Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Fragment-Wise Interpretability in Graph Neural Networks via Molecule Decomposition and Contribution Analysis

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper introduces SEAL, a graph neural network whose prediction for a molecule is, by construction, the sum of separate contributions from chemically meaningful fragments, and argues these contributions are faithful explanations because

desk verdict SEAL is a clean, useful fragment-additive GNN with a real claim to novelty and an honest synthetic benchmark; the main gap is that 'causal fragment influence' is asserted rather than tested, and the real-data fidelity protocol flatters SEAL. read the letter →

arxiv 2508.15015 v1 pith:ZWACN2K2 submitted 2025-08-20 cs.LG

Sebastian Musia{l} , Bartosz Zieliński , Tomasz Danel This is my paper
classification cs.LG
keywords graphneuralnetworksmolecularpropertypredictioninterpretabilityfragmentcontributionsattributionexplainableAIsubstructureexplanationmessage-passingregularization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SEAL is a graph neural network for molecular property prediction that rewrites the prediction head so the output is, by construction, the sum of per-fragment contributions plus a bias. Each fragment contribution is produced by an MLP over the sum-pooled representations of the fragment's atoms, so no separate explainer is needed: the score the model uses is the explanation. To keep those contributions local and chemically meaningful, SEAL-GCN layers use separate weights for edges inside a fragment and edges between fragments, with an L1 penalty that shrinks the inter-fragment weights. The paper argues that this design yields explanations that are faithful by construction and competitive in accuracy, and supports this with synthetic benchmarks, fidelity tests on real-world datasets, and a user study in which chemists preferred SEAL's explanations in 14 of 19 compounds.

What carries the argument

The load-bearing object is the fragment-contribution decomposition: molecule, BRICS-style fragments, sum-pooled fragment representations, fragment scores c_i, and prediction as an exact sum. The new layer SEAL-GCN carries the argument: it applies separate weight matrices to same-fragment neighbors (W_intra) and cross-fragment neighbors (W_inter), and an L1 penalty on W_inter forces the model to encode as much signal as possible locally inside fragments. The hyperparameter λ controls the strength of this locality pressure; larger λ means less inter-fragment information flow and, the paper finds, cleaner explanations, at some cost on tasks that genuinely need long-range information.

What would settle it

Build a synthetic task whose ground truth is a known nonlinear interaction between two distant fragments, for example a positive label only when both fragments appear together. If SEAL's explanation quality on positive examples stays near chance (SE close to 0.5) or its prediction accuracy falls far below GIN, then the fragment-additivity premise fails exactly where inter-fragment interactions carry the signal.

Watch

Extended reading notes

Core claim

The central claim is that a GNN can be built so that its prediction is literally a sum of fragment-level numbers, and those numbers are the explanation. After message passing, each fragment's atoms are sum-pooled and passed through an MLP to give c_i; the prediction is y-hat = sum of c_i plus a bias. Because there is no graph-level readout that mixes all atoms before the final layer, attributing the prediction to fragments is not a post-hoc guess but a restatement of the model's arithmetic. The architecture, SEAL-GCN, enforces locality by giving intrafragment and interfragment edges separate weights and adding an L1 penalty on interfragment weights; the authors hypothesize that for many chem

Load-bearing premise

The load-bearing premise is that molecular properties, and the networks that predict them, are dominated by fragment-local signals, so suppressing inter-fragment message passing does not destroy prediction accuracy or explanation truth; the boron and largest-ring failures in the paper show places where this premise breaks.

Editorial extensions

If this is right

  • For fragment-local properties, explanations are exact by construction: the highlighted fragment scores are the arithmetic components of the predicted value, so no separate explainer is required and there is no approximation step to argue about.
  • The model can trade a small amount of accuracy for interpretability: SEAL reaches AUROC 0.85 on hERG and 0.81 on CYP2C9 versus 0.86 for GIN, with substantially better fidelity scores.
  • Fragment contributions give chemists a direct handle for hypothesis generation: a predicted property can be decomposed into which fragment adds or subtracts, and those numbers can be ranked, compared across molecules, or used to propose modifications.
  • The λ selection protocol, which picks the highest regularizer not significantly worse than the best model by a Wilcoxon signed-rank test, provides a principled way to choose how much locality to enforce on a given task.
  • On tasks requiring cross-fragment interactions, the same mechanism predicts where SEAL will fail: the boron and largest-ring synthetic tasks degrade precisely because the ground-truth substructure is spread over larger motifs that the chosen fragmentation cannot cleanly isolate.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The additivity premise is itself testable and marks the paper's implicit boundary: one could construct a benchmark where the true label is a nonlinear function of two distant fragments, and SEAL's explanation quality should collapse; the paper's own boron and largest-ring results point in that direction.
  • Because fragment definitions determine what a 'contribution' means, the same architecture with finer or coarser fragmentation would produce different explanations; a chemical-discovery tool built on this idea should probably expose fragmentation as a user choice rather than a fixed preprocessing step.
  • The exact-sum head makes SEAL a natural base for counterfactual reasoning, since replacing or removing a fragment changes the prediction by exactly the change in that fragment's contribution; the paper does not pursue this, but it could support a generative optimization loop.
  • When the optimal λ is near zero, the task is dominated by long-range interactions and fragment-level reasoning may be inappropriate; SEAL's own regularization search could therefore double as a diagnostic for whether a dataset is fragment-local.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces SEAL, a graph neural network for molecular property prediction that decomposes molecules into BRICS fragments, applies a modified GCN layer with separate intra-fragment and inter-fragment weights, and produces predictions as the sum of per-fragment contributions (Eqs. 1–2). An L1 penalty on inter-fragment weights (Eq. 4) is intended to reduce cross-fragment information leakage and make the fragment contributions more local and interpretable. The method is evaluated on the B-XAIC synthetic benchmark, three TDC real-world datasets (hERG, CYP2C9, AqSolDB), and a user study with chemistry experts. The authors report competitive predictive accuracy and claim state-of-the-art explanation fidelity and human-aligned interpretability.

Significance. The idea of making a GNN's prediction an additive sum of fragment-level contributions is attractive: if the contributions are genuinely local and chemically meaningful, SEAL would provide built-in, directly usable fragment explanations without post-hoc explainers. The synthetic benchmark with ground-truth substructures is an appropriate test bed, and the paper is commendably honest about the boron and rings-max failures. The code is released, and the user study, despite its reporting gaps, addresses an important dimension rarely evaluated. However, the paper's central interpretability claim is currently not established: additive decomposition alone does not make the contributions causal or fragment-local, and the headline real-world fidelity comparison uses different masking protocols for SEAL than for the baselines.

major comments (4)
  1. [Real-world datasets, Metrics; Table 4; Appendix C] The fidelity comparison is protocol-asymmetric. The Metrics text states 'For all models, we mask node features at the input level, ensuring a fair comparison,' yet the SEAL rows in Table 4 exactly match the 'SEAL-mask' rows in Appendix C (e.g., hERG λ=10⁻⁴: 0.63/0.09/0.71/0.07/0.78/0.05 in both Table 4 and Table 14), meaning SEAL is evaluated by zeroing its fragment contributions c_i, not by masking input features. Since the prediction is defined as ŷ = Σ c_i + b (Eq. 2), zeroing c_i is a built-in linear sensitivity operation, which inflates positive fidelity and deflates negative fidelity by construction. The claim that SEAL 'outperforms all other methods in both positive and negative fidelity' on hERG and CYP2C9 is therefore not supported by the current evaluation. The authors should either re-run all methods under the same input-masking protocol or justify contribution masking with ap
  2. [Fragment Contributions/SEAL-GCN, Eqs. (2)–(4); Tables 1–2, 6–7] The central claim that c_i is a 'causal influence' or fragment-local explanation is not established. Eq. (2) is an algebraic decomposition, not a causal statement. c_i is computed from a fragment representation that is the sum-pooled output of multiple GNN layers, and with nonzero inter-fragment weights W_inter, h_i depends on the entire molecule. The L1 penalty in Eq. (4) does not zero W_inter unless λ is infinite, and the selected λ is small or zero for several tasks (λ=0 for PAINS, λ=10⁻⁴ for hERG and AqSol, λ=10⁻³ for rings-count). The synthetic results show exactly the failure mode: when the ground-truth substructure does not align with the BRICS fragmentation, SEAL's SE drops to 0.34 on rings-max and 0.88 on boron (Table 2). To support the locality/causal reading, the authors should report the actual magnitudes or sparsity of W_inter after training, compare zeroing c_i with removin
  3. [User study] The user study is reported without essential experimental detail. The text does not state the number of participants, the total number of votes, or any statistical test comparing SEAL to the other methods. The sentence 'Deconvolution and IntegratedGradients were chosen 5 times, and InputXGradients was selected 3 times' is ambiguous and does not permit the reader to assess whether SEAL's apparent preference is significant or due to chance. For a central claim about 'human-aligned interpretability,' the authors should report participant count, recruitment procedure, per-method vote totals, and a significance test (e.g., per-question or per-participant mixed-effects model).
  4. [Abstract and Discussion; Table 2] The abstract's claim that SEAL 'outperforms other explainability methods in both quantitative attribution metrics' is too broad given the paper's own results. On the rings-max task, SEAL's SE is 0.34, while GuidedBackprop reaches 0.67 and Saliency 0.55; on boron, SEAL (0.88) is also below several gradient baselines (1.00). The text in the results section acknowledges the boron and rings-max difficulties, but the abstract, introduction, and discussion generalize the outperformance claim. The authors should qualify the claim and report whether the overall advantage is statistically significant across tasks, or restrict the claim to the tasks where SEAL is indeed strong (PAINS, rings-count, indole, X, P).
minor comments (5)
  1. [Methods, Eq. (1)] The fragment representation is described as sum pooling followed by LayerNorm, but LayerNorm is not shown in Eq. (1); please align the equation with the text.
  2. [Related Work / Baselines] The baseline list says 'Deconvolution (Mahendran and Vedaldi 2016), (Shrikumar et al. 2016), and Guided Backpropagation' — the Shrikumar method name is missing (likely DeepLIFT).
  3. [Appendix C] The naming convention for masking strategies ('mask-abs', 'mask', 'abs', 'zero') is confusing and appears to mix 'zeroing features' with 'masking contributions.' Please define each term explicitly and use consistent terminology in the main text.
  4. [Metrics, real-world datasets] For the regression task (AqSolDB), the fidelity metric is defined only for classification ('proportion of times the predicted class changes'). Table 13 reports numerical values for regression without defining the metric; please provide the regression-specific definition.
  5. [Figure 2] The 'Baseline' row is not named per task; since different tasks have different best-performing baselines, please identify which method is shown for each column.

Circularity Check

1 steps flagged · score 6.0 of 10

Real-data fidelity claims reduce to the defining equation y-hat = sum c_i + b by construction; independent synthetic and human evaluations keep the paper from being fully circular.

  1. self definitional [Results — real-world datasets, Metrics; Appendix C (ablation on masking strategy); Eq. (2)]
    "Unlike standard explainers that only operate on node masks, our model allows for masking specific contribution scores directly at the level of the model's architecture by setting ci = 0 for a given fragment."

    Eq. (2) defines the prediction as y-hat = sum_i c_i + b. The fidelity evaluation for SEAL then 'masks' by setting c_i = 0, i.e., deleting terms from the very sum that defines the prediction. Positive fidelity ('mask the most important nodes and check how the prediction has changed') and negative fidelity ('retain the most important nodes and check if the prediction is the same') are therefore algebraic consequences of the additive definition: zeroing a term changes the sum by exactly that term, and retaining the dominant terms leaves the sum nearly unchanged. These numbers do not independently test whether c_i corresponds to a chemically meaningful or counterfactual causal effect, and comparing them with baselines that mask input node features is a protocol mismatch. The headline 'outperfo

full rationale

The paper's central identity (Eq. 2) makes fragment contributions add up to the prediction by definition; the real-data fidelity numbers in Table 4 are computed for SEAL by zeroing those same contributions (Appendix C 'SEAL-mask'), so the high positive and near-zero negative fidelity largely follow from the defining equation, not from external evidence about chemical causality. This is the core circular step. However, the paper also provides independent evidence that is not circular: the synthetic B-XAIC benchmark uses external ground-truth substructures, and SEAL's explanation scores there are falsifiable (e.g., failure on boron and rings-max shows the evaluation can reject the method); the user study with domain experts is an external human judgment independent of the model's additive parameterization. The B-XAIC benchmark is co-authored by one of the present authors, but it is a dataset with pre-defined ground truth rather than a uniqueness theorem imported to force a conclusion, so it does not by itself raise the circularity score. The causal-influence interpretation of c_i is an additional unverified assumption (with lambda = 0 or small values, inter-fragment weights are not zero, so c_i can depend on the whole molecule), but that is a correctness/assumption gap rather than a distinct circular reduction. Overall: one central evaluation metric reduces by construction to the defining equation, giving partial circularity (6), while the synthetic benchmark and user study keep the paper from being entirely self-referential.

Assumptions & free parameters 1 free parameters · 4 assumptions · 1 invented entities

The central claim rests on the fragment-additivity hypothesis, the validity of the B-XAIC ground truth, and the acceptability of the fidelity protocol. The only explicitly fitted hyperparameter driving the reported explanation quality is lambda, chosen per task on the validation set. No new physical entities are introduced; the constructed 'fragment contribution' is the paper's main object and has an external check via the synthetic ground truth.

free parameters (1)
  • lambda (inter-fragment weight regularization strength) = per task/dataset: rings-count 1e-3, rings-max 2, X 2, P 2, B 2, indole 1e-4, PAINS 0, hERG 1e-4, CYP2C9 2, AqSol 1e-4
    Selected on the validation set per task using the Wilcoxon 'highest lambda not significantly worse than best' rule (Methods, Eq. 4). It directly controls the claimed interpretability/performance trade-off, and the reported explanation scores are conditioned on this selection.
assumptions (4)
  • domain assumption Molecular property predictions can be decomposed into additive, fragment-local contributions; inter-fragment interactions are secondary.
    Introduction: 'We hypothesize that GNNs operating on molecular graphs decomposed into fragments are more interpretable while preserving sufficient expressivity... many chemical prediction tasks inherently depend more on the presence and identity of specific fragments than on complex interactions between them.' This is the load-bearing premise for interpreting c_i as meaningful explanations.
  • domain assumption B-XAIC ground-truth labels are a valid gold standard for explanation quality.
    The synthetic evaluation uses B-XAIC (Proszewska, Danel, and Rymarczyk 2025) as ground truth; that benchmark is co-authored by a co-author of this paper.
  • ad hoc to paper Fidelity by masking is a valid and fair proxy for explanation quality across methods.
    Real-world evaluation compares SEAL, which masks its own architectural contributions (Appendix C), against baselines, which mask input node features. The paper partially acknowledges this in the Discussion but the headline Table 4 does not flag the asymmetry.
  • standard math Standard GNN and MLP machinery (sum pooling, LayerNorm, MLP readout) preserves the additive decomposition.
    Used throughout Methods, Eqs. (1)-(2).
invented entities (1)
  • Fragment contribution c_i independent evidence
    purpose: The explanation unit: each fragment's scalar contribution to the prediction such that y-hat = sum(c_i) + bias.
    A constructed quantity with no physical status. It is checked against external ground truth (B-XAIC SE metric) and expert judgment (user study), giving it a falsifiable handle outside the model, though it is not an independent physical entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fragment-Wise Interpretability in Graph Neural Networks via Molecule Decomposition and Contribution Analysis." pith.science (2026). https://pith.science/paper/ZWACN2K2

@misc{pith2026250815015,
  author       = {Pith},
  title        = {Pith review of: Fragment-Wise Interpretability in Graph Neural Networks via Molecule Decomposition and Contribution Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZWACN2K2}},
  note         = {Machine review of arXiv:2508.15015}
}
read the original abstract

Graph neural networks have demonstrated remarkable success in predicting molecular properties by leveraging the rich structural information encoded in molecular graphs. However, their black-box nature reduces interpretability, which limits trust in their predictions for important applications such as drug discovery and materials design. Furthermore, existing explanation techniques often fail to reliably quantify the contribution of individual atoms or substructures due to the entangled message-passing dynamics. We introduce SEAL (Substructure Explanation via Attribution Learning), a new interpretable graph neural network that attributes model predictions to meaningful molecular subgraphs. SEAL decomposes input graphs into chemically relevant fragments and estimates their causal influence on the output. The strong alignment between fragment contributions and model predictions is achieved by explicitly reducing inter-fragment message passing in our proposed model architecture. Extensive evaluations on synthetic benchmarks and real-world molecular datasets demonstrate that SEAL outperforms other explainability methods in both quantitative attribution metrics and human-aligned interpretability. A user study further confirms that SEAL provides more intuitive and trustworthy explanations to domain experts. By bridging the gap between predictive performance and interpretability, SEAL offers a promising direction for more transparent and actionable molecular modeling.

Figures

Figures reproduced from arXiv: 2508.15015 by the authors.

Figure 1
Figure 1. Overview of SEAL comprising a message passing stage [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Node-level explanation examples for selected synthetic compounds from the B-XAIC dataset. Each column corre [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Effect of regularization on explanation quality in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (26 more)
Figure 4
Figure 4. Figure 4: Node-level explanation examples for selected [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: All compounds and visualizations that were used [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Node-level explanation examples from user study, The red color indicates that the highlighted atoms had a positive [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Node-level explanation examples from user study, The red color indicates that the highlighted atoms had a positive [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Node-level explanation examples from user study, The red color indicates that the highlighted atoms had a positive [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Node-level explanation examples from user study, The red color indicates that the highlighted atoms had a positive [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Node-level explanation examples from user study, The red color indicates that the highlighted atoms had a positive [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Node-level explanation examples from user study, The red color indicates that the highlighted atoms had a positive [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: Node-level explanation examples from user study, The red color indicates that the highlighted atoms had a positive [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: Node-level explanation examples of the SEAL method and Ground-Truth evaluated on the Boron (B) task for the [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]
Figure 14
Figure 14. Figure 14: Node-level explanation examples of the SEAL method evaluated on the Boron (B) task for the negative target class. [PITH_FULL_IMAGE:figures/full_fig_p026_14.png]
Figure 15
Figure 15. Figure 15: Node-level explanation examples of the SEAL method and Ground-Truth evaluated on Halogens (X) task for the [PITH_FULL_IMAGE:figures/full_fig_p027_15.png]
Figure 16
Figure 16. Figure 16: Node-level explanation examples of the SEAL method evaluated on the Halogens (X) task for the negative target [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]
Figure 17
Figure 17. Figure 17: Node-level explanation examples of the SEAL method and Ground-Truth evaluated on the rings-max task for the [PITH_FULL_IMAGE:figures/full_fig_p028_17.png]
Figure 18
Figure 18. Figure 18: Node-level explanation examples of the SEAL method evaluated on the rings-max task for the negative target class. [PITH_FULL_IMAGE:figures/full_fig_p028_18.png]
Figure 19
Figure 19. Figure 19: Node-level explanation examples of the SEAL method and Ground-Truth evaluated on the rings-count task for the [PITH_FULL_IMAGE:figures/full_fig_p029_19.png]
Figure 20
Figure 20. Figure 20: Node-level explanation examples of the SEAL method evaluated on the rings-count task for the negative target class. [PITH_FULL_IMAGE:figures/full_fig_p029_20.png]
Figure 21
Figure 21. Figure 21: Node-level explanation examples of the SEAL method and Ground-Truth evaluated on the PAINS task for the [PITH_FULL_IMAGE:figures/full_fig_p030_21.png]
Figure 22
Figure 22. Figure 22: Node-level explanation examples of the SEAL method evaluated on the PAINS task for the negative target class. [PITH_FULL_IMAGE:figures/full_fig_p030_22.png]
Figure 23
Figure 23. Figure 23: Node-level explanation examples of the SEAL method and Ground-Truth evaluated on the indole task for the positive [PITH_FULL_IMAGE:figures/full_fig_p031_23.png]
Figure 24
Figure 24. Figure 24: Node-level explanation examples of the SEAL method evaluated on the indole task for the negative target class. [PITH_FULL_IMAGE:figures/full_fig_p031_24.png]
Figure 25
Figure 25. Figure 25: Node-level explanation examples of the SEAL method and Ground-Truth evaluated on the Phosphorus (P) task [PITH_FULL_IMAGE:figures/full_fig_p032_25.png]
Figure 26
Figure 26. Figure 26: Node-level explanation examples of the SEAL method evaluated on the Phosphorus (P) task for the negative target [PITH_FULL_IMAGE:figures/full_fig_p032_26.png]
Figure 27
Figure 27. Figure 27: Node-level explanation examples of the SEAL method evaluated on the hERG dataset. The red color indicates that [PITH_FULL_IMAGE:figures/full_fig_p033_27.png]
Figure 28
Figure 28. Figure 28: Node-level explanation examples of the SEAL method evaluated on the CYP2C9 dataset. The red color indicates [PITH_FULL_IMAGE:figures/full_fig_p034_28.png]
Figure 29
Figure 29. Figure 29: Node-level explanation examples of the SEAL method evaluated on the aqueous solubility dataset. The red color [PITH_FULL_IMAGE:figures/full_fig_p035_29.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 33 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    L.; Kiros, J

    Ba, J. L.; Kiros, J. R.; and Hinton, G. E. 2016. Layer normalization. arXiv preprint arXiv:1607.06450

    arXiv 2016

  4. [4]

    Cao, P.-Y.; He, Y.; Cui, M.-Y.; Zhang, X.-M.; Zhang, Q.; and Zhang, H.-Y. 2024. Group graph: a molecular graph representation with enhanced performance, efficiency and interpretability. Journal of Cheminformatics, 16(1): 133

    work page 2024

  5. [5]

    Degen, J.; Wegscheid-Gerlach, C.; Zaliani, A.; and Rarey, M. 2008. On the art of compiling and using'drug-like'chemical fragment spaces. ChemMedChem, 3(10): 1503

    work page 2008

  6. [6]

    C.; Boukouvalas, Z.; Fuge, M

    Elton, D. C.; Boukouvalas, Z.; Fuge, M. D.; and Chung, P. W. 2019. Deep learning for molecular design—a review of the state of the art. Molecular Systems Design & Engineering, 4(4): 828--849

    work page 2019

  7. [7]

    Feng, Q.; Liu, N.; Yang, F.; Tang, R.; Du, M.; and Hu, X. 2022. DEGREE : Decomposition Based Explanation for Graph Neural Networks. In International Conference on Learning Representations

    work page 2022

  8. [8]

    S.; Riley, P

    Gilmer, J.; Schoenholz, S. S.; Riley, P. F.; Vinyals, O.; and Dahl, G. E. 2020. Message passing neural networks. In Machine learning meets quantum physics, 199--214. Springer

    work page 2020

Show all 44 references
  1. [9]

    Gui, S.; Yuan, H.; Wang, J.; Lao, Q.; Li, K.; and Ji, S. 2023. Flowx: Towards explainable graph neural networks via message flows. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7): 4567--4578

    2023

  2. [10]

    Henderson, R.; Clevert, D.-A.; and Montanari, F. 2021. Improving molecular graph neural network explainability with orthonormalization and induced sparsity. In International Conference on Machine Learning, 4203--4213. PMLR

    2021

  3. [11]

    W.; Xiao, C.; Sun, J.; and Zitnik, M

    Huang, K.; Fu, T.; Gao, W.; Zhao, Y.; Roohani, Y.; Leskovec, J.; Coley, C. W.; Xiao, C.; Sun, J.; and Zitnik, M. 2021. Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development. arXiv preprint arXiv:2102.09548

    2021 arXiv

  4. [12]

    Karim, A.; Lee, M.; Balle, T.; and Sattar, A. 2021. CardioTox net: a robust predictor for hERG channel blockade based on deep learning meta-feature ensembles. Journal of Cheminformatics, 13(1): 60

    2021

  5. [13]

    Kipf, T. 2016. Semi-Supervised Classification with Graph Convolutional Networks. arXiv preprint arXiv:1609.02907

    2016 arXiv

  6. [14]

    K.; Han, Y.-K.; Kang, J.; Kang, C.-J.; and Son, Y

    Lee, S.; Park, H.; Choi, C.; Kim, W.; Kim, K. K.; Han, Y.-K.; Kang, J.; Kang, C.-J.; and Son, Y. 2023. Multi-order graph attention network for water solubility prediction and interpretation. Scientific Reports, 13(1): 957

    2023

  7. [15]

    Mahendran, A.; and Vedaldi, A. 2016. Salient deconvolutional networks. In European conference on computer vision, 120--135. Springer

    2016

  8. [16]

    Maziarka, .; Majchrowski, D.; Danel, T.; Gai \'n ski, P.; Tabor, J.; Podolak, I.; Morkisz, P.; and Jastrz e bski, S. 2024. Relative molecule self-attention transformer. Journal of Cheminformatics, 16(1): 3

    2024

  9. [17]

    Murcko, M. A. 1995. Computational methods to predict binding free energy in ligand-receptor complexes. Journal of medicinal chemistry, 38(26): 4953--4967

    1995

  10. [18]

    A.; and Campillo, N

    Ponzoni, I.; P \'a ez Prosper, J. A.; and Campillo, N. E. 2023. Explainable artificial intelligence: A taxonomy and guidelines for its application to drug discovery. Wiley Interdisciplinary Reviews: Computational Molecular Science, 13(6): e1681

    2023

  11. [19]

    Proszewska, M.; Danel, T.; and Rymarczyk, D. 2025. B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data. arXiv preprint arXiv:2505.22252

    2025 arXiv

  12. [20]

    Rong, Y.; Bian, Y.; Xu, T.; Xie, W.; Wei, Y.; Huang, W.; and Huang, J. 2020. Self-supervised graph transformer on large-scale molecular data. Advances in neural information processing systems, 33: 12559--12571

    2020

  13. [21]

    Rymarczyk, D.; Dobrowolski, D.; and Danel, T. 2023. ProGReST : Prototypical graph regression soft trees for molecular property prediction. In Proceedings of the 2023 SIAM International Conference on Data Mining (SDM), 379--387. SIAM

    2023

  14. [22]

    Shrikumar, A.; Greenside, P.; Shcherbina, A.; and Kundaje, A. 2016. Not just a black box: Learning important features through propagating activation differences. arXiv preprint arXiv:1605.01713

    2016 arXiv

  15. [23]

    Simonyan, K.; Vedaldi, A.; and Zisserman, A. 2014. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

    2014

  16. [24]

    C.; Khetan, A.; and Er, S

    Sorkun, M. C.; Khetan, A.; and Er, S. 2019. AqSolDB, a curated reference set of aqueous solubility and 2D descriptors for a diverse set of compounds. Scientific data, 6(1): 143

    2019

  17. [25]

    T.; Dosovitskiy, A.; Brox, T.; and Riedmiller, M

    Springenberg, J. T.; Dosovitskiy, A.; Brox, T.; and Riedmiller, M. 2014. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806

    2014 arXiv

  18. [26]

    Sundararajan, M.; Taly, A.; and Yan, Q. 2017. Axiomatic attribution for deep networks. In International conference on machine learning, 3319--3328. PMLR

    2017

  19. [27]

    Talanquer, V. 2022. The complexity of reasoning about and with chemical representations. Jacs Au, 2(12): 2658--2669

    2022

  20. [28]

    Tanaka, A.; Okamoto, H.; and Bersohn, M. 2010. Construction of functional group reactivity database under various reaction conditions automatically extracted from reaction database in a synthesis design system. Journal of chemical information and modeling, 50(3): 327--338

    2010

  21. [29]

    R.; Brackman, E

    Tuttle, M. R.; Brackman, E. M.; Sorourifar, F.; Paulson, J.; and Zhang, S. 2023. Predicting the solubility of organic energy storage materials based on functional group identity and substitution pattern. The Journal of Physical Chemistry Letters, 14(5): 1318--1325

    2023

  22. [30]

    P.; Lloyd, D

    Veith, H.; Southall, N.; Huang, R.; James, T.; Fayne, D.; Artemenko, N.; Shen, M.; Inglese, J.; Austin, C. P.; Lloyd, D. G.; et al. 2009. Comprehensive characterization of cytochrome P450 isozyme selectivity across chemical libraries. Nature biotechnology, 27(11): 1050--1055

    2009

  23. [31]

    Veli c kovi \'c , P.; Cucurull, G.; Casanova, A.; Romero, A.; Lio, P.; and Bengio, Y. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903

    2017 arXiv

  24. [32]

    von Korff, M.; and Sander, T. 2006. Toxicity-indicating structural patterns. Journal of chemical information and modeling, 46(2): 536--544

    2006

  25. [33]

    Vu, M.; and Thai, M. T. 2020. Pgm-explainer: Probabilistic graphical model explanations for graph neural networks. Advances in neural information processing systems, 33: 12225--12235

    2020

  26. [34]

    Wang, J.; Min, Y.; Li, M.; and Wu, J. 2025. FragFormer: A Fragment-based Representation Learning Framework for Molecular Property Prediction. Transactions on Machine Learning Research

    2025

  27. [35]

    Wieder, O.; Kohlbacher, S.; Kuenemann, M.; Garon, A.; Ducrot, P.; Seidel, T.; and Langer, T. 2020. A compact review of molecular property prediction with graph neural networks. Drug Discovery Today: Technologies, 37: 1--12

    2020

  28. [36]

    Wu, Z.; Wang, J.; Du, H.; Jiang, D.; Kang, Y.; Li, D.; Pan, P.; Deng, Y.; Cao, D.; Hsieh, C.-Y.; et al. 2023. Chemistry-intuitive explanation of graph neural networks for molecular property prediction with substructure masking. Nature communications, 14(1): 2585

    2023

  29. [37]

    Xiong, Z.; Wang, D.; Liu, X.; Zhong, F.; Wan, X.; Li, X.; Li, Z.; Luo, X.; Chen, K.; Jiang, H.; et al. 2019. Pushing the boundaries of molecular representation for drug discovery with the graph attention mechanism. Journal of medicinal chemistry, 63(16): 8749--8760

    2019

  30. [38]

    Xu, K.; Hu, W.; Leskovec, J.; and Jegelka, S. 2018. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826

    2018 arXiv

  31. [39]

    Ying, Z.; Bourgeois, D.; You, J.; Zitnik, M.; and Leskovec, J. 2019. Gnnexplainer: Generating explanations for graph neural networks. Advances in neural information processing systems, 32

    2019

  32. [40]

    Yuan, H.; Yu, H.; Wang, J.; Li, K.; and Ji, S. 2021. On explainability of graph neural networks via subgraph explorations. In International conference on machine learning, 12241--12252. PMLR

    2021

  33. [41]

    Zhang, X.; Xu, Y.; He, W.; Guo, W.; and Cui, L. 2023. A comprehensive review of the oversmoothing in graph neural networks. In CCF Conference on Computer Supported Cooperative Work and Social Computing, 451--465. Springer

    2023

  34. [42]

    Zhang, Z.; Liu, Q.; Wang, H.; Lu, C.; and Lee, C. 2022. Protgnn: Towards self-explaining graph neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 36, 9127--9135

    2022

  35. [43]

    Zhang, Z.; Liu, Q.; Wang, H.; Lu, C.; and Lee, C.-K. 2021. Motif-based graph self-supervised learning for molecular property prediction. Advances in Neural Information Processing Systems, 34: 15870--15882

    2021

  36. [44]

    Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; and Torralba, A. 2016. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2921--2929

    2016

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.