Pith. sign in

REVIEW 3 major objections 4 minor 53 references

Combining Graph Neural Networks and Mixed Integer Linear Programming for Molecular Inference under the Two-Layered Model

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper encodes a graph neural network's prediction as a mixed-integer linear program, making GNN-based inverse molecular design exact under the two-layered model, so feasible solutions are guaranteed chemical graphs with property…

desk verdict New MILP for GNN-based inverse design under the two-layered model, but the appendix omits the MLP head, so the documented formulation doesn't enforce the property range. read the letter →

arxiv 2507.03920 v1 pith:P3XG26KY submitted 2025-07-05 cs.LG physics.chem-ph

Jianshen Zhu , Naveed Ahmed Azam , Kazuya Haraguchi , Liang Zhao , Tatsuya Akutsu This is my paper
classification cs.LGphysics.chem-ph MSC 90C1192E10
keywords inverseQSAR/QSPRgraphneuralnetworksmixedintegerlinearprogrammingtwo-layeredmodelchemicalinferenceQM9datasetmol-inferLeakyReLUlinearization
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

This paper claims that the inverse QSAR/QSPR problem can be solved exactly when the prediction function is a graph neural network, provided the network is designed to match the two-layered model of chemical graphs. The authors embed a purpose-built GNN, 2L-GNN, into the mol-infer framework by writing a mixed-integer linear program (MILP) that simulates every layer of the network, including the LeakyReLU activations, and simultaneously encodes the structural constraints of the target molecule class. A feasible solution of this MILP is a chemical graph that satisfies the declared topology and whose predicted property value lies in the desired range; an infeasible program certifies that no such graph exists. Experiments on QM9 show that the GNN improves prediction quality for several properties that were hard for descriptor-based mol-infer models, and that molecules with up to 20 non-hydrogen atoms can be inferred within an hour.

What carries the argument

The central object is the 2L-GNN architecture paired with the MILP $\mathcal{M}_{\mathrm{GNN}}$. 2L-GNN is a three-layer message-passing network that operates only on interior vertices of the two-layered chemical graph, aggregates neighbor messages through an affine map, applies LeakyReLU with slope $\alpha = 0.1$, pools the final node features, and finishes with a small fully connected head. The MILP reproduces this computation with indicator variables $\delta^\tau$ and big-$M$ bounds $M_\ell$, so that the piecewise-linear LeakyReLU becomes linear constraints (68)--(81), and the structural variables (path lengths, fringe-tree choices, element assignments, bond multiplicities) provide the initial node features. This lets the inverse problem be posed as a single feasibility program whose variables simultaneously determine the graph and its GNN prediction.

What would settle it

Construct (or randomly sample) a molecule that satisfies a topological specification $\sigma$ and run both the direct 2L-GNN predictor and the MILP $\mathcal{M}_{\mathrm{GNN}}$ on it; if any hidden-layer activation exceeds the training-data maximum $M_\ell$, or the MILP's predicted value disagrees with the direct network forward pass, the exactness claim is falsified. A cheaper check is to solve the MILP for a target range that the GNN assigns to a hand-built out-of-distribution molecule and see whether the program reports infeasible despite the molecule being a valid member of $\mathcal{G}_\sigma$.

Watch

Extended reading notes

Core claim

Phase 2 of mol-infer-GNN solves the feasibility problem of finding $C^* \in \mathcal{G}_\sigma$ with $\eta(C^*) \in [y_*, y^*]$ by solving the MILP $\mathcal{M}_{\mathrm{GNN}}(g,x,y;C_1,C_2)$. Appendix C.8 gives the constraint set, which linearizes the LeakyReLU of every neuron with big-$M$ variables and couples the network's messages to the graph-structural variables of the two-layered model. The paper's claim is that this formulation is exact: any feasible solution corresponds to a genuine chemical graph that respects the seed graph, fringe-tree choices, and valence conditions, and whose GNN-predicted property value is provably inside the target interval, while infeasibility of the MILP is a certificate that the specified graph class contains no desired molecule. If this holds, it is the first inverse QSAR/QSPR method that combines a GNN learner with the flexible topological specifications of the two-layered model and retains both optimality and exactness guarantees.

Load-bearing premise

The MILP linearization is only guaranteed to reproduce the GNN exactly if every candidate chemical graph's neuron activations stay within the bounds $M_\ell$ observed on the training data; a graph with larger activations would be wrongly excluded or mis-encoded.

Editorial extensions

If this is right

  • Properties that were poorly predicted by descriptor-based mol-infer models, such as electric dipole moment (mu), HOMO, LUMO, and gap, obtain substantially better R2 values with 2L-GNN, which translates into higher-quality inferred molecules.
  • The MILP formulation inherits the two-layered model's flexibility: by changing the seed graph and fringe-tree set, the same code can target monocyclic, bicyclic, or QM9-like molecules, a generality that earlier GNN-MILP approaches lacked.
  • When the MILP is infeasible, the framework outputs a certificate that no chemical graph in the specified class meets the target range, a guarantee not offered by generative-model approaches.
  • The computational experiments show that increasing the GNN width from 16 to 32 roughly doubles the MILP difficulty, so the tractable inference size (up to 20 heavy atoms) is sensitive to the network's architecture.

Reading between the lines

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

  • A straightforward extension would recompute the big-$M$ bounds $M_\ell$ from the entire class $\mathcal{G}_\sigma$ (or tighten them with interval arithmetic per instance) instead of from training data; the paper's exactness claim currently depends on training-set maxima bounding all candidate graphs.
  • The same big-$M$ linearization strategy transfers to other piecewise-linear activations and to GNN readouts that are linear, suggesting a whole family of 'exactly invertible' GNN architectures for molecular design.
  • The reported disagreement between the GNN's predictions and the Psi4/PySCF DFT values indicates that the bottleneck for practical design is the prediction function's accuracy rather than the exactness of the encoding; pairing 2L-GNN with a post-hoc DFT filter or a hybrid descriptor-GNN head, as the authors hint, is a natural next step.
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

3 major / 4 minor

Summary. The paper proposes mol-infer-GNN, an extension of the two-phase mol-infer framework that replaces handcrafted descriptors with a purpose-built graph neural network, 2L-GNN, for inverse QSAR/QSPR under the two-layered model. Phase 1 trains 2L-GNN on the QM9 dataset for eight properties; Phase 2 solves a mixed-integer linear program MGNN(g,x,y;C1,C2) that is claimed to simulate the trained 2L-GNN and the topological/chemical constraints of the two-layered model, thereby certifying existence or infeasibility of a chemical graph in G_sigma with predicted property in [y*,y*]. Experiments report Phase 1 accuracy, Phase 2 solve counts for five instances and three properties, and DFT validation of selected inferred molecules with PySCF and Psi4.

Significance. If the MILP formulation were exact, this would be a valuable advance: it would be the first GNN-based inverse QSAR/QSPR method under the flexible two-layered model with mathematical guarantees of optimality and exactness, including infeasibility certificates. The manuscript has several strengths: the two-layered structural constraints are given in full in the appendix, the GNN architecture is simple enough to be linearly encoded, the authors provide public code and data, and the external DFT validation is an independent check on the quality of inferred molecules. The paper also honestly reports low solve rates and acknowledges discrepancies between predicted and computed property values. However, the central exactness claim is not currently supported by the written formulation, because the documented MILP stops before the prediction head and the target-range constraint.

major comments (3)
  1. [Section 3.1 and Appendix C.8, Eq. (81)] The formulation as written does not enforce the target-range constraint eta(C*) in [y*,y*]. Section 3.1 defines 2L-GNN as a message-passing encoder followed by a fully connected ReLU network whose output is y* = eta(C), but Appendix C.8, introduced as the complete constraint set simulating 2L-GNN, ends at Eq. (81) with the computation of the representation vector theta_C(p). There is no scalar variable for the prediction y, no constraints for the fully connected head, and no constraint linking theta_C to the interval [y*,y*]. Consequently, a feasible solution of MGNN(g,x,y;C1,C2) as documented only guarantees consistency of the encoder and the structural constraints; it does not guarantee eta(C*) in [y*,y*], and infeasibility does not certify that no such graph exists. This is not a mere presentation gap: the central exactness and infeasibility claims in Section 3.2 depend on these omitted constraints. The authors should add the head variables and constraints, or explicitly narrow the claims to the encoder readout.
  2. [Appendix C.8, definition of M_l] The big-M bounds M_l are defined as upper bounds on the maximum activation values over the training data D_pi, and the linearization constraints (68)-(81) assume these bounds hold for every candidate chemical graph in Phase 2. A target graph whose activations exceed M_l would be incorrectly constrained, so the exactness claim depends on training-set bounds generalizing to the entire search space G_sigma. This is a load-bearing correctness assumption separate from the GNN's predictive accuracy. The authors should either prove finite bounds over G_sigma from the bounded input features and weights, or state this as a restriction on the class of graphs for which the formulation is exact.
  3. [Tables 3-4 and Section 4.2] The experimental support for the feasibility/infeasibility claim is limited. In Tables 3 and 4, zero instances are certified infeasible across all 150 formulations for each model, so the infeasibility-certification capability is never demonstrated. The reported success counts are 43/150 (28.7%) for 2L-GNN16 and 20/150 (13.3%) for 2L-GNN32, and the subsequent quality analysis in Tables 5 and 6 selects only 12 solved instances. Phase 1 also reports single runs without error bars, and Table 2 compares against mol-infer baselines trained on a 1000-molecule subset while 2L-GNN uses 110,000 training molecules. These choices should be acknowledged more transparently, and the claims about practical applicability should be tempered accordingly.
minor comments (4)
  1. [Abstract and Section 1] There is a typo in the introduction: 'temeprature' should be 'temperature'.
  2. [Section 4.1, property list] The bullet for Lumo reads 'Lumo energy of lowest occupied molecular orbital (meV);' and is missing the word 'highest' for the HOMO analog; also the formatting should be consistent with the other bullets.
  3. [Appendix C.8, Eq. (81)] The variables tau_ftr(p) and delta_tau_ftr(p) are used in Eq. (81) but are not declared in the variable list; please add their definitions to make the constraint set self-contained.
  4. [Appendix C.8, section heading] The heading 'Calculating the (ell + 1)-th vector from the ell-th vecctor' contains a typo ('vecctor' should be 'vector').

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the inverse problem is framed against an independently trained 2L-GNN and externally validated with DFT software; the main caveat is an undocumented MLP-head constraint in Appendix C.8, which is an omitted proof rather than an equivalence-by-construction step.

full rationale

The derivation chain is not circular. Phase 1 trains 2L-GNN on QM9 data, and Phase 2 constructs an MILP whose feasible solutions are intended to satisfy C in G_sigma and eta(C) in [y*, y*]. The target is stated as a constraint on the trained eta, not as a quantity fitted inside the MILP, so no fitted parameter is renamed as a prediction. The constraints in Appendix C.8 simulate the message-passing layers and graph-level pooling of 2L-GNN using big-M linearizations of LeakyReLU; the constants M_l are defined as upper bounds on training-data activations. This is a generalization assumption that may fail out-of-distribution, but it is not circular because the assumption is not constructed from the target range or from the existence of the solution. The paper's reliance on the authors' prior mol-infer framework and two-layered model is framework reuse, not a self-citation chain that forces the central result, and there is no imported uniqueness theorem. The independent PySCF/Psi4 DFT validation provides external evidence for the inferred molecules. The one serious caveat is that Appendix C.8, which the paper introduces as the complete set of constraints for MGNN(x,y;C1), stops at the representation vector theta_C (Eq. 81); no variables or constraints are shown for the fully connected ReLU head that maps theta_C to the scalar prediction eta(C) or for the target-range condition eta(C) in [y*, y*]. That is an omitted proof of the exactness claim, not a circular reduction: the formulation as printed may not enforce the target range, but nothing in the paper defines eta in terms of the MILP solution or fits the MILP to its own output.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The framework relies on a trained GNN with fitted weights, learned fringe-tree embeddings, and hand-chosen hyperparameters, plus bounds derived from training data. It also assumes the QM9 labels, the two-layered model validity constraints, and the DFT validation software are correct.

free parameters (4)
  • 2L-GNN weights and biases = not listed in paper
    Learned on QM9 training set; includes W(ℓ), B(ℓ), wC, and bias terms in Appendix C.8.
  • Secondary fringe-tree GNN embedding θψ = 8-dimensional vector per fringe-tree
    Learned feature vector used in initial node features and treated as a constant in the MILP.
  • Hyperparameters L=3, Khid=16 or 32, KC=32, KF=8, Knode=15 = 2L-GNN16 and 2L-GNN32
    Chosen by preliminary experiments to balance learning accuracy and MILP solve time.
  • Big-M bounds Mℓ = not given numerically
    Computed from training data to linearize LeakyReLU; if too small, constraints may exclude valid inferred graphs.
assumptions (5)
  • domain assumption QM9 dataset property values are reliable labels for training and testing
    Training and validation use DFT-computed QM9 values without independent experimental verification.
  • domain assumption The two-layered model and topological specification constraints from prior work correctly characterize all chemical graphs in G(GC,σint,σce)
    The paper relies on the validity of prior mol-infer formulations for chemical graph constraints.
  • domain assumption The trained GNN generalizes to unseen chemical graphs satisfying the topological specification
    Phase 2 searches over graphs not necessarily in the training distribution.
  • domain assumption Big-M linearization exactly reproduces LeakyReLU behavior when Mℓ bounds hold
    The MILP encoding in Appendix C.8 assumes the bounds from training data are valid for all candidate graphs.
  • domain assumption DFT at B3LYP/6-31G(2df,p) as implemented in PySCF and Psi4 is a meaningful external validation
    The paper compares inferred molecules against these quantum chemistry codes to assess quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Combining Graph Neural Networks and Mixed Integer Linear Programming for Molecular Inference under the Two-Layered Model." pith.science (2026). https://pith.science/paper/P3XG26KY

@misc{pith2026250703920,
  author       = {Pith},
  title        = {Pith review of: Combining Graph Neural Networks and Mixed Integer Linear Programming for Molecular Inference under the Two-Layered Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3XG26KY}},
  note         = {Machine review of arXiv:2507.03920}
}
read the original abstract

Recently, a novel two-phase framework named mol-infer for inference of chemical compounds with prescribed abstract structures and desired property values has been proposed. The framework mol-infer is primarily based on using mixed integer linear programming (MILP) to simulate the computational process of machine learning methods and describe the necessary and sufficient conditions to ensure such a chemical graph exists. The existing approaches usually first convert the chemical compounds into handcrafted feature vectors to construct prediction functions, but because of the limit on the kinds of descriptors originated from the need for tractability in the MILP formulation, the learning performances on datasets of some properties are not good enough. A lack of good learning performance can greatly lower the quality of the inferred chemical graphs, and thus improving learning performance is of great importance. On the other hand, graph neural networks (GNN) offer a promising machine learning method to directly utilize the chemical graphs as the input, and many existing GNN-based approaches to the molecular property prediction problem have shown that they can enjoy better learning performances compared to the traditional approaches that are based on feature vectors. In this study, we develop a molecular inference framework based on mol-infer, namely mol-infer-GNN, that utilizes GNN as the learning method while keeping the great flexibility originated from the two-layered model on the abstract structure of the chemical graph to be inferred. We conducted computational experiments on the QM9 dataset to show that our proposed GNN model can obtain satisfying learning performances for some properties despite its simple structure, and can infer small chemical graphs comprising up to 20 non-hydrogen atoms within reasonable computational time.

Figures

Figures reproduced from arXiv: 2507.03920 by the authors.

Figure 1
Figure 1. An illustration of the two-phase framework [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An illustration of the 2L-model for a chemical graph [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. An illustration of 2-fringe-trees C[ui ], u ∈ [1, 28] of the example C depicted in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: An illustration of the two-phase GNN-based molecular infer [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: An illustration of the graph neural network [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Illustrations of the seed graph Gi C of instance Ii , i ∈ [1, 5], respectively. The edges in E(≥2) are depicted with dotted lines, the edges in E(≥1) are depicted with dashed lines, the edges in E(0/1) are depicted in gray bold lines and the edges in E(=1) are depicted…
Figure 7
Figure 7. Figure 7: Illustrations of the inferred chemical graphs. (a)-(l) co [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: (a) An illustration of a seed graph GC with r(GC) = 5 where the vertices in VC are depicted with gray circles, the edges in E(≥2) are depicted with dotted lines, the edges in E(≥1) are depicted with dashed lines, the edges in E(0/1) are depicted with gray bold lines an…
Figure 9
Figure 9. Figure 9: The interior Hint of chemical graph C with hCi in [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: (a) A cyclical-base S = Hint − S u∈{u5,u18,u22} (V (Qu) \ {u}) of the interior Hint in [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]
Figure 11
Figure 11. Figure 11: An illustration of a graph H∗ that is obtained from the seed graph GC in [PITH_FULL_IMAGE:figures/full_fig_p027_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 49 canonical work pages

  1. [1]

    https://pubchem.ncbi.nlm .nih.gov/source/11933

    Annotations from HSDB (on pubchem). https://pubchem.ncbi.nlm .nih.gov/source/11933. Accessed on May 16th, 2025

    work page 2025

  2. [2]

    N. A. Azam, R. Chiewvanichakorn, F. Zhang, A. Shurbevski, H. N agamochi, and T. Akutsu. A novel method for the inverse QSAR/QSPR based on artificial neura l networks and mixed inte- ger linear programming with guaranteed admissibility. In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies - BIOINFORMAT...

    work page 2020

  3. [3]

    N. A. Azam, J. Zhu, K. Haraguchi, L. Zhao, H. Nagamochi, and T. Akutsu. Molecular design based on artificial neural networks, integer programming and grid neighbor search. In 2021 IEEE International Conference on Bioinformatics and Biome dicine (BIBM) , pages 360–363. IEEE, 2021

    work page 2021

  4. [4]

    N. A. Azam, J. Zhu, Y. Sun, Y. Shi, A. Shurbevski, L. Zhao, H. Na gamochi, and T. Akutsu. A novel method for inference of acyclic chemical compounds with bou nded branch-height based on artificial neural networks and integer programming. Algorithms for Molecular Biology , 16:1–39, 2021

    work page 2021

  5. [5]

    Bojchevski, O

    A. Bojchevski, O. Shchur, D. Z¨ ugner, and S. G¨ unnemann. Ne tGAN: Generating graphs via random walks. In J. Dy and A. Krause, editors, Proceedings of the 35th International Conference on Machine Learning , volume 80 of Proceedings of Machine Learning Research , pages 610–619. PMLR, 10–15 Jul 2018

    work page 2018

  6. [6]

    W. Bort, D. Mazitov, D. Horvath, F. Bonachera, A. Lin, G. Marc ou, I. Baskin, T. Madzhidov, and A. Varnek. Inverse QSAR: Reversing descriptor-driven pred iction pipeline using attention- based conditional variational autoencoder. Journal of Chemical Information and Modeling , 62(22):5471–5484, 11 2022

    work page 2022

  7. [7]

    H. Cai, H. Zhang, D. Zhao, J. Wu, and L. Wang. FP-GNN: a versat ile deep learning architec- ture for enhanced molecular property prediction. Briefings in Bioinformatics , 23(6):bbac408, 09 2022

    work page 2022

  8. [8]

    N. D. Cao and T. Kipf. MolGAN: An implicit generative model for small molecular graphs. arXiv:1805.11973, 2022

    arXiv 2022

Show all 53 references
  1. [9]

    Cheng, Y

    Y. Cheng, Y. Gong, Y. Liu, B. Song, and Q. Zou. Molecular design in drug discovery: a comprehensive review of deep generative models. Briefings in Bioinformatics , 22(6):bbab344, 08 2021

    2021

  2. [10]

    Cherkasov, E

    A. Cherkasov, E. N. Muratov, D. Fourches, A. Varnek, I. I. Baskin, M. Cronin, J. Dearden, P. Gramatica, Y. C. Martin, R. Todeschini, et al. QSAR modeling: wher e have you been? where are you going to? Journal of Medicinal Chemistry , 57(12):4977–5010, 2014

    2014

  3. [11]

    Duvenaud, D

    D. Duvenaud, D. Maclaurin, J. Aguilera-Iparraguirre, R. G´ om ez-Bombarelli, T. Hirzel, A. Aspuru-Guzik, and R. P. Adams. Convolutional networks on gra phs for learning molecu- lar fingerprints. In Proceedings of the 29th International Conference on Neural Information Processin...

    2015

  4. [12]

    Fey and J

    M. Fey and J. E. Lenssen. Fast graph representation learning with PyTorch Geometric. In ICLR Workshop on Representation Learning on Graphs and Mani folds, 2019

    2019

  5. [13]

    Gasteiger, S

    J. Gasteiger, S. Giri, J. T. Margraf, and S. G¨ unnemann. Fast and uncertainty-aware directional message passing for non-equilibrium molecules. arXiv:2011.14115, 20 22

    2011 arXiv

  6. [14]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl. Neural message passing for quantum chemistry. In D. Precup and Y. W. Teh, editors, Proceedings of the 34th International Conference on Machine Learning , volume 70 of Proceedings of Machine Learning Research...

    2017

  7. [15]

    G´ omez-Bombarelli, J

    R. G´ omez-Bombarelli, J. N. Wei, D. Duvenaud, J. M. Hern´ ande z-Lobato, B. S´ anchez- Lengeling, D. Sheberla, J. Aguilera-Iparraguirre, T. D. Hirzel, R. P . Adams, and A. Aspuru- Guzik. Automatic chemical design using a data-driven continuous re presentation of molecules. AC...

    2018

  8. [16]

    R. Ido, N. A. Azam, J. Zhu, H. Nagamochi, and T. Akutsu. A dyn amic programming algorithm for generating chemical isomers based on frequency vectors. Scientific Reports , 15(1):22214, 2025

    2025

  9. [17]

    Ikebata, K

    H. Ikebata, K. Hongo, T. Isomura, R. Maezono, and R. Yoshid a. Bayesian molecular design with a chemical language model. Journal of Computer-aided Molecular Design , 31:379–391, 2017

    2017

  10. [18]

    R. Ito, N. A. Azam, C. Wang, A. Shurbevski, H. Nagamochi, and T. Akutsu. A novel method for the inverse QSAR/QSPR to monocyclic chemical compounds base d on artificial neural networks and integer programming. In Advances in Computer Vision and Computational Biology: Proceedings...

  11. [19]

    H. Kaneko. Molecular descriptors, structure generation, an d inverse QSAR/QSPR based on SELFIES. ACS Omega, 8(24):21781–21786, 06 2023

    2023

  12. [20]

    S. Kim, J. Chen, T. Cheng, A. Gindulyte, J. He, S. He, Q. Li, B. A. Shoemaker, P. A. Thiessen, B. Yu, et al. PubChem 2023 update. Nucleic Acids Research , 51(D1):D1373–D1380, 2023

    2023

  13. [21]

    T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. arXiv:1609.02907, 2017

    2017 arXiv

  14. [22]

    M. J. Kusner, B. Paige, and J. M. Hern´ andez-Lobato. Gramm ar variational autoencoder. In International Conference on Machine Learning , pages 1945–1954. PMLR, 2017

    1945

  15. [23]

    Y.-C. Lo, S. E. Rensi, W. Torng, and R. B. Altman. Machine learnin g in chemoinformatics and drug discovery. Drug Discovery Today , 23(8):1538–1546, 2018

    2018

  16. [24]

    C. Lu, Q. Liu, C. Wang, Z. Huang, P. Lin, and L. He. Molecular pro perty prediction: a mul- tilevel quantum interactions modeling perspective. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Inn ovative Applications of Artificial In...

    2019

  17. [25]

    McDonald, C

    T. McDonald, C. Tsay, A. M. Schweidtmann, and N. Yorke-Smith . Mixed-integer optimisa- tion of graph neural networks for computer-aided molecular desig n. Computers & Chemical Engineering, 185:108660, 2024

    2024

  18. [26]

    A. Micheli. Neural network for graphs: A contextual constru ctive approach. IEEE Transac- tions on Neural Networks , 20(3):498–511, 2009

    2009

  19. [27]

    Miyao, H

    T. Miyao, H. Kaneko, and K. Funatsu. Inverse QSPR/QSAR ana lysis for chemical structure generation (from y to x). Journal of Chemical Information and Modeling , 56(2):286–299, 2016

    2016

  20. [28]

    Morgan and R

    D. Morgan and R. Jacobs. Opportunities and challenges for mac hine learning in materials science. Annual Review of Materials Research , 50(Volume 50, 2020):71–103, 2020

    2020

  21. [29]

    Ramakrishnan, P

    R. Ramakrishnan, P. O. Dral, M. Rupp, and O. A. Von Lilienfeld. Qu antum chemistry structures and properties of 134 kilo molecules. Scientific data , 1(1):1–7, 2014

    2014

  22. [30]

    Rupakheti, A

    C. Rupakheti, A. Virshup, W. Yang, and D. N. Beratan. Strate gy to discover diverse optimal molecules in the small molecule universe. Journal of Chemical Information and Modeling , 55(3):529–537, 2015

    2015

  23. [31]

    K. T. Sch¨ utt, F. Arbabzadah, S. Chmiela, K. R. M¨ uller, and A. Tkatchenko. Quantum- chemical insights from deep tensor neural networks. Nature Communications , 8(1):13890, 2017

    2017

  24. [32]

    C. Shi, M. Xu, Z. Zhu, W. Zhang, M. Zhang, and J. Tang. GraphA F: a flow-based autore- gressive model for molecular graph generation. arXiv:2001.09382, 2020

    2001 arXiv

  25. [33]

    Y. Shi, J. Zhu, N. A. Azam, K. Haraguchi, L. Zhao, H. Nagamoch i, and T. Akutsu. An inverse QSAR method based on a two-layered model and integer programmin g. International Journal of Molecular Sciences , 22(6):2847, 2021

    2021

  26. [34]

    Shino and H

    Y. Shino and H. Kaneko. Improving molecular design with direct inv erse analysis of QSAR/QSPR model. Molecular Informatics, 44(1):e202400227, 2025

    2025

  27. [35]

    M. I. Skvortsova, I. I. Baskin, O. L. Slovokhotova, V. A. Paly ulin, and N. S. Zefirov. Inverse problem in QSAR/QSPR studies for the case of topological indexes ch aracterizing molecular shape (kier indices). Journal of Chemical Information and Computer Sciences , 33(4):630–634, 1993

    1993

  28. [36]

    D. G. A. Smith, L. A. Burns, A. C. Simmonett, R. M. Parrish, M. C . Schieber, R. Galvelis, P. Kraus, H. Kruse, R. Di Remigio, A. Alenaizan, A. M. James, S. Leh tola, J. P. Misiewicz, M. Scheurer, R. A. Shaw, J. B. Schriber, Y. Xie, Z. L. Glick, D. A. Sir ianni, J. S. O’Brien, ...

    2020

  29. [37]

    Q. Sun, T. C. Berkelbach, N. S. Blunt, G. H. Booth, S. Guo, Z. L i, J. Liu, J. D. McClain, E. R. Sayfutyarova, S. Sharma, S. Wouters, and G. K.-L. Chan. P ySCF: the python-based simulations of chemistry framework. WIREs Computational Molecular Science , 8(1):e1340, 2018. 2LMM ...

    2018

  30. [38]

    Q. Sun, X. Zhang, S. Banerjee, P. Bao, M. Barbry, N. S. Blunt , N. A. Bogdanov, G. H. Booth, J. Chen, Z.-H. Cui, J. J. Eriksen, Y. Gao, S. Guo, J. Hermann, M. R. Hermes, K. Koh, P. Ko- val, S. Lehtola, Z. Li, J. Liu, N. Mardirossian, J. D. McClain, M. Motta , B. Mussard, H. Q....

    2020

  31. [39]

    Tanaka, J

    K. Tanaka, J. Zhu, N. A. Azam, K. Haraguchi, L. Zhao, H. Naga mochi, and T. Akutsu. An inverse QSAR method based on decision tree and integer program ming. In Intelligent Computing Theories and Application: 17th International Co nference, ICIC 2021, Shenzhen, China, August 12–...

    2021

  32. [40]

    I. V. Tetko and O. Engkvist. From big data to artificial intelligenc e: chemoinformatics meets new challenges. Journal of Cheminformatics , 12:1–3, 2020

    2020

  33. [41]

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu. A comp rehensive survey on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems , 32(1):4–24, 2021

    2021

  34. [42]

    Z. Wu, B. Ramsundar, E. N. Feinberg, J. Gomes, C. Geniesse, A . S. Pappu, K. Leswing, and V. Pande. MoleculeNet: a benchmark for molecular machine learning. Chemical Science , 9(2):513–530, 2018

    2018

  35. [43]

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka. How powerful are g raph neural networks? In International Conference on Learning Representations , 2019

    2019

  36. [44]

    Zhang, J

    F. Zhang, J. Zhu, R. Chiewvanichakorn, A. Shurbevski, H. Nag amochi, and T. Akutsu. A new approach to the design of acyclic chemical compounds using ske leton trees and integer linear programming. Applied Intelligence, 52(15):17058–17072, 2022

    2022

  37. [45]

    Zhang, J

    S. Zhang, J. S. Campos, C. Feldmann, F. Sandfort, M. Mathea , and R. Misener. Augmenting optimization-based molecular design with graph neural networks. Computers & Chemical Engineering, 186:108684, 2024

    2024

  38. [46]

    Zhang, Y

    S. Zhang, Y. Liu, and L. Xie. A universal framework for accura te and efficient geometric deep learning of molecular systems. Scientific Reports, 13(1):19171, 2023

    2023

  39. [47]

    J. Zhu. Novel Methods for Chemical Compound Inference Based on Mach ine Learning and Mixed Integer Linear Programming . PhD thesis, Kyoto University, 9 2023

    2023

  40. [48]

    J. Zhu, N. A. Azam, S. Cao, R. Ido, K. Haraguchi, L. Zhao, H. N agamochi, and T. Akutsu. Quadratic descriptors and reduction methods in a two-layered mod el for compound inference. Frontiers in Genetics , 15:1483490, 2025

    2025

  41. [49]

    J. Zhu, N. A. Azam, K. Haraguchi, L. Zhao, H. Nagamochi, and T . Akutsu. An inverse QSAR method based on linear regression and integer programming. Frontiers in Bioscience- Landmark, 27(6):188, 2022. 2LMM GNN: September 23, 2025 22

    2022

  42. [50]

    J. Zhu, N. A. Azam, K. Haraguchi, L. Zhao, H. Nagamochi, and T . Akutsu. Molecular design based on integer programming and splitting data sets by hype rplanes. IEEE/ACM Transactions on Computational Biology and Bioinformatics , 21(5):1529–1541, 2024

    2024

  43. [51]

    J. Zhu, K. Haraguchi, H. Nagamochi, and T. Akutsu. Adjustive linear regression and its application to the inverse QSAR. In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies - BIOIN FORMATICS, pages 144–151. INSTICC, SciTe...

    2022

  44. [52]

    J. Zhu, C. Wang, A. Shurbevski, H. Nagamochi, and T. Akutsu. A novel method for inference of chemical compounds of cycle index two with desired properties ba sed on artificial neural networks and integer programming. Algorithms, 13(5):124, 2020. 2LMM GNN: September 23, 2025 23 ...

    2020

  45. [53]

    In the four instances Ii,i = 1, 2, 3, 4, the following specifications in ( σint,σ ce) are common. Set Λ := Λ(π) for a given property π∈{ Homo, Lumo, Gap}, set Λ int dg to be the set of all possible symbols in Λ × [1, 4] that appear in the data set Dπ and set Γ int to be the set...

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.