acceptodds
Under review as a conference paper at ICLR 2027

ShellGatedGNN: Long-Range Message Passing for 2D Molecular Electronic Property Prediction

Abstract

Message passing neural networks (MPNNs) provide a general framework for molecular graph modeling. To capture long-range structural information in a molecule, the standard approach is to stack many layers of 1-hop message passing; however, such propagation does not capture complex topology precisely and is prone to oversmoothing as it gets deeper. We propose SGGNN (ShellGatedGNN), a long-range message-passing architecture. It extends classical gated message passing from a single edge to the triple (sender, sender's environment, receiver), where the sender's environment is the structure around the sender. We divide the nodes into shells by their distance to the receiver, a common grouping strategy, so that the nodes in one shell are at the same distance. This shelling is permutation-equivariant, so SGGNN can use a concatenation instead of a sum in the node update, which keeps more information. On QM9 and QM40, SGGNN attains the best performance among the strong baselines on all six electronic properties. On the HOMO-LUMO gap regression task of the large-scale PCQM4Mv2 benchmark, it outperforms the message-passing baselines. Our code is available at the anonymous repository https://anonymous.4open.science/r/SGGNN-114514A.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Molecular Property Prediction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Tj28OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ShellGatedGNN for predicting electronic properties from 2D molecular graphs. The model partitions nodes into shells according to shortest-path distance and constructs receiver-relative sender environments using parent-edge features, same-shell edge features, and accumulated parent states. These environments enter both the message and gate computations, and the shell outputs are concatenated to update node representations. The paper reports improvements across six properties on QM9 and QM40, a validation MAE of 0.0957 eV on PCQM4Mv2, and depth-sweep and component-ablation experiments.

Critical strengths

The reported improvements on QM9 and QM40 cover multiple targets. SGGNN achieves the lowest reported MAE on all six properties in both datasets. Relative to the strongest reported baseline on each dataset, std-MIX improves by approximately 12.4% and 3.9%, respectively. These results provide meaningful support for the effectiveness of the complete architecture in the evaluated settings.

The receiver-relative environment and accumulated parent states constitute a concrete architectural design. They provide a structured way to use local connectivity to modulate multi-hop communication. The equations, complete forward-pass algorithm, and component experiments make the design inspectable. Training on PCQM4Mv2 also demonstrates the feasibility of handling a large number of samples.

Critical weaknesses

The PCQM4Mv2 comparison omits stronger public baselines. SGGNN reports a validation MAE of 0.0957, while publicly available results include 0.0937 for GraphGPS-small and 0.0901 for CoAtGIN-tiny [1, 2]. Neither appears in Table 3. GraphGPS is evaluated on QM9 and QM40, but its omission from the large-scale comparison is not adequately explained.

The novelty relative to existing shell-based methods is unclear. AIMNet-X2D already aggregates shells and concatenates their representations [3]. KP-GNN also incorporates structural information within hop-based neighborhoods, but is neither discussed nor compared [4]. The paper needs to distinguish its contribution more clearly from these existing designs.

The ablations do not establish the gate’s independent contribution. Every configuration in Table 4 retains the gate. The experiments also do not separate the benefit of supplying environmental information to the message from supplying it to the gate. Consequently, the specific value of environment-conditioned gating remains unresolved.

The long-range modeling claim lacks targeted validation. The paper does not analyze how its advantage varies with molecular size or graph distance. Evaluation on substantially larger molecular graphs, such as Peptides-struct [5], or a controlled size- and distance-based analysis within the existing tasks would more directly test this claim.

References

[1] Rampášek, L., et al. (2022). Recipe for a General, Powerful, Scalable Graph Transformer. Advances in Neural Information Processing Systems, 35. Official PCQM4Mv2 results.

[2] Zhang, X., et al. (2022). CoAtGIN: Marrying Convolution and Attention for Graph-based Molecule Property Prediction. IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp. 374–379. Official OGB-LSC leaderboard.

[3] Nandakumar, R. (2025). AIMNET-X2D: Universal and Scalable Framework for Chemical Property Transfer Learning. Master’s thesis, Carnegie Mellon University. 2025 implementation of shell aggregation and concatenation.

[4] Feng, J., et al. (2022). How Powerful are K-hop Message Passing Graph Neural Networks. Advances in Neural Information Processing Systems, 35.

[5] Dwivedi, V. P., et al. (2022). Long Range Graph Benchmark. Advances in Neural Information Processing Systems, 35, Datasets and Benchmarks Track.

Questions for authors

Why were GraphGPS and other stronger public PCQM4Mv2 results not included? Can the authors expand the comparison, documenting input information, parameter counts, training budgets, and measured runtime and memory usage, and clarify whether SGGNN offers a competitive accuracy–cost trade-off?

What specifically does the method add beyond AIMNet-X2D’s existing shell aggregation and concatenation, and beyond KP-GNN? Can the authors compare removing the gate, or removing environmental information only from the gate, while controlling model capacity and keeping the message inputs unchanged?

Does the advantage persist when the task places greater demands on long-range communication? Please provide targeted validation on larger molecular graphs, such as Peptides-struct, or evaluate performance by molecular size and graph distance within the paper’s task scope, reporting well-tuned baselines and computational costs. These experiments would test the central mechanism more directly than adding unrelated datasets.

3 comments

We are grateful for the detailed comments. Each point is addressed below, together with the corresponding revisions.

W1 & Q1 — Stronger public baselines on PCQM4Mv2

Question: GraphGPS (0.0937) and CoAtGIN-tiny (0.0901) report better PCQM4Mv2 MAE than ours (0.0957) but do not appear in the comparison; why, and is SGGNN competitive in performance-versus-cost terms?

Reply. We understand the concern that stronger public baselines are absent from the comparison. GraphGPS and CoAtGIN-tiny are hybrids of a graph transformer and message passing, which is a different class of method from pure message passing. The submitted abstract already noted that the comparison on the PCQM4Mv2 dataset is between message-passing methods. In this revision we expanded Appendix B.1 into a delimitation of the four families — message passing, graph transformers, 3D methods and their hybrids.

Cross-class comparison is nevertheless important — we already include GraphGPS in the QM9 and QM40 comparisons — so we have added GraphGPS and CoAtGIN-tiny to the large-scale table. This does not affect our claim of being state of the art within the same class of methods.

We additionally report hardware, training time, parameter count and input type, so that the cost-versus-performance trade-off is visible. Table 1 is the new table. All baseline figures in that table are taken from the official OGB website. Note: 0.0937 is GraphGPS-small; the table lists the stronger GraphGPS entry on OGB, 0.0852.

Table 1 — PCQM4Mv2 validation MAE (eV, lower is better). A dash means the work reports no training time.

MethodInputParamsHardwareTimeMAE
GCN2D1.96M1× RTX 2080--0.1379
GIN2D3.76M1× RTX 2080--0.1195
GCN-virtual2D4.85M1× RTX 2080--0.1153
GIN-virtual2D6.66M1× RTX 2080--0.1083
SGGNN (ours)2D8.56M1× RTX 5070 Ti laptop (12 GB)17.9 h0.0957
GraphGPS2D13.81M1× A100--0.0852
CoAtGIN-tiny2D6.40M1× A40--0.0901
HFAGNN2D+3D3.95M1× TITAN V--0.1005

W2 & Q2 — Novelty over AIMNet-X2D and KP-GNN

Question: AIMNet-X2D already aggregates and concatenates shell representations and KP-GNN already uses hop-wise structural information, yet neither is discussed or compared; what remains beyond them?

Reply. We understand the concern about novelty, since the original submission did not discuss these two multi-hop GNNs in detail. Our novel elements are the following.

1. AIMNet-X2D. Grouping shells by BFS distance is not new in multi-hop GNNs, but there are three differences from AIMNet-X2D. First, in how a shell representation is obtained: AIMNet-X2D maps each shell through its own linear layer and merges the shells by a learned weighted sum, whereas we gate each node pair on the sender's environment and normalize within the shell. Second, AIMNet-X2D does not use edge features in message passing but only nodes, whereas our model uses edge features directly through the sender-environment gate. Third, in what a shell contains: AIMNet-X2D only groups the k-hop nodes into a shell, whereas we group the k-hop nodes together with their environment. AIMNet-X2D does concatenate its shells; the concatenation of the node update is therefore not our contribution either, and the core innovation remains the sender-environment-controlled gating that aggregates the shells. We revised the presentation in the contributions of the introduction (Section 1), added the distinction from these two methods in the related work (Section 2), and added the capacity-matched sum-versus-concatenation ablation to Appendix D.6.

2. KP-GNN. Its core contribution is that, beyond k-hop aggregation, it uses a structural summary of the induced subgraph on the receiver's k-th hop (edge-type counts and node configurations of that subgraph), encoded and added to that hop's aggregated output. The core contribution of our method is to modulate the aggregation of the k-th hop message with a sender-centred environment combined with a gate. The way environment information is used clearly differs between the two.

3. KP-GNN as a baseline. KP-GNN is a classic multi-hop method, and it was not included in the original submission; this deserves attention, so we trained it as a baseline on QM9 and QM40 and added its results to the QM9 and QM40 comparison tables of the main text. Table 2 lists the strongest baselines after adding KP-GNN; our model remains ahead.

To compare cost and performance we also report the training time on QM9 and QM40 on the same hardware (RTX 5070 Ti laptop, 12 GB), in Table 2. Our model performs better than KP-GNN; the training time is 32.3 vs 288.4 min on QM40 and 97.0 vs 67.7 min on QM9. Our method is multi-hop, yet on QM40 it reaches the state of the art with about the training time of the strongest one-hop baseline, GatedGCN+ (32.3 vs 27.3 min), while its training cost is far below that of the two multi-hop methods KP-GIN (VN) (288.4 min) and DuAxMPNN (218.1 min).

Table 2 — Strongest baselines on QM9 and QM40: test std-MIX (three runs, sample std in parentheses) and training time on the full training set (median of three runs).

MethodQM9 std-MIXQM40 std-MIXQM9 train (min)QM40 train (min)
GatedGCN+0.1037 (0.0002)0.2012 (0.0004)17.127.3
KP-GIN (VN)0.0992 (0.0004)0.2056 (0.0002)67.7288.4
DuAxMPNN0.0933 (0.0005)0.2104 (0.0061)173.2218.1
SGGNN (ours)0.0817 (0.0004)0.1933 (0.00003)97.032.3

W3 & Q3 — Does the ablation isolate the gate?

Question: Every configuration in the ablation keeps the gate, and the benefit of giving the environment to the message is not separated from giving it to the gate; what is the value of environment-conditioned gating?

Reply. We thank the reviewer for this insightful comment. Separating the role of the gate from that of the environment information is indeed valuable, and in response we have designed a dedicated ablation for exactly this question and added it to the ablation study of the main text (Section 4.4). Table 3 reports the new ablation, with the parameter count matched across all five rows; the result is that both the gate and the environment information are positive contributors.

Table 3 — Gate and environment entry points, matched parameters (QM9, Dmax=4; three runs, sample std in parentheses).

Variantenv → messageenv → gatestd-MIXabs. Δ vs both-offrel. Δ
both off××0.0899 (0.0003)——
full model✓✓0.0816 (0.0001)0.0083 (0.0003)9.3 (0.3)
message only✓×0.0823 (0.0002)0.0076 (0.0005)8.4 (0.6)
no gate (g=1)✓–0.0851 (0.0002)0.0048 (0.0004)5.3 (0.5)
gate only×✓0.0859 (0.0002)0.0040 (0.0004)4.5 (0.4)

All five rows are matched in parameter count. The full-model row was re-trained inside this ablation campaign; its 0.0816 reproduces the main-table value of 0.0817 to within the reported spread.


W4 & Q4 — Targeted validation of the long-range claim

Question: The long-range claim is not analysed against molecular size or graph distance; does the advantage persist when longer-range communication matters more?

Reply. We appreciate the opportunity to clarify this point, and we regret that our wording invited this reading. The claim we make is not that long-range modelling has a stronger effect on larger molecules; it is that our long-range message-passing model substantially improves performance on molecular property prediction, and that its architecture genuinely reaches further structure than conventional one-hop message passing. Our experiments show that the advantage is present across molecular sizes: the molecules of QM9 are clearly smaller than those of QM40, and we are also ahead on QM9 (Table 2).

Your question points to an aspect we are glad to examine further, namely how the gain of long-range message passing relates to graph size. We therefore re-ran inference with the already-trained SGGNN and baseline checkpoints on the QM9 and QM40 test splits and recorded, for every molecule, its error, number of nodes and BFS diameter, and plotted the resulting distributions. The result is that our advantage is present across molecular sizes. We also observed one curious phenomenon: the prediction errors of the different models show an interesting consistency in how they distribute over node count and BFS diameter, and on QM40 the prediction difficulty increases with molecular size. We have added this to Appendix D.5.

Reviewer_Hq1cOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a new graph neural network architecture, the ShellGatedGNN (SGGNN) and proposes that this architecture is especially well suited to model long range structural information in 2D molecular graphs. To achieve this not only „one-hop“ messages are included in each message passing step. Instead for each receiver, all sender nodes are grouped by their BFS distance into shells. Messages are then accumulated first in each of these shells and only afterwards mixed into the receiver node features. Additionally the messages of each sender node do not only depend on the local node features, but also on the environment of the sender node.

Overall, while the proposed method is novel and interesting, it seems like the given comparisons to existing methods are too sparse too support the claims.

Critical strengths

To my knowledge the presented method is novel and an interesting approach in itself. Passing long range information in graph neural networks is a very active field with important applications and novel ideas that tackle this problem are definitely interesting for the community.

Critical weaknesses

It should be stated much more clearly on what basis models are excluded and included in the provided comparisons in Table 1, 2 and 3. As it stands the provided comparisons make it seem like the introducde method reaches a new state of the art on QM9, QM40 and PCQM4Mv2. However there are many better models on these datasets (see for example Gotennet for QM9 and the leaderboard on https://ogb.stanford.edu/docs/lsc/leaderboards/ for PCQM4Mv2). At multiple points it is said that SGGNN is compared to models of the same type. What does that mean in detail? And why should other models be excluded here?

Using only the 2D graph and no atom positions is justified by arguing that accurate conformers and therefore atom positions are expensive to generate. While this is true, most of the best models listed in the leaderboard on https://ogb.stanford.edu/docs/lsc/leaderboards/ seem to use an inexpensive method to generate the conformer information and use these positions as input. These methods perform better on PCQM4Mv2 as the model presented here. It is not clear why cheap position information should not be used.

Questions for authors
  • Could a detailed justification for the choice of baseline methods (and the exclusion of methods in that context) be given?
  • Would an evaluation of the method on some other not molecule related problem be possible? Perhaps a setting where no information like atom positions exist would be informative?
  • It seems a bit odd that the discussion is placed in the Appendix. Could this be moved to the main body, maybe moving ablations to the appendix?
  • Why only evaluate on the 6 given properties on QM9 and QM40? Why not on the rest?
6 comments

We sincerely appreciate the constructive feedback. Our answers follow the order of your questions below.

W1 & Q1 — On what basis are models included in or excluded from the comparison, and what "same type of model" means

Question: The three comparison tables should state more explicitly on what basis a model is included in or excluded from the comparison offered, since many better models exist on these datasets (e.g. GOTENNet on QM9, and the PCQM4Mv2 leaderboard). The replies repeatedly say that SGGNN is compared with models of the same type. What exactly does that mean, and why should other models be excluded?

Reply. We understand the concern: the class of method our approach belongs to is worth stating explicitly. The paper now delineates the difference from the other classes of model in more detail in Appendix B.1, which distinguishes the model families. GOTENNet, which you mention for QM9, is mainly an equivariant method among 3D methods, and most of the methods at the top of the PCQM4Mv2 leaderboard are graph-transformer methods or hybrid methods; neither is the same class as our purely message-passing method.


W2 & Q2 — Why the cheap 3D information used by most top leaderboard models is not used here

Question: Most of the best models on the leaderboard appear to use cheaply generated junction information as input, and these methods do better on PCQM4Mv2. It is not clear why this cheap positional information should not be used.

Reply. This 3D information is in general continuous and needs a 3D GNN, or geometric GNN, to process continuous spatial information; GOTENNet, which you mention for QM9, is of this type. Our model is conventional message passing and is not of the same type, and conventional message passing cannot use such continuous spatial information effectively. Exploiting cheap molecular 3D spatial information is outside the scope of our discussion. We do, however, consider comparison across method classes important, so we have added some graph-transformer methods and 3D-GNN methods from the leaderboard to the large-scale comparison table below as controls. Adding them clearly does not affect our state-of-the-art position in the message-passing setting.

Table 1 — PCQM4Mv2 validation MAE (eV, lower is better). The upper block holds 2D message-passing models, sorted by MAE; the lower block is not 2D message passing and is listed for reference only (HFAGNN also uses 3D conformations). All entries are single models; parameter counts, hardware and training time are as reported by the original work or the submission, and a dash means no wall-clock time is reported.

MethodInputParamsHardwareTimeMAE
GCN2D1.96M1× RTX 2080--0.1379
GIN2D3.76M1× RTX 2080--0.1195
GCN-virtual2D4.85M1× RTX 2080--0.1153
GIN-virtual2D6.66M1× RTX 2080--0.1083
SGGNN (ours)2D8.56M1× RTX 5070 Ti laptop (12 GB)17.9 h0.0957
GraphGPS2D13.81M1× A100--0.0852
CoAtGIN-tiny2D6.40M1× A40--0.0901
HFAGNN2D+3D3.95M1× TITAN V--0.1005

Q3 — Evaluation on non-molecular problems, and a setting without atom positions

Question: Can the method be evaluated on other, non-molecule-related problems? A setting without information such as atom positions would perhaps be more informative?

Reply (1). Our claims concern molecular property prediction, and we have not extended the evaluation to non-molecule-related problems. Our architecture was nevertheless designed with generality in mind and can generalise to other undirected graphs with edge features and node features, provided those features are discrete; we say so in Appendix A.2, where continuous node or edge attributes are listed as unvalidated. We do not guarantee, however, that it retains its performance advantage on other non-molecule-related problems; we can only guarantee its generality. We ran experiments on four graph-classification datasets, and the ten-fold cross-validation results are given in Table 2.

Table 2 — Transfer to non-molecular graphs. Ten-fold cross-validated accuracy on four TUDataset benchmarks; one shared protocol for all three models (width 64, 3 layers, 100 epochs, batch size 32, per-fold model selection on a held-out 10% of the training split, no task-specific tuning). SGGNN keeps the architecture and depth setting of the molecular experiments (Dmax=4, L=3) with only the width reduced, and consumes the node and edge labels supplied by TU. Standard deviations over the ten folds are in parentheses.

Dataset#graphsN maxSGGNNGCNGIN
MUTAG188280.8406 (0.0603)0.6865 (0.0944)0.8249 (0.1010)
PTC_MR344640.5961 (0.0870)0.5518 (0.0581)0.5869 (0.0392)
BZR_MD306330.6892 (0.0799)0.6570 (0.0754)0.6637 (0.0803)
DHFR_MD393390.6667 (0.0554)0.6794 (0.0116)0.6407 (0.0646)

Reply (2). We do not use atom-position information: all of our experiments use only the 2D molecular graph as stored in the SMILES string, and contain no spatial information of any kind, including atom positions. We have added the input type to the PCQM4Mv2 comparison table above.


Q4 — Placement of the discussion

Question: It seems a little odd that this discussion is placed in the appendix. Could the backbone be moved into the main body, and perhaps the ablation moved to the appendix?

Reply. We understand the concern: introducing the relevant background is important for a deeper understanding of the paper, but the main text is limited in length, and the ablation section contains the experimental evidence that is extremely important to our main contribution, so it is difficult for us to move all of the discussion into the main text. As an alternative, we have strengthened the delineation from some closely related methods in the related work (Section 2).


Q5 — Why only the six properties of QM9 and QM40

Question: Why is the evaluation done only on the six given properties of QM9 and QM40, and not on the others?

Reply. The title and the experiments of the submitted version already indicate that this work focuses on the prediction of molecular electronic properties and does not regress every property of QM9 and QM40. The six tasks of our experiments correspond to electronic properties: the HOMO–LUMO gap in particular has very strong application value and can be used for property prediction of organic photovoltaic devices and for structure-based virtual screening in drug discovery (see the PCQM4Mv2 website). Most of the remaining properties are thermodynamic, which is outside the scope of this study. We have added this to the limitations (Appendix A.1): the targets are the electronic properties only, and extending the model to the regression of thermal properties is listed as future work.

Reviewer_Hq1cOpenReview ↗

Response to rebuttal

Thank you for the detailed answers and thorough revision!

The more explicit distinction between model classes and the openly stated limitation of restricting to six electronic properties are significant improvements.

However, the baseline selection remains a concern. Baselines are excluded from main tables with the explanation that they belong to a different model class (for example Table 1 in the revised manuscript). Without a clear disadvantage of such models, it is not clear why a comparison to them would be inappropriate. The claim that conventional message passing cannot use continuous spatial information is not convincing, as distances are routinely used as edge features in many GNNs. It is also not clear why Table 2 in the rebuttal is titled "Transfer to non-molecular graphs", although the datasets are molecular.

The detailed answer is appreciated, but for the reasons above I will keep the rating at 2.

Reviewer_2ezwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ShellGatedGNN (SGGNN), a message-passing architecture for predicting electronic properties from 2D molecular graphs. For each receiver, nodes are grouped into shells according to shortest-path distance. Messages and elementwise gates incorporate a sender-environment representation consisting of parent-edge and same-shell edge feature sums, together with accumulated parent ledger states. Messages are aggregated within each shell using gate normalization, and the shell representations are concatenated with the receiver’s features before a residual node update. This design aims to capture long-range interactions and receiver-relative topology while retaining shell-specific information before projection. The paper also provides permutation-equivariance proofs and analyses of receptive-field coverage and computational complexity.

Experiments cover six electronic properties on QM9 and QM40 and HOMO–LUMO gap prediction on PCQM4Mv2. SGGNN achieves the lowest reported mean MAE on all six properties among the evaluated baselines on QM9 and QM40. On PCQM4Mv2, it obtains a validation MAE of 0.0957 eV, outperforming the four reported GCN/GIN baselines, including their virtual-node variants. Depth sweeps indicate that most improvements in the aggregate metric are obtained by a shell budget of approximately four. Ablation studies support the contributions of the sender-environment components and shell-wise concatenation.

Critical strengths
  1. A coherent architecture for incorporating structural context into long-range message passing. The main methodological strength is the integration of BFS shells with a receiver-relative description of the sender’s environment. Parent-edge and same-shell edge summaries distinguish different types of local connectivity, while accumulated parent ledger states provide additional context. Conditioning both messages and gates on this environment gives the model a concrete mechanism for incorporating topology into multi-hop interactions. The shell representations retain their distance-specific identities through concatenation before projection, and the complete update preserves permutation equivariance. This design is consequential because it provides a structured approach to long-range molecular modeling using only 2D graph inputs, with clearly specified components and a complete forward algorithm.

  2. Consistent empirical improvements across multiple molecular prediction settings. On QM9 and QM40, SGGNN achieves the lowest reported mean MAE on all six electronic properties among the evaluated baselines. Relative to the strongest respective baselines in Tables 1–2, std-MIX decreases from 0.0933 to 0.0817 on QM9 and from 0.2012 to 0.1933 on QM40, corresponding to approximately 12.4% and 3.9% reductions. Shared data splits and input features, together with results over three runs, strengthen the evidence for these improvements. The QM9 ablations provide supporting evidence for the usefulness of the environment components and shell concatenation, while the PCQM4Mv2 validation MAE of 0.0957 eV demonstrates applicability at a much larger dataset scale. The consistency across targets and molecular sizes makes the empirical contribution relevant beyond a single benchmark result.

Critical weaknesses
  1. The experiments do not sufficiently separate the benefits of the proposed architecture from differences in model capacity and optimization. Shared data splits, input features, and hidden width are useful controls, but the manuscript does not provide a sufficiently complete comparison of parameter counts, training schedules, and tuning budgets for the strongest baselines. Increasing the shell budget also increases the number of depth-specific parameters and the size of the fusion projection. Furthermore, the implementation and parameter count of the summation variant in the concatenation ablation are not sufficiently specified. The zero-filling strategy appropriately retains parameter blocks for the environment ablations, but does not by itself resolve these other capacity questions. These gaps limit confidence in attributing the improvements specifically to longer-range structural information and preservation of shell identities, which are central to the methodological contribution.

  2. The practical trade-off between predictive accuracy and computational cost remains insufficiently characterized. Appendix B.9 identifies an O(Dmax⁡N3C)O(D_{\max}N^3C) cost for the current dense edge-sum computation, together with additional computation over shell pairs and parent ledger states. However, the paper does not report comparative training time, inference throughput, or peak memory usage. The PCQM4Mv2 experiment demonstrates that the model can be trained on a large collection of molecules, but does not establish the computational cost of its accuracy gains relative to the main baselines. This matters for assessing the practical value of the method, particularly given the motivation of large-scale molecular screening.

Questions for authors
  1. Clarification of baseline comparability. Could the authors summarize the parameter counts, layer counts, training schedules, early-stopping criteria, and hyperparameter-selection budgets for the strongest baselines, particularly DuAxMPNN and GatedGCN+? Were their training objectives and checkpoint-selection criteria aligned with the six-property evaluation? References to the exact released configurations would also help assess the comparability of the reported results.

  2. Clarification and supporting evidence for the ablations. How exactly is the projection implemented in the “sum” variant of Table 4, and how does its parameter count compare with the concatenation variant? Since increasing Dmax⁡D_{\max} also increases model capacity, do existing results—or a targeted comparison with approximately matched parameter counts—show that the improvements from greater shell range or concatenation persist after controlling for capacity? Such evidence would directly strengthen the attribution of the gains to the proposed mechanisms.

  3. Evidence for the practical computational trade-off. Could the authors provide representative training time or throughput, inference throughput, and peak memory measurements for SGGNN and a strong message-passing baseline on the same hardware, with batch sizes specified, on QM9 or QM40? Please also clarify whether BFS distances and static edge sums are cached or recomputed, and whether their cost is included in the measurements. These results would help determine the computational price of the reported accuracy improvements.

2 comments

Thank you for your comments. We have considered each point and answer them one by one below.

W1 & Q1 — Baseline comparability: parameters, layers, schedules, tuning budget

Question: Which exact configurations do the strongest baselines (DuAxMPNN, GatedGCN+) use — parameters, layers, training schedule, early-stopping criterion, hyper-parameter budget — and is the training objective and checkpoint selection the same as for the six properties?

Reply. Spelling these details out is also what makes the comparison in our experiments verifiable and persuasive. The table below gives the parameter count and the number of layers of the two strongest baselines, DuAxMPNN and GatedGCN+. The training schedule uses the official default hyperparameters of the baselines, with no early stopping, and a fixed number of epochs with the best validation epoch selected. Neither our model nor the baselines performed any hyperparameter selection for the task. Our model and all baselines use the same six-property objective. We additionally ran the capacity-matched GatedGCN+ experiment on QM9 and QM40 (Table 1).

Table 1 — Baselines and SGGNN on QM9 and QM40 (three runs; sample std in parentheses).

MethodLayersParamsQM9 train (min)QM9 std-MIXQM40 train (min)QM40 std-MIX
GatedGCN+41.40M17.10.1037 (0.0002)27.30.2012 (0.0004)
GatedGCN+* (capacity-matched)268.66M60.70.0936 (0.0014)122.10.1998 (0.0005)
DuAxMPNN59.81M173.20.0933 (0.0005)218.10.2104 (0.0061)
SGGNN (ours)38.55M97.00.0817 (0.0004)32.30.1933 (0.00003)

This table has been added to Appendix D.8. After controlling the parameter count, GatedGCN+ improves noticeably, but it does not change the state-of-the-art position of our model.


W2 & Q2 — (a) the sum variant: implementation and parameter count; (b) is the shell-range gain capacity-driven?

Question: How is the projection of the "sum" variant implemented and how do its parameters compare with concatenation; does the gain of more shell range or of concatenation survive when capacity is controlled?

Reply (a). We agree that the ablation of the concatenation versus the summation node update needs a matched parameter count to keep the attribution correct: as implemented, the "sum" variant and the concatenation variant differ in their parameters, so we added a parameter-matched node-update ablation, now in Appendix D.6. Once the parameters are matched, the concatenation variant is still clearly better than the "sum" variant, so the conclusion that concatenation brings a larger improvement is unchanged.

Table 2 — Node update at matched capacity (QM9, Dmax=4, H=256, test std-MIX; Δ against concat on unrounded run means).

ReadoutFusionParamsstd-MIXΔ
concat[(Dmax+1)H] → H8,551,1100.0817 (0.0004)—
sumH → H7,764,6780.0854 (0.0002)+4.4%
sum+wideH → ⌊(Dmax+1)H/2⌋ → H8,556,8700.0846 (0.0002)+3.4%

Reply (b). For the depth sweep we ran an additional parameter-matched experiment, but the full-scale experiment clearly cannot be finished within the discussion period. As a substitute, we changed H from 256 to 64, and on both datasets we drew a smaller sample by gap-stratified sampling so that the experiment could be completed within the discussion period and demonstrate the trend; it has been added to Appendix D.7. The results show that, after controlling the capacity, the larger shell range keeps improving on QM9 from Dmax=1 to 4 and hardly improves after 4, which is the same phenomenon as in our original experiment, whereas on QM40 the improvement continues all the way to Dmax=9, which differs from the best Dmax=4 of the original experiment, but this phenomenon instead further confirms that more shells bring more improvement.

Table 3 — Parameter-matched depth sweep (test std-MIX, mean of three runs).

Dataset123456789
QM90.20540.15490.14380.14080.14010.14010.13990.14020.1399
QM400.30370.26490.25410.24600.24110.23620.23410.23290.2320

W3 & Q3 — Cost of the performance gain: training, inference, memory, caching

Question: What do SGGNN and strong message-passing baselines cost (training time, inference throughput, peak memory) on the same hardware; are the BFS distances and the static edge sums cached or recomputed, and are their costs inside the reported numbers?

Reply (a). Measuring the cost does provide an additional check of our model's competitiveness beyond performance: we added the training time and the inference throughput of our model and of the strong message-passing baselines on QM9 and QM40 on the same hardware, an RTX 5070 Ti laptop, and added them to Appendix D.4.

Table 4 — QM9 (the four lowest-std-MIX methods of the main comparison).

MethodTrain minTrain GBInfer molecule/sInfer GB
GatedGCN+17.10.2624,0000.13
KP-GIN (VN)67.71.76,0001.2
DuAxMPNN173.21.55,1000.22
SGGNN (ours)97.00.887,0000.97

Table 5 — QM40 (16,296 test molecules).

MethodTrain minTrain GBInfer molecule/sInfer GB
KP-GIN (VN)288.47.51,273 (bs=32)8.5
GIN+14.90.2826,0000.16
GatedGCN+27.30.6018,0000.28
SGGNN (ours)32.32.42,2000.66

Note: for KP-GIN (VN) only the two inference cells (1,273 molecule/s and 8.5 GB) are measured at batch size 32, because it does not fit at batch size 128 (pool 27.7 GB); its throughput is therefore not comparable with the other rows. Its training cells (288.4 min, 7.5 GB) are at batch size 128.

Reply (b). For the PCQM4Mv2 we added the training time, the training hardware and the input information, which shows that we can train on millions of molecules in 17.9 h on a consumer laptop gaming GPU, that we are in a state-of-the-art position among message-passing methods (Table 6), and that we are competitive in both cost and performance. GraphGPS and CoAtGIN-tiny are hybrids of a graph transformer and message passing, and HFAGNN is a hybrid of a 3D GNN and message passing; these three are listed as controls outside the method class. All baseline figures in the table are taken from the official OGB website.

Table 6 — PCQM4Mv2 validation MAE (eV, lower is better). A dash means the work reports no training time.

MethodInputParamsHardwareTimeMAE
GCN2D1.96M1× RTX 2080--0.1379
GIN2D3.76M1× RTX 2080--0.1195
GCN-virtual2D4.85M1× RTX 2080--0.1153
GIN-virtual2D6.66M1× RTX 2080--0.1083
SGGNN (ours)2D8.56M1× RTX 5070 Ti laptop (12 GB)17.9 h0.0957
GraphGPS2D13.81M1× A100--0.0852
CoAtGIN-tiny2D6.40M1× A40--0.0901
HFAGNN2D+3D3.95M1× TITAN V--0.1005

Reply (c). The BFS distances are obtained in preprocessing and are not recomputed, so they are not included in the training time or in the inference throughput. We additionally measured the cost of the BFS distances: about 2 minutes of single-thread CPU time on QM9 and about 3 minutes on QM40, which is negligible against the tens of minutes of GPU time in the table. The two edge sums are not cached in preprocessing either: they are computed inside every forward pass, so their cost is already included in the training time and in the inference throughput. We also measured the cost of the dense edge sums, and the result shows that the cubic term is not a bottleneck at the molecule sizes we have tested; this has been added to Appendix D.10.

Table 7 — Wall-clock share of the two edge sums (Dmax=4, L=3, H=256, batch size 128).

DatasetTwo edge sumsOne forwardShare of forwardOne training stepShare of step
QM9 (N=12)0.52 ms10.49 ms4.9%31.94 ms1.6%
QM40 (N=37)0.44 ms48.37 ms0.9%108.93 ms0.4%

Comments on the paper

1 comment

Dear Area Chair,

As the discussion deadline approaches, the authors and Reviewer Hq1c remain in disagreement on two points, and we therefore request your clarification of the evaluation standard.

Reviewer Hq1c's requests.

  1. Add 3D models to the baselines of the main tables (Tables 1 and 2), for example GOTENNet on QM9.
  2. "The claim that conventional message passing cannot use continuous spatial information is not convincing, as distances are routinely used as edge features in many GNNs."

Our position.

  1. A 3D model should not be added to the baselines of Tables 1 and 2; this lies beyond the setting of this paper, whose inputs are 2D molecular graphs.
  2. We did not write that conventional message passing cannot use continuous spatial information; our formulation in the previous response was that conventional message passing cannot use it effectively. The reviewer's quotation drops the qualifier and turns a statement about effectiveness into one about capability. Conventional message passing is also a subset of GNNs, so evidence about "many GNNs" does not by itself say anything about conventional message passing. Moreover, a distance is still 3D information, and this paper takes only 2D molecular graphs as input.

Our specific reasons.

  1. A 3D model requires additional 3D coordinates, whereas Tables 1 and 2 on QM9 and QM40 are 2D-input experiments: the title scopes the task to 2D molecular property prediction, and the introduction separates methods by their input, either a 3D conformer or the 2D molecular graph. Adding a 3D model would therefore push the input beyond the setting of this paper and break the fair protocol of these experiments.
  2. Encoding distances as edge features also requires additional 3D coordinates: a bond length is not a property of the 2D graph, and in the benchmarks used here it is taken from the DFT-optimised 3D conformer supplied with the dataset. It would likewise go beyond that setting and break the fair protocol.
  3. To our knowledge, models that consume 3D coordinates are rarely used in settings whose inputs are 2D molecular graphs, since no 3D information is available there.

Our request. Could you clarify

(i) Must a model that consumes 3D coordinates appear in the main tables (Tables 1 and 2)?

(ii) Do you agree with Reviewer Hq1c's second point?

Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.