NGTM: Substructure-based Neural Graph Topic Model for Interpretable Graph Generation

Dazhong Shen; Ying Sun; Yuanxin Zhuang

REVIEW 5 major objections 5 minor 64 references

NGTM: Substructure-based Neural Graph Topic Model for Interpretable Graph Generation

T0 review · 5 major / 5 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 NGTM models graph generation as sampling from latent substructure topics, claiming competitive fidelity plus explicit interpretability and fine-grained control.

desk verdict Novel topic-based graph generation, but the biological control claim is unsupported because the model generates only unlabeled adjacency matrices. read the letter →

arxiv 2507.13133 v1 pith:EWF7XLIQ submitted 2025-07-17 cs.LG

Yuanxin Zhuang , Dazhong Shen , Ying Sun This is my paper

classification cs.LG

keywords neuraltopicmodelgraphgenerationinterpretablesubstructuretopicscontrollablevariationalautoencodermolecularmodeling

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 proposes NGTM, a neural graph topic model that treats graph generation as a three-stage process: sample a topic mixture, sample substructures from topic-specific distributions, and assemble them under a global structural constraint. The central claim is that this transparent pipeline matches state-of-the-art black-box generators in quality, as measured by MMD and F1 metrics on four datasets, while uniquely making the generative decisions interpretable and adjustable. If the claim holds, users could steer generated molecular graphs toward desired structural and biological properties simply by changing topic proportions, without the usual fidelity loss. The paper shows learned topics align with known chemical classes and that topic manipulation shifts predicted carcinogenicity in the expected direction.

What carries the argument

The machinery is a conditional variational autoencoder with topic-structured priors and a soft substructure-assembly operator. A topic-proportion vector $\theta = \mathrm{softmax}(z_\theta)$ is sampled from a Gaussian; for each of W substructures, a topic $c_w$ is drawn multinomially and a substructure embedding $z_w$ is sampled from the selected topic's Gaussian $\mathcal{N}(\mu_k,\sigma_k)$; the structure decoder maps $z_w$ into a subgraph adjacency matrix. Assembly uses a mapping network that produces soft assignment matrices $m_w$, updating the graph as $A_w = A_{w-1} + m_w^\top A_{s_w} m_w$, conditioned on a global structure vector $g$. Training maximizes an ELBO whose reconstruction term is the micro-macro loss from GraphVAE-MM, plus KL divergences and an orthogonality regularizer that keeps topic means apart.

What would settle it

Take one generated or training graph, fix the learned topic and global variables, and run NGTM's forward pass on the same graph with two different node orderings; if the micro-macro reconstruction loss differs by more than numerical precision, the permutation-invariance claim is false and the model's assembly training is order-dependent. A cheaper check is to retrain NGTM on the same data with randomly permuted node orderings and see whether the reported MMD/F1 scores degrade.

Watch

Extended reading notes

Core claim

NGTM's core discovery is that a topic-model inductive bias can be embedded in a graph variational autoencoder without sacrificing fidelity. Graphs are represented as mixtures of K latent topics; each topic gives a Gaussian distribution over substructure embeddings, and sampled embeddings are decoded into adjacency matrices of subgraphs. A global structural vector g guides soft mapping matrices that sequentially merge substructures into the full graph. The paper reports that NGTM rivals or outperforms baselines including GraphRNN, GRAN, BiGG, DiGress, G2PT, and ConStruct across MUTAG, Lobster, PTC, and ogbg-molbbbp, and demonstrates that learned topics correspond to interpretable motifs—chains, rings, fused cycles—whose weight manipulation changes both structural statistics and the probability a generated graph is classified as carcinogenic.

Load-bearing premise

The load-bearing premise is that the reconstruction loss is insensitive to the order nodes are listed in even when substructures are merged through learned soft maps; the paper states this invariance but offers no proof or ablation, so if the loss does shift under node reordering the training signal for assembly would be misaligned and the reported generation quality could rest on a false assumption.

Editorial extensions

If this is right

If the central claim is right, interpretable graph generation need not trade fidelity: a topic-structured model can sit alongside black-box autoregressive and diffusion baselines on standard benchmarks.
Learned topics can serve as a semantic control panel: increasing a topic's weight produces predictable changes in diameter, modularity, density, and class probability, so users can steer generation toward target structural profiles.
Because substructures are generated from continuous topic-conditioned distributions rather than a fixed fragment vocabulary, the model can produce novel as well as known motifs, extending substructure-based molecular generation beyond predefined building blocks.
The ablation results imply global structural guidance and sequential assembly are necessary components; removing either degrades realism, which suggests that interpretable compositional generation relies on global context, not just local motif selection.

Reading between the lines

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

The topic-model analogy suggests a natural extension the paper leaves implicit: hierarchical topics, where substructures themselves are composed of smaller motifs, could give multi-scale interpretability and is directly testable by stacking NGTM on substructure-level topic mixtures.
Because the paper only measures controllability by post-hoc class shifts, a stronger test would be to train NGTM jointly with a property predictor or an explicit constraint on topic proportions, turning its demonstrated correlations into a goal-directed generation optimization.
The absence of a quantitative topic-coherence metric means one could test whether the learned topics are more than visually plausible by measuring within-topic substructure overlap and cross-topic separation in embedding space.
If the permutation-invariance premise fails, a likely fix is graph-matching or set-pooling during assembly, which would preserve the topic narrative while making the reconstruction objective genuinely order-free.

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.

Desk Editor's Note

read the letter

Colleague — quick take on 2507.13133. The core idea is genuinely novel: casting graph generation as a topic model over substructures, where topics are distributions over structural motifs and graphs are assembled from sampled substructures under a global structural variable. I haven't seen that exact combination. The three-stage generative narrative (topic mixture, substructure selection, assembly) is transparent, and the authors actually lean into interpretability rather than treating it as an afterthought. The class-wise structural comparison in Figure 3 is a thoughtful evaluation, and the topic manipulation experiments in Figures 5–7 show coherent internal structure: changing topic weights shifts metrics and classifier outputs in consistent directions. That is real evidence the topics carry semantics.

The soft spots are not trivial. The most serious: the model operates on adjacency matrices only. No node or edge attributes are generated anywhere in the method. Yet the PTC class-probability experiment feeds generated graphs to a GIN classifier trained on real labeled molecules. The paper never specifies how atom types are assigned to the unlabeled generated graphs. If the assignment is arbitrary or heuristic, the classifier outputs in Figure 7 don't reflect biological activity. The abstract's promise to 'induce biological properties' is therefore unsupported as written. The reader's stress-test concern is on point.

There's also a direct hyperparameter contradiction: Section 2.2.3 says K=10, W=30, while Appendix Table 2 lists K=5, W=20 for every dataset, including PTC. The topic analysis uses K=5, but it's unclear which configuration produced Table 1. No code, no error bars, no seeds, so the competitive numbers are hard to verify.

The permutation-invariance claim for the micro-macro loss is asserted without proof or ablation. The soft mapping matrix assembly could introduce node-ordering biases. That's a real unknown, though the empirical results suggest the model works well enough in practice.

The phrase 'causal relationship' in Section 2.4.2 overreaches; they're perturbing model inputs, not doing causal inference. A wording fix.

Bottom line: novel, well-motivated, and the internal experiments are credible. But the biological-control claim depends on an unstated atom-type assignment, and the reproducibility issues are easy to fix. I would send it to review, because a serious referee could help the authors close the gap, and the topic-modeling direction is worth exploring. Would I cite it now? Not yet. Reading-group? Maybe — it's a good discussion piece for graph generation and interpretability.

Referee Report

5 major / 5 minor

Summary. The paper proposes NGTM, a conditional VAE-style generative model in which graphs are represented as mixtures of K latent topics, each topic defining a Gaussian distribution over substructure embeddings. Generation proceeds by sampling a topic proportion vector θ and a global structural vector g, then sampling W substructures from topic-conditioned distributions and assembling them into an adjacency matrix through learned soft mapping matrices. The manuscript reports generation quality on MUTAG, Lobster, PTC, and Ogbg-molbbbp, claims parity or superiority over a range of graph-generation baselines, and supports interpretability through topic visualizations, topic-weight manipulation experiments, and a PTC class-probability study using a GIN classifier. The central claim is that NGTM matches state-of-the-art generation quality while providing explicit local and global interpretability and controllability.

Significance. If the central claims hold, NGTM would be a genuinely useful contribution: it is one of few graph generators whose building blocks and high-level semantic variables are visible in the generative process, and the topic-weight manipulation experiments offer a concrete mechanism for steering structural statistics. The idea of importing neural topic modeling into generative graph modeling is original and well-motivated, and the paper includes a comparison with several strong baselines. The significance is conditional, however, on two unresolved issues: the method as described generates only unlabeled adjacency matrices, which undercuts the biological-property claims, and the quantitative evaluation is presented without error bars, multiple seeds, or code, so the competitive-quality claims are not yet verifiable. The paper does not provide machine-checked proofs or reproducibility artifacts; its strengths at this stage are conceptual and qualitative.

major comments (5)

[§2.2.3, Appendix Table 2, §2.4] The experimental configuration is internally inconsistent. Section 2.2.3 states 'We fix the number of topics K = 10, the number of sampled substructures W = 30, and set the substructure size n to the average number of nodes in each dataset,' and the sentence is then repeated verbatim. Appendix Table 2 reports K = 5, W = 20, and n = 8/10/20/20 for the four datasets, and Section 2.4 says that a model trained with five latent topics was used for the topic analysis. These choices are not interchangeable for interpreting the parameter-sensitivity study in Figure 2, the main results in Table 1, or the topic-manipulation results in Figures 5 and 7. Please report the exact hyperparameters used for each experiment and reconcile the contradiction.
[§2.4.3, §4.2, Figure 7] The PTC class-probability experiment is not supported by the method as described. The generative process in Section 4.2 produces only adjacency matrices, with updates A_w = A_{w-1} + m_w^T A_s m_w; there is no node or edge attribute distribution anywhere in the generative model, the loss, or the assembly procedure. A GIN classifier trained on real PTC graphs uses atom and bond features, yet the manuscript never states how atom types are assigned to NGTM-generated graphs before classification. Without that step, the class probabilities in Figure 7 are not interpretable as biological activity predictions, and the abstract's promise to 'induce biological properties through topic-level adjustments' is unsubstantiated. The authors should either extend the model to generate labeled molecular graphs and evaluate under that protocol, or explicitly restrict the claim to structural surrogates and revise the corresponding statements.
[§4.3.1] The statement that the micro-macro reconstruction loss 'ensures permutation invariance and improves robustness to node ordering' is unsupported. The loss is adopted from GraphVAE-MM, but in NGTM the reconstruction must be evaluated through the soft mapping matrices m_w and the update A_w = A_{w-1} + m_w^T A_s m_w. Whether this composition is invariant to permutations of the nodes within a substructure or to the ordering of the W substructures is a nontrivial property that is neither proved nor ablated. If the training objective is not aligned with the actual assembly operation, the reconstruction signal could be misdirected. Please provide an analysis or an ablation, and if the invariance holds only approximately, state the limitation explicitly.
[Table 1, §2.3.1, §2.2] The quantitative evaluation is not sufficiently robust to support the 'competitive generation quality' claim. Table 1 reports a single value per metric per model with no error bars, no multiple-seed statistics, and no significance tests; the text does not state how many random seeds were used. Metrics such as MMD and F1 are noisy on datasets of the size of PTC (344 graphs), and the only reference point is the 50/50 split row. In addition, no code is provided, so the reader cannot verify the pipeline or the ablation variants. Please report means and standard deviations over at least three seeds, clearly state whether baseline numbers were re-run or taken from prior papers, and provide code or a detailed reproducibility statement.
[§2.3.2, Figure 2] The parameter-sensitivity experiment is difficult to interpret given the hyperparameter contradiction. Figure 2 varies K from 1 to 30 and W from 20 to 50, but the main configuration is stated as K = 10, W = 30 in Section 2.2.3 and as K = 5, W = 20 in Appendix Table 2. It is unclear which of the other hyperparameters (α_MM, β_KL, etc.) were held fixed during the sensitivity runs and whether those values came from Table 2 or from Section 2.2.3. Please specify the sensitivity protocol precisely so that the claimed trends are reproducible.

minor comments (5)

[§2.2.3] The sentence fixing K = 10, W = 30, and n is duplicated verbatim; please remove the duplicate.
[Appendix A.1] The appendix refers to 'the IToG framework'; this should read 'the NGTM framework.'
[Table 1] The row labeled '50/50 split' should be clearly marked as a reference or ideal-score row, with a footnote explaining that it follows O'Bray et al.; as presented, it could be mistaken for another baseline model.
[§2.4.1, Figure 4] The 'decorated molecules' in Figure 4 are obtained by decorating sampled skeletons using PTC hypotheses; please clarify whether this decoration step is automated or manually curated, since the strength of the topic-semantics claim depends on the objectivity of this procedure.
[§6] The Data Availability section lists datasets but not code; please add a code availability statement or explicitly state that code will be released.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: NGTM's topic-conditioned VAE is self-contained, and its interpretability claims rest on post hoc empirical probes rather than on the model's own fitted quantities.

full rationale

The paper's derivation chain is a standard conditional VAE: a topic proportion θ and global variable g are sampled from priors, substructure latents are decoded to adjacency matrices, and assembly is performed via learned soft mapping matrices (Sec. 4.2). The reconstruction loss is the micro-macro loss borrowed from GraphVAE-MM; no equation in the paper defines a prediction in terms of the same fitted quantity it claims to predict. The topic-manipulation experiments (Secs. 2.4.2–2.4.3) are post hoc validations, not derivations: the model is trained without PTC labels, so the observed correlations between topic weights, structural metrics, and GIN classifier outputs are emergent rather than forced by the loss. The GIN classifier is an external probe trained on the same dataset, which is a standard and legitimate evaluation design; it does not make the claim circular, though the paper would be stronger if it specified how atom types are supplied to the classifier for unlabeled generated graphs. There are no load-bearing self-citations: the reference list contains no works by the present authors, and the only imported component (the micro-macro loss, ref. 26) is used as a loss term, not as a premise that defines the target result. The 'first framework' novelty statement is a literature claim, not a circular argument. Overall, no step reduces by construction to its inputs.

Assumptions & free parameters 8 free parameters · 5 assumptions · 2 invented entities

The central claim depends on the assumptions that the reconstruction loss is a faithful surrogate for likelihood, that graphs decompose into a fixed number of substructures, and that the learned topics carry semantic meaning. The free parameters K, W, n and the loss weights are chosen by hand or tuned on validation data, with the main text and appendix contradicting each other on K and W.

free parameters (8)

Number of topics K = 5 in Appendix Table 2; 10 claimed in Section 2.2.3
Chosen by hand; the main text says K=10 while the appendix lists K=5 for all datasets. This directly affects semantic granularity and all reported quality metrics.
Number of substructures W = 20 in Appendix Table 2; 30 claimed in Section 2.2.3
Chosen by hand; controls how many substructures are sampled per graph. The discrepancy between main text and appendix affects the reproducibility of every experiment.
Substructure node size n = 8, 10, 20, 20 for MUTAG, Lobster, PTC, Ogbg-molbbbp
Set per dataset, described in Section 2.2.3 as the average number of nodes, but the appendix values do not match the averages of all datasets.
Loss weight alpha_MM = 80, 20, 10, 200 per dataset
Tuned on validation set; balances the micro-macro reconstruction loss against the KL and orthogonality terms.
Loss weight beta_KL = 90, 10, 1, 50 per dataset
Tuned on validation set; weight for the topic proportion KL divergence.
Loss weight delta_KL = 15, 0.5, 1, 5 per dataset
Tuned on validation set; weight for the global structure variable KL divergence.
Loss weight gamma_KL = 10, 1, 1, 4 per dataset
Tuned on validation set; weight for the topic-specific substructure KL divergences.
Loss weight omega_ortho = 10, 1.5, 1, 5 per dataset
Tuned on validation set; weight for the topic orthogonality regularization.

assumptions (5)

domain assumption The micro-macro reconstruction loss is a valid surrogate for the negative log-likelihood of the graph under the model
Section 4.3.1 uses the micro-macro loss as the reconstruction term in the ELBO, but no explicit likelihood p(G|Z) is ever defined. The paper assumes a distance-based loss behaves like a log-likelihood for training.
domain assumption Graphs can be decomposed into a fixed number W of substructures of at most n nodes, each generated from a topic-specific Gaussian and assembled with a soft mapping matrix
The entire generative story in Section 2.1 and Section 4.2 relies on this decomposition. If real graphs cannot be represented this way, generation quality degrades and the interpretability claims lose meaning.
domain assumption The learned topic means and covariances in the latent space correspond to semantically coherent structural families
Section 2.4 interprets sampled substructures as chemical scaffolds and links topics to carcinogenicity. This semantic alignment is the paper's core interpretability claim, assumed to arise from the CVAE training rather than being guaranteed by any formal property.
domain assumption The softmax-normalized mapping matrix provides a differentiable, permutation-aware way to merge substructures into the current graph
Section 4.2 defines the assembly update A_w = A_{w-1} + m_w^T A_{s_w} m_w, but no proof is given that this soft matching correctly aligns node identities or preserves graph structure during training.
standard math Reparameterization trick and amortized variational inference are applicable to this model
The paper uses standard reparameterization for Gaussian latents and factorized variational posteriors, which are routine tools in VAE literature.

invented entities (2)

Latent topics
purpose: Define Gaussian distributions over substructure latent embeddings, acting as semantic classes for graph motifs
Latent variables internal to the model; they have no falsifiable handle outside the model itself. Their 'meaning' is asserted through visualization and correlation with a PTC-trained classifier.
Global structural vector g
purpose: Captures high-level graph properties such as density and connectivity to guide substructure assembly
A latent variable sampled from a Gaussian; its effect is shown only through the model's own behavior, and no external verification is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of NGTM: Substructure-based Neural Graph Topic Model for Interpretable Graph Generation." pith.science (2026). https://pith.science/paper/EWF7XLIQ

@misc{pith2026250713133,
  author       = {Pith},
  title        = {Pith review of: NGTM: Substructure-based Neural Graph Topic Model for Interpretable Graph Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EWF7XLIQ}},
  note         = {Machine review of arXiv:2507.13133}
}

read the original abstract

Graph generation plays a pivotal role across numerous domains, including molecular design and knowledge graph construction. Although existing methods achieve considerable success in generating realistic graphs, their interpretability remains limited, often obscuring the rationale behind structural decisions. To address this challenge, we propose the Neural Graph Topic Model (NGTM), a novel generative framework inspired by topic modeling in natural language processing. NGTM represents graphs as mixtures of latent topics, each defining a distribution over semantically meaningful substructures, which facilitates explicit interpretability at both local and global scales. The generation process transparently integrates these topic distributions with a global structural variable, enabling clear semantic tracing of each generated graph. Experiments demonstrate that NGTM achieves competitive generation quality while uniquely enabling fine-grained control and interpretability, allowing users to tune structural features or induce biological properties through topic-level adjustments.

Figures

Figures reproduced from arXiv: 2507.13133 by the authors.

**Figure 2.** Impact of key parameters on the PTC dataset. (a) Effect of substructure count (W). (b) Effect of topic count (K). (c) Effect of substructure node size (n). Notably, our evaluations show that NGTM is capable of robustly capturing both local structural motifs and global topological characteristics, essential for realistic graph synthesis. This is particularly significant in chemically relevant scenarios such as MUTAG,… view at source ↗

**Figure 3.** Comparison of Structural Metric Effect Sizes Between Real and Generated Graphs across Models [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Sampled substructure skeletons (top row) from each NGTM topic, and corresponding chemically plausible molecules (bottom row) generated by decorating the skeletons based on PTC dataset hypotheses. patterns observed in real data, closely matching both the direction and m…

**Figure 5.** Figure 5: Topic-wise impacts on structural metrics evaluated through systematic topic weight manipulation. 2.4.1 Topic Visualization We conduct a topic visualization analysis for a better understanding of the semantic meanings captured by the learned topics. We first randomly sa…

**Figure 6.** Figure 6: Structural changes induced by progressively increasing the contribution of Topic 1 and Topic 3 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Effect of progressively increasing individual topic weights on the class probability of generated graphs in the PTC dataset. Solid lines represent Class 0 (non-carcinogenic) and dashed lines represent Class 1 (carcinogenic). Figures 6(a) and 6(b) illustrate concrete ex…

**Figure 8.** Figure 8: Interpretable Comparison of Carcinogenic and Non-carcinogenic Graphs Generated by NGTM. 2.5 Case Study: Interpretable Comparison of Carcinogenic and Non-carcinogenic Graphs [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

**Figure 9.** Figure 9: The Overview of NGTM Architecture. Decoder Module: Each sampled substructure vector z k=cw is concatenated with its topic one-hot vector and decoded as follows: sw = StructureDecoder(z k=cw ⊕ok=cw ). This conditioning ensures that each generated substructure is semanti…

**Figure 10.** Figure 10: shows representative graphs generated by different models across four datasets, alongside randomly selected test samples. For each model, 20 samples are generated, and visually representative examples are shown for comparison. On the Lobster dataset, NGTM is able to g…

**Figure 11.** Figure 11: Effect of targeted topic weight manipulation on structural properties of generated graphs. 17/17 [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 48 canonical work pages

[1]

Leveraging domain motif assembler for multi-objective, multi-domain and explainable molecular design (2025)

Tian, Y .et al. Leveraging domain motif assembler for multi-objective, multi-domain and explainable molecular design (2025)

work page 2025
[2]

& Cao, Y

Wu, W., Wen, C., Yuan, Q., Chen, Q. & Cao, Y . Construction and application of knowledge graph for construction accidents based on deep learning. Engineering, construction and architectural management 32, 1097–1121 (2025)

work page 2025
[3]

& Sperduti, A

Rigoni, D., Navarin, N. & Sperduti, A. Rgcvae: Relational graph conditioned variational autoencoder for molecule design. Machine Learning 114, 47 (2025)

work page 2025
[4]

& Gautam, R

Kumar, A., Hora, H., Rohilla, A., Kumar, P. & Gautam, R. Explainable artificial intelligence (xai) for healthcare: Enhancing transparency and trust. In International Conference on Cognitive Computing and Cyber Physical Systems , 295–308 (Springer, 2025)

work page 2025
[5]

Zheng, L. et al. Application scenario-oriented molecule generation platform developed for drug discovery. Methods 222, 112–121 (2024)

work page 2024
[6]

& Kim, B

Doshi-Velez, F. & Kim, B. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608 (2017)

arXiv 2017
[7]

& M¨uller, H

Holzinger, A., Langs, G., Denk, H., Zatloukal, K. & M¨uller, H. Causability and explainability of artificial intelligence in medicine. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 9, e1312 (2019)

work page 2019
[8]

& Guan, C

Tjoa, E. & Guan, C. A survey on explainable artificial intelligence (xai): Toward medical xai. IEEE transactions on neural networks and learning systems 32, 4793–4813 (2020)

work page 2020

Show all 64 references

[9]

& Zhao, L

Guo, X. & Zhao, L. A systematic survey on deep generative models for graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 5370–5390 (2022)

2022
[10]

Zhang, M. et al. A survey on graph diffusion models: Generative ai in science for molecule, protein and material. arXiv preprint arXiv:2304.01565 (2023)

2023 arXiv
[11]

Liu, C. et al. Generative diffusion models on graphs: Methods and applications. arXiv preprint arXiv:2302.02591 (2023)

2023 arXiv
[12]

Chen, X. et al. Graph generative pre-trained transformer. arXiv preprint arXiv:2501.01073 (2025)

2025 arXiv
[13]

& Zhu, F

Wang, J. & Zhu, F. Exselfrl: An exploration-inspired self-supervised reinforcement learning approach to molecular generation. Expert Systems with Applications 260, 125410 (2025)

2025
[14]

& Wang, Z

Hou, D., Gao, C., Li, X. & Wang, Z. Dag-aware variational autoencoder for social propagation graph generation. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 8508–8516 (2024)

2024
[15]

Xu, C., Deng, X., Lu, Y . & Yu, P. Generation of molecular conformations using generative adversarial neural networks. Digital Discovery 4, 161–171 (2025)

2025
[16]

& Frossard, P

Madeira, M., Vignac, C., Thanou, D. & Frossard, P. Generative modelling of structurally constrained graphs. Advances in Neural Information Processing Systems 37, 137218–137262 (2024)

2024
[17]

& Ohue, M

Kengkanna, A. & Ohue, M. Enhancing property and activity prediction and interpretation using multiple molecular graph representations with mmgx. Communications Chemistry 7, 74 (2024)

2024
[18]

& Jaakkola, T

Jin, W., Barzilay, R. & Jaakkola, T. Junction tree variational autoencoder for molecular graph generation. In International conference on machine learning, 2323–2332 (PMLR, 2018)

2018
[19]

& Liu, Y

Kong, X., Huang, W., Tan, Z. & Liu, Y . Molecule generation by principal subgraph mining and assembling.Advances in Neural Information Processing Systems 35, 2550–2563 (2022)

2022
[20]

& Jaakkola, T

Jin, W., Barzilay, R. & Jaakkola, T. Multi-objective molecule generation using interpretable substructures. In International conference on machine learning, 4849–4859 (PMLR, 2020). 13/17

2020
[21]

& Stuehmer, J

Stoehr, N., Yilmaz, E., Brockschmidt, M. & Stuehmer, J. Disentangling interpretable generative parameters of random and real-world graphs. arXiv preprint arXiv:1910.05639 (2019)

2019 arXiv
[22]

Guo, X. et al. Interpretable deep graph generation with node-edge co-disentanglement. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 1697–1707 (2020)

2020
[23]

M., Ng, A

Blei, D. M., Ng, A. Y . & Jordan, M. I. Latent dirichlet allocation. Journal of machine Learning research 3, 993–1022 (2003)

2003
[24]

& Yan, X

Sohn, K., Lee, H. & Yan, X. Learning structured output representation using deep conditional generative models.Advances in neural information processing systems 28 (2015)

2015
[25]

& Komodakis, N

Simonovsky, M. & Komodakis, N. Graphvae: Towards generation of small graphs using variational autoencoders. Le Centre pour la Communication Scientifique Directe - HAL - Diderot,Le Centre pour la Communication Scientifique Directe - HAL - Diderot (2018)

2018
[26]

Zahirnia, K., Schulte, O., Naddaf, P. & Li, K. Micro and macro level graph modeling for graph variational auto-encoders (2022)

2022
[27]

& Leskovec, J

You, J., Ying, R., Ren, X., Hamilton, W. & Leskovec, J. Graphrnn: Generating realistic graphs with deep auto-regressive models. International Conference on Machine Learning,International Conference on Machine Learning (2018)

2018
[28]

Liao, R. et al. Efficient graph generation with graph recurrent attention networks. Advances in neural information processing systems 32 (2019)

2019
[29]

& Schuurmans, D

Dai, H., Nazi, A., Li, Y ., Dai, B. & Schuurmans, D. Scalable deep generative modeling for sparse graphs.International Conference on Machine Learning,International Conference on Machine Learning (2020)

2020
[30]

Vignac, C. et al. Digress: Discrete denoising diffusion for graph generation. arXiv preprint arXiv:2209.14734 (2022)

2022 arXiv
[31]

Golomb, S. W. Polyominoes: puzzles, patterns, problems, and packings, vol. 16 (Princeton University Press, 1996)

1996
[32]

K., Lopez de Compadre, R

Debnath, A. K., Lopez de Compadre, R. L., Debnath, G., Shusterman, A. J. & Hansch, C. Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds. correlation with molecular orbital energies and hydrophobicity. Journal of medicinal chemistry 34, 78...

1991
[33]

D., Kramer, S

Toivonen, H., Srinivasan, A., King, R. D., Kramer, S. & Helma, C. Statistical evaluation of the predictive toxicology challenge 2000–2001. Bioinformatics 19, 1183–1193 (2003)

2003
[34]

& Jegelka, S

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

2018 arXiv
[35]

Hu, W. et al. Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems 33, 22118–22133 (2020)

2020
[36]

& Taylor, G

Thompson, R., Knyazev, B., Ghalebi, E., Kim, J. & Taylor, G. W. On evaluation metrics for graph generative models. arXiv preprint arXiv:2201.09871 (2022)

2022 arXiv
[37]

& Borgwardt, K

O’Bray, L., Horn, M., Rieck, B. & Borgwardt, K. Evaluation metrics for graph generative models: Problems, pitfalls, and practical solutions. Learning,Learning (2021)

2021
[38]

& Cook, D

Gonzalez, J., Holder, L. & Cook, D. J. Application of graph-based concept learning to the predictive toxicology domain. In Proceedings of the Predictive Toxicology Challenge Workshop(2001)

2001
[39]

Swamidass, S. J. et al. Kernels for small molecules and the prediction of mutagenicity, toxicity and anti-cancer activity. In ISMB (Supplement of Bioinformatics), 359–368 (2005)

2005
[40]

K., Kuznetsov, S

Blinova, V ., Dobrynin, D., Finn, V . K., Kuznetsov, S. O. & Pankratova, E. Toxicology analysis by means of the jsm-method. Bioinformatics 19, 1201–1207 (2003)

2003
[41]

& Kipf, T

De Cao, N. & Kipf, T. Molgan: An implicit generative model for small molecular graphs. arXiv preprint arXiv:1805.11973 (2018)

2018 arXiv
[42]

& Zhang, W

Wang, Z., Wallace, C., Bifet, A., Yao, X. & Zhang, W. Fairness-aware graph generative adversarial networks. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 259–275 (Springer, 2023)

2023
[43]

Kong, L. et al. Autoregressive diffusion model for graph generation. In International conference on machine learning, 17391–17408 (PMLR, 2023)

2023
[44]

Bu, J., Mehrab, K. S. & Karpatne, A. Let there be order: Rethinking ordering in autoregressive graph generation. arXiv preprint arXiv:2305.15562 (2023). 14/17

2023 arXiv
[45]

Luo, Y ., Yan, K. & Ji, S. Graphdf: A discrete flow model for molecular graph generation. InInternational conference on machine learning, 7192–7203 (PMLR, 2021)

2021
[46]

& Polykovskiy, D

Kuznetsov, M. & Polykovskiy, D. Molgrow: A graph normalizing flow for hierarchical molecular generation. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, 8226–8234 (2021)

2021
[47]

& Kim, W

Cho, H., Jeong, M., Jeon, S., Ahn, S. & Kim, W. H. Multi-resolution spectral coherence for graph generation with score-based diffusion. Advances in Neural Information Processing Systems 36 (2024)

2024
[48]

& Cosmo, L

Minello, G., Bicciato, A., Rossi, L., Torsello, A. & Cosmo, L. Graph generation via spectral diffusion. arXiv preprint arXiv:2402.18974 (2024)

2024 arXiv
[49]

& Wang, F

Zang, C. & Wang, F. Moflow: an invertible flow model for generating molecular graphs. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 617–626 (2020)

2020
[50]

Brockschmidt, M., Allamanis, M., Gaunt, A. L. & Polozov, O. Generative code modeling with graphs. arXiv preprint arXiv:1805.08490 (2018)

2018 arXiv
[51]

& Song, L

Dai, H., Tian, Y ., Dai, B., Skiena, S. & Song, L. Syntax-directed variational autoencoder for structured data.arXiv preprint arXiv:1802.08786 (2018)

2018 arXiv
[52]

Li, J. et al. Dirichlet graph variational autoencoder. Advances in Neural Information Processing Systems 33, 5274–5283 (2020)

2020
[53]

& Zhao, L

Du, Y ., Guo, X., Shehu, A. & Zhao, L. Interpretable molecular graph generation via monotonic constraints. InProceedings of the 2022 SIAM International Conference on Data Mining (SDM), 73–81 (SIAM, 2022)

2022
[54]

Unsupervised learning by probabilistic latent semantic analysis

Hofmann, T. Unsupervised learning by probabilistic latent semantic analysis. Machine learning 42, 177–196 (2001)

2001
[55]

& Seung, H

Lee, D. & Seung, H. S. Algorithms for non-negative matrix factorization. Advances in neural information processing systems 13 (2000)

2000
[56]

& Blunsom, P

Miao, Y ., Yu, L. & Blunsom, P. Neural variational inference for text processing. InInternational conference on machine learning, 1727–1736 (PMLR, 2016)

2016
[57]

& Blunsom, P

Miao, Y ., Grefenstette, E. & Blunsom, P. Discovering discrete latent topics with neural variational inference. In International conference on machine learning, 2410–2419 (PMLR, 2017)

2017
[58]

& Hovy, D

Bianchi, F., Terragni, S. & Hovy, D. Pre-training is a hot topic: Contextualized document embeddings improve topic coherence. arXiv preprint arXiv:2004.03974 (2020)

2020 arXiv
[59]

& Lin, W

Long, Q., Jin, Y ., Song, G., Li, Y . & Lin, W. Graph structural-topic neural network. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 1065–1073 (2020)

2020
[60]

& Nie, J.-Y

Xie, Q., Zhu, Y ., Huang, J., Du, P. & Nie, J.-Y . Graph neural collaborative topic model for citation recommendation.ACM Transactions on Information Systems (TOIS) 40, 1–30 (2021)

2021
[61]

K., Raiko, T., Maaløe, L., Sønderby, S

Sønderby, C. K., Raiko, T., Maaløe, L., Sønderby, S. K. & Winther, O. Ladder variational autoencoders. Advances in neural information processing systems 29 (2016)

2016
[62]

Kingma, D. P. & Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

2013 arXiv
[63]

M., Kucukelbir, A

Blei, D. M., Kucukelbir, A. & McAuliffe, J. D. Variational inference: A review for statisticians. Journal of the American statistical Association 112, 859–877 (2017)

2017
[64]

L., Kiros, J

Ba, J. L., Kiros, J. R. & Hinton, G. E. Layer normalization. arXiv preprint arXiv:1607.06450 (2016). 15/17 A Appendix A.1 Setup and Hyperparameters The IToG framework employs three encoders, each utilizing a two-layer Graph Convolutional Network (GCN) followed by a graph-level...

2016 arXiv

Pith tools

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

[1] [1]

Leveraging domain motif assembler for multi-objective, multi-domain and explainable molecular design (2025)

Tian, Y .et al. Leveraging domain motif assembler for multi-objective, multi-domain and explainable molecular design (2025)

work page 2025

[2] [2]

& Cao, Y

Wu, W., Wen, C., Yuan, Q., Chen, Q. & Cao, Y . Construction and application of knowledge graph for construction accidents based on deep learning. Engineering, construction and architectural management 32, 1097–1121 (2025)

work page 2025

[3] [3]

& Sperduti, A

Rigoni, D., Navarin, N. & Sperduti, A. Rgcvae: Relational graph conditioned variational autoencoder for molecule design. Machine Learning 114, 47 (2025)

work page 2025

[4] [4]

& Gautam, R

Kumar, A., Hora, H., Rohilla, A., Kumar, P. & Gautam, R. Explainable artificial intelligence (xai) for healthcare: Enhancing transparency and trust. In International Conference on Cognitive Computing and Cyber Physical Systems , 295–308 (Springer, 2025)

work page 2025

[5] [5]

Zheng, L. et al. Application scenario-oriented molecule generation platform developed for drug discovery. Methods 222, 112–121 (2024)

work page 2024

[6] [6]

& Kim, B

Doshi-Velez, F. & Kim, B. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608 (2017)

arXiv 2017

[7] [7]

& M¨uller, H

Holzinger, A., Langs, G., Denk, H., Zatloukal, K. & M¨uller, H. Causability and explainability of artificial intelligence in medicine. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 9, e1312 (2019)

work page 2019

[8] [8]

& Guan, C

Tjoa, E. & Guan, C. A survey on explainable artificial intelligence (xai): Toward medical xai. IEEE transactions on neural networks and learning systems 32, 4793–4813 (2020)

work page 2020