Pith. sign in

REVIEW 4 major objections 4 minor 41 references

DeepScaffold: a comprehensive tool for scaffold-based de novo drug discovery using deep learning

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A scaffold-first graph generator promises exact scaffold preservation in every output molecule.

desk verdict A useful scaffold-growing generator with a credible evaluation story, but the advertised guarantee that outputs always contain the input scaffold is not actually enforced by the described action set. read the letter →

arxiv 1908.07209 v4 pith:5X3NLD32 submitted 2019-08-20 q-bio.QM cs.LG

Yibo Li , Jianxing Hu , Yanxing Wang , Jielong Zhou , Liangren Zhang , Zhenming Liu This is my paper · ORCID
classification q-bio.QMcs.LG
keywords scaffold-baseddrugdiscoverydenovomoleculedesigngraphgenerativemodelvariationalautoencodercyclicskeletonclassicalscaffoldpharmacophorequerymoleculardocking
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

DeepScaffold tries to show that de novo molecule design can be anchored to a user-chosen scaffold instead of starting from an empty graph. The model accepts three kinds of queries: cyclic skeletons (ring-only graphs), classical BM-scaffolds with atom and bond types, and scaffolds with extra constraints on side-chain properties. Its central promise is that every generated molecule contains the queried scaffold as an exact substructure, and that the growth rules learned from training data transfer to scaffolds never seen during training. That matters because scaffold-based design is a standard medicinal-chemistry workflow, and earlier generative models either did not guarantee scaffold preservation or handled only one scaffold type.

What carries the argument

The load-bearing mechanism is the sequential graph-expansion decoder. It starts from the molecular graph of the scaffold and at each step chooses one of three actions — append a new atom, connect two existing atoms, or stop — and no action in the described repertoire deletes or alters scaffold bonds, so the scaffold is meant to be an exact substructure of every completed molecule. Two auxiliary components extend the range of queries: a graph-convolutional variational autoencoder that maps cyclic skeletons to complete classical scaffolds by sampling atom and bond types, and a rule-based pharmacophore filter that keeps only generated molecules whose side chains satisfy user-specified size, hydrogen-bond-donor, and hydrogen-bond-acceptor constraints. The growth model itself is a 20-layer dense graph convolutional network with virtual bonds and virtual ring/linker nodes, trained by likelihood maximization with importance sampling.

What would settle it

Instrument the decoder to log all actions during sampling: if a 'connect two existing atoms' action ever joins two atoms of the input scaffold, or any scaffold bond is altered, the exact-substructure guarantee fails. A more direct test is a subgraph isomorphism check between the queried scaffold and each generated molecule; any output that does not contain the scaffold as an exact substructure disproves the claim.

Watch

Extended reading notes

Core claim

On its own terms, DeepScaffold's contribution is a generative architecture in which molecule construction is a graph-growing process seeded by the scaffold itself. A graph neural network is trained to extend the scaffold by appending atoms, connecting two existing atoms, or terminating, and the full molecule is built by repeated application of those actions. Because the design starts from the scaffold and never removes or changes scaffold atoms and bonds, the scaffold remains an exact substructure of every output molecule. For cyclic skeletons, a separate variational autoencoder first decorates the ring-only graph with atom and bond types to produce a classical scaffold, which then feeds the same growth process; side-chain queries on size, hydrogen-bond donors, and acceptors are applied as a post-filter. Across the test scaffolds the paper reports 98.9% average validity, outputs that are often more structurally diverse than the corresponding test molecules, and leave-one-out reproduction of known GPCR actives and drugs from privileged scaffolds.

Load-bearing premise

The exact-substructure guarantee rests on the assumption that the 'connect two existing atoms' action will never join two scaffold atoms or change a scaffold bond, yet no masking or constraint enforcing that is described.

Editorial extensions

If this is right

  • A chemist can hand the model a ring-only cyclic skeleton and receive a complete scaffold with atom and bond types filled in, then have full molecules grown from it.
  • Because generation starts from the scaffold graph itself, the model can in principle accept scaffolds that never appeared in training, not just memorized ones.
  • Ten thousand samples per privileged GPCR scaffold reproduced 5–16% of known actives and 13–25% of known drugs in leave-one-out tests, suggesting practical scaffold-based hit discovery.
  • Docking scores of generated molecules against DRD2 matched the test-set distribution and, for two of the three case-study scaffolds, beat random database molecules.
  • The observed positive correlation between scaffold size and validity and negative correlation with uniqueness means benchmarks for such models should compare per-scaffold metrics on a common scaffold set, not aggregate averages.

Reading between the lines

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

  • If the missing connect-action constraint were made explicit and proven, the exact-substructure guarantee would become a formal invariant; until then, a skeptical reader should treat it as a design intent rather than a demonstrated theorem.
  • The same scaffold-seeded growth procedure could be applied to other graph-generation tasks that require a fixed subgraph to survive, such as constructing molecules around a covalent warhead or a conserved pharmacophore.
  • The paper's own bad-case analysis implies that literature-derived test sets are biased toward single-assay chemistry; constructing a deliberately unbiased scaffold benchmark would more directly test generalization than MMD against those sets.
  • The larger scatter in property distributions for generated molecules, read together with the bias analysis, suggests the model may be exploring legitimate chemical space rather than overfitting the side-chain preferences of a few publications—an interpretation that could be tested by prospective synthesis or by scoring on a diverse external library.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper presents DeepScaffold, a graph-based generative model for scaffold-directed de novo drug design. The system combines three components: a VAE-based model that assigns atom and bond types to cyclic skeletons (producing classical scaffolds), a scaffold-based molecule generator that grows side chains from a given scaffold graph, and a side-chain property filter based on pharmacophore-like queries. The authors evaluate the scaffold-diversification model on holdout scaffolds from ChEMBL and on three privileged GPCR scaffolds in a leave-one-out setting, reporting validity, uniqueness, property distributions, internal diversity, MMD, active-molecule reproduction rates, and DRD2 docking scores. The central claim is that the method guarantees that generated molecules match the input scaffold query and generalizes to new scaffolds.

Significance. If the scaffold-preservation guarantee were established, DeepScaffold would be a practically valuable contribution, extending earlier scaffold-based graph generative work by supporting multiple scaffold definitions, including cyclic skeletons and side-chain property constraints. The paper also contains a thoughtful treatment of evaluation methodology, including the use of a U-statistic for internal diversity and a bad-case analysis attributing high MMD scores to publication bias in ChEMBL. These are genuine strengths. However, the headline guarantee is not supported by the described architecture or by the reported metrics, as detailed in the major comments. The manuscript is therefore of interest to the molecular generative-modeling community but requires substantial revision before the central claim can be accepted.

major comments (4)
  1. [§2.4, Figure 7b] The central guarantee that the output molecule always contains the input scaffold is not enforced by the described action set. The three allowed actions are 'append a new atom', 'connecting two existing nodes', and 'terminate'. The paper does not state that the connect action is masked so that at least one endpoint is a newly added atom, nor that bonds between two original scaffold atoms are prohibited. If such a bond is sampled, the scaffold is topologically altered and the output no longer contains the query as a substructure. Since the guarantee is load-bearing for the paper's main claim, please either describe the explicit masking/constraint mechanism or, if none exists, add a post-hoc substructure check and weaken the claim accordingly.
  2. [§3.3.1, Eq. (10)] The reported average validity of 98.9% is computed using RDKit Chem.Sanitize, which checks chemical validity only, not whether the generated molecule contains the input scaffold as a substructure. Therefore the validity, uniqueness, and docking results do not provide evidence for the scaffold-match guarantee. A per-scaffold substructure-containment rate should be reported alongside validity to justify the 'guarantee' language.
  3. [§2.3, Table 1] The cyclic-skeleton-to-scaffold component achieves only 82.5% validity and 61.0% reconstruction with the best β-VAE setting, yet no end-to-end evaluation of the full pipeline (cyclic skeleton → classical scaffold → final molecule) is presented. The abstract and introduction claim generation from cyclic skeletons as one of the model's capabilities, so the absence of an end-to-end metric leaves this aspect only partially validated. Please report the fraction of generated molecules that pass sanitization and contain the target skeleton.
  4. [§2.4.2, Table 2] The leave-one-out validation is performed on only three hand-picked privileged scaffolds, and no uncertainty estimates or statistical tests are reported. This is a very limited basis for the claim that the model generalizes to new scaffolds outside the training set. Please provide confidence intervals for the three cases or extend the leave-one-out evaluation to a larger random sample of scaffolds.
minor comments (4)
  1. [§3.3.1, Figure 9a] The text states that the average validity is 98.9%, while the caption of Figure 9a reports 98.8%; please make these consistent.
  2. [§1, reference 5] The sentence 'tools for discovery of privileged scaffolds (such as CSE6 and5)' appears to have a malformed citation; please fix the reference formatting.
  3. [§2.5] The pharmacophore-based filter is described only briefly; please specify how hydrogen-bond donor/acceptor counts are defined and how the side-chain size is measured, since these queries are part of the claimed functionality.
  4. [§3.3.6, Figure 13] The caption of Figure 13 lists active molecule reproduction rates that differ in precision from Table 2 (e.g., '5.17%' vs '0.0517'); please ensure consistent formatting between figure captions and tables.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the scaffold-based generator reuses the authors' prior graph-generative architecture, but the paper's central claims are independently evaluated by holdout/leave-one-out splits and external docking.

full rationale

DeepScaffold's scaffold-based generator is adapted from the authors' previous work [18] (e.g., 'The generation process of side-chain follows largely from our previous work.18'), but that citation supplies architecture and training details rather than the paper's conclusions. The scaffold-based capability is established by independent evaluation: holdout and leave-one-out splits exclude scaffold-molecule pairs from training, and docking against DRD2 is an external computational assay, so no fitted parameter is repackaged as a prediction. The CSK-to-scaffold VAE is trained with a reconstruction objective and evaluated on a held-out 20% split. The pharmacophore side-chain filter is an explicit post-processing step, not a hidden input. The only notable risk, namely that the 'connect two existing nodes' action could join two scaffold atoms and thereby undermine the advertised scaffold-match guarantee, is a correctness or implementation concern rather than circularity: nothing in the derivation assumes the conclusion, and no equation or metric reduces to its own input by construction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The main free parameters are hyperparameters of the neural models, not physical constants. The central claims rest on a set of domain assumptions about chemical space, the generative action space, and the validity of docking as a bioactivity proxy.

free parameters (5)
  • beta in beta-VAE = 0.5 (also 0.1 tested)
    The VAE that converts cyclic skeletons to classical scaffolds is trained with a beta-VAE objective; beta=0.5 gives the highest scaffold validity (82.5%) while beta=0.1 gives higher reconstruction (91.3%). Model performance depends on this hand-chosen hyperparameter.
  • latent variable dimension = 10
    Dimension of latent variables for each node and edge in the skeleton-to-scaffold VAE. The authors state they tried more but found little improvement.
  • importance sampling k and alpha = k=5, alpha=0.5
    Hyperparameters for sampling in the scaffold-based molecule generator, following the authors' previous work [18]. These affect generation diversity and validity.
  • QED threshold for dataset filtering = 0.5
    Molecules with QED below 0.5 are excluded from training, shaping the learned chemical space.
  • sampling mixture weight = 0.5
    Mini-batches mix scaffold-first and molecule-first sampling with equal weight; the authors note this weight could be tuned.
assumptions (6)
  • standard math VAE evidence lower bound is a valid training objective for the skeleton-to-scaffold model (standard variational inference).
    Used in Section 2.3.2, eq. 2. No proof needed.
  • standard math U-statistics internal diversity and MMD with Tanimoto kernel are unbiased, valid distributional distance metrics.
    Used in Section 2.6.2; the paper correctly notes the U-statistic correction.
  • domain assumption ChEMBL molecules with QED>0.5 are a representative training distribution for drug-like chemical space.
    The model learns to generate molecules from this filtered distribution; if the distribution is unrepresentative, generated molecules may not be drug-like.
  • domain assumption The defined action space (append atom, connect two atoms, terminate) can express all relevant side-chain modifications of a scaffold.
    The generator can only produce molecules reachable by these actions; complex side chains that require simultaneous multi-bond changes may be unreachable.
  • domain assumption A scaffold remains unmodified during generation (guarantee of scaffold matching).
    No explicit enforcement is described; the paper claims the guarantee based on starting graph and action set.
  • domain assumption Docking score is a valid proxy for bioactivity against DRD2.
    Used in Section 3.3.8 to claim generated molecules are bioactive; docking is a computational approximation with well-known inaccuracies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepScaffold: a comprehensive tool for scaffold-based de novo drug discovery using deep learning." pith.science (2026). https://pith.science/paper/5X3NLD32

@misc{pith2026190807209,
  author       = {Pith},
  title        = {Pith review of: DeepScaffold: a comprehensive tool for scaffold-based de novo drug discovery using deep learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5X3NLD32}},
  note         = {Machine review of arXiv:1908.07209}
}
read the original abstract

The ultimate goal of drug design is to find novel compounds with desirable pharmacological properties. Designing molecules retaining particular scaffolds as the core structures of the molecules is one of the efficient ways to obtain potential drug candidates with desirable properties. We proposed a scaffold-based molecular generative model for scaffold-based drug discovery, which performs molecule generation based on a wide spectrum of scaffold definitions, including BM-scaffolds, cyclic skeletons, as well as scaffolds with specifications on side-chain properties. The model can generalize the learned chemical rules of adding atoms and bonds to a given scaffold. Furthermore, the generated compounds were evaluated by molecular docking in DRD2 targets and the results demonstrated that this approach can be effectively applied to solve several drug design problems, including the generation of compounds containing a given scaffold and de novo drug design of potential drug candidates with specific docking scores. Finally, a command line interface is created.

Figures

Figures reproduced from arXiv: 1908.07209 by the authors.

Figure 1
Figure 1. An overview of DeepScaffold. DeepScaffold is a comprehensive tool for scaffold [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The basic components for DeepScaffold. As shown in the figure, DeepScaffold is [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. a-d. The distribution of several important molecular properties (a for MW, b for [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Formal charges are removed and explicit hydrogen atoms are fixed after removing [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The architecture of the model to transform CSKs to classicial scaffolds. Under [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: The architecture of the graph convolutional neural network (GNN). Before feeding [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: The generation process. a. Unconditional generative models (such as MolMP) starts molecule generation from empty graph G0 = (,). b. DeepScaffold starts molecule generation directly from scaffolds. 2.4.2 Leave-one-out validation Leave-one-out validation is carried out i…
Figure 8
Figure 8. Figure 8: Generated samples from DeepScaffold. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 9
Figure 9. Figure 9: It is immediately noticed that the result in [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 9
Figure 9. Figure 9: Sample validity and uniqueness. a. The average validity (98.8%) and uniqueness (69.1%) among all scaffolds. b. The relationship between the molecular weight of each scaffold and the (logit) precent-valid output. c. The relationship between molecular weight for each sca…
Figure 10
Figure 10. Figure 10: Comparing the distribution of several molecular properties (MW, LogP and [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]
Figure 11
Figure 11. Figure 11: Internal diversity and MMD. a. Comparing the internal diversity between gener￾ated molecules and testset molecules. Scaffolds with more than 400 corresponding molecules (that is, |M(s)| > 400) are highlighted in blue. b. Scaffolds with high output diversity. c. The di…
Figure 12
Figure 12. Figure 12: Demonstration of several bad-cases drawn from the test set scaffolds. [PITH_FULL_IMAGE:figures/full_fig_p029_12.png]
Figure 13
Figure 13. Figure 13: Generated samples and reproduced active molecules. This figure shows the [PITH_FULL_IMAGE:figures/full_fig_p031_13.png]
Figure 14
Figure 14. Figure 14: The distribution of side chain properties at each location for scaffold [PITH_FULL_IMAGE:figures/full_fig_p033_14.png]
Figure 15
Figure 15. Figure 15: The distribution of docking scores. This plot shows the distribution of docking [PITH_FULL_IMAGE:figures/full_fig_p034_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 37 canonical work pages

  1. [1]

    Computational Exploration of Molecular Scaffolds in Medicinal Chemistry

    Hu, Y.; Stumpfe, D.; Bajorath, J. Computational Exploration of Molecular Scaffolds in Medicinal Chemistry. J Med Chem 2016, 59, 4062--76

    work page 2016

  2. [2]

    W.; Murcko, M

    Bemis, G. W.; Murcko, M. A. The properties of known drugs. 1. Molecular frameworks. J Med Chem 1996, 39, 2887--93

    work page 1996

  3. [3]

    Barreiro, E. J. Privileged Scaffolds in Medicinal Chemistry; Royal Society of Chemistry: London, 2015

    work page 2015

  4. [4]

    Mining for bioactive scaffolds with scaffold networks: improved compound set enrichment from primary screening data

    Varin, T.; Schuffenhauer, A.; Ertl, P.; Renner, S. Mining for bioactive scaffolds with scaffold networks: improved compound set enrichment from primary screening data. J Chem Inf Model 2011, 51, 1528--38

    work page 2011

  5. [5]

    Identification of Bioactive Scaffolds Based on QSAR Models

    Nakagawa, T.; Miyao, T.; Funatsu, K. Identification of Bioactive Scaffolds Based on QSAR Models. Mol Inform 2018, 37

    work page 2018

  6. [6]

    N.; Zhang, J.-H.; Raman, P.; Ertl, P.; Schuffenhauer, A

    Varin, T.; Gubler, H.; Parker, C. N.; Zhang, J.-H.; Raman, P.; Ertl, P.; Schuffenhauer, A. Compound set enrichment: a novel approach to analysis of primary HTS data. Journal of chemical information and modeling 2010, 50, 2067--2078

    work page 2010

  7. [7]

    J.; Janes, J.; Su, A

    Wilkens, S. J.; Janes, J.; Su, A. I. HierS: hierarchical scaffold clustering using topological chemical graphs. Journal of medicinal chemistry 2005, 48, 3182--3193

    work page 2005

  8. [8]

    A.; Schuffenhauer, A.; Scheck, M.; Wetzel, S.; Casaulta, M.; Odermatt, A.; Ertl, P.; Waldmann, H

    Koch, M. A.; Schuffenhauer, A.; Scheck, M.; Wetzel, S.; Casaulta, M.; Odermatt, A.; Ertl, P.; Waldmann, H. Charting biologically relevant chemical space: a structural classification of natural products (SCONP). Proc Natl Acad Sci 2005, 102, 17272--7

    work page 2005

Show all 41 references
  1. [9]

    A.; Waldmann, H

    Schuffenhauer, A.; Ertl, P.; Roggo, S.; Wetzel, S.; Koch, M. A.; Waldmann, H. The scaffold tree--visualization of the scaffold universe by hierarchical scaffold classification. J Chem Inf Model 2007, 47, 47--58

    2007

  2. [10]

    Scaffold Hunter: a comprehensive visual analytics framework for drug discovery

    Schafer, T.; Kriege, N.; Humbeck, L.; Klein, K.; Koch, O.; Mutzel, P. Scaffold Hunter: a comprehensive visual analytics framework for drug discovery. J Cheminform 2017, 9, 28

    2017

  3. [11]

    Deep learning

    LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436--44

    2015

  4. [12]

    Hu, Z.; Yang, Z.; Salakhutdinov, R.; Xing, E. P. On Unifying Deep Generative Models. International Conference for Learning Representations

  5. [13]

    Computer-based de novo design of drug-like molecules

    Schneider, G.; Fechner, U. Computer-based de novo design of drug-like molecules. Nat Rev Drug Discov 2005, 4, 649--63

    2005

  6. [14]

    Segler, M. H. S.; Kogej, T.; Tyrchan, C.; Waller, M. P. Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks. ACS Cent Sci 2018, 4, 120--131

    2018

  7. [15]

    N.; Duvenaud, D.; Hernandez-Lobato, J

    Gomez-Bombarelli, R.; Wei, J. N.; Duvenaud, D.; Hernandez-Lobato, J. M.; Sanchez-Lengeling, B.; Sheberla, D.; Aguilera-Iparraguirre, J.; Hirzel, T. D.; Adams, R. P.; Aspuru-Guzik, A. Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules. ACS Cent...

    2018

  8. [16]

    Molecular de-novo design through deep reinforcement learning

    Olivecrona, M.; Blaschke, T.; Engkvist, O.; Chen, H. Molecular de-novo design through deep reinforcement learning. Journal of Cheminformatics 2017, 9, 48

    2017

  9. [17]

    SMILES, a chemical language and information system

    Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. Journal of Chemical Information and Modeling 1988, 28, 31--36

    1988

  10. [18]

    Multi-objective de novo drug design with conditional graph generative model

    Li, Y.; Zhang, L.; Liu, Z. Multi-objective de novo drug design with conditional graph generative model. J Cheminform 2018, 10, 33

    2018

  11. [19]

    GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders

    Simonovsky, M.; Komodakis, N. GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders. arXiv:1610.02920 2017,

    2017 arXiv

  12. [20]

    Learning Deep Generative Models of Graphs

    Li, Y.; Vinyals, O.; Dyer, C.; Pascanu, R.; Battaglia, P. Learning Deep Generative Models of Graphs. arXiv:1803.03324 2017,

    2017 arXiv

  13. [21]

    L.; Leskovec, J

    You, J.; Ying, R.; Ren, X.; Hamilton, W. L.; Leskovec, J. GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models. International Conference on Machine Learning

  14. [22]

    Junction Tree Variational Autoencoder for Molecular Graph Generation

    Jin, W.; Barzilay, R.; Jaakkola, T. Junction Tree Variational Autoencoder for Molecular Graph Generation. International Conference on Machine Learning

  15. [23]

    MolGAN: An implicit generative model for small molecular graphs

    De Cao, N.; Kipf, T. MolGAN: An implicit generative model for small molecular graphs. arXiv:1805.11973 2018,

    2018 arXiv

  16. [24]

    Latent Molecular Optimization for Targeted Therapeutic Design

    Aumentado-Armstrong, T. Latent Molecular Optimization for Targeted Therapeutic Design. arXiv:1809.02032 2018,

    2018 arXiv

  17. [25]

    QBMG: quasi-biogenic molecule generator with deep recurrent neural network

    Zheng, S.; Yan, X.; Gu, Q.; Yang, Y.; Du, Y.; Lu, Y.; Xu, J. QBMG: quasi-biogenic molecule generator with deep recurrent neural network. J Cheminform 2019, 11, 5

    2019

  18. [26]

    Lim, J.; Hwang, S.-Y.; Kim, S.; Moon, S.; Kim, W. Y. Scaffold-based molecular design using graph generative model. arXiv:1905.13639 2019,

    1905 arXiv

  19. [27]

    J.; Bento, A

    Gaulton, A.; Bellis, L. J.; Bento, A. P.; Chambers, J.; Davies, M.; Hersey, A.; Light, Y.; McGlinchey, S.; Michalovich, D.; Al-Lazikani, B.; Overington, J. P. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res 2012, 40, D1100--7

    2012

  20. [28]

    R.; Paolini, G

    Bickerton, G. R.; Paolini, G. V.; Besnard, J.; Muresan, S.; Hopkins, A. L. Quantifying the chemical beauty of drugs. Nat Chem 2012, 4, 90--8

    2012

  21. [29]

    Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions

    Ertl, P.; Schuffenhauer, A. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. J Cheminform 2009, 1, 8

    2009

  22. [30]

    P.; Welling, M

    Kingma, D. P.; Welling, M. Auto-encoding variational bayes. International Conference for Learning Representations

  23. [31]

    R.; Vilnis, L.; Vinyals, O.; Dai, A

    Bowman, S. R.; Vilnis, L.; Vinyals, O.; Dai, A. M.; Jozefowicz, R.; Bengio, S. Generating sentences from a continuous space. SIGNLL Conference on Computational Natural Language Learning

  24. [32]

    beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework

    Higgins, I.; Matthey, L.; Pal, A.; Burgess, C.; Glorot, X.; Botvinick, M.; Mohamed, S.; Lerchner, A. beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework. ICLR 2017, 2, 6

    2017

  25. [33]

    ChemGAN challenge for drug discovery: can AI reproduce natural chemical diversity? arXiv:1708.08227 2017,

    Benhenda, M. ChemGAN challenge for drug discovery: can AI reproduce natural chemical diversity? arXiv:1708.08227 2017,

    2017 arXiv

  26. [34]

    Mmd gan: Towards deeper understanding of moment matching network

    Li, C.-L.; Chang, W.-C.; Cheng, Y.; Yang, Y.; Póczos, B. Mmd gan: Towards deeper understanding of moment matching network. Advances in Neural Information Processing Systems. pp 2203--2213

  27. [35]

    H.; Taylor, G

    Jiwoong Im, D.; Ma, A. H.; Taylor, G. W.; Branson, K. Quantitatively Evaluating GANs With Divergences Proposed for Training. International Conference on Learning Representations

  28. [36]

    Wasserstein GAN

    Arjovsky, M.; Chintala, S.; Bottou, L. Wasserstein GAN. International Conference on Learning Representations

  29. [37]

    Frechet ChemNet Distance: A Metric for Generative Models for Molecules in Drug Discovery

    Preuer, K.; Renz, P.; Unterthiner, T.; Hochreiter, S.; Klambauer, G. Frechet ChemNet Distance: A Metric for Generative Models for Molecules in Drug Discovery. J Chem Inf Model 2018, 58, 1736--1741

    2018

  30. [38]

    Gower, J. C. A general coefficient of similarity and some of its properties. Biometrics 1971, 857--871

    1971

  31. [39]

    Madras, B. K. History of the discovery of the antipsychotic dopamine D2 receptor: a basis for the dopamine hypothesis of schizophrenia. Journal of the History of the Neurosciences 2013, 22, 62--78

    2013

  32. [40]

    P.; Salimans, T.; Duan, Y.; Dhariwal, P.; Schulman, J.; Sutskever, I.; Abbeel, P

    Chen, X.; Kingma, D. P.; Salimans, T.; Duan, Y.; Dhariwal, P.; Schulman, J.; Sutskever, I.; Abbeel, P. Variational lossy autoencoder. International Conference on Learning Representations

  33. [41]

    6DTpDQ 2(C

    Kingma, D. P.; Salimans, T.; Jozefowicz, R.; Chen, X.; Sutskever, I.; Welling, M. Improved variational inference with inverse autoregressive flow. Advances in Neural Information Processing Systems. pp 4743--4751 mcitethebibliography article.tex000066400000000000000000000070271...

    1999

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.