- 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
- 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.
- 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.
- 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
- 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.
- 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
- 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.
- 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
Response to Reviewer JbyJ - Part II
- W2. Discarded-edge information in the first 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 message-passing layers () 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 ) further studies the contribution of the dense prefix. Moving the explainer from to 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 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 –% 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 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 is weighted by , while at inference it is retained if . With a large sharpness , rapidly approaches inside the learned radius and 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 on the first test molecules twice, changing only the edge pruning method. This comparison isolates the effect of the soft-to-hard transition raised by the reviewer.
| Property | Soft MAE std | Hard MAE std | Pruned / retained soft weight (%) |
|---|---|---|---|
| (D) | |||
| (bohr³) | |||
| (eV) | |||
| (eV) |
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 . Moreover, the total soft-gate weight assigned to pruned edges is at most % of the weight assigned to retained edges. In other words, the total amount of information exchanged through pruned edges during message passing is negligible (%) 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.
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 () 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.
| Model | Edges kept | Inference, ms per batch (backbone / E2L) | Training, s per epoch (backbone / E2L) |
|---|---|---|---|
| SchNet, (batch size 256) | % | 11.1 / 10.6 | 19.8 / 22.0 |
| DimeNet, (batch size 128) | % | 116.6 / 76.2 | 800 / 809 |
| GearNet-Edge, , (batch size 8) | % | 303.2 / 215.6 | 2,236 / 2,328 |
Inference. E2L reduces inference time by % on SchNet, % on DimeNet, and % 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 s per epoch on SchNet (%), s on DimeNet (%), and s on GearNet-Edge (%). 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.
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.
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.
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 50% edge retention, E2L achieves an α MAE of 0.29, compared with 4.53 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 279 to 234 and computational cost from 7.229 to 6.067 GFLOPs (approximately 16% 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.