acceptodds
Under review as a conference paper at ICLR 2027

From Reliable Explanations to Better Learning: Bridging the OOD Gap in 3D GNN Explanation

Abstract

Graph neural network (GNN) explainers aim to identify subgraphs that preserve model predictions and have been widely studied on 3D graphs. However, two fundamental questions remain open: (1) Are these explanations reliable? and (2) What benefits can they bring to model learning? In this paper, we show that these two questions are inherently connected through an overlooked issue in 3D GNNs: explanatory subgraphs are fundamentally out-of-distribution (OOD). Due to the cutoff-based construction of 3D proximity graphs, existing explainers break the underlying geometric constraints, producing subgraphs that violate the training distribution. As a result, the predictor behaves unreliably on such inputs, leading to uninformative or even misleading explanations. To address this, we propose E2L, an explanation-aware framework that bridges the OOD gap by aligning the training and explanation distributions. By integrating the explainer into the learning process, the model is directly optimized on the explanatory distribution, ensuring that explanations are in-distribution and reliable for the predictor. Building on this foundation, we further bridge the gap between explanation and representation learning, demonstrating that reliable explanations naturally induce compact and expressive subgraph structures that provide both efficiency and expressivity benefits for representation learning. Experiments on both small and large-scale 3D molecular graph benchmarks show that E2L overcomes the OOD issue, improves explanation reliability, and achieves substantial efficiency and expressivity gains. The code is released at https://anonymous.4open.science/r/E2L-28EE.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
GNN explaination, Out-of-distribution problem, 3D graph learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_JbyJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes E2L, an explainability-oriented learning framework, to address the issue that conventional post-hoc explainers induce out-of-distribution (OOD) shifts when removing edges from 3D proximity graphs constructed using cutoff thresholds, thereby compromising the reliability of pretrained predictors on the resulting subgraphs. By jointly training the explainer and predictor, E2L learns node-adaptive radii to generate compact and sparse subgraphs, mitigating distribution shifts while improving the predictive efficiency and representational capacity of 3D graph neural networks (GNNs). The authors conduct experiments on the QM9, GEOM-Drugs, and EC-Protein datasets using multiple GNN backbones and evaluate the effectiveness of the proposed method in terms of predictive accuracy and computational efficiency.

Critical strengths
  1. The paper identifies a practically relevant failure mode in explainability for 3D graph neural networks: arbitrary edge removal can violate the construction rules of proximity graphs based on distance cutoffs. Linking this structural mismatch to the reliability of explanations provides a novel perspective on the problem.
  2. The proposed method offers a simple parameterization of node-wise radii, enabling adaptive adjustment of receptive fields while preserving the aggregation mechanism of the backbone network. The differentiable edge-gating mechanism supports joint optimization without requiring a separate graph generative model or additional loss functions, demonstrating a degree of methodological novelty.
  3. The experimental evaluation covers molecular property prediction and protein-related tasks, involving multiple GNN backbones and several explainability baselines. The inclusion of multiple comparative experiments helps investigate whether the performance gains stem merely from graph sparsification or from the learned heterogeneous connectivity structures.
Critical weaknesses
  1. Simply emphasizing that previous explanation-aware learning methods primarily target 2D graphs is insufficient to clearly establish the novelty of E2L in 3D settings. The paper should further distinguish its design choices, including node selection, radius constraints, and geometric perturbations, and explain why these differences enable E2L to address limitations that existing methods cannot adequately resolve.
  2. The method performs message passing on the full graph for the first (t) layers before generating a sparse explanation graph from intermediate representations. Thus, information from removed edges may already be encoded in node representations and influence subsequent predictions, making it difficult to attribute low prediction error solely to the sparse structure. Although Appendix F.3 presents E2L-FT as supplementary evidence, its additional fine-tuning complicates the comparison. Further quantification of the full-graph prefix's contribution is needed to establish the explanation structure's independent predictive value
  3. E2L uses continuous sigmoid edge gates during training but applies hard thresholding to construct sparse graphs at inference, resulting in a mismatch between the two computational settings. Although Appendix H demonstrates stability across different sharpness settings, it does not evaluate prediction discrepancies between soft gating and hard pruning using the same trained model.
  4. The paper primarily measures efficiency using edge counts and GFLOPs, which may not fully reflect actual runtime costs. E2L still requires full-graph message passing, radius prediction, and geometric computations on sparse graphs. The authors should report end-to-end inference latency and training time, clarifying whether these additional costs are included.
Questions for authors

See Critical Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer JbyJ - Part I

We thank the reviewer for their detailed feedback. We have revised the paper accordingly and address each comment below.

  • W1. Insufficient 2D–3D distinction; node selection, radius constraints, and geometric perturbations; limitations overcome.

In Sec. 2 and Appendix C, we have already clarified how E2L differs from existing 2D and 3D explanation-aware methods in both problem formulation and explanation setting. Prior work includes prototype-based reasoning (ProtGNN), stochastic attention (GSAT), and 3D geometric interpretation through learnable perturbations (LRI). Neither the primary 2D methods nor existing 3D approaches target the cutoff-induced train–explanation mismatch studied here. Many also adopt different explanation settings, rather than requiring prediction from a compact extracted subgraph.

E2L directly addresses the explanation-induced OOD gap through distribution alignment. This principle is independent of the particular explainer and provides a framework for incorporating compatible differentiable 3D explainers into predictor learning. E2L combines three concrete design choices:

(1) Edge selection that preserves input identity. Removing atom or residue nodes changes the represented molecular or protein instance. We therefore retain every node and select only interaction edges. This differs from 3DGraphX’s motif-based node selection [1] and LRI-Bernoulli’s node-removal mechanism [2], which do not adopt an edge-only explanation setting that retains the complete input entity set.

(2) Geometry-constrained, differentiable selection. Node-wise radii enforce distance-ordered neighborhoods under a fixed total radius budget. Differentiable gates allow the predictor to learn directly from the induced interactions using the original prediction objective, without an additional sparsity regularizer (Eqs. 7–9). GSAT [3] and GMT [4] instead learn attention-based subgraph distributions with information-bottleneck regularization; their selection mechanisms do not enforce this geometric ordering or fixed radius budget.

(3) Consistent computation on the original geometry. Coordinates remain fixed, and edge selection also controls the associated angular and line-graph interactions. After pruning, geometric features and interaction indices are recomputed on the retained structure. It is different from LRI [2] which changes the original geometric information.

Beyond E2L’s technical advances over existing 2D and 3D explanation-aware learning methods discussed above, we emphasize that directly adapting 2D methods to 3D is nontrivial. The experiments below demonstrate these challenges and verify E2L’s benefits in both explanation reliability and predictor learning. E2L outperforms GSAT–SchNet, a 2D explanation-aware method adapted to the 3D setting, on all four QM9 targets at comparable sparsity (Appendix F.4, Table 11). For example, at approximately 5050% edge retention, E2L achieves an αα MAE of 0.290.29, compared with 4.534.53 for the best reported GSAT configuration. The geometric consistency enables E2L not only to produce reliable explanatory subgraphs but also to improve the accuracy–efficiency trade-off of 3D predictors, whereas the adapted GSAT does not achieve them in our experiments. For instance, Table 8 demonstrates that E2L reduces DimeNet’s average edge count from 279279 to 234234 and computational cost from 7.2297.229 to 6.0676.067 GFLOPs (approximately 1616% reductions in both), while maintaining comparable MAE across all four QM9 targets. We have expanded the methodological comparison with existing explanation-aware methods in Appendix C.

AuthorsOpenReview ↗

Response to Reviewer JbyJ - Part II

  • W2. Discarded-edge information in the first tt layers; E2L-FT fine-tuning confound; dense-prefix contribution and independent subgraph value.

There is an ignored principle: to decide which interactions to retain, any instance-level explainer needs information from the full graph. Post-hoc explainers [5-7] also compute their masks from full-graph representations of the frozen predictor. In E2L, the first tt message-passing layers (t≥1t ≥ 1) provide the geometric context that lets the explainer assign different radii to atoms of the same element in different environments. We agree that a later removed edge can therefore still influence the prediction through the dense prefix. This is unavoidable for an instance-level explainer, and the question is how much of the predictive performance is attributable to the prefix. Two results in the paper quantify this.

Table 5 (ablation on tt) further studies the contribution of the dense prefix. Moving the explainer from t=3t=3 to t=1t=1 changes the MAE trivially. With a single dense layer, which can only encode one-hop information of the removed edges, E2L retains almost all of its accuracy, so the sparse structure, not the dense prefix, carries the prediction. We have clarified the role of the dense prefix in the ablation study in Sec. 5.4.

Table 10 (E2L-FT) tests the selected subgraph alone. In E2L-FT, we fix each extracted GexpG_{exp} and fine-tune SchNet on it using the original atom features and coordinates, with every message-passing layer restricted to retained edges and no reuse of dense-prefix representations. With 7171–7272% of the edges, this subgraph-only predictor matches or improves the MAE of SchNet trained on uniformly smaller cutoff graphs with more edges. The fine-tuning here further proves that the subgraph obtained by E2L carries expressive connections for GNN prediction. The experiment also includes the comparison between predictors of the same architecture that see the full graph versus GexpG_{exp} only, which is already the cleanest test of the independent predictive value of the explanatory structures.

  • W3. Soft–hard pruning computational mismatch.

By design, our sharp sigmoid gates make soft pruning a close differentiable approximation to hard pruning. During training, each edge (i,j)(i,j) is weighted by wij=σ(k(ri−dij))w_{ij}=σ(k(r_i-d_{ij})), while at inference it is retained if dij≤rid_{ij}≤ r_i. With a large sharpness kk, wijw_{ij} rapidly approaches 11 inside the learned radius and 00 outside it, preserving retained interactions while effectively suppressing pruned ones. This enables continuous optimization during training and a natural transition to sparse hard-pruned computation at inference.

We directly tested the predictive consistency of this transition using the same trained checkpoints. For each of the four QM9 properties, we evaluated the same SchNet checkpoint at budget b=0.4b=0.4 on the first 1,0001{,}000 test molecules twice, changing only the edge pruning method. This comparison isolates the effect of the soft-to-hard transition raised by the reviewer.

PropertySoft MAE ±± stdHard MAE ±± stdPruned / retained soft weight (%)
μμ (D)0.020±0.0210.020 ± 0.0210.020±0.0210.020 ± 0.0210.100.10
αα (bohr³)0.080±0.0850.080 ± 0.0850.080±0.0840.080 ± 0.0840.100.10
εHOMOε_{HOMO} (eV)0.045±0.0440.045 ± 0.0440.045±0.0440.045 ± 0.0440.100.10
εLUMOε_{LUMO} (eV)0.040±0.0390.040 ± 0.0390.040±0.0390.040 ± 0.0390.100.10

Here, std denotes the standard deviation of absolute prediction errors across molecules. The last column reports the total soft weight on pruned edges divided by that on retained edges, pooled across the evaluated molecules.

Hard pruning leaves average MAE unchanged on every target, while the error standard deviation changes by at most 0.0010.001. Moreover, the total soft-gate weight assigned to pruned edges is at most 0.100.10% of the weight assigned to retained edges. In other words, the total amount of information exchanged through pruned edges during message passing is negligible (0.100.10%) compared with that exchanged through retained edges. These results indicate that soft gating effectively suppresses interactions along the edges removed by hard pruning. We have added this experiment to Sec. 5.5 (Table 6) in the revision.

AuthorsOpenReview ↗

Response to Reviewer JbyJ - Part III

  • W4. Edge counts/GFLOPs versus runtime; end-to-end inference and training time.

End-to-end wall-clock comparison. We measured runtime on a single NVIDIA RTX A6000 (Appendix F.2, Table 9), evaluating SchNet and DimeNet on QM9 (μμ) and GearNet-Edge on EC-Protein. Inference is timed per batch at the training batch size over the full test split; training time covers one full epoch over the training split.

ModelEdges keptInference, ms per batch (backbone / E2L)Training, s per epoch (backbone / E2L)
SchNet, b=0.4b=0.4 (batch size 256)7272%11.1 / 10.619.8 / 22.0
DimeNet, b=0.5b=0.5 (batch size 128)4343%116.6 / 76.2800 / 809
GearNet-Edge, b=0.8b=0.8, (batch size 8)6565%303.2 / 215.62,236 / 2,328

Inference. E2L reduces inference time by 4.54.5% on SchNet, 34.634.6% on DimeNet, and 28.928.9% on GearNet-Edge, including explanation-generation overhead. The larger gains on DimeNet and GearNet-Edge reflect reductions in higher-order geometric feature computation.

Training. The training cost of E2L stays at the level of the original training. E2L adds 2.22.2 s per epoch on SchNet (11.111.1%), 99 s on DimeNet (1.11.1%), and 9292 s on GearNet-Edge (4.14.1%). During training, the soft gates reweight interactions on the original graph instead of deleting edges, so the sparse-computation savings primarily arise at inference. Training retains the backbone’s original computations while adding radius prediction and differentiable gating. We have added these measurements and their analysis to Appendix F.2 (Table 9).

[1] Liu et al. 3DGraphX: Explaining 3D Molecular Graph Models via Incorporating Chemical Priors. KDD, 2025.

[2] Miao et al. Interpretable Geometric Deep Learning via Learnable Randomness Injection. ICLR, 2023.

[3] Miao et al. Interpretable and Generalizable Graph Learning via Stochastic Attention Mechanism. ICML, 2022.

[4] Chen et al. How interpretable are interpretable graph neural networks? ICML, 2024.

[5] Ying et al. Gnnexplainer: Generating explanations for graph neural networks. NeurIPS, 2019.

[6] Luo et al. "Parameterized explainer for graph neural network." NeurIPS, 2020.

[7] Qu et al. RISE: Radius of Influence based Subgraph Extraction for 3D Molecular Graph Explanation. ICML, 2025.

Reviewer_JbyJOpenReview ↗

Response to Authors

Thank you for your response and clarifications, which have addressed most of my concerns. However, after considering the comments from the other reviewers, I will keep my current score unchanged at this stage.

Reviewer_X8WrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Explanation to Learning (E2L), a framework for improving the reliability of explanations in 3D graph neural networks (GNNs) by addressing the distribution shift induced by post hoc subgraph explanations. E2L introduces a differentiable, node adaptive radius mechanism to select explanatory subgraphs and jointly trains the explainer and predictor end-to-end, allowing the predictor to operate on the learned sparse graphs. Experiments on QM9 and GEOM-Drugs evaluate explanation reliability for molecular property prediction, while experiments on EC-Protein investigate predictive performance and computational efficiency. The authors report results compared with existing post hoc explainers, as well as improvements in prediction accuracy and computational efficiency on the protein classification task.

Critical strengths
  1. The paper is generally well-written and clearly organized, making it easy to follow the main ideas and technical contributions.
Critical weaknesses
  1. The faithfulness of the explanatory subgraphs is not sufficiently established. Before the explainer is applied, the preceding GNN layers perform message passing on the original, dense proximity graph, allowing node representations to encode information propagated through the entire input graph. The explainer then generates a sparse explanatory subgraph based on these representations. However, edges excluded from the explanatory subgraph may have already contributed to the node representations used by the downstream predictor. Consequently, the final prediction may still depend on information propagated through edges that are absent from the explanatory subgraph, meaning that the prediction is not necessarily based solely on the selected explanatory structure. This raises concerns about whether the explanatory subgraph faithfully captures the information underlying the model's prediction?

  2. E2L aims to mitigate the OOD issue by jointly training the predictor with the explainer, such that the predictor learns to make predictions on explainer-generated subgraphs. However, this strategy adapt the predictor to the explainer induced graph distribution, the resulting explanatory subgraphs may still deviate from the original graph distribution and thus remain out-of-distribution with respect to it. Therefore, it remains unclear whether E2L genuinely resolves the OOD issue or merely adapts the predictor to a potentially shifted graph distribution.

  3. The relationship between E2L and existing self-explainable GNN methods remains unclear. The authors should clarify how E2L differs from existing approaches and include representative self-explainable GNN methods as baselines to better establish its novelty and effectiveness.

  4. The experiments are conducted exclusively on molecular and protein datasets, which limits the evidence for the generalizability of E2L to broader 3D graph learning tasks. The authors should consider evaluating the proposed method on additional 3D graph benchmarks from different domains to demonstrate its applicability beyond molecular property prediction and protein classification.

Questions for authors

See the weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer X8Wr - Part I

We appreciate the reviewer’s thoughtful comments. We have revised the paper accordingly and respond to each point below.

  • W1. Discarded-edge information through the dense prefix; prediction beyond selected edges; subgraph faithfulness.

There is an ignored principle: a fundamental requirement of instance-level subgraph explanation is access to the original graph to determine which interactions to retain. Post-hoc explainers [1-3] also use information from the full graph to select an explanatory subgraph. In E2L, the first tt message-passing layers provide the geometric context that lets the explainer assign different radii to different atoms. We agree that in E2L, a discarded edge can still influence its prediction through the dense prefix. However, this is unavoidable and needed by explainers.

To isolate the effect of the message passing on discarded edges, in Appendix F.3 (Table 10), we directly test whether the selected subgraph alone supports accurate prediction. In E2L-FT, we fix each extracted GexpG_{exp} and fine-tune SchNet on it. SchNet starts from the original atom features and coordinates, and every layer passes messages only along retained edges. With 7171–7272% of the original edges, this subgraph-only predictor achieves lower MAE than standard E2L on all four QM9 targets. It demonstrates that the explanatory subgraph captures expressive information to support the model's prediction. We have clarified the role of the dense prefix in Sec. 5.4 and the subgraph-only evaluation in Appendix F.3 (Table 10).

  • W2. Shifted explanatory distribution; OOD resolution versus predictor adaptation.

There might be a misunderstanding about what constitutes OOD. OOD is defined relative to the predictor's training distribution. Let DD denote the original cutoff-graph distribution and DEθD_{E_θ} the distribution induced by explainer EθE_θ. Post-hoc explanation evaluates a predictor trained on DD on explanatory inputs from DEθD_{E_θ}. E2L explicitly optimizes prediction on explainer-induced inputs. Consequently, DEθ≠DD_{E_θ}≠D does not itself imply an OOD mismatch for E2L: adapting the predictor to the explanatory distribution is precisely how this mismatch is addressed.

In E2L, predictor adaptation is only a part of how we resolve the OOD issue. Table 10 in the revision isolates what adaptation alone achieves. Even after fine-tuning the backbone on post-hoc explainer-induced graphs, E2L still outperforms this adaptation-only setting. For example, on QM9 αα, fine-tuning SchNet on fixed RISE explanations reduces MAE from 1.431.43 to 0.190.19, whereas E2L achieves 0.150.15 while retaining fewer edges (7272% versus 7777%). This demonstrates that E2L’s joint optimization of explanatory structures and the predictor provides further gains beyond solely adapting the predictor.

AuthorsOpenReview ↗

Response to Reviewer X8Wr - Part II

  • W3. Differences from self-explainable GNNs; novelty and effectiveness; representative baselines.

There is a misunderstanding about E2L's contribution. E2L’s central contribution is to bridge post-hoc 3D explanation and representation learning by reformulating the explanation target and addressing the explanation-induced OOD mismatch, enabling post-hoc explainers to support accurate prediction and sparse computation. We shift from explaining a fixed pretrained predictor to learning compact computation structures that support accurate prediction. This reformulation allows joint optimization of the explainer and predictor on the induced explanatory distribution, directly addressing the mismatch between predictor training and explanation evaluation. To the best of our knowledge, no prior instance-level self-explainable 3D GNN provides an equivalent framework for resolving this mismatch. LRI [5] studies node-wise existence and location importance through learnable perturbations, while GMT [7] develops intrinsically interpretable GNN architectures for subgraph multilinear approximation. In particular, GMT-sam requires repeated subgraph sampling and a separate classifier-retraining stage. Our E2L uses the original prediction objective under a sparsity budget, without repeated subgraph sampling, while preserving the original nodes and backbone aggregation. The fixed-graph adaptation and subgraph-only controls further demonstrate that the selected structures retain predictive value beyond predictor adaptation and information propagated through the dense prefix. We have clarified these differences in Sec. 2 and Appendix C.

Transferring 2D explanation methods to 3D is nontrivial, as the selection mechanism must account for geometry-induced connectivity and distance-dependent interactions. To illustrate these challenges empirically, we adapt GSAT [4], an explanation-aware method developed for 2D graphs, to the 3D setting. The comparison in Appendix F.4 directly tests whether GSAT reproduces E2L's benefits. Using the same SchNet backbone, E2L outperforms both GSAT configurations on all four QM9 targets. For instance, on αα, E2L achieves MAE 0.290.29 versus GSAT's 4.534.53, retaining 4747% versus 5050% of edges. Thus, directly adapting 2D explanation-aware methods does not recover E2L's accuracy–sparsity trade-off. This supports the technical value of E2L's RISE-based instance in this work.

  • W4. Molecular/protein-only evaluation; limited generalizability; additional 3D benchmarks from other domains.

Compared with prior 3D explanation studies, our evaluation provides broader learning-target coverage by including both regression and classification, while spanning graph scales from small molecules to large proteins. We study four continuous molecular properties on QM9, conformer-energy regression on the larger GEOM-Drugs molecules, and discrete enzyme-function classification on EC-Protein. RISE [1] evaluates molecular-property and conformer-energy regression on QM9 and GEOM; 3DGraphX [6] focuses on post-hoc molecular graph explanation using chemical motif priors only on QM9. LRI [5] evaluates the performance on classification tasks, while explicitly leaving regression for future work. In comparison, E2L demonstrates its benefits across continuous physical-property prediction and discrete protein-function classification, with graph scales ranging from small molecules to macromolecules. These results establish applicability across multiple prediction settings and graph scales.

Extending the evaluation to macroscopic domains requires a compatible setup: (1) data modeled as cutoff-based 3D proximity graphs and (2) 3D GNN benchmarking which serves as the premise of 3D GNN explanation. For example, large point-cloud benchmarks often use kk-NN graphs or voxel representations to manage their large node counts, so their standard pipelines require adaptation to our cutoff-based setting. We therefore respectfully ask that our experimental breadth across graph scales and learning targets be recognized when assessing coverage, rather than judging it solely by the molecular and protein domains.

[1] Qu et al. RISE: Radius of Influence based Subgraph Extraction for 3D Molecular Graph Explanation. ICML, 2025.

[2] Ying et al. Gnnexplainer: Generating explanations for graph neural networks. NeurIPS, 2019.

[3] Luo et al. "Parameterized explainer for graph neural network." NeurIPS, 2020.

[4] Miao, et al. Interpretable and Generalizable Graph Learning via Stochastic Attention Mechanism. ICML, 2022.

[5] Miao, et al. Interpretable Geometric Deep Learning via Learnable Randomness Injection. ICLR, 2023.

[6] Liu, et al. 3DGraphX: Explaining 3D Molecular Graph Models via Incorporating Chemical Priors. KDD, 2025.

[7] Chen et al. How interpretable are interpretable graph neural networks? ICML, 2024.

AuthorsOpenReview ↗

Kindly Reminder to Reviewer X8Wr

Dear Reviewer X8Wr,

We sincerely appreciate your constructive feedback. We posted our detailed rebuttal four days ago, and with only three days remaining in the discussion period, we would like to highlight several important clarifications and additional experiments addressing your concerns.

We believe W1 and W2 partly arise from misunderstandings of E2L's explanation objective and OOD formulation. For W1, the dense prefix provides context for subgraph selection, while our additional subgraph-only E2L-FT experiment demonstrates that the selected structures independently support accurate prediction without dense-prefix representations. For W2, OOD is defined relative to the predictor's training distribution, not necessarily the original graph distribution. Adapting the predictor to the explanatory distribution is precisely how E2L resolves this mismatch, and our additional adaptation controls demonstrate further benefits from jointly learning explanatory structures instead of only predictor adaption.

Regarding W3 (novelty), we clarify E2L's methodological differences from existing self-explainable GNNs and provide new GSAT–SchNet comparisons demonstrating E2L's superiority at comparable sparsity.

Regarding W4 (generalizability), we further clarify our evaluation scope across regression and classification tasks, from small molecules to large protein graphs, and the corresponding applicability of our cutoff-based formulation.

Together, these clarifications and new experimental results provide stronger evidence for E2L's distinct contributions and are important for correctly understanding its novelty, explanation effectiveness, and practical value. Notably, Reviewer JbyJ has acknowledged that our rebuttal addressed most of their concerns, but maintained their current score after considering other reviewers' comments. Since your review shares important concerns regarding the independent predictive value of explanations and E2L's distinction from existing methods, your updated assessment is particularly important to this discussion and to a fair evaluation of our contributions.

Could you please revisit our point-by-point response and let us know whether it resolves your concerns? If any points remain open, we would welcome the opportunity to address them. We look forward to hearing from you.

Sincerely,

Authors

Thank you for the detailed responses. Some of my original concerns remain insufficiently addressed. Rewardning W4, the authors should include at least one 3D proximity graph benchmark from a field outside of traditional molecules/proteins to validate the effectiveness of the proposed method. Rewarding W2, this method essentially adapts the predictor to the explainer induced distribution rather than addressing the distribution shift relative to the original graph distribution. I think the claims regarding OOD mitigation should be more carefully qualified. Overall, I would like to maintain my original score.

Reviewer_3tjYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This works identified the OOD problems in 3D GNNs, which may potentially cause unreliable explanation. They propose E2L to address this issue, which jointly optimize the explainer and predictor.

Critical strengths

I like the motivation and problem formulation. This paper identifies an overlooked problem and propose an method to address it. The OOD issue is a very interesting problem to solve. The proposed methods integrates explanation into the learning process is also intuitive and straight-forward.

Critical weaknesses
  1. I'm wondering does the OOD issue really solved? The paper claims that E2L eliminates the distribution mismatch by training hte predictor joinly with the explainer. However, during training, E2L employs continuous sigmoid-based edge weights, meaning that the original graph connectivity is still preserved but with different edge weight. During inference, the discrete edge pruning is employed. So, the training and inference graph distribution are not necessarily identical, potentially reintroducing the very OOD issue that E2L aims to resolve. Although Appendix H evaluates different sigmoid sharpness parameters, it does not directly compare model predictions using continuous versus discrete explanatory graphs. An experiment evaluating both representations with the same trained predictor would be necessary to substantiate the claimed distribution alignment. However, in theory, I think the OOD still exists even though the exprimental result only shows a minimal difference, meaning that the motivation is not aligned with what is actually achieved.

  2. Does E2L really do a better job on explanation? In Tables 1–3, the baseline explainers operate on a pretrained, frozen predictor, whereas E2L jointly optimizes the explainer and predictor. Therefore, the reported improvements may largely arise from the predictor's adaptation to the explanatory distribution rather than the effectiveness of the extracted subgraphs themselves. I think the author should evaluate the existing post-hoc baseline explainers on the E2L trained predictor. Such a comparison, with a clearly defined common predictor and evaluation protocol, would help disentangle the contributions of predictor adaptation and explanatory subgraph selection. For now, Tables 1-3's cannot show anything about E2L's expalantion effectiveness.

Questions for authors

see the weakness above

8 comments
AuthorsOpenReview ↗

Response to Reviewer 3tjY - Part I

We thank the reviewer for their careful reading and constructive feedback. We have revised our manuscript accordingly and provide our point-by-point responses below.

  • W1. Soft–hard OOD gap; insufficient sharpness ablation; same-predictor comparison; theoretical mismatch despite small empirical differences.

To clarify this point, we first briefly explain how E2L reduces the post-hoc OOD gap, and then test its soft-to-hard approximation in a new experiment. The experiment shows how closely the soft gates approximate hard pruning, and proves that there isn't a newly introduced OOD issue in E2L.

The two representations can theoretically differ, but our sharp soft gate is designed to closely approximate hard pruning, limiting the discrepancy between training and inference. E2L reduces the post-hoc OOD gap by training the predictor on the explanatory distribution. During training, edge (i,j)(i,j) receives weight wij=σ(k(ri−dij))w_{ij}=\sigma(k(r_i-d_{ij})). At inference, the edge is retained exactly when dij≤rid_{ij}\leq r_i. With the sharpness kk, weights approach zero or one rapidly away from the boundary.

We isolate the effect of such theoretically-existing difference on the first 1,000 molecules in the QM9 test split. For each of four properties, we run the same trained SchNet checkpoint at budget b=0.4b=0.4 twice, changing only the edge pruning method. We report the MAE and the standard deviation. The final column additionally gives the ratio of total soft weight on pruned edges to that on retained edges.

PropertySoft MAE ±\pm stdHard MAE ±\pm stdPruned / retained soft weight (%)
μ\mu (D)0.020±0.0210.020 \pm 0.0210.020±0.0210.020 \pm 0.0210.100.10
α\alpha (bohr³)0.080±0.0850.080 \pm 0.0850.080±0.0840.080 \pm 0.0840.100.10
ϵHOMO\epsilon_{\mathrm{HOMO}} (eV)0.045±0.0440.045 \pm 0.0440.045±0.0440.045 \pm 0.0440.100.10
ϵLUMO\epsilon_{\mathrm{LUMO}} (eV)0.040±0.0390.040 \pm 0.0390.040±0.0390.040 \pm 0.0390.100.10

Here, std denotes the standard deviation of absolute prediction errors across molecules. Across all four properties, the total soft-gate weight assigned to pruned edges is at most 0.100.10% of that assigned to retained edges. In other words, the total amount of information exchanged through pruned edges during message passing is negligible (0.100.10%) compared with that exchanged through retained edges. Consistently, soft gating and hard pruning yield nearly identical MAE and error standard deviation, with differences of at most 0.0010.001 across all four properties. Together, these results demonstrate that soft gating closely approximates hard pruning, effectively aligning the training and inference computations. Consequently, hard pruning merely removes edges whose contributions have already been suppressed during training, introducing no meaningful additional OOD discrepancy in predictive behavior. We have added this experiment to Sec. 5.5, Table 6 in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 3tjY - Part II

  • W2. Frozen baselines versus joint training; predictor adaptation versus subgraph quality; insufficient explanation evidence; common-predictor evaluation.

There might be a misunderstanding about E2L’s explanation objective. Post-hoc explainers identify subgraphs that explain a frozen predictor’s output on the original graph. E2L reformulates explanation as identifying subgraphs having strong expressivity on a pre-defined GNN architecture. This reformulation directly addresses the OOD problem identified in Sec. 4.1, where post-hoc explainers evaluate subgraphs using a predictor trained on full proximity graphs. Predictor adaptation is thus an intended part of E2L’s solution to remove such an OOD issue in post-hoc explainers.

Applying post-hoc explainers to E2L can cause confusion about the explanation target. Since E2L’s downstream predictor already operates on GexpG_{\mathrm{exp}}, are we explaining the full model’s prediction on the original graph GG, or are we seeking a further explanation within GexpG_{\mathrm{exp}}? Despite this ambiguity, we directly address the reviewer’s concern by comparing subgraphs selected by E2L and by post-hoc explainers using the same frozen E2L predictor.

Specifically, on the first 1,000 QM9 test molecules, we freeze each E2L–SchNet checkpoint (b=0.4b=0.4; the same checkpoints as in W1) and use the molecular property as the explanation target. Each baseline uses its method-specific objective and retains exactly the same number of edges as E2L for each molecule. This fixes predictor parameters, input representations, and sparsity while varying the selected explanatory subgraphs.

Explainers with frozen E2L predictorμ\muα\alphaϵHOMO\epsilon_{\mathrm{HOMO}}ϵLUMO\epsilon_{\mathrm{LUMO}}
E2L0.0200.0200.0800.0800.0450.0450.0400.040
GNNExplainer0.0680.0680.2380.2380.1110.1110.1070.107
PGExplainer0.0740.0740.3560.3560.1440.1440.1420.142
LRI-Bernoulli0.7870.7870.2960.2960.0760.0760.0840.084

E2L retains its advantage with the predictor and per-molecule edge counts fixed. All three post-hoc alternatives yield higher MAE on every target. Thus, reusing the adapted predictor alone does not recover E2L’s performance; its learned explanatory subgraphs remain essential. These results demonstrate that predictor adaptation alone cannot account for E2L’s advantage.

Table 10 in the paper provides a complementary adaptation control by fine-tuning the predictor on fixed explanatory graphs. Even after this adaptation, E2L’s explanations support more accurate prediction. For example, on QM9 α\alpha, fine-tuning on fixed E2L explanations (E2L-FT) achieves an MAE of 0.120.12, compared with 0.190.19 for fine-tuning on RISE explanations, while retaining fewer edges (7272% vs. 7777%). This comparison demonstrates that E2L’s advantage extends beyond predictor adaptation and supports the predictive value of its selected structures. We have clarified this distinction in Appendix F.3, in the analysis of Table 10.

AuthorsOpenReview ↗

Kindly Reminder to Reviewer 3tjY

Dear Reviewer 3tjY,

We sincerely appreciate your thoughtful review. We posted our detailed rebuttal four days ago, and with only three days remaining in the discussion period, we would like to highlight the key clarifications and additional experiments that directly address your concerns.

Regarding W1 (soft–hard OOD gap), our new controlled experiment shows that soft gating and hard pruning yield identical MAE across all four QM9 targets using the same trained checkpoints. This provides direct evidence that the theoretical soft–hard discrepancy introduces no OOD gap in predictive performance.

Regarding W2 (explanation effectiveness), we believe there is a misunderstanding of E2L's explanation objective: E2L seeks task-relevant explanatory structures rather than explaining a fixed pretrained predictor. Importantly, our new experiment with a shared frozen predictor and matched sparsity demonstrates that E2L consistently outperforms three post-hoc explainers, confirming that its benefits extend beyond predictor adaptation.

These clarifications and additional experiments directly substantiate our central contributions and are important for correctly understanding and evaluating E2L's explanation reliability and effectiveness. Notably, Reviewer JbyJ has acknowledged that our rebuttal addressed most of their concerns, but maintained their current score after considering other reviewers' comments. Since your review raises similar questions about soft–hard consistency and the independent effectiveness of explanatory structures, your updated assessment is particularly important to this discussion and to a fair evaluation of our contributions.

Could you please revisit our point-by-point response and let us know whether it resolves your concerns? If any points remain open, we would welcome the opportunity to address them. We look forward to hearing from you.

Sincerely,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.