Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

QCA-MolGAN: Quantum Circuit Associative Molecular GAN with Multi-Agent Reinforcement Learning

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

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

Pith's one-line read A quantum circuit prior trained to imitate discriminator features, combined with multi-agent reward shaping, can steer molecular generation toward targeted druglike properties.

desk verdict New architecture combo, but the diversity claim is untested—no classical baseline, and their own MARL results show mode collapse. read the letter →

arxiv 2509.05051 v1 pith:JZ37W6UX submitted 2025-09-05 quant-ph cs.LG

Aaron Mark Thomas , Yu-Cheng Chen , Hubert Okadome Valencia , Sharu Theresa Jose , Ronin Wu This is my paper
classification quant-phcs.LG
keywords quantumcircuitBornmachinemoleculargenerationgenerativeadversarialnetworkmulti-agentreinforcementlearningdrug-likenessassociativeQM9hybridquantum-classical
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

The paper introduces QCA-MolGAN, a hybrid quantum-classical generative model for small drug-like molecules. It replaces the GAN's usual fixed latent prior with a quantum circuit Born machine (QCBM), trained to match high-level feature activations from the discriminator's middle layer, so the prior acts as an associative memory. A set of reinforcement-learning agents, one per chemical property (drug-likeness, LogP, synthetic accessibility), collectively reward the generator and steer it to optimize the properties together. On a QM9 subset, single-objective runs push their target property up while keeping validity and novelty near perfect, and joint MARL achieves the best macro average across properties (0.721). The paper's headline claim is that the trained QCBM prior significantly enhances generative diversity, although the authors also report a sharp uniqueness drop under MARL consistent with mode collapse.

What carries the argument

The load-bearing objects are (1) the QCBM prior—a parameterized circuit whose Born-rule probabilities define a discrete latent distribution, trained with the associative loss of Eq. 6 to match the discriminator's bottleneck activations—and (2) the multi-agent reward network, which predicts QED, LogP, and SA and feeds their weighted sum back as the generator's RL reward (Eq. 9). These sit inside a Wasserstein GAN with gradient penalty (Eq. 7), combined in the total objective Eq. 10.

What would settle it

Take the exact QCA-MolGAN setup and replace the trained QCBM with a classical Gaussian or MLP prior of the same latent dimension, leaving generator, discriminator, and MARL unchanged. If diversity and the 0.721 macro average are reproduced, then the QCBM associative prior is not doing the work.

Watch

Extended reading notes

Core claim

At the center of the proposal is a quantum circuit Born machine that defines a distribution over n-bit strings; samples from that distribution are fed as latent codes into a Wasserstein GAN generator for molecular graphs. The QCBM is trained not to match the data directly but to match the distribution of activations at a deep layer of the discriminator, giving the generator a learned high-level feature prior. Around this, a multi-agent reinforcement-learning module trains one predictor per target property; once trained, a weighted sum of their predictions is maximized as the generator's reward. On the QM9 subset the authors report near-perfect validity and novelty in all configurations, targ

Load-bearing premise

The claim that the QCBM prior enhances diversity rests on the assumption that matching the discriminator's middle-layer activations gives the generator a useful latent feature distribution; because no classical-prior ablation is run, the improvement could come entirely from the RL rewards or other architecture choices.

Editorial extensions

If this is right

  • A QCBM prior targeted at a single property moves that property in the intended direction without sacrificing chemical validity (≥99.7%) or novelty (≥99.4%).
  • Joint multi-agent optimization outperforms each single-objective run on average property alignment, reaching a macro average of 0.721.
  • The reported uniqueness collapse to 4.5% under MARL means the joint reward currently finds a few high-scoring molecular modes, so improving uniqueness is the immediate next bottleneck.
  • Because the quantum and classical training loops are decoupled, the circuit can be sampled on NISQ devices while classical updates happen separately.
  • If the associative QCBM training is doing the claimed work, the same Eq. 6-style recipe could be transplanted to other graph-generation GANs.

Reading between the lines

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

  • The paper does not ablate the QCBM against a Gaussian or classical learned prior, so the diversity claim cannot yet be attributed to the quantum prior itself; the same result might come from the RL reward shaping or the WGAN-GP architecture.
  • A natural test is to swap in a classical MLP or Gaussian prior of identical latent dimension; if diversity and the macro average stay the same, the associative QCBM training is not load-bearing.
  • If scaled to the full QM9 set, the 16-qubit, 2-layer circuit may be too small to represent the feature distribution, so the quantum advantage is more likely to appear in tasks with higher-dimensional or strongly correlated latent spaces.
  • Annealing gamma from WGAN dominance to RL dominance, or using dynamic reward weights, could recover uniqueness while keeping property alignment; this is directly testable with the authors' setup.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes QCA-MolGAN, a hybrid quantum-classical GAN for small-molecule generation. A QCBM supplies a discrete latent prior, trained via an associative adversarial objective (Eq. 6) to match discriminator-layer activations. The generator is trained with a WGAN loss and a multi-agent reinforcement-learning reward combining QED, LogP, and SA. Experiments on a QM9 subset report near-perfect validity and novelty under different reward objectives, with the MARL variant achieving the highest average property score, but also collapsing uniqueness to 4.5%.

Significance. The idea of using a QCBM as an associative prior is original and, if validated, would be a useful hybrid quantum-classical building block. The experimental section is readable and uses RDKit as an external property oracle, which gives the property scores independent grounding. The authors are also transparent about mode collapse and explicitly list the missing classical baseline as future work. However, the paper does not provide the evidence needed to support its central diversity claim: there is no comparison to MolGAN or to a classical prior, and the only reported comparison is among QCBM-based RL objectives. As a result, the significance of the contribution cannot currently be assessed.

major comments (4)
  1. [Section I; Section IV.B, Table I] The first contribution claims that the trained QCBM prior 'significantly enhances the generative diversity of the typical MolGAN architecture.' Table I compares only QCBM-based priors under different RL objectives; there is no MolGAN baseline and no classical (Gaussian/MLP) prior ablation. The conclusion explicitly defers benchmarking against a classical MLP variant to future work. Thus the central diversity claim is not tested. Moreover, the MARL row shows uniqueness of 4.5±0.6%, which the text itself calls 'a clear sign that the GAN mechanism has undergone mode collapse'; the simultaneous diversity of 1.000±0.001 is then computed over a collapsed set of a few unique molecules and does not measure distributional diversity.
  2. [Section III.A, Eq. (6)] The associative training objective is under-specified. p_l(z) is described as the distribution of the activations of the l-th deep latent layer of the discriminator, while q_theta(z) is a discrete distribution over {0,1}^n. Unless the activations are binarized or otherwise quantized, the cross-entropy/log-likelihood in Eq. (6) is not a well-defined density-matching objective. No such discretization is described. Since L_AAN is the only mechanism by which the QCBM prior is connected to the GAN, this gap prevents attributing any observed behavior to the QCBM.
  3. [Section IV.A; Section IV.B] The evaluation protocol is not fully specified. Section IV.B states that metrics are evaluated 'when the drug candidacy score is at a maximum,' but no definition of this score or selection rule is given. If this selects a favorable epoch per model, the reported property scores are optimistic and the comparison across objectives is not on equal footing. The authors should report full training curves or define a fixed evaluation epoch.
  4. [Section III.D; Section IV.A] The combined loss is L_AAN + gamma L_WGAN + (1−gamma) L_MARL, but the experimental section says 'subsequently optimising only the RL component (γ=0.0).' This is confusing: if γ=0.0, the WGAN term is zero during the reported final training. The paper does not clarify which terms are active when the reported metrics are obtained, making it difficult to attribute the results to the GAN/QCBM component rather than to reward maximization alone.
minor comments (5)
  1. [Author affiliations] Typo: 'Departement' should be 'Department'.
  2. [Abstract/Introduction] The estimate '10 23 −10 100' should be typeset with superscripts as 10^23–10^100.
  3. [Figure 2] The notation in the figure (e.g., 'Rl,1,2xx') is hard to parse; the caption also calls the R_xx gates 'controlled' although R_xx is a two-qubit rotation. Please clarify.
  4. [Section III.A, Eq. (6)] The objective is written as max_theta of an expectation, but the optimization should be argmax; the loss being minimized is the negative log-likelihood.
  5. [Section IV.A] Provide more reproducibility details for the RL reward networks and the SPSA optimization of the QCBM (learning rate, number of SPSA iterations, gradient estimation method).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the diversity claim is under-supported, but no derivation reduces to its own inputs.

full rationale

The paper's derivation chain is not circular. The QCBM prior is trained to the discriminator's bottleneck distribution via Eq. (6), but this is a co-adaptive training objective, not a definition of the reported evaluation metrics. The property scores (QED, SA, LogP) are computed by the external RDKit package, independent of the model's internal training signals. The RL agents are fit to those RDKit properties (Eq. 8) and then used as rewards (Eq. 9); reporting improved scores on the optimized properties is the intended objective, not a hidden equivalence. The diversity claim in Contribution 1 is empirically under-supported: the paper provides no classical-prior baseline and Table I itself reports mode collapse under MARL (uniqueness 4.5±0.6%), but that is an evidentiary gap or correctness concern, not circularity. References [35], [38], and [41] are external prior works, and no load-bearing argument reduces to a self-citation. No circular step can be exhibited under the required standard of showing a specific reduction of a claimed result to its own inputs.

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

The central results rest on domain assumptions about molecular property proxies and about the value of matching a discriminator bottleneck, plus several hand-set hyperparameters (reward weights, gamma, lambda, QCBM size). No new physical entities are introduced.

free parameters (5)
  • MARL reward weights (w_QED=0.4, w_LogP=0.3, w_SA=0.3) = 0.4, 0.3, 0.3
    Hand-chosen following Tang et al. [41]; directly determines the aggregate reward in Eq. (9) and therefore the balance of properties, a central output of the paper.
  • gamma in combined loss = 0.0 after pretraining
    Controls WGAN versus RL weighting in Eq. (10); setting to zero removes the WGAN term from the generator objective, yet the paper still describes WGAN training.
  • gradient penalty coefficient lambda = 10
    Standard WGAN-GP hyperparameter, but affects training stability and hence all reported metrics.
  • QCBM qubits/layers/shots = 16 qubits, 2 layers, 1000 shots
    Architecture choices for the latent prior; no sensitivity analysis is provided, so results may depend on these values.
  • clipping constant epsilon in L_AAN = not specified
    Clipped log-likelihood in Eq. (6) prevents collapse; the value is not reported, leaving the prior training objective underspecified.
assumptions (5)
  • standard math Born's rule defines the QCBM output distribution q_theta(z) = |<z|Psi(theta)>|^2.
    Invoked in Eq. (2), Section III.A; foundational quantum mechanics.
  • domain assumption Molecular graph representation (A, X) and RDKit-computed QED, LogP, SA scores are valid ground-truth rewards for drug-likeness.
    Used to train RL agents and evaluate results (Section II and IV.A); if these proxies are flawed, the property alignment results are flawed.
  • domain assumption The discriminator's deep bottleneck layer distribution p_l(z) captures high-level features useful as an associative target.
    Core premise of the associative adversarial network; introduced in Section III.A and used in Eq. (6). No empirical validation that these features are meaningful for molecular structures.
  • domain assumption WGAN-GP gives a stable training objective for molecular graphs.
    Adopted from Gulrajani et al. [40] in Eq. (7); standard but unverified for this discrete graph generator.
  • domain assumption QM9 5k subset is representative for small drug-like molecule generation.
    Dataset choice in Section IV.A; limited to molecules with at most nine heavy atoms and four element types.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QCA-MolGAN: Quantum Circuit Associative Molecular GAN with Multi-Agent Reinforcement Learning." pith.science (2026). https://pith.science/paper/JZ37W6UX

@misc{pith2026250905051,
  author       = {Pith},
  title        = {Pith review of: QCA-MolGAN: Quantum Circuit Associative Molecular GAN with Multi-Agent Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JZ37W6UX}},
  note         = {Machine review of arXiv:2509.05051}
}
read the original abstract

Navigating the vast chemical space of molecular structures to design novel drug molecules with desired target properties remains a central challenge in drug discovery. Recent advances in generative models offer promising solutions. This work presents a novel quantum circuit Born machine (QCBM)-enabled Generative Adversarial Network (GAN), called QCA-MolGAN, for generating drug-like molecules. The QCBM serves as a learnable prior distribution, which is associatively trained to define a latent space aligning with high-level features captured by the GANs discriminator. Additionally, we integrate a novel multi-agent reinforcement learning network to guide molecular generation with desired targeted properties, optimising key metrics such as quantitative estimate of drug-likeness (QED), octanol-water partition coefficient (LogP) and synthetic accessibility (SA) scores in conjunction with one another. Experimental results demonstrate that our approach enhances the property alignment of generated molecules with the multi-agent reinforcement learning agents effectively balancing chemical properties.

Figures

Figures reproduced from arXiv: 2509.05051 by the authors.

Figure 1
Figure 1. Pipeline of Associative adversarial quantum MolGAN with multi￾agent RL. We first sample bitstring samples from the computational basis of a quantum circuit born machine, which acts as the GAN’s prior distribution. The generator then generates candidate molecule graphs, guided by the multi-agent RL network to maximise several target properties. At the end of an epoch, the prior is trained to match a bottleneck layer … view at source ↗
Figure 2
Figure 2. QCBM: Unitary learning layer composed of a rotation layer with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Multi Agent Reinforcement Learning Network: A series of RL [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Implementations of Quantum and Classical Topology-Aligned Architectures for Molecular Property Prediction

    cs.LG 2026-07 conditional novelty 6.0 of 10

    A quantum circuit and a classical network with identical topology-shaped 64-parameter design perform comparably on QM9 classification, indicating the inductive bias rather than quantum substrate drives parameter efficiency.

  2. Rank-Refined Quantum-Behaved Particle Swarm Optimization for Quantum Molecular Generation

    quant-ph 2026-07 conditional novelty 5.5 of 10

    RR-QPSO raises the validity–uniqueness product of 9-heavy-atom quantum molecular generation from 0.902 (BO) to 0.942 by rank-refined mean-best and fitness-guided swarm updates.

Reference graph

Works this paper leans on

44 extracted references · 34 canonical work pages · cited by 2 Pith papers

  1. [1]

    Improving drug candidates by design: a focus on physicochemical properties as a means of improving compound dis- position and safety,

    N. A. Meanwell, “Improving drug candidates by design: a focus on physicochemical properties as a means of improving compound dis- position and safety,”Chemical research in toxicology, vol. 24, no. 9, pp. 1420–1456, 2011

    work page 2011

  2. [2]

    Deep reinforcement learning for de novo drug design,

    M. Popova, O. Isayev, and A. Tropsha, “Deep reinforcement learning for de novo drug design,”Science advances, vol. 4, no. 7, p. eaap7885, 2018

    work page 2018

  3. [3]

    Inverse design in search of materials with target function- alities,

    A. Zunger, “Inverse design in search of materials with target function- alities,”Nature Reviews Chemistry, vol. 2, no. 4, p. 0121, 2018

    work page 2018

  4. [4]

    Virtual compound libraries in computer-assisted drug discovery,

    N. van Hilten, F. Chevillard, and P. Kolb, “Virtual compound libraries in computer-assisted drug discovery,”Journal of chemical information and modeling, vol. 59, no. 2, pp. 644–651, 2019

    work page 2019

  5. [5]

    Principles of early drug discovery,

    J. P. Hughes, S. Rees, S. B. Kalindjian, and K. L. Philpott, “Principles of early drug discovery,”British journal of pharmacology, vol. 162, no. 6, pp. 1239–1249, 2011

    work page 2011

  6. [6]

    Applications of machine learning in drug discovery and development,

    J. Vamathevan, D. Clark, P. Czodrowski, I. Dunham, E. Ferran, G. Lee, B. Li, A. Madabhushi, P. Shah, M. Spitzer,et al., “Applications of machine learning in drug discovery and development,”Nature reviews Drug discovery, vol. 18, no. 6, pp. 463–477, 2019

    work page 2019

  7. [7]

    Machine learning in drug discovery: a review,

    S. Dara, S. Dhamercherla, S. S. Jadav, C. M. Babu, and M. J. Ahsan, “Machine learning in drug discovery: a review,”Artificial intelligence review, vol. 55, no. 3, pp. 1947–1999, 2022

    work page 1947

  8. [8]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems, vol. 27, 2014

    2014

Show all 44 references
  1. [9]

    Auto-encoding variational bayes,

    D. P. Kingma, “Auto-encoding variational bayes,”arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  2. [10]

    Recurrent neural networks,

    L. R. Medsker, L. Jain,et al., “Recurrent neural networks,”Design and Applications, vol. 5, no. 64-67, p. 2, 2001

    2001

  3. [11]

    Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules,

    D. Weininger, “Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules,”Journal of chemical information and computer sciences, vol. 28, no. 1, pp. 31–36, 1988

    1988

  4. [12]

    Quantitative structure- activity relationship methods: Perspectives on drug discovery and tox- icology,

    R. Perkins, H. Fang, W. Tong, and W. J. Welsh, “Quantitative structure- activity relationship methods: Perspectives on drug discovery and tox- icology,”Environmental Toxicology and Chemistry, vol. 22, no. 8, pp. 1666–1679, 2003

    2003

  5. [13]

    Molgan: An implicit generative model for small molecular graphs,

    N. De Cao and T. Kipf, “Molgan: An implicit generative model for small molecular graphs,”arXiv preprint arXiv:1805.11973, 2018

    2018 arXiv

  6. [14]

    Temporally unstructured quantum computation,

    D. Shepherd and M. J. Bremner, “Temporally unstructured quantum computation,”Proceedings of the Royal Society A: Mathematical, Phys- ical and Engineering Sciences, vol. 465, no. 2105, pp. 1413–1439, 2009

    2009

  7. [15]

    Quantum computing in the nisq era and beyond,

    J. Preskill, “Quantum computing in the nisq era and beyond,”Quantum, vol. 2, p. 79, 2018

    2018

  8. [16]

    Quantum machine learning,

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,”Nature, vol. 549, no. 7671, pp. 195–202, 2017

    2017

  9. [17]

    Quantum computational advantage with a programmable photonic processor,

    L. S. Madsen, F. Laudenbach, M. F. Askarani, F. Rortais, T. Vincent, J. F. Bulmer, F. M. Miatto, L. Neuhaus, L. G. Helt, M. J. Collins,et al., “Quantum computational advantage with a programmable photonic processor,”Nature, vol. 606, no. 7912, pp. 75–81, 2022

    2022

  10. [18]

    Quantum generative models for small molecule drug discovery,

    J. Li, R. O. Topaloglu, and S. Ghosh, “Quantum generative models for small molecule drug discovery,”IEEE transactions on quantum engineering, vol. 2, pp. 1–8, 2021

    2021

  11. [19]

    Hybrid quantum-classical machine learning for generative chemistry and drug design,

    A. Gircha, A. Boev, K. Avchaciov, P. Fedichev, and A. Fedorov, “Hybrid quantum-classical machine learning for generative chemistry and drug design,”Scientific Reports, vol. 13, no. 1, p. 8250, 2023

    2023

  12. [20]

    Ex- ploring the advantages of quantum generative adversarial networks in generative chemistry,

    P.-Y . Kao, Y .-C. Yang, W.-Y . Chiang, J.-Y . Hsiao, Y . Cao, A. Aliper, F. Ren, A. Aspuru-Guzik, A. Zhavoronkov, M.-H. Hsieh,et al., “Ex- ploring the advantages of quantum generative adversarial networks in generative chemistry,”Journal of Chemical Information and Modeling, ...

    2023

  13. [21]

    Hybrid quantum cycle generative ad- versarial network for small molecule generation,

    M. Anoshin, A. Sagingalieva, C. Mansell, D. Zhiganov, V . Shete, M. Pflitsch, and A. Melnikov, “Hybrid quantum cycle generative ad- versarial network for small molecule generation,”IEEE Transactions on Quantum Engineering, 2024

    2024

  14. [22]

    Quantum machine learning for classical data,

    L. Wossnig, “Quantum machine learning for classical data,”arXiv preprint arXiv:2105.03684, 2021

    2021 arXiv

  15. [23]

    A review on mode collapse reducing gans with gan’s algorithm and theory,

    S. Tomar and A. Gupta, “A review on mode collapse reducing gans with gan’s algorithm and theory,”GANs for Data Augmentation in Healthcare, pp. 21–40, 2023

    2023

  16. [24]

    L- molgan: An improved implicit generative model for large molecular graphs,

    Y . Tsujimoto, S. Hiwa, Y . Nakamura, Y . Oe, and T. Hiroyasu, “L- molgan: An improved implicit generative model for large molecular graphs,” 2021

    2021

  17. [25]

    A comprehensive survey on graph neural networks,

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, “A comprehensive survey on graph neural networks,”IEEE transactions on neural networks and learning systems, vol. 32, no. 1, pp. 4–24, 2020

    2020

  18. [26]

    Graph neural networks: A review of methods and applications,

    J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, and M. Sun, “Graph neural networks: A review of methods and applications,” AI open, vol. 1, pp. 57–81, 2020

    2020

  19. [27]

    Modeling relational data with graph convolutional networks,

    M. Schlichtkrull, T. N. Kipf, P. Bloem, R. Van Den Berg, I. Titov, and M. Welling, “Modeling relational data with graph convolutional networks,” inThe semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedings 15, pp. 593–607,...

    2018

  20. [28]

    Games of gans: Game-theoretical models for generative adversarial networks,

    M. Mohebbi Moghaddam, B. Boroomand, M. Jalali, A. Zareian, A. Daeijavad, M. H. Manshaei, and M. Krunz, “Games of gans: Game-theoretical models for generative adversarial networks,”Artificial Intelligence Review, vol. 56, no. 9, pp. 9771–9807, 2023

    2023

  21. [29]

    Exploration in deep rein- forcement learning: A survey,

    P. Ladosz, L. Weng, M. Kim, and H. Oh, “Exploration in deep rein- forcement learning: A survey,”Information Fusion, vol. 85, pp. 1–22, 2022

    2022

  22. [30]

    Deterministic policy gradient algorithms,

    D. Silver, G. Lever, N. Heess, T. Degris, D. Wierstra, and M. Riedmiller, “Deterministic policy gradient algorithms,” inInternational conference on machine learning, pp. 387–395, Pmlr, 2014

    2014

  23. [31]

    Continuous control with deep reinforcement learning,

    T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y . Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforcement learning,”arXiv preprint arXiv:1509.02971, 2015

    2015 arXiv

  24. [32]

    Quantifying the chemical beauty of drugs,

    G. R. Bickerton, G. V . Paolini, J. Besnard, S. Muresan, and A. L. Hopkins, “Quantifying the chemical beauty of drugs,”Nature chemistry, vol. 4, no. 2, pp. 90–98, 2012

    2012

  25. [33]

    Lipophilicity profiles: theory and measurement,

    J. Comer and K. Tam, “Lipophilicity profiles: theory and measurement,” Pharmacokinetic optimization in drug research: Biological, physico- chemical, and computational strategies, pp. 275–304, 2001

    2001

  26. [34]

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

    P. Ertl and A. Schuffenhauer, “Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions,”Journal of cheminformatics, vol. 1, pp. 1–11, 2009

    2009

  27. [35]

    Associative adversarial networks,

    T. Arici and A. Celikyilmaz, “Associative adversarial networks,”arXiv preprint arXiv:1611.06953, 2016

    2016 arXiv

  28. [36]

    A learning algorithm for boltzmann machines,

    D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, “A learning algorithm for boltzmann machines,”Cognitive science, vol. 9, no. 1, pp. 147–169, 1985

    1985

  29. [37]

    Quantum-assisted associative adversarial network: Applying quantum annealing in deep learning,

    M. Wilson, T. Vandal, T. Hogg, and E. G. Rieffel, “Quantum-assisted associative adversarial network: Applying quantum annealing in deep learning,”Quantum Machine Intelligence, vol. 3, no. 1, p. 19, 2021

    2021

  30. [38]

    Generation of high-resolution handwritten digits with an ion-trap quantum computer,

    M. S. Rudolph, N. B. Toussaint, A. Katabarwa, S. Johri, B. Peropadre, and A. Perdomo-Ortiz, “Generation of high-resolution handwritten digits with an ion-trap quantum computer,”Physical Review X, vol. 12, no. 3, p. 031010, 2022

    2022

  31. [39]

    Comparing the effects of boltzmann machines as associative memory in generative adversarial networks between classical and quantum samplings,

    M. Urushibata, M. Ohzeki, and K. Tanaka, “Comparing the effects of boltzmann machines as associative memory in generative adversarial networks between classical and quantum samplings,”Journal of the Physical Society of Japan, vol. 91, no. 7, p. 074008, 2022

    2022

  32. [40]

    Improved training of wasserstein gans,

    I. Gulrajani, F. Ahmed, M. Arjovsky, V . Dumoulin, and A. C. Courville, “Improved training of wasserstein gans,”Advances in neural information processing systems, vol. 30, 2017

    2017

  33. [41]

    Molecular generative adversarial network with multi-property optimization,

    H. Tang, C. Li, S. Kamei, Y . Yamanishi, and Y . Morimoto, “Molecular generative adversarial network with multi-property optimization,”arXiv preprint arXiv:2404.00081, 2024

    2024 arXiv

  34. [42]

    Quantum chemistry structures and properties of 134 kilo molecules,

    R. Ramakrishnan, P. O. Dral, M. Rupp, and O. A. V on Lilienfeld, “Quantum chemistry structures and properties of 134 kilo molecules,” Scientific data, vol. 1, no. 1, pp. 1–7, 2014

    2014

  35. [43]

    Enumeration of 166 billion organic small molecules in the chemical universe database gdb-17,

    L. Ruddigkeit, R. Van Deursen, L. C. Blum, and J.-L. Reymond, “Enumeration of 166 billion organic small molecules in the chemical universe database gdb-17,”Journal of chemical information and model- ing, vol. 52, no. 11, pp. 2864–2875, 2012

    2012

  36. [44]

    Efficient global optimization using spsa,

    J. L. Maryak and D. C. Chin, “Efficient global optimization using spsa,” inProceedings of the 1999 American Control Conference (Cat. No. 99CH36251), vol. 2, pp. 890–894, IEEE, 1999

    1999

Pith tools

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