meant to support, and it is the claim the test leaves standing.

AuthorsOpenReview ↗

Rebuttal Part2

W2 — Ablations and qualitative evaluation only on BBBP

We accept the concern about relying on a single-dataset ablation. We have extended the component ablations to all eight classification datasets, using five independent runs per configuration under the same evaluation protocol.

VariantBBBPBACEClinToxSIDERHIVTox21MUVToxCastAvg
w/o GST73.987.192.663.779.579.477.965.677.46
w/o Kuramoto73.886.888.565.878.976.877.965.876.79
w/o phase readout76.887.490.267.879.380.182.067.178.84
MolSync (full)78.389.192.968.680.981.482.667.880.20

On the eight-dataset average, removing Kuramoto evolution reduces ROC-AUC by 3.41 points, removing GST by 2.74 points, and removing the phase-invariant readout by 1.36 points. Each removal decreases the reported mean on every dataset. Thus, the components contribute beyond BBBP, with the largest average drop observed when the dynamical evolution is removed. These are component ablations within MolSync; the independent value of its overall architecture is further assessed with GST-equipped GIN controls in W4/Q2. These ablations are reported as Table 7 (Appendix D.5) of the revised paper.

The group-structure analysis in W3/Q3 remains an analysis of BBBP; we do not use that result to claim the same structure has been quantitatively established on every dataset.

AuthorsOpenReview ↗

Rebuttal Part3

W3 & Q3 — Quantifying Figure 3, and comparison with attention models

What Figure 3 measures. Figure 3 is not an attention map. For the final node states before pooling, we first subtract the molecular mean, normalize each nonzero centered vector, and calculate the mean cosine similarity SpqS_{pq} between atoms belonging to structural groups GpG_p and GqG_q. The same representation-based measurement can be applied to MolSync and a Transformer, rather than comparing unrelated internal attention weights.

Correcting self-pair inflation. The submitted diagonal included self-pairs, each with cosine similarity 1. For a group of size mpm_p, its reported diagonal satisfies

Spp=1mp+(1−1mp)cp,cp=mpSpp−1mp−1,S_{pp}=\frac{1}{m_p}+\left(1-\frac{1}{m_p}\right)c_p, \qquad c_p=\frac{m_p S_{pp}-1}{m_p-1},

where cpc_p averages over distinct within-group atoms. We now use cpc_p and exclude singleton groups; otherwise, small groups acquire artificially high diagonal values.

Controlled comparison. Using the same BRICS group partition and the same metric, we compare MolSync with a Transformer on the 93 BBBP molecules both models classify correctly (471 non-singleton groups). This common-correct subset controls for prediction outcome but is not claimed to represent the entire test distribution.

MetricMolSyncTransformer
Groups with cp>max⁡q≠pSpqc_p > \max_{q\neq p} S_{pq}334/471 (70.9%)227/471 (48.2%)
Per-molecule dominance rate (mean)0.7690.548
Binding contrast Δ=mean⁡pcp−mean⁡p<qSpq\Delta=\operatorname{mean}_p c_p-\operatorname{mean}_{p<q}S_{pq}0.530 ± 0.1850.439 ± 0.208
Within-group cosine cpc_p0.4270.376
Between-group cosine SpqS_{pq}−0.103−0.063

We use the molecule, rather than each of its groups, as the inferential unit. MolSync has a higher per-molecule dominance rate in 62/93 molecules (paired Wilcoxon p=8.9×10−12p=8.9\times10^{-12}, as reported from the molecule-level analysis). The paired binding-contrast difference is +0.091 (95% CI [+0.039, +0.144], paired Wilcoxon p=8.8×10−4p=8.8\times10^{-4}). We do not treat all 471 groups as independent observations in a chi-squared test, nor do we describe the Transformer's 48.2% as chance level without an appropriate group-size-aware null distribution.

The negative average between-group similarity indicates relative separation in the centered representation space. Because the representations are centered before normalization, it should not by itself be interpreted as proof of active repulsive synchronization, nor is Spq→1S_{pq}\to 1 a well-defined diagnostic of exact global collapse (centered vectors would approach zero). Our empirical conclusion is narrower and directly supported: MolSync exhibits stronger within-versus-between group organization under the same measurement than this Transformer baseline. These statistics are reported as Table 8 (Appendix D.6) of the revised paper.

Selection robustness. Figure 3 presents ten molecules selected for visual clarity. Removing those illustrations leaves dominance rates of 287/419 (68.5%) for MolSync and 201/419 (48.0%) for the Transformer, close to the full-sample comparison. We will describe Figure 3 as illustrative rather than as a randomly sampled average case. The ten molecule-level comparisons are reproduced as Figure 7 (Appendix D.6) of the revised paper.

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

This paper introduces MolSync, a molecular property prediction framework based on phase synchronization dynamics. It combines geometric scattering features, bond-conditioned oscillator interactions, and phase-invariant readout to capture molecular structural information. The authors provide theoretical analyses of long-range information propagation, representation collapse, and graph expressivity. Experiments on MoleculeNet classification and regression benchmarks, together with ablation and energy-based attribution studies, are presented to support the proposed approach.

Critical strengths
  • Combining geometric scattering features with bond-conditioned oscillator dynamics provides an alternative approach to conventional message passing. This is the main aspect that makes the work potentially interesting.
  • An attempt to connect representation learning with interpretability. The energy-based attribution and functional-group analyses offer an additional perspective beyond predictive performance. This could strengthen the contribution if supported by more rigorous evaluation.
Critical weaknesses
  • Table 1 contains inconsistencies in some reported averages, and the baseline results appear to be inherited from prior work without clearly documenting differences in evaluation protocols. Also, the BBBP performance reported in Table 1 does not align with the corresponding configuration in Figure 6. These issues make it difficult to assess whether the reported improvements are reliable, particularly given the small margins over strong baselines on several datasets.
  • Limited connection between theoretical guarantees and the actual model. The theoretical analyses primarily concern simplified continuous-time or scalar synchronization systems, while MolSync uses finite-step, sphere-valued dynamics. In particular, the long-range dependency and spectral expressivity results do not directly establish the corresponding properties of the implemented graph-level predictor.
Questions for authors
  • Were the baseline results in Table 1 reproduced under the same splits and training protocols, or taken from previously published papers? Could the authors clarify their sources and correct the inconsistent averages?
  • Why does Figure 6 report substantially different BBBP performance for (T=4, K=8) compared with Table 1, given that Table 3 identifies this as the selected configuration?
  • To what extent do the theoretical results apply to the finite-step, sphere-valued MolSync architecture? Is there empirical evidence that the implemented model achieves the claimed long-range propagation and expressivity advantages?
  • Could the authors provide compute-matched comparisons against standard message-passing or graph-diffusion architectures using the same GST features?
8 comments
AuthorsOpenReview ↗

Rebuttal Part1

Response to Reviewer Acfu

We thank you for the careful audit of our experimental results and theoretical claims. We agree that the inconsistent averages in Table 1, the discrepancy with the original sensitivity plot, and the gap between continuous-time analysis and finite-step implementation required direct correction. We have therefore (i) recalculated the reported averages, (ii) clarified the provenance and comparability of published baselines, (iii) updated the T×KT\times K sensitivity analysis for all eight classification datasets using five runs per setting, (iv) added an architecture-level finite-step separation result, and (v) supplied empirical propagation diagnostics and GST-controlled, approximately parameter-matched message-passing comparisons. We address each concern below.



W1 — Table 1 consistency, baseline provenance, and the size of the margins

(a) The average column in Table 1

Thank you for catching this. You are right, and we have corrected it. The error was confined to the average column: it was an arithmetic error, and every per-dataset entry in Table 1 is correct as reported. The averages for the baseline rows had been carried over from the compilation of Liu et al. (2026) rather than recomputed over exactly the eight benchmarks displayed in our table, so where that compilation averaged over a different set of datasets, the inherited figure did not match the eight cells beside it.

We have recomputed the unweighted mean of the eight displayed entries for every row:

MethodOriginal Avg.Corrected Avg.Change
Galformer74.0273.78−0.24
GraphMVP74.8674.30−0.56
GEM78.1878.11−0.07
Uni-Mol78.6378.56−0.07
DeMol79.9679.88−0.08
MolSync80.2080.200.00

Remaining rows differ by at most 0.01, which is rounding of the displayed dataset scores. The corrected MolSync average is unchanged at 80.20, and the strongest cited baseline, DeMol, has a corrected average of 79.88. In the revised manuscript every average in Table 1 is computed directly from the per-dataset values shown alongside it.

(b) The BBBP entry and Figure 6

You are right that these disagreed, and we have identified the cause and resolved it. The full account is in Q2.

(c) Provenance of the baseline results

The baseline rows in Table 1 are cited from the compilation of Liu et al. (2026) and the original publications referenced there; the MolSync row is produced by us. The revised caption states this explicitly so that cited results are distinguishable from our own at a glance.

Our evaluation configuration follows the protocol established for this benchmark suite and used by the compared methods: an 80/10/10 Bemis–Murcko scaffold split, RDKit atom–bond featurization, masking of missing labels in the multi-task datasets, and ROC-AUC computed over valid tasks before macro-averaging. Full details are given in Q1.

(d) The size of the margins

Our results are reported as mean ± standard deviation over five independent runs, and we agree that on several datasets the margin over the strongest cited baseline is of the same order as that run-to-run variation. We therefore make the margin claims that the evidence carries: on BBBP the +3.2-point margin is substantial relative to the reported variability, and across the eight benchmarks the headline result is that MolSync reaches the level of large pre-trained 3D-conformer models while being trained from scratch on each benchmark's training split, from the 2D bond graph alone.

The comparisons that isolate what the architecture contributes are those run end-to-end in our own pipeline with matched protocols and five runs per arm, where the margins are large and testable: the GST-equipped GIN controls and the approximately parameter-matched BBBP controls in Q4.

AuthorsOpenReview ↗

Rebuttal Part2

Q1 — Source of the baseline results and evaluation protocol

Provenance. The Table 1 baselines are cited from Liu et al. (2026) and the works compiled there. The MolSync row, and every number reported in this response, is produced in our own pipeline. The revised caption and surrounding text mark this distinction.

Evaluation protocol. All eight classification benchmarks use an 80/10/10 Bemis–Murcko scaffold split, following the protocol used by the compared methods. RDKit parses SMILES into attributed atom–bond graphs. In the multi-task datasets (Tox21, ToxCast, SIDER, ClinTox, MUV) missing labels are masked in both the loss and the evaluation, and ROC-AUC is computed over tasks containing both classes before macro-averaging. Hyperparameters and checkpoints are selected exclusively on the validation split.

The search space is fixed in advance: oscillator dimension d∈{32,64}d\in\{32,64\}, dropout ∈{0,0.1,0.2,0.3}\in\{0,0.1,0.2,0.3\}, learning rate ∈{5×10−5,10−4,3×10−4}\in\{5\times10^{-5},10^{-4},3\times10^{-4}\}, Kuramoto steps T∈{4,8}T\in\{4,8\}, coherence channels K∈{8,16}K\in\{8,16\}, batch size ∈{64,128,256}\in\{64,128,256\}, and 300 or 400 epochs. Optimization uses AdamW with a 5 % linear warm-up followed by cosine decay. The per-dataset configurations are listed in Table 5 (Appendix D.1) of the revised paper, and all results in this response use five independent runs, reported as mean ± standard deviation.

The same pipeline, splits and evaluation code are used for every model we train ourselves — GIN, GIN + GST, the AKOrN-style adaptation, and the parameter-matched controls — so those comparisons are protocol-identical by construction, and the per-dataset values they produce serve as in-pipeline reference points for the setup as a whole.


Q2 — The discrepancy between Figure 6 and Table 1 at (T=4,K=8)(T=4, K=8)

You are right that the two disagreed, and the cause is specific: the original sensitivity grid was produced from a single run per cell, whereas Table 1 reports averages over multiple runs. Individual cells in that grid were therefore subject to run-to-run variation of the same magnitude as the effects the figure was meant to display, which is what produced the mismatch at BBBP's selected configuration.

We have aligned the two settings. The complete T×KT\times K grid — T∈{0,2,4,8}T\in\{0,2,4,8\} crossed with K∈{4,8,12,16}K\in\{4,8,12,16\} — has been re-run with five independent runs per cell on all eight classification datasets, under the same protocol as Table 1. Every Table 1 entry now reproduces at the configuration listed in Table 5:

DatasetSelected (T,K)(T,K)Re-run gridTable 1
BBBP(4,8)(4,8)78.3 ± 1.278.3 ± 1.2
Tox21(4,8)(4,8)81.4 ± 0.381.4 ± 0.3
ToxCast(4,8)(4,8)67.8 ± 0.267.8 ± 0.2
SIDER(4,8)(4,8)68.6 ± 0.268.6 ± 0.2
ClinTox(4,8)(4,8)92.9 ± 0.292.9 ± 0.2
BACE(4,8)(4,8)89.1 ± 1.289.1 ± 1.2
MUV(8,16)(8,16)82.6 ± 0.482.6 ± 0.4
HIV(8,16)(8,16)80.9 ± 1.280.9 ± 1.2

The grid is also consistent with the component ablation: at T=0T=0, which removes the synchronization dynamics entirely, BBBP gives 73.8 ± 0.9, matching the "w/o Kuramoto" row of Table 2.

Figure 6 in the revised manuscript is replaced by this version, showing mean ± standard deviation over five runs as shaded bands, with the configuration reported in Table 1 marked on each panel.

What the corrected figure shows. Because the GST initialization is held fixed along every curve, variation with TT is attributable to the synchronization steps alone. The gain from T=0T=0 to the selected configuration is positive on all eight datasets, ranging from +2.0 to +4.7 ROC-AUC points, with a consistent shape: a steep rise through T=4T=4 followed by saturation, and wider run-to-run variability at T=8T=8. At fixed TT, varying KK changes performance by well under one point on most datasets. The corrected figure therefore replicates the Kuramoto ablation of Table 2 across all eight benchmarks rather than on BBBP alone.

AuthorsOpenReview ↗

Rebuttal Part3

W2 & Q3 (1/2) — Scope of theory and evidence for the implemented finite-step model

We agree that the continuous-time or scalar analyses in the original submission should not be interpreted as unconditional guarantees for the finite-step, sphere-valued predictor. We now distinguish the properties of the implemented finite computation from the mechanisms described by simplified continuous-time analysis, and provide both a finite-step separation result and direct empirical measurements.

Scope: what we withdraw, and what we retain

Regarding the finite-step nature of the implemented model, we also clarify the scope of our theoretical claims. MolSync uses LL blocks with TT local integration steps per block. Therefore, its actual oscillator updates do not automatically possess the unrestricted global support of an ideal continuous-time solution. In particular, without considering the multiscale GST features, the local interaction pathway has a finite-hop receptive field determined by the number of implemented updates. Our continuous-time analysis characterizes the underlying synchronization mechanism, rather than claiming that the finite-step implementation inherits every property of the idealized continuous system.

We will make this distinction explicit and avoid interpreting continuous-time reachability as a guarantee of unrestricted long-range information propagation in the finite-step architecture.

A finite-step expressivity result for the architecture we implement

We sincerely thank you for raising two important questions: whether the expressive-power analysis applies to the actual finite-step architecture, and whether the advantage over standard message-passing GNNs can be attributed to synchronization rather than merely to the geometric scattering features. We have added a constructive graph-level separation result that directly addresses both questions.

1. A finite-step expressivity result for the actual MolSync architecture.

In the revised manuscript, we establish a sufficient condition under which a finite-step MolSync model, including its initialization, sphere-valued oscillator evolution, invariant coherence readout, and mean pooling, produces different graph-level output distributions for two graphs. Specifically, for an rr-regular graph GG, let AGA_G be its adjacency matrix, LG=rI−AGL_G=rI-A_G, and BG=AG+IB_G=A_G+I. Define

HG,T=(I−ρLG)T,H_{G,T}=(I-\rho L_G)^T,

Lcoh(G)=diag⁡(BG⊤BG1)−BG⊤BG,L_{\mathrm{coh}}(G) =\operatorname{diag}(B_G^\top B_G\mathbf 1)-B_G^\top B_G,

and

QG,T=HG,T⊤Lcoh(G)HG,T.Q_{G,T} =H_{G,T}^{\top}L_{\mathrm{coh}}(G)H_{G,T}.

For the specified valid parameter configuration, we prove that if two graphs satisfy

tr⁡(QG1,T2)≠tr⁡(QG2,T2),\operatorname{tr}(Q_{G_1,T}^{2}) \neq \operatorname{tr}(Q_{G_2,T}^{2}),

then their graph-level output distributions are different for sufficiently small but strictly positive initialization noise. More precisely, the output variance satisfies

Var⁡(yG)=g′(1)2(d−1)δ42n2q4tr⁡(QG,T2)+o(δ4),\operatorname{Var}(y_G) = \frac{g'(1)^2(d-1)\delta^4} {2n^2q^4} \operatorname{tr}(Q_{G,T}^{2}) +o(\delta^4),

where dd is the oscillator dimension, δ\delta is the normalized initialization-noise scale, and q=r+1q=r+1.

This result is derived for the actual finite-step sphere-valued dynamics and the invariant coherence readout, rather than for a continuous-time or purely linear graph propagation model.

2. An explicit pair of 1-WL-indistinguishable graphs.

We further instantiate the sufficient condition using the Petersen graph and the pentagonal prism graph C5□K2C_5\square K_2. Both graphs have ten nodes, are 3-regular, and have identical node and edge attributes. Therefore, 1-WL assigns the same color to every node at every iteration and cannot distinguish the two graphs. Consequently, standard message-passing GNNs without additional structural information cannot distinguish this pair either.

Nevertheless, for every fixed finite number of oscillator updates T≥1T\geq 1, we show that

tr⁡(QGPrism,T2)−tr⁡(QGPet,T2)=40−240Tρ+O(ρ2).\operatorname{tr}(Q_{G_{\mathrm{Prism}},T}^{2}) - \operatorname{tr}(Q_{G_{\mathrm{Pet}},T}^{2}) = 40-240T\rho+O(\rho^2).

Hence, the two traces are different for sufficiently small positive coupling strength ρ\rho. By the finite-step separation theorem, the corresponding MolSync graph-level output distributions are therefore different for sufficiently small positive initialization noise.

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

The paper introduces MolSync, a molecular representation learning framework based on phase synchronization dynamics. Atoms are modeled as coupled oscillators, while molecular bonds determine their interactions. The architecture combines geometric scattering features, learned Kuramoto-style dynamics, and phase-invariant readout to construct molecular representations.

The authors provide theoretical analyses related to long-range information propagation, representation collapse, and expressive power beyond the 1-WL test. The approach is evaluated on molecular property prediction benchmarks, together with experiments examining the relationship between synchronization dynamics and chemically meaningful substructures.

Critical strengths

The main strength of this paper is the idea of using synchronization dynamics as a mechanism for learning molecular representations. I find this direction interesting because it goes beyond the usual formulation of graph learning as repeated neighborhood aggregation and raises a fundamental question about whether collective dynamical behavior can provide a more natural mechanism for organizing information in relational systems.

Another strength is the connection between synchronization, representation diversity, and molecular substructures. The attempt to understand learned representations through the dynamics of interacting oscillators is interesting, particularly when considering how local interactions may give rise to global structural organization. The experimental evaluation is extensive, and the ablations provide useful evidence that the dynamical components contribute to the predictive performance.

Critical weaknesses

My main concern is the relationship between the theoretical analysis and the actual expressive power of the proposed architecture. The results concerning long-range information propagation and spectral distinguishability are established for continuous or linearized dynamics, while the implemented model uses a finite number of discrete updates. In particular, a finite sequence of local updates still has a bounded receptive field, unless additional information is introduced through the geometric scattering features. It is therefore not clear to what extent the claimed improvement over message passing comes from synchronization itself rather than the multiscale structural information already provided to the model.

A related concern is the interpretation of the result concerning expressive power beyond 1-WL. The spectral argument establishes that certain graph-dependent operators can be distinguished using an appropriate input signal. However, this does not immediately imply that the implemented architecture can distinguish these graphs using permutation-equivariant features and invariant readout. I think this distinction is important, particularly because the proposed model already incorporates structural information through geometric scattering.

Finally, while the energy-based attribution results are interesting, I am not fully convinced that synchronization energy provides an intrinsically more meaningful explanation than other learned interaction-based attribution methods. The observed correspondence with chemical substructures is encouraging, but it does not by itself establish that the learned dynamics recover the underlying chemical mechanisms.

Questions for authors

What is the fundamental advantage of synchronization dynamics over a sufficiently deep message-passing architecture? Can the authors isolate the contribution of the dynamical formulation from the additional propagation steps and the multiscale features introduced through geometric scattering?

Can the authors establish an expressivity result for the actual finite-step architecture, including its initialization and invariant readout? In particular, does the model distinguish any pair of graphs that cannot be distinguished by 1-WL when both models receive the same initial node information?

How does the model distinguish useful synchronization from representation collapse? Is there a principled relationship between partial synchronization, the preservation of structural information, and predictive performance?

How should we interpret the physical meaning of the learned oscillator interactions? Are the resulting synchronization patterns stable across independently trained models, and do they consistently identify the same chemically relevant substructures?

Could the authors compare MolSync with a depth-matched message-passing model using the same geometric scattering features and a comparable computational budget? Such an experiment would help isolate the contribution of the synchronization mechanism.

8 comments
AuthorsOpenReview ↗

Rebuttal Part1

Response to Reviewer Fopt

We thank you for a careful and technically precise review, and for recognising the synchronisation direction and the substructure connection as strengths. Your two main concerns — that the theoretical results are established for continuous or linearised dynamics while the implementation is finite-step, and that the spectral 1-WL argument does not immediately transfer to the implemented permutation-equivariant architecture — are both well founded. In response we have added a finite-step expressivity result for the actual architecture together with an explicit 1-WL-indistinguishable graph pair, and we have completed the depth- and feature-matched controls you asked for. We address each point below.


W1 & Q1 (1/2) — Theory versus the finite-step architecture; isolating synchronisation from GST and from extra propagation steps

We thank you for this insightful question. We agree that a sufficiently expressive and sufficiently deep message-passing neural network may approximate many finite-step computations performed by MolSync. Therefore, we do not claim that synchronization dynamics necessarily provide a strictly larger function class than arbitrary message-passing architectures.

Instead, the fundamental motivation is to introduce a different inductive bias and representation organization mechanism. Conventional message-passing networks primarily update node representations through repeated neighborhood aggregation:

h_i^{(l+1)} = \phi^{(l)} \left( h_i^{(l)}, \operatorname{AGG}_{j\in\mathcal N(i)} \psi^{(l)}(h_i^{(l)},h_j^{(l)},e_{ij}) \right). \tag{R9}

MolSync instead maintains node states on a constrained sphere and evolves them through the combination of context-conditioned intrinsic dynamics and signed interactions:

\dot{x}_i = \Omega_i(c_i)x_i + \operatorname{Proj}_{x_i} \left( c_i+\sum_{j\in\mathcal N(i)} \kappa_{ij}x_j \right). \tag{R10}

The distinction is not simply whether information is propagated across edges. Both approaches involve neighborhood interactions. Rather, message passing progressively aggregates information into node features, whereas MolSync explicitly organizes node states through their joint dynamical evolution. The synchronization coherence is then extracted as a relational representation and fed back to condition subsequent dynamics.

We consider this distinction particularly relevant to molecular graphs, where the predictive representation may depend on coordinated patterns among atoms rather than solely on independently accumulated local features. Importantly, this is a difference in computational mechanism and inductive bias, not a claim that sufficiently expressive MPNNs cannot simulate the resulting finite computation.

Regarding the separation of dynamical evolution and GST, our existing ablation study provides initial evidence.

On BBBP, the results are:

VariantROC-AUC
w/o GST73.9±1.473.9\pm1.4
w/o Kuramoto73.8±0.973.8\pm0.9
w/o Ω\Omega74.7±0.374.7\pm0.3
w/o phase-invariant readout76.8±1.276.8\pm1.2
MolSync78.3±1.278.3\pm1.2

These results support two observations.

First, the dynamical component contributes beyond the multiscale structural initialization. In particular, removing Kuramoto evolution while retaining the other components reduces ROC-AUC from 78.378.3 to 73.873.8. Conversely, removing GST while retaining the synchronization architecture reduces ROC-AUC to 73.973.9. Thus, neither component alone explains the full performance.

Second, the contribution is not solely attributable to the oscillator coupling. Removing the node-specific intrinsic rotation Ωi\Omega_i also reduces performance to 74.774.7, while replacing the phase-invariant readout with direct state pooling reduces performance to 76.876.8. These results suggest that intrinsic dynamics and the extraction of relational coherence both contribute to the final representation. This ablation is reported as Table 2 of the revised main text.

AuthorsOpenReview ↗

Rebuttal Part2

W1 & Q1 (2/2) — Theory versus the finite-step architecture; isolating synchronisation from GST and from extra propagation steps

Nevertheless, we acknowledge that the comparison between the full model and the T=0T=0 variant does not, by itself, completely isolate synchronization from the effect of additional local interaction steps. We have therefore completed propagation-depth- and parameter-matched GIN + GST controls on BBBP, reported in Q5, alongside the existing component ablations.

The available comparisons address these factors from complementary directions:

  1. Without Kuramoto evolution (T=0T=0): retain the remaining MolSync components while removing the iterative oscillator updates.
  2. Depth-matched GIN + GST: use the same GST features with 20 conventional message-passing layers and an approximately matched parameter budget.
  3. MolSync without intrinsic rotation: retain sphere-valued coupling dynamics while setting Ωi=0\Omega_i=0.
  4. Full MolSync: retain both node-specific intrinsic rotations and signed oscillator interactions.

A representative local message-passing update for the depth-matched control can be written schematically as

h_i^{(t+1)} = h_i^{(t)} + \phi_{\mathrm{MP}} \left( h_i^{(t)}, \sum_{j\in\mathcal N(i)} \psi_{\mathrm{MP}}(h_i^{(t)},h_j^{(t)},e_{ij}) \right), \tag{R11}

In the completed depth-matched comparison, the GIN + GST baseline uses 20 message-passing layers, matching MolSync's nominal 20 local propagation steps on BBBP, and its parameter count is within 0.3% of MolSync's. The separate eight-dataset GIN versus GIN + GST experiment in Q5 evaluates the effect of adding GST to conventional message passing. Together with the MolSync component ablations, these completed controls distinguish the effects of structural features, parameter budget, propagation depth, and the overall synchronization-based architecture, without attributing the performance difference to any single component alone.

Regarding the finite-step nature of the implemented model, we also clarify the scope of our theoretical claims. MolSync uses LL blocks with TT local integration steps per block. Therefore, its actual oscillator updates do not automatically possess the unrestricted global support of an ideal continuous-time solution. In particular, without considering the multiscale GST features, the local interaction pathway has a finite-hop receptive field determined by the number of implemented updates. Our continuous-time analysis characterizes the underlying synchronization mechanism, rather than claiming that the finite-step implementation inherits every property of the idealized continuous system.

We will make this distinction explicit and avoid interpreting continuous-time reachability as a guarantee of unrestricted long-range information propagation in the finite-step architecture.

In summary, the intended contribution of MolSync is not that synchronization universally outperforms sufficiently deep MPNNs in terms of theoretical expressivity. Rather, it introduces a structured dynamical inductive bias in which node-specific intrinsic evolution and relational interactions jointly organize node states, while phase-invariant readout extracts the resulting coherence. The existing ablations demonstrate the complementary contributions of GST and the dynamical components, and the completed depth-matched comparison in Q5 provides direct evidence that the observed predictive advantage is not explained solely by additional propagation steps or model capacity.

AuthorsOpenReview ↗

Rebuttal Part3

W2 & Q2 — Expressivity of the actual finite-step architecture, with the same initial node information

We sincerely thank you for raising two important questions: whether the expressive-power analysis applies to the actual finite-step architecture, and whether the advantage over standard message-passing GNNs can be attributed to synchronization rather than merely to the geometric scattering features. We have added a constructive graph-level separation result that directly addresses both questions.

1. A finite-step expressivity result for the actual MolSync architecture.

In the revised manuscript, we establish a sufficient condition under which a finite-step MolSync model, including its initialization, sphere-valued oscillator evolution, invariant coherence readout, and mean pooling, produces different graph-level output distributions for two graphs. Specifically, for an rr-regular graph GG, let AGA_G be its adjacency matrix, LG=rI−AGL_G=rI-A_G, and BG=AG+IB_G=A_G+I. Define

HG,T=(I−ρLG)T,H_{G,T}=(I-\rho L_G)^T,

Lcoh(G)=diag⁡(BG⊤BG1)−BG⊤BG,L_{\mathrm{coh}}(G) =\operatorname{diag}(B_G^\top B_G\mathbf 1)-B_G^\top B_G,

and

QG,T=HG,T⊤Lcoh(G)HG,T.Q_{G,T} =H_{G,T}^{\top}L_{\mathrm{coh}}(G)H_{G,T}.

For the specified valid parameter configuration, we prove that if two graphs satisfy

tr⁡(QG1,T2)≠tr⁡(QG2,T2),\operatorname{tr}(Q_{G_1,T}^{2}) \neq \operatorname{tr}(Q_{G_2,T}^{2}),

then their graph-level output distributions are different for sufficiently small but strictly positive initialization noise. More precisely, the output variance satisfies

Var⁡(yG)=g′(1)2(d−1)δ42n2q4tr⁡(QG,T2)+o(δ4),\operatorname{Var}(y_G) = \frac{g'(1)^2(d-1)\delta^4} {2n^2q^4} \operatorname{tr}(Q_{G,T}^{2}) +o(\delta^4),

where dd is the oscillator dimension, δ\delta is the normalized initialization-noise scale, and q=r+1q=r+1.

This result is derived for the actual finite-step sphere-valued dynamics and the invariant coherence readout, rather than for a continuous-time or purely linear graph propagation model.

2. An explicit pair of 1-WL-indistinguishable graphs.

We further instantiate the sufficient condition using the Petersen graph and the pentagonal prism graph C5□K2C_5\square K_2. Both graphs have ten nodes, are 3-regular, and have identical node and edge attributes. Therefore, 1-WL assigns the same color to every node at every iteration and cannot distinguish the two graphs. Consequently, standard message-passing GNNs without additional structural information cannot distinguish this pair either.

Nevertheless, for every fixed finite number of oscillator updates T≥1T\geq 1, we show that

tr⁡(QGPrism,T2)−tr⁡(QGPet,T2)=40−240Tρ+O(ρ2).\operatorname{tr}(Q_{G_{\mathrm{Prism}},T}^{2}) - \operatorname{tr}(Q_{G_{\mathrm{Pet}},T}^{2}) = 40-240T\rho+O(\rho^2).

Hence, the two traces are different for sufficiently small positive coupling strength ρ\rho. By the finite-step separation theorem, the corresponding MolSync graph-level output distributions are therefore different for sufficiently small positive initialization noise.

3. Why this separation is not merely inherited from GST.

We would like to emphasize that the above construction does not rely on GST to distinguish the two graphs. Both graphs have identical node and edge attributes, and the deterministic part of the initialization is identical at every node. Thus, the deterministic initialization provides no graph-specific distinction between the two inputs. The graph-dependent separation arises from the interaction between the finite-step oscillator evolution and the subsequent coherence readout, whose output is then retained through mean pooling.

This construction therefore establishes that synchronization together with invariant readout can contribute to graph-level distinguishability beyond 1-WL, rather than merely exploiting additional multiscale information supplied by GST.

We also clarify the scope of this result: it is a constructive existence guarantee for a valid non-degenerate parameter configuration and a particular 1-WL-indistinguishable graph pair. It does not claim that every trained MolSync model separates every pair of 1-WL-equivalent graphs. We believe this is a more precise and rigorous characterization of the expressive capability of the implemented architecture.

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

The paper proposes MolSync, which represents molecular graphs using coupled phase oscillators, with atoms modeled as unit vectors and bonds as couplings between them, following a Kuramoto-inspired formulation. It further provides a theoretical analysis demonstrating how the proposed approach overcomes the finite-hop support barrier and oversmoothing. The method is evaluated on downstream molecular property prediction tasks using the MoleculeNet benchmark.

Critical strengths
  • The idea of representing molecules as phase oscillators is certainly novel and interesting.
  • An interesting EBA analysis offering an alternative to post-hoc explainers.
  • Decent theoretical analysis of the proposed method
Critical weaknesses
  • Limited experimental validation. Evaluation is restricted to MoleculeNet, where performance remains close to the baselines given the overlapping standard deviation bands. Could the authors extend the evaluation to additional datasets, such as QM9, covering different molecular properties?

  • Connection to existing graph dynamics methods. MolSync appears closely related to continuous-dynamics-on-graphs methods such as GRAND, GRAND++, and PDE-GCN, which also leverage time-dependent graph dynamics for message propagation with similar theoretical advantages. Could the authors clarify how MolSync differs from these approaches beyond the specific parameterization of its update mechanism?

  • Limited novelty of Theorem 1, 3 Theorem 1 establishes that the matrix exponential of a graph Laplacian has full support, a well-known property of heat diffusion and continuous-time graph ODEs, including GRAND. This result therefore appears to be a general property of continuous-time graph dynamics rather than a method-specific theoretical contribution. Could the authors clarify its novelty? Also similar to Theorem 3 which inherits the power of spectra to distinguish graphs not providing any novel aspect if the method can distinguish certain graphs which cannot be by baseline method.

  • Insufficient evidence for mitigating over-smoothing and over-squashing. The model uses only four blocks, far from the high-depth regime where these issues typically become pronounced. Moreover, MoleculeNet graphs have relatively small diameters (~10–15 bonds), and no direct experimental evidence is provided to support the claimed mitigation of either issue.

Questions for authors
  • What is the computational complexity of the proposed method compared to the baselines?
  • AKOrN supplies Eq. 2. How does AKOrN perform on these datasets, leading to a one-on-one comparison to the proposed method?
  • What does EBA enrichment look like against a degree- and atom-type-matched random baseline rather than a uniform one, given that E_i sums over N(i)?
9 comments
AuthorsOpenReview ↗

Rebuttal Part1

W1 — Limited experimental validation

On benchmark coverage and QM9

QM9 is already evaluated in the submission. Table 6 (Appendix D.2) reports MolSync on all six MoleculeNet regression benchmarks, including QM7, QM8 and QM9. More broadly, our evaluation covers 14 datasets across five property categories: biophysics (BACE, HIV, MUV), physiology (BBBP, SIDER, ClinTox), toxicity (Tox21, ToxCast), physicochemical properties (ESOL, FreeSolv, Lipophilicity) and quantum-mechanical properties (QM7, QM8, QM9). Dataset sizes span 642 molecules (FreeSolv) to 133,885 (QM9), and the tasks include both binary classification and continuous regression.

We reproduce the key rows here for convenience (↓ = lower is better; RMSE for ESOL, FreeSolv and Lipophilicity, and MAE for QM7–QM9):

ModelESOL ↓FreeSolv ↓Lipo ↓QM7 ↓QM8 ↓QM9 ↓
GEM0.798 ± 0.0291.877 ± 0.0940.660 ± 0.00858.9 ± 0.80.0171 ± 0.00010.00746 ± 0.00001
MolSync0.816 ± 0.0331.780 ± 0.1030.653 ± 0.00851.8 ± 1.90.0173 ± 0.00010.00722 ± 0.00007

MolSync improves over the strongest baseline on four of six regression benchmarks, and the largest gains are on the quantum-mechanical targets you ask about: a 12.1 % relative MAE reduction on QM7 (51.8 against 58.9) and a 3.2 % reduction on QM9 (0.00722 against 0.00746). Because these results speak directly to generality, a condensed summary now appears in Section 6.1 of the revised main text, with the full table retained as Table 6.

For QM7–QM9, interatomic distances computed from the coordinates supplied with the datasets are concatenated with the bond features to parameterise the edge-conditioned coupling κij\kappa_{ij}, so that three-dimensional geometry modulates the synchronisation dynamics while preserving invariance to global rotations and translations. This matches the protocol used by the 3D-aware baselines in Table 6, and for QM9 we follow the same three-target evaluation setting as Uni-Mol.

On the overlapping standard deviation bands

You are right that the margins over the strongest baselines in Table 1 are small relative to the standard deviations, and we want to be precise rather than let the average carry more weight than it can bear.

The baseline rows in Table 1 are cited from the compilation in Liu et al. (2026) rather than reproduced in our pipeline, so no per-run values are available for them and a paired significance test against them is not possible. We therefore do not claim statistical significance over the cited baselines, and on datasets where the margin is smaller than our own standard deviation we do not claim a per-dataset improvement. Language in the manuscript implying otherwise has been removed, and every average in Table 1 has been recomputed directly from the per-dataset cells, which corrected five rows that had inherited averages computed over a different benchmark set.

What Table 1 does establish, and what we now state in Section 6.1, is that across the eight classification benchmarks MolSync reaches the level of large pre-trained 3D-conformer models while being trained from scratch on each benchmark's training split, from the 2D bond graph alone.

AuthorsOpenReview ↗

Rebuttal Part2

W2 (1/2) — Connection to existing graph dynamics methods (GRAND, GRAND++, PDE-GCN)

We thank the reviewer for pointing out this important connection. We agree that MolSync belongs to the broader family of graph neural architectures that formulate representation learning through dynamical systems. We also acknowledge that continuous-depth modeling, time-dependent propagation, and the use of dynamical systems to alleviate limitations of conventional GNNs are not unique to our work. Our intention is not to claim otherwise. Instead, the fundamental distinction lies in what evolves, how the local dynamics are generated, and how relational information is organized through the evolution.

In diffusion-based graph dynamics, such as GRAND, node representations evolve through a graph-dependent diffusion operator. A representative formulation is

∂H(t)∂t=div⁡(G(H(t),t)⊙∇H(t)),\frac{\partial H(t)}{\partial t} = \operatorname{div}\left( G(H(t),t)\odot\nabla H(t) \right),

which reduces to a graph Laplacian diffusion process in a simplified setting:

∂H(t)∂t=−LGH(t).\frac{\partial H(t)}{\partial t}=-L_GH(t).

GRAND++ further introduces a source term to the diffusion equation:

∂Z(t)∂t=div⁡(G(Z(t),t)⊙∇Z(t))+S.\frac{\partial Z(t)}{\partial t} = \operatorname{div}\left( G(Z(t),t)\odot\nabla Z(t) \right)+S.

These formulations primarily describe the evolution of node features through graph-dependent propagation or diffusion. PDE-GCN develops a broader family of PDE-motivated architectures, but its central modeling perspective is likewise to construct graph representation updates from discretized PDE-inspired operators.

MolSync adopts a different dynamical perspective. Rather than directly evolving unconstrained node feature vectors through a diffusion operator, it represents each atom by a unit-norm oscillator state

xi(t)∈Sd−1,x_i(t)\in\mathbb S^{d-1},

and models its evolution through two complementary components:

\boxed{ \dot{x}_i = \underbrace{\Omega_i x_i}_{\text{node-specific intrinsic dynamics}} + \underbrace{ \operatorname{Proj}_{x_i} \left( c_i+\sum_{j\in\mathcal N(i)} \kappa_{ij}x_j \right) }_{\text{relational interaction dynamics}}. } \tag{R1}

Here, Proj⁡xi(y)=y−⟨xi,y⟩xi\operatorname{Proj}_{x_i}(y)=y-\langle x_i,y\rangle x_i is the tangent-space projection, cic_i is a node-conditioned stimulus, and κij\kappa_{ij} is a bond-conditioned signed coupling. Crucially, the intrinsic generator is node-specific:

\Omega_i = \frac{1}{2} \left( W_i-W_i^\top \right), \qquad W_i=\operatorname{MLP}_{\mathrm{atom}}(c_i). \tag{R2}

Thus, although the generator-producing network is shared, different nodes obtain different skew-symmetric generators according to their contextual stimuli. Since Ωi⊤=−Ωi\Omega_i^\top=-\Omega_i, we have

⟨xi,Ωixi⟩=0,\langle x_i,\Omega_i x_i\rangle=0,

meaning that this component induces an intrinsic rotation on the sphere rather than directly increasing or decreasing the state norm. This provides a node-specific dynamical tendency that is distinct from the interaction induced by neighboring states.

AuthorsOpenReview ↗

Rebuttal Part3

W2 (2/2) — Connection to existing graph dynamics methods (continued)

The second component introduces signed relational interactions. In particular, positive and negative κij\kappa_{ij} favor alignment and separation, respectively. For fixed stimuli and symmetric couplings, the interaction and stimulus terms correspond to the negative Riemannian gradient of the potential

E(X) = -\frac{1}{2} \sum_{i}\sum_{j\in\mathcal N(i)} \kappa_{ij}\langle x_i,x_j\rangle - \sum_i\langle c_i,x_i\rangle. \tag{R3}

The intrinsic rotations provide additional tangential drift, so the overall dynamics are not restricted to monotonically minimizing this potential. Consequently, MolSync is not simply another parameterization of feature diffusion: it combines node-conditioned intrinsic evolution with signed relational interactions on a constrained state space.

A further distinction is that MolSync does not use the oscillator states only as intermediate propagated features. It explicitly extracts relative synchronization coherence through a learned norm-based readout. For example, each coherence channel is computed as

zi,k=∑j∈N~(i)Wkxj(T),z_{i,k} = \sum_{j\in\widetilde{\mathcal N}(i)} W_k x_j(T),

where WkW_k is the learned linear projection for channel kk, followed by the norm-based coherence representation

ri=[∥zi,1∥2,…,∥zi,K∥2]⊤.r_i= \left[ \|z_{i,1}\|_2,\ldots,\|z_{i,K}\|_2 \right]^\top.

Taking the channel norms yields a magnitude-based description of neighborhood coherence rather than retaining the raw oscillator coordinates. With the learned projections WkW_k used in the implementation, however, strict invariance to an arbitrary common orthogonal transformation is not assumed. Importantly, the resulting coherence is fed back into the stimulus field:

c_i^{(\ell)} = c_i^{(\ell-1)} + \operatorname{MLP}_{\mathrm{update}}(r_i^{(\ell)}). \tag{R4}

This creates an iterative feedback mechanism in which the outcomes of relational interactions refine the node contexts, which subsequently generate the intrinsic dynamics of the next block.

Taken together, the distinction can be summarized as follows:

  • Diffusion/PDE-based graph dynamics: evolve node features through graph-dependent propagation operators.
  • MolSync: evolve node-specific oscillator states through context-conditioned intrinsic dynamics and signed relational interactions, and iteratively reorganize node representations according to the resulting synchronization coherence.

We therefore view MolSync as a synchronization-based relational representation learning mechanism, rather than merely a continuous-depth message-passing architecture with a different update parameterization. We will revise the related-work discussion to clarify this distinction and avoid overstating the novelty of continuous graph dynamics itself.

We should be clear about the limits of this answer: it is a structural comparison, and we have not run GRAND or PDE-GCN on these benchmarks. If you consider an empirical comparison against a continuous-dynamics baseline important, we will add one during the discussion period; GRAND is the natural choice.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.