Pith. sign in

REVIEW 2 major objections 6 minor 37 references

A collaborative constrained graph diffusion model for the generation of realistic synthetic molecules

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A constrained graph diffusion model achieves 100% valid molecule generation by swapping bonds instead of adding noise.

desk verdict A genuinely constrained diffusion method with a real benchmarking caveat: the formula is drawn from the training set, which inflates the distribution-matching claims. read the letter →

arxiv 2505.16365 v1 pith:3XLAAQG5 submitted 2025-05-22 cs.LG cs.AIphysics.comp-phq-bio.QM

Manuel Ruiz-Botella , Marta Sales-Pardo , Roger Guimerà This is my paper
classification cs.LGcs.AIphysics.comp-phq-bio.QM
keywords moleculargraphgenerationdiscretediffusionmodelconstraineddoubleedgeswappingvalenceconstraintsGuacaMolbenchmarkpropertydistributionsTuring-liketest
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

CoCoGraph is a discrete graph diffusion model for molecule generation whose central mechanism is to noise and denoise molecules by double edge swapping, which preserves each atom's valence and the molecular formula at every step. The paper claims this makes every generated molecule chemically valid by construction rather than by learned filtering, and it reports 100% validity on the GuacaMol benchmark, 99.9% uniqueness, and 98.5% novelty while using an order of magnitude fewer parameters than JTVAE and DiGress. It further claims that the generated molecules' distributions over 36 physicochemical properties more closely match real molecules than either baseline on most properties, and that the 534K-parameter BASE model already beats both baselines on GuacaMol's KL divergence. The efficiency enabled an 8.2 million molecule database and a Turing-like test in which trained chemists identified the real molecule only 62% of the time, near random chance. The paper's central claim is that high-fidelity molecular generation does not need to learn chemical validity; constraints in the generative process make validity cheap and make realistic exploration of chemical space feasible at scale.

What carries the argument

The load-bearing mechanism is valence-preserving double edge swapping: at each diffusion step the model deletes two existing bonds $(i,j)$ and $(k,l)$ and creates bonds $(i,k)$ and $(j,l)$, so each atom loses one bond and gains one, keeping the degree sequence, valence, and molecular formula unchanged. Iterating this swap drives the molecular graph toward the Molloy-Reed distribution, the maximum-entropy distribution over graphs with a fixed degree sequence. The denoiser is a small graph neural network with enhanced graph-isomorphism message-passing layers that scores all feasible quadruplets of bonds for the swap to revert, while a second time network estimates the normalized diffusion time of the current graph. During sampling, the estimated time replaces the true step counter, and the final molecule is chosen as the trajectory graph predicted to lie closest to time zero. Because the double edge swap guarantees validity by construction, the neural networks spend their capacity on structural patterns rather than on chemical rules.

What would settle it

Re-run the GuacaMol and 36-property comparisons with CoCoGraph sampling molecular formulas from a learned or uniform prior rather than from the database, leaving the baselines unchanged; if heavy atom count, molecular weight, exact mass, and valence electron distributions fall to parity with DiGress, the realism claim rests on formula inheritance rather than on the constrained diffusion itself.

Watch

Extended reading notes

Core claim

The central discovery is that chemical validity can be moved out of the model's parameters and into the diffusion dynamics. CoCoGraph never adds or removes atoms and never changes any atom's degree, because each noising step deletes two bonds AB and CD and forms AC and BD, so every intermediate and final graph has the same molecular formula and valence sequence as a real seed. The denoiser therefore only has to choose which double edge swap to undo, and a second time model predicts how far a graph is from a clean state so that sampling can stop at the trajectory point closest to a valid molecule. On the GuacaMol benchmark the paper reports 100% validity, 99.9% uniqueness, 98.5% novelty, and KL divergence scores of 96.0 for the BASE model and 96.7 for the fingerprint-enhanced FPS model, versus 92.6 for DiGress and 47.3 for JTVAE. Across 36 cheminformatics properties, CoCoGraph has a smaller Jensen-Shannon distance than DiGress on 23 properties and than JTVAE on 33 properties; the smaller BASE model still outperforms both comparators on most properties.

Load-bearing premise

The load-bearing premise is that sampling begins by drawing a molecular formula directly from the real database and the constrained diffusion never changes it; if formulas were sampled from the model's own prior instead, the reported realism advantage on size-related properties could shrink substantially.

Editorial extensions

If this is right

  • Every molecule output by the model satisfies valence constraints, so no rejection sampling or validity repair step is needed.
  • With 534K parameters the BASE model attains a GuacaMol KL divergence of 96.0, roughly halving DiGress's typical KL divergence value, implying that benchmark-level realism can be bought with a much smaller architecture.
  • The distributional improvements extend beyond the ten GuacaMol descriptors to 36 properties, with the largest gains in topological, electronic, and drug-likeness descriptors.
  • The same model can produce a database of roughly 7.6 million novel, unique, valid molecules, and expert chemists identify the real molecule only 62% overall, and at chance level for acyclic and predominantly aliphatic molecules.
  • Because the molecular formula and valence sequence are fixed throughout generation, the method can be combined with a separate formula sampler when the formula itself is part of the design space.

Reading between the lines

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

  • Editorial inference: because sampling starts from a molecular formula drawn from the real database and the diffusion never changes it, size-related descriptors such as heavy atom count, molecular weight, exact mass, and valence electron count are matched by construction; comparing against baselines conditioned on the same formula would isolate how much of the realism gain comes from the diffusion
  • Editorial inference: the rule of choosing the trajectory graph with the smallest predicted time acts as a learned self-consistency check; ablating this selection rule and comparing property distributions would reveal whether the time model mainly helps choose among denoising trajectories or also improves per-step predictions.
  • Editorial inference: because each swap preserves the degree sequence, the model can only rearrange bonds within a fixed valence pattern; structures requiring unusual valences or formula changes are out of reach unless a separate formula generator feeds it. A natural test is whether the 8.2 million molecule database covers chemical space beyond scaffolds that already share those formulas in the tra
  • Editorial inference: the $O(n^4)$ cost per diffusion step and the 5-to-70-atom cutoff limit the current implementation to small and medium molecules; scaling to larger systems will require sparse or approximate selection among candidate bond quadruplets.
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

2 major / 6 minor

Summary. The paper introduces CoCoGraph, a discrete graph diffusion model for molecules in which the noising and denoising processes use double edge swaps that preserve the molecular formula and degree sequence, guaranteeing chemical validity by construction. A second 'time model' is trained to predict the denoising progress and is used during sampling to condition the diffusion model and to select the most plausible graph along the trajectory. On the GuacaMol benchmark, the authors report 100% validity, high uniqueness and novelty, and a KL divergence score above DiGress and JTVAE. They also evaluate 36 physicochemical properties, report better Jensen-Shannon distances on 23 of 36 properties versus DiGress and 33 of 36 versus JTVAE (FPS model), and present a Turing-like test in which 102 experts identify real versus generated molecules with 62% accuracy. Code, model weights, and a database of 8.2 million generated molecules are made available.

Significance. The double-edge-swap constrained diffusion is a clean and credible mechanism for guaranteed-valid generation with a fixed molecular formula, and the collaborative time model is an interesting inference-time innovation. If the distributional claims survive a properly matched benchmark, the model could be a useful lightweight alternative for formula-conditional molecule generation. The open-sourced code, weights, and large generated database are valuable assets for the community. However, the main benchmark and property-distribution claims are currently overstated because the sampling protocol injects the empirical molecular-formula distribution directly into the generated molecules, making the comparison to unconstrained baselines not apples-to-apples.

major comments (2)
  1. [Methods, 'Sampling of new molecules'; Discussion] The sampling protocol states that 'a molecular formula is selected directly from the database,' and the double-edge-swap diffusion preserves the molecular formula and degree sequence throughout the trajectory. Every generated molecule therefore inherits its formula from the empirical training distribution, and any property that is a deterministic function of the formula (molecular weight, exact mass, heavy atom count, number of valence electrons, N-O count, and several chi/VSA descriptors) has a marginal that matches the training set by construction, independent of what the model learned. The largest reported improvements over DiGress occur on exactly these descriptors (e.g., Fig. 3a heavy atom count JS 0.016 vs 0.081; Fig. 3b valence electrons 0.017 vs 0.074). Thus the claims of 'beating the state-of-the-art' on GuacaMol (Table 1) and of producing distributions 'more closely matching real molecules' (Abstract) are not supported for unconstrained generation; they hold only for the conditional task of generating an isomer for a database-supplied formula. The Discussion flags the fixed formula as a limitation, but the benchmark conclusions are not qualified accordingly. The authors should either re-run the evaluation with molecular formulas drawn from a learned or population prior, or explicitly re-frame the contribution as formula-conditional generation and compare against formula-conditioned baselines.
  2. [Table ED1; Fig. 3k] The aggregate 36-property comparison is inflated by the same confound. Even setting aside formula-determined descriptors, Table ED1 shows that CoCoGraph is worse than DiGress on several structural descriptors (N-H/OH count 0.186 vs 0.026; H-bond donors 0.187 vs 0.025; aromatic rings 0.089 vs 0.020; max absolute partial charge 0.110 vs 0.075). The 23/36 and 33/36 win counts over DiGress and JTVAE should therefore be re-computed separately for formula-determined and structure-determined properties. A fair comparison also requires that comparators be conditioned on the same molecular-formula distribution (for example, by providing the formula as input or by sampling formulas from the same database), so that the comparison isolates the models' ability to generate plausible structures rather than their ability to reproduce the training formula distribution.
minor comments (6)
  1. [Abstract and main text] The phrase '8.2M million' (Abstract) and '8.2 millions' (Discussion) is redundant; use '8.2 million' consistently.
  2. [Main text] There are several typos: 'juntion tree VAE' should be 'junction tree VAE'; 'Kullback-Liebler' should be 'Kullback-Leibler'; 'constrains' should be 'constraints'; 'the number fo atoms' in the Fig. 4c caption should be 'the number of atoms'.
  3. [Methods, 'Molecular data processing' v. 'Model training'] The dataset size is inconsistent: 'Molecular data processing' reports that after filtering 'approximately 1.67 million molecules remained', but 'Model training' begins with 'Our training dataset consisted of 2.25 million molecules'. Please clarify which number was used for training.
  4. [Table 3 and 'Node-level features'] Table 3 lists 'Hydrogen neighbors' as a node feature, but the 'Node-level features' text enumerates only element type, cycle participation, non-hydrogen neighbors, and bridge count. Add the hydrogen-neighbor feature to the text or remove it from the table.
  5. [Fig. 5 and architecture text] The text says EnhancedGINE layers produce 124-dimensional node embeddings, while Fig. 5b shows a '1x62' tensor after the global mean pool in the time model. Please clarify the dimensions so the figure and text agree.
  6. [Fig. ED2 caption] The caption refers to 'CocoGraph' and 'CocoGraph FPS' while the rest of the paper uses 'CoCoGraph'; please standardize the spelling.

Circularity Check

1 steps flagged · score 6.0 of 10

CoCoGraph samples each molecular formula directly from the training database and the double-edge-swap diffusion preserves it, so molecular-weight, heavy-atom, and atom-count distributions match real molecules by construction; the reported GuacaMol and 36-property advantage over DiGress and JTVAE is therefore partly forced by input copying.

  1. self definitional [Main text, 'A collaborative constrained diffusion model...' paragraph; Methods, 'Sampling of new molecules']
    "First, we use a discrete process that involves double edge swapping and constrains each atom to always have the correct valence, maintaining other chemical properties such as molecular weight, number of atoms, number of bonds, and molecular formula. ... Initially, a molecular formula is selected directly from the database."

    The molecular formula is drawn from the training database and the double-edge-swap diffusion provably preserves the formula and degree sequence, so formula-determined properties (molecular weight, exact mass, heavy atom count, valence-electron count, N/O count) inherit the database marginals by construction rather than being learned. The GuacaMol KL score and the 36-property JS tables include these exact size/composition descriptors (e.g., MW JS 0.014 vs 0.078 for DiGress; heavy atoms JS 0.016 vs 0.081; valence electrons 0.017 vs 0.074), so the headline distributional advantage over DiGress and JTVAE partially reduces to copying the empirical formula distribution.

full rationale

The core validity guarantee and the graph-structure generation are independent achievements: 100% validity follows from the designed double-edge-swap constraint, and structural/topological matching (Bertz, aromatic rings, QED, Balaban) is genuinely learned, so not all comparisons are circular. However, the paper's broadest distributional claims—'beats the state-of-the-art on the most comprehensive existing benchmark' and 'distributions more closely matching real molecules'—are substantially inflated on size/composition descriptors because the sampling protocol fixes each generated molecule's formula to an empirical formula from the database. The paper itself concedes this design limitation in the Discussion. There is no load-bearing self-citation chain; the circularity is a construction/input-copying issue rather than a citation issue. Score 6 reflects partial circularity: several reported benchmark properties reduce by construction, while substantial independent content remains.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical particles, forces, or conserved quantities. Its assumptions are modeling choices: the definition of chemical validity as valence preservation, the use of real molecular formulas as fixed inputs, and the standard maximum-entropy guarantee for edge-swap random graphs. The main free parameters are sampling and training hyperparameters that are chosen by hand rather than derived.

free parameters (3)
  • Diffusion step count = about 25% of bonds
    Chosen by hand to determine how many double edge swaps fully randomize a molecule; this controls the noising depth and sampling length.
  • Sampling selection threshold = 95%, decremented by 5%
    Manual heuristic for choosing which double edge swap to apply during denoising; affects diversity and trajectory selection.
  • Loss weighting coefficients = not reported
    The three binary cross-entropy losses and the time-loss are described as weighted and masked, but the actual weights are not specified in the paper.
assumptions (3)
  • domain assumption Preserving each atom's degree (total bond order) under double edge swaps is sufficient to guarantee chemical validity.
    The paper equates validity with valence preservation plus connectivity; other chemical validity aspects such as aromaticity, formal charge, and stereochemistry are not enforced by the construction, as seen in the F_t feasibility conditions in Methods.
  • domain assumption A molecular formula may be taken directly from the training database and treated as a fixed input for generation.
    The sampling procedure selects a formula from the database and the diffusion process preserves it by construction, which drives the benchmark property distributions.
  • standard math Repeated double edge swaps converge to the Molloy-Reed maximum-entropy distribution over graphs with fixed degree sequence.
    The paper relies on classical results on degree-preserving random graph rewiring, citing references 33, 34, and 39.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A collaborative constrained graph diffusion model for the generation of realistic synthetic molecules." pith.science (2026). https://pith.science/paper/3XLAAQG5

@misc{pith2026250516365,
  author       = {Pith},
  title        = {Pith review of: A collaborative constrained graph diffusion model for the generation of realistic synthetic molecules},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3XLAAQG5}},
  note         = {Machine review of arXiv:2505.16365}
}
read the original abstract

Developing new molecular compounds is crucial to address pressing challenges, from health to environmental sustainability. However, exploring the molecular space to discover new molecules is difficult due to the vastness of the space. Here we introduce CoCoGraph, a collaborative and constrained graph diffusion model capable of generating molecules that are guaranteed to be chemically valid. Thanks to the constraints built into the model and to the collaborative mechanism, CoCoGraph outperforms state-of-the-art approaches on standard benchmarks while requiring up to an order of magnitude fewer parameters. Analysis of 36 chemical properties also demonstrates that CoCoGraph generates molecules with distributions more closely matching real molecules than current models. Leveraging the model's efficiency, we created a database of 8.2M million synthetically generated molecules and conducted a Turing-like test with organic chemistry experts to further assess the plausibility of the generated molecules, and potential biases and limitations of CoCoGraph.

Figures

Figures reproduced from arXiv: 2505.16365 by the authors.

Figure 1
Figure 1. Constrained collaborative graph diffusion model, CoCoGraph. a, Constrained diffusion process. We introduce noise in the molecular graph by swapping two chemical bonds at each step. We then train diffusion and time models to revert this process. b, Diffusion model. At each step, it receives molecular features and the timestep as an input and assigns a score to all possibilities of edge swaps. c, Time model. It receiv… view at source ↗
Figure 2
Figure 2. Performance comparison on GuacaMol benchmark properties. a-f, Distributions of six molecular properties: a, molecular weight; b, molecular logP; c, internal similarity; d, Bertz complexity index; e, number of aromatic rings; and f, number of H-bond donors. For each property, the distribution of values calculated for molecules generated by CoCoGraph (black line) is compared to that of the original molecules (green di… view at source ↗
Figure 3
Figure 3. Detailed performance comparison on a subset of 36 chemical properties. a-j, Distributions of ten molecular properties: a, heavy atom count; b, number of valence electrons; c, NOCount; d, Balaban’s J Index; e, number of H acceptors; f, ring count; g, topological polar surface area (TPSA); h, quantitative estimate of drug-likeness (QED); i, maximum absolute partial charge; and j, NHOHCount. For each property, the dist… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance in the Turing-like test. We assess the performance of participants in the Turing-like test by computing their accuracy at correctly identifying the original, non-generated molecule over all attempts. Error bars represent the standard error of the mean calcu…
Figure 5
Figure 5. Figure 5: Architecture of CoCoGraph components. a, The diffusion model processes the molecular graph through a sequence of EnhancedGINE layers, the embedding of pairs of nodes are concatenated with edge properties and processed through two feedforward modules to predict the prob…
Figure 6
Figure 6. Figure 6: Architecture of CoCoGraph FPS (fingerprint-enhanced) components. a, The diffusion model processes the molecular graph through a sequence of EnhancedGINE layers. Then, the embedding of pairs of nodes are concatenated with edge and graph properties and processed through …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 12 canonical work pages

  1. [2]

    & Ranganathan, R

    Menon, D. & Ranganathan, R. A generative approach to materials discovery, design, and optimization. ACS Omega7, 25958–25973, 10.1021/acsomega.2c03264 (2022)

    work page doi:10.1021/acsomega.2c03264 2022

  2. [3]

    Alkhatib, I. I. I., Albà, C. G., Darwish, A. S., Llovell, F . & Vega, L. F . Searching for sustainable refrigerants by bridging molecular modeling with machine learning. Ind. & Eng. Chem. Res. 61, 7414–7429, 10.1021/ acs.iecr.2c00719 (2022)

    work page 2022

  3. [4]

    & Wang, B

    Young, A., Röst, H. & Wang, B. Tandem mass spectrum prediction for small molecules using graph transformers (2024)

    work page 2024

  4. [5]

    Ackloo, S. et al. Cache (critical assessment of computational hit-finding experiments): A public–private partnership benchmarking initiative to enable the development of computational methods for hit-finding (2022)

    work page 2022

  5. [6]

    G., Madzhidov, T

    Polishchuk, P . G., Madzhidov, T . I. & Varnek, A. Estimation of the size of drug-like chemical space based on gdb-17 data. J. Comput. Mol. Des.27, 675–679, 10.1007/s10822-013-9672-4 (2013)

    doi:10.1007/s10822-013-9672-4 2013

  6. [7]

    Kutchukian, P . S. & Shakhnovich, E. I.De novodesign: balancing novelty and confined chemical space. Expert. Opin. on Drug Discov.5, 789–812, 10.1517/17460441.2010.497534 (2010)

    work page arXiv 2010

  7. [8]

    T ., Davies, D

    Butler, K. T ., Davies, D. W ., Cartwright, H., Isayev, O. & Walsh, A. Machine learning for molecular and materials science. Nature559, 547–555, 10.1038/s41586-018-0337-2 (2018)

    doi:10.1038/s41586-018-0337-2 2018

  8. [9]

    Reinvent 2.0: An ai tool for de novo drug design

    Blaschke, T .et al. Reinvent 2.0: An ai tool for de novo drug design. J. Chem. Inf. Model. 60, 5918–5922, 10.1021/acs.jcim.0c00915 (2020). 10.Bilodeau, C., Jin, W ., Jaakkola, T ., Barzilay , R. & Jensen, K. F . Generative models for molecular discovery: Recent advances and challenges. Wiley Interdiscip. Rev. Comput. Mol. Sci. 12, e1608, 10.1002/WCMS. 160...

    work page doi:10.1021/acs.jcim.0c00915 2020

Show all 37 references
  1. [11]

    & Jaakkola, T

    Jin, W ., Barzilay , R. & Jaakkola, T . Junction tree variational autoencoder for molecular graph generation (2019). 1802.04364

    2019 arXiv

  2. [12]

    Cao, N. D. & Kipf, T . Molgan: An implicit generative model for small molecular graphs (2022). 1805.11973. 13.Mercado, R. et al. Graph networks for molecular design (2020). 10.26434/chemrxiv.12843137.v1

    2022 arXiv

  3. [14]

    & Schuurmans, D

    Dai, H., Nazi, A., Li, Y., Dai, B. & Schuurmans, D. Scalable deep generative modeling for sparse graphs (2020). 2006.15502. 15.Liao, R. et al. Efficient graph generation with graph recurrent attention networks (2020). 1910.00760

    2020 arXiv

  4. [16]

    You, J., Ying, R., Ren, X., Hamilton, W . L. & Leskovec, J. Graphrnn: Generating realistic graphs with deep auto-regressive models (2018). 1802.08773

    2018 arXiv

  5. [17]

    & Hwang, S

    Lee, S., Jo, J. & Hwang, S. J. Exploring chemical space with score-based out-of-distribution generation (2023). 2206.07632

    2023 arXiv

  6. [18]

    A., Maheswaranathan, N

    Sohl-Dickstein, J., Weiss, E. A., Maheswaranathan, N. & Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics (2015). 1503.03585. 19/28 19.Ho, J., Jain, A. & Abbeel, P . Denoising diffusion probabilistic models (2020). 2006.11239

    2015 arXiv

  7. [20]

    Song, Y. et al. Score-based generative modeling through stochastic differential equations (2021). 2011. 13456

    2021

  8. [21]

    Yang, L. et al. Diffusion models: A comprehensive survey of methods and applications. ACM Comput. Surv. 56, 10.1145/3626235 (2023). 22.Ramesh, A. et al. Zero-shot text-to-image generation (2021). 2102.12092

    2023 arXiv

  9. [23]

    & Chen, M

    Ramesh, A., Dhariwal, P ., Nichol, A., Chu, C. & Chen, M. Hierarchical text-conditional image generation with clip latents (2022). 2204.06125

    2022 arXiv

  10. [24]

    D., Ho, J., Tarlow, D

    Austin, J., Johnson, D. D., Ho, J., Tarlow, D. & van den Berg, R. Structured denoising diffusion models in discrete state-spaces. ArXivabs/2107.03006(2021)

    2021 arXiv

  11. [25]

    G., Vignac, C

    Hoogeboom, E., Satorras, V . G., Vignac, C. & Welling, M. Equivariant diffusion for molecule generation in 3d (2022). 2203.17003

    2022 arXiv

  12. [26]

    & Hwang, S

    Jo, J., Lee, S. & Hwang, S. J. Score-based generative modeling of graphs via the system of stochastic differential equations (2022). 2202.02514

    2022 arXiv

  13. [27]

    K., Martinkus, K., Perraudin, N

    Haefeli, K. K., Martinkus, K., Perraudin, N. & Wattenhofer, R. Diffusion models for graphs benefit from discrete state spaces (2023). 2210.01549

    2023 arXiv

  14. [28]

    & Frossard, P

    Qin, Y., Vignac, C. & Frossard, P . Sparse training of discrete diffusion models for graph generation (2024). 2311.02142. 29.Vignac, C. et al. Digress: Discrete denoising diffusion for graph generation (2023). 2209.14734

    2024 arXiv

  15. [30]

    Huang, H., Sun, L., Du, B. & Lv, W . Conditional diffusion based on discrete graph structures for molecular graph generation (2023). 2301.00427

    2023 arXiv

  16. [31]

    Yang, L. et al. Graphusion: Latent diffusion for graph generation. IEEE Transactions on Knowl. Data Eng. 36, 6358–6369, 10.1109/TKDE.2024.3389783 (2024)

    2024

  17. [32]

    & Liu, S

    Du, Y ., Fu, T ., Sun, J. & Liu, S. Molgensurvey: A systematic survey in machine learning models for molecule design (2022). 2203.14500

    2022 arXiv

  18. [33]

    & Sneppen, K

    Maslov, S. & Sneppen, K. Specificity and stability in topology of protein networks. Science 296, 910–913, https://doi.org/10.1126/science.1065103 (2002)

    2002 doi

  19. [34]

    Milo, R., Kashtan, N., Itzkovitz, S., Newman, M. E. J. & Alon, U. On the uniform generation of random graphs with prescribed degree sequences, https://doi.org/10.48550/arXiv.cond-mat/0312028 (2004). cond-mat/0312028

    2004 doi arXiv

  20. [35]

    R., Mezei, T

    Kharel, S. R., Mezei, T . R., Chung, S., Erd˝os, P . L. & Toroczkai, Z. Degree-preserving network growth. Nat. Phys.18, 100–106 (2022)

    2022

  21. [36]

    & Frossard, P

    Madeira, M., Vignac, C., Thanou, D. & Frossard, P . Generative modelling of structurally constrained graphs (2024). 2406.17341

    2024 arXiv

  22. [37]

    D., Mathieu, E

    Fishman, N., Klarner, L., Bortoli, V . D., Mathieu, E. & Hutchinson, M. Diffusion models for constrained domains (2024). 2304.05364

    2024 arXiv

  23. [38]

    & de Bortoli, V

    Fishman, N., Klarner, L., Mathieu, E., Hutchinson, M. & de Bortoli, V . Metropolis sampling for constrained diffusion models (2023). 2307.05439

    2023 arXiv

  24. [39]

    & Reed, B

    Molloy, M. & Reed, B. A critical point for random graphs with a given degree sequence. Random Struct. & Algorithms6, 161–180, https://doi.org/10.1002/rsa.3240060204 (1995)

    1995 doi

  25. [40]

    Brown, N., Fiscato, M., Segler, M. H. & Vaucher, A. C. Guacamol: Benchmarking models for de novo molecular design. J. Chem. Inf. Model.59, 1096–1108, 10.1021/acs.jcim.8b00839 (2019). 20/28 41.Landrum, G. et al. RDKit: Open-source cheminformatics. https://www.rdkit.org (Zenodo, 2025)

    2019 doi

  26. [42]

    Artificial intelligence needs a scientific method-driven reset

    Nunes Amaral, L. Artificial intelligence needs a scientific method-driven reset. Nat. Phys. 20, 523–524, 10.1038/s41567-024-02403-5 (2024). 43.Ren, P . et al. A survey of deep active learning (2021). 2009.00236

    2024 arXiv

  27. [44]

    & Klimov, O

    Schulman, J., Wolski, F ., Dhariwal, P ., Radford, A. & Klimov, O. Proximal policy optimization algorithms (2017). 1707.06347

    2017 arXiv

  28. [45]

    & Coley, C

    Bohde, M., Manjrekar, M., Wang, R., Ji, S. & Coley, C. W . Diffms: Diffusion generation of molecules conditioned on mass spectra (2025). 2502.09571. 46.Hu, W . et al. Strategies for pre-training graph neural networks (2020). 1905.12265

    2025 arXiv

  29. [47]

    Morgan, H. L. The generation of a unique machine description for chemical structures—A technique developed at Chemical Abstracts Service. J. Chem. Documentation 5, 107–113, 10.1021/c160017a018 (1965). Acknowledgements This research was supported by project PID2022-142600NB-I00...

    1965 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.