Enhancing Molecular Design through Graph-based Topological Reinforcement Learning

Xiangyu Zhang

REVIEW 3 major objections 8 minor 15 references

Enhancing Molecular Design through Graph-based Topological Reinforcement Learning

T0 review · 3 major / 8 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read GraphTRL claims that a graph-and-topology state representation improves reinforcement learning for drug-like molecule design.

desk verdict The paper's one good idea—tracking atom-type changes via SYBYL types—is buried under an untested binding-affinity claim and thin baselines that don't justify the conclusion. read the letter →

arxiv 2411.14726 v1 pith:6UFAD6WS submitted 2024-11-22 cs.LG q-bio.BM

Xiangyu Zhang This is my paper · ORCID

classification cs.LGq-bio.BM

keywords reinforcementlearningmolecularoptimizationpersistenthomologymultiscaleweightedcoloredgraphsdrugdiscoveryQEDpenalizedlogP

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes GraphTRL, a reinforcement-learning method that gives the agent a structural view of a molecule by concatenating three state components: multiscale weighted colored graphs (MWCG), which encode atom-type-specific interactions by distance; persistent images, which summarize the molecule's shape across scales; and molecular fingerprints. The claim is that this graph-and-topology state makes a dueling deep Q-network better at finding drug-like molecules than earlier methods—MolDQN, JT-VAE, GCPN, and ORGAN—while keeping every generated molecule chemically valid. On the reported benchmark, which starts from 800 ZINC molecules, GraphTRL reaches penalized logP 11.89 and QED 0.951, slightly above the best prior scores of 11.84 and 0.948. The abstract goes further, claiming that the structural state also improves binding-affinity prediction; the experiments in the paper, however, measure drug-likeness rather than affinity directly.

What carries the argument

The load-bearing machinery is the state representation, not the policy. MWCG builds element-specific subgraphs whose edges are weighted by a generalized exponential radial basis function $\Phi_E(\|r_i-r_j\|;\eta_{kk'}) = e^{-(\|r_i-r_j\|/\eta_{kk'})^\kappa}$, with SYBYL atom types as vertex colors; persistent homology converts the molecule's multi-scale connective structure into persistent images, a vector-valued density representation. These are concatenated with molecular fingerprints to form the MDP state, and a dueling deep Q-network estimates $Q(s_i,a_i)=V(s_i)+A(s_i,a_i)$, separating state value from action advantage. The reward functions are what make the method usable: a constrained reward $R_1$ optimizes penalized logP while penalizing violation of a Tanimoto similarity floor $\delta$ and a Betti-number floor $\varepsilon$, and an alternative reward $R_2$ balances a target Betti number against similarity to the starting molecule.

What would settle it

Re-running MolDQN-bootstrap and GraphTRL from the same 800 ZINC molecules with identical reward coefficients $\lambda,\delta,\varepsilon$, identical action sets, and identical seeds would settle the performance claim: if the gap between 11.89 and 11.84 penalized logP (or 0.951 and 0.948 QED) is within run-to-run standard deviation, the claimed improvement is falsified. For the abstract's binding-affinity claim, docking or measuring the affinity of GraphTRL-generated versus baseline-generated molecules against a fixed protein target would provide a direct test.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that persistent homology and multiscale weighted colored graphs can be folded directly into a reinforcement-learning state so that the agent's decisions reflect both local atom environments and global molecular topology. GraphTRL represents each molecule as the concatenation of MWCG features built from SYBYL atom types and a generalized-exponential distance kernel, a persistent-image vector from the molecule's filtration, and molecular fingerprints. The MDP acts by atom addition, bond addition, and bond removal, and the reward function combines penalized logP with penalties for violating a Tanimoto-similarity threshold $\delta$ and a Betti-number threshold $\varepsilon$, so the agent is pushed to optimize logP while preserving the scaffold. The experiments report top scores of 11.89 penalized logP and 0.951 QED with 100% validity, the best in the comparison table. The paper's abstract additionally claims better binding-affinity prediction, an outcome not directly measured in the presented experiments.

Load-bearing premise

The performance comparison assumes the Table 1 baseline scores were produced under the same protocol as GraphTRL—same action space, reward coefficients, starting molecules, and random seeds—but the paper reports them as fixed numbers and does not state that all methods were re-run under identical conditions.

Editorial extensions

If this is right

Because the constrained reward $R_1$ can be swapped for the target reward $R_2$, the same trained pipeline addresses both lead optimization (high penalized logP, preserved scaffold) and targeted generation (prescribed Betti number or molecular weight).
SYBYL atom-type coloring makes the state sensitive to bond-order changes, so a valid action such as converting a $C.1$ to a $C.2$ atom changes the features; the agent can learn the chemical consequences of its own edits.
The 100% chemical validity reported for GraphTRL means the atom/bond action space plus structural state avoids the invalid-molecule failures of sequence-based generators, lowering the cost of exploration.
On the reported numbers, the added topological features do not sacrifice drug-likeness: GraphTRL's QED and penalized logP exceed those of all compared baselines, including the previous best MolDQN-bootstrap.

Reading between the lines

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

The abstract's binding-affinity claim is not tested in the experiments, which optimize penalized logP and QED; a direct follow-up would be to train a supervised affinity predictor on the same MWCG and persistent-image features and then score GraphTRL-generated molecules against a protein–ligand benchmark.
Table 1 reports top-three scores rather than distributions over runs, so a meaningful next step is to report means and standard deviations across seeds; if the 0.05 penalized logP and 0.003 QED gaps over MolDQN-bootstrap fall inside run-to-run noise, the practical advantage would shrink.
The MWCG kernel width $\eta_{kk'}$, exponent $\kappa$, and distance cutoff $c$ are tunable per target; this suggests a natural extension in which the state representation is specialized to a binding site's geometry, which could make the structural reward more informative than generic drug-likeness.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's my take on GraphTRL.

The genuinely useful part is small but real: the author uses SYBYL atom types inside the MWCG state so that the RL agent can literally see when an action changes an atom's hybridization state. That's a sensible way to incorporate structural awareness into a MolDQN-style optimizer, and I haven't seen that exact combination before. The state-construction pipeline is described clearly, and the choice of dueling DQN is reasonable.

What's new is an extension, not a new paradigm. The MDP, actions, and reward structure are borrowed from MolDQN and JT-VAE, and the descriptors (MWCG, persistent images, fingerprints) all come from prior work. That's fine, but it means the contribution should be framed as 'does this state representation help?' rather than 'a new framework.'

The problems start in the abstract. It claims GraphTRL 'outperforms existing methods in binding affinity prediction,' but there is no binding-affinity experiment anywhere in the paper. No docking scores, no PDBbind, no IC50, no comparison to the affinity predictors cited in [6,7]. The experiment section reports only penalized logP and QED. That gap isn't a missing robustness check; it's the central claim left completely untested. Even if Table 1 were perfectly controlled, it would say nothing about binding affinity.

The Table 1 comparison is also shaky on its own terms. The baselines are reported as fixed numbers, and the paper doesn't state whether they were re-run under identical action spaces, reward coefficients, similarity thresholds, or seeds. With no standard deviations, a 0.05 gain in penalized logP and 0.003 in QED can't be distinguished from noise. And there's no code or hyperparameter list, so the empirical claim isn't reproducible.

On the plus side, there's no circularity problem: the benchmarks are external, and the results are not derived from fitted parameters. The method section is readable and the idea of tracking atom-type changes is actually clever.

Who should read it? Someone working on descriptor engineering for RL-based molecular design might find the state representation worth a look. But as submitted, the paper does not support its headline claim. I would desk reject it in its current form, and tell the authors to either add a real binding-affinity evaluation or rewrite the abstract to match the optimization results. If they do that, it could become a modest workshop paper.

Recommendation: don't send to peer review as is. Ask for major revision first.

Referee Report

3 major / 8 minor

Summary. The paper proposes GraphTRL, a reinforcement learning framework for molecular optimization that augments SMILES-based states with multiscale weighted colored graphs (MWCG) and persistent images derived from persistent homology. The agent uses a dueling DQN with actions for atom addition, bond addition, and bond removal, and reward functions based on penalized logP, synthetic accessibility, Betti numbers, and Tanimoto similarity. Section 3 reports an experiment on 800 ZINC molecules, comparing top-scoring molecules against JT-VAE, GCPN, ORGAN, and MolDQN variants in terms of penalized logP and QED. The abstract claims that GraphTRL 'outperforms existing methods in binding affinity prediction,' but no binding-affinity metric or benchmark appears in the paper.

Significance. The combination of topological/graph features with RL for molecular generation is plausible and builds on established feature-engineering work in binding-affinity prediction. However, the paper does not substantiate the central claim, and the optimization gains are marginal (0.05 in penalized logP and 0.003 in QED over MolDQN-bootstrap) with no statistical analysis. The manuscript is too incomplete to establish a nontrivial advance, even if the underlying idea has merit.

major comments (3)

[Abstract; Section 3, Table 1] The abstract states that GraphTRL 'outperforms existing methods in binding affinity prediction,' yet the only experimental evaluation, Table 1, reports penalized logP and QED. No binding-affinity experiment appears anywhere: there is no docking score, no PDBbind or other affinity benchmark, no IC50/Kd prediction, and no comparison to EISA-Score [6] or PDFL-ML [7], which are cited as affinity-prediction methods. Moreover, neither reward function in Section 2.5 includes an affinity term. Thus the paper's central claim is completely unaddressed by its experiments; even a perfectly controlled replication of Table 1 would not support the abstract's binding-affinity assertion.
[Section 3, Table 1] The comparison to baselines is not established as a controlled experiment. The text says the setup is 'inspired by the experimental setup of [13] and [5]' but does not state that JT-VAE, GCPN, ORGAN, or the MolDQN variants were re-run under the same action space, reward coefficients (lambda, delta, epsilon), molecule subsets, or random seeds. The reported advantages of GraphTRL (11.89 vs. 11.84 in penalized logP; 0.951 vs. 0.948 in QED) are within typical run-to-run variability for such methods, and no standard deviations, confidence intervals, or number of seeds are provided. Without this information, the table cannot be interpreted as evidence of superiority.
[Section 2, Eqs. (4), (5), (7); Section 3] The manuscript omits many implementation details needed to reproduce or assess the method. The MWCG parameters (kernel width eta_kk', distance cutoff c, exponent kappa), the persistent-image grid and resolution, the dueling DQN architecture, learning rate, exploration schedule, and the reward coefficients lambda, delta, epsilon, and w are never specified. Since the central claim is empirical, the absence of these details makes the reported results non-reproducible and leaves open the possibility that the performance depends sensitively on choices not described in the paper.

minor comments (8)

[Throughout] The text uses 'SMILE' instead of 'SMILES' in Sections 2.3 and 3; please correct throughout.
[Eq. (4)] The exponent is written with both kappa and k: the definition 'kappa > 0' is inconsistent with the formula using k. Please clarify the notation and define kappa (or k) explicitly.
[Section 2.5.1, Eq. (7)] The term B(m, m0) is used to represent topological complexity or a Betti-number difference, but its precise definition is not given. Please define how the Betti number of a molecule is computed and how the difference between m and m0 is measured.
[Table 1] The acronym QED is not defined in the text; please provide the full term (quantitative estimate of drug-likeness) and a citation on first use.
[Table 1] For MolDQN-twosteps, the penalized logP entries are listed as '-'. Please state explicitly why these values are missing (e.g., not reported in the original reference) rather than leaving the reader to infer it.
[Section 2.4] The phrase 'duel deep neural network' should be 'dueling deep neural network'.
[Figure 4] The caption says the figure shows a flowchart of state construction, but the figure is not described in the text. Please add a short explanation of the components and how they are concatenated.
[Section 3] The phrase 'the model was run on each molecule for a single episode' is ambiguous. Please clarify whether one episode means one optimization trajectory of a fixed maximum number of steps and, if so, what that maximum is.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GraphTRL's reported optima are produced by its own reward-driven RL loop against external baselines, and the abstract's binding-affinity claim is unsupported by Table 1 but that is missing evidence, not circularity.

full rationale

The derivation chain is operationally self-contained: the state space is built from MWCG, persistent images, and molecular fingerprints; the policy is a dueling DQN; the rewards R1 and R2 are explicit functions of Penalized logP, Tanimoto similarity, and Betti number; and the evaluation compares the resulting molecules against external baselines (JT-VAE, GCPN, ORGAN, MolDQN). No parameter is fitted to the reported benchmark and then renamed as a prediction, no load-bearing premise is imported from the author's own prior citations, and no uniqueness theorem is invoked to force the chosen representation. The closest thing to a self-referential step is that Penalized logP appears both inside R1 and as a headline metric in Table 1, meaning the top Penalized logP values partly reflect the optimization target rather than an independent prediction; however, that is the standard objective of the benchmark and does not make the comparison circular. The abstract's stronger claim that GraphTRL 'outperforms existing methods in binding affinity prediction' is not backed by any binding-affinity metric in the experiments, since Table 1 reports only Penalized logP and QED; this is a validation gap and a correctness risk, but not a circular reduction of the derivation to its own inputs. Therefore the circularity score is 0.

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

The paper contributes no new mathematical or physical entity. It assembles existing MWCG, persistent homology, and fingerprint descriptors into an RL state and tunes the reward function, with all key coefficients and hyperparameters undisclosed.

free parameters (4)

Reward coefficients lambda, delta, epsilon in Eq. (7) = not reported
R1(m) uses lambda to balance penalized logP against Tanimoto similarity threshold delta and Betti number threshold epsilon; these values control which optimized molecules are accepted and must be set by the user. The paper never reports them.
MWCG kernel parameters kappa and eta_kk' (Eq. 4) = not reported
The generalized exponential kernel defines multiscale edge weights and requires choosing the scale parameter eta per atom pair and exponent kappa; these shape the state features and are not specified.
Persistent image grid parameters = not reported
Converting persistence diagrams to persistent images requires a resolution, weighting function, and smoothing width; none are given.
Dueling DQN hyperparameters = not reported
Learning rate, discount factor, exploration schedule, network width, batch size, and training steps are absent; these directly affect the reported optimization quality.

assumptions (4)

domain assumption MWCG with SYBYL atom types and generalized exponential kernels is a valid state representation for RL (Eqs. 1-5)
The paper imports MWCG from prior work and assumes the resulting features reflect chemically meaningful modifications after each action; no ablation or validation of the representation is provided.
domain assumption Persistent homology invariants, encoded as persistent images, add useful structural signal beyond molecular fingerprints
Section 2.2 lists general advantages of topology but no experiment isolates the contribution of persistent images to the RL policy.
domain assumption Penalized logP, QED, Betti number, and Tanimoto similarity are valid reward and constraint targets for drug-likeness
Eqs. (6)-(8) adopt these metrics from prior work; no justification links them to binding affinity, despite the abstract's affinity claim.
domain assumption The benchmark protocol matches [5,13] so external baseline numbers are directly comparable
Section 3 states the setup is 'inspired by' [13] and [5], but does not report re-running baselines or identical seeds, action sets, and reward coefficients.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Enhancing Molecular Design through Graph-based Topological Reinforcement Learning." pith.science (2026). https://pith.science/paper/6UFAD6WS

@misc{pith2026241114726,
  author       = {Pith},
  title        = {Pith review of: Enhancing Molecular Design through Graph-based Topological Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6UFAD6WS}},
  note         = {Machine review of arXiv:2411.14726}
}

read the original abstract

The generation of drug-like molecules is crucial for drug design. Existing reinforcement learning (RL) methods often overlook structural information. However, feature engineering-based methods usually merely focus on binding affinity prediction without substantial molecular modification. To address this, we present Graph-based Topological Reinforcement Learning (GraphTRL), which integrates both chemical and structural data for improved molecular generation. GraphTRL leverages multiscale weighted colored graphs (MWCG) and persistent homology, combined with molecular fingerprints, as the state space for RL. Evaluations show that GraphTRL outperforms existing methods in binding affinity prediction, offering a promising approach to accelerate drug discovery.

Figures

Figures reproduced from arXiv: 2411.14726 by the authors.

**Figure 2.** Example of Persistent Images 2.3 State Space and Action Space For each SMILE, the Markov decision process (MDP) is denoted as MDP(S, A, Psa, R). S denotes the state space, the concatenation of multiscale weighted colored graph feature, persistent image (in the form of a feature vector), and molecular fingerprint. A denotes the actions space, in which each action a represent the modifications given SMILE. Actions fal… view at source ↗

**Figure 3.** Visualization of Action: performing sequential actions including atom addition, bond addition, and bond removal on the structure COc1cc2c(c1OC)CC([NH3+])C2 to explore its optimization potential 3 [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: State Construction 2.4 Q-value Function We utilize a duel deep neural network to approximate the Q function, Q(si , ai). The dueling network structure separates the estimation of the state value V and the advantage for each action, expressed as Q (si , ai) = V (si) +A …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 14 canonical work pages

[6]

Eisa-score: Element interactive surface area score for protein–ligand binding affinity prediction

Md Masud Rana and Duc Duy Nguyen. Eisa-score: Element interactive surface area score for protein–ligand binding affinity prediction. Journal of Chemical Information and Modeling, 62 (18):4329–4341, 2022. 1 5 Enhancing Molecular Design through Graph-based Topological Reinforcement Learning

work page 2022
[7]

Persistent Directed Flag Laplacian (PDFL)-Based Machine Learning for Protein-Ligand Binding Affinity Prediction

Mushal Zia, Benjamin Jones, Hongsong Feng, and Guo-Wei Wei. Persistent directed flag laplacian (pdfl)-based machine learning for protein-ligand binding affinity prediction. arXiv preprint arXiv:2411.02596, 2024. 1

work page Pith review arXiv 2024
[13]

Junction tree variational autoencoder for molecular graph generation

Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Junction tree variational autoencoder for molecular graph generation. In International conference on machine learning, pages 2323–2332. PMLR, 2018. 4, 5

work page 2018
[5]

Optimization of molecules via deep reinforcement learning

Zhenpeng Zhou, Steven Kearnes, Li Li, Richard N Zare, and Patrick Riley. Optimization of molecules via deep reinforcement learning. Scientific reports, 9(1):10752, 2019. 1, 5

work page 2019
[1]

Deep reinforcement learning for multiparameter optimization in de novo drug design

Niclas Ståhl, Goran Falkman, Alexander Karlsson, Gunnar Mathiason, and Jonas Bostrom. Deep reinforcement learning for multiparameter optimization in de novo drug design. Journal of chemical information and modeling, 59(7):3166–3176, 2019. 1

work page 2019
[2]

Molecular de-novo design through deep reinforcement learning

Marcus Olivecrona, Thomas Blaschke, Ola Engkvist, and Hongming Chen. Molecular de-novo design through deep reinforcement learning. Journal of cheminformatics, 9:1–14, 2017

work page 2017
[3]

Objective-reinforced generative adversarial networks (organ) for sequence generation models

Gabriel Lima Guimaraes, Benjamin Sanchez-Lengeling, Carlos Outeiral, Pedro Luis Cunha Farias, and Alán Aspuru-Guzik. Objective-reinforced generative adversarial networks (organ) for sequence generation models. arXiv preprint arXiv:1705.10843, 2017

arXiv 2017
[4]

Reinforced adversarial neural computer for de novo molecular design

Evgeny Putin, Arip Asadulaev, Yan Ivanenkov, Vladimir Aladinskiy, Benjamin Sanchez- Lengeling, Alán Aspuru-Guzik, and Alex Zhavoronkov. Reinforced adversarial neural computer for de novo molecular design. Journal of chemical information and modeling, 58(6):1194–1204,

work page

Show all 15 references

[8]

A review of geometric, topological and graph theory apparatuses for the modeling and analysis of biomolecular data

Kelin Xia and Guo-Wei Wei. A review of geometric, topological and graph theory apparatuses for the modeling and analysis of biomolecular data. arXiv preprint arXiv:1612.01735, 2016. 2

2016 arXiv
[9]

Geometric graph learning with extended atom-types features for protein-ligand binding affinity prediction

Md Masud Rana and Duc Duy Nguyen. Geometric graph learning with extended atom-types features for protein-ligand binding affinity prediction. arXiv e-prints, pages arXiv–2301, 2023. 2

2023
[10]

Pedro J Ballester, Adrian Schreyer, and Tom L Blundell. Does a more precise chemical description of protein–ligand complexes lead to more accurate prediction of binding affinity? Journal of chemical information and modeling, 54(3):944–955, 2014. 2

2014
[11]

Fast and anisotropic flexibility-rigidity index for protein flexibility and fluctuation analysis

Kristopher Opron, Kelin Xia, and Guo-Wei Wei. Fast and anisotropic flexibility-rigidity index for protein flexibility and fluctuation analysis. The Journal of chemical physics, 140(23), 2014. 2

2014
[12]

A value-based deep reinforcement learning model with human expertise in optimal treatment of sepsis

XiaoDan Wu, RuiChang Li, Zhen He, TianZhi Yu, and ChangQing Cheng. A value-based deep reinforcement learning model with human expertise in optimal treatment of sepsis. NPJ Digital Medicine, 6(1):15, 2023. 4

2023
[14]

Mol-cyclegan: a generative model for molecular optimization

Łukasz Maziarka, Agnieszka Pocha, Jan Kaczmarczyk, Krzysztof Rataj, Tomasz Danel, and Michał Warchoł. Mol-cyclegan: a generative model for molecular optimization. Journal of Cheminformatics, 12(1):2, 2020. 4

2020
[15]

Zinc: a free tool to discover chemistry for biology

John J Irwin, Teague Sterling, Michael M Mysinger, Erin S Bolstad, and Ryan G Coleman. Zinc: a free tool to discover chemistry for biology. Journal of chemical information and modeling, 52 (7):1757–1768, 2012. 5 6

2012

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [6]

Eisa-score: Element interactive surface area score for protein–ligand binding affinity prediction

Md Masud Rana and Duc Duy Nguyen. Eisa-score: Element interactive surface area score for protein–ligand binding affinity prediction. Journal of Chemical Information and Modeling, 62 (18):4329–4341, 2022. 1 5 Enhancing Molecular Design through Graph-based Topological Reinforcement Learning

work page 2022

[2] [7]

Persistent Directed Flag Laplacian (PDFL)-Based Machine Learning for Protein-Ligand Binding Affinity Prediction

Mushal Zia, Benjamin Jones, Hongsong Feng, and Guo-Wei Wei. Persistent directed flag laplacian (pdfl)-based machine learning for protein-ligand binding affinity prediction. arXiv preprint arXiv:2411.02596, 2024. 1

work page Pith review arXiv 2024

[3] [13]

Junction tree variational autoencoder for molecular graph generation

Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Junction tree variational autoencoder for molecular graph generation. In International conference on machine learning, pages 2323–2332. PMLR, 2018. 4, 5

work page 2018

[4] [5]

Optimization of molecules via deep reinforcement learning

Zhenpeng Zhou, Steven Kearnes, Li Li, Richard N Zare, and Patrick Riley. Optimization of molecules via deep reinforcement learning. Scientific reports, 9(1):10752, 2019. 1, 5

work page 2019

[5] [1]

Deep reinforcement learning for multiparameter optimization in de novo drug design

Niclas Ståhl, Goran Falkman, Alexander Karlsson, Gunnar Mathiason, and Jonas Bostrom. Deep reinforcement learning for multiparameter optimization in de novo drug design. Journal of chemical information and modeling, 59(7):3166–3176, 2019. 1

work page 2019

[6] [2]

Molecular de-novo design through deep reinforcement learning

Marcus Olivecrona, Thomas Blaschke, Ola Engkvist, and Hongming Chen. Molecular de-novo design through deep reinforcement learning. Journal of cheminformatics, 9:1–14, 2017

work page 2017

[7] [3]

Objective-reinforced generative adversarial networks (organ) for sequence generation models

Gabriel Lima Guimaraes, Benjamin Sanchez-Lengeling, Carlos Outeiral, Pedro Luis Cunha Farias, and Alán Aspuru-Guzik. Objective-reinforced generative adversarial networks (organ) for sequence generation models. arXiv preprint arXiv:1705.10843, 2017

arXiv 2017

[8] [4]

Reinforced adversarial neural computer for de novo molecular design

Evgeny Putin, Arip Asadulaev, Yan Ivanenkov, Vladimir Aladinskiy, Benjamin Sanchez- Lengeling, Alán Aspuru-Guzik, and Alex Zhavoronkov. Reinforced adversarial neural computer for de novo molecular design. Journal of chemical information and modeling, 58(6):1194–1204,

work page