MolEditRL: Structure-Preserving Molecular Editing via Discrete Diffusion and Reinforcement Learning

Dazhong Shen; Ying Sun; Yuanxin Zhuang

REVIEW 4 major objections 6 minor 43 references

MolEditRL: Structure-Preserving Molecular Editing via Discrete Diffusion and Reinforcement Learning

T0 review · 4 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 MolEditRL claims that discrete graph diffusion pretraining followed by KL-regularized reinforcement learning yields state-of-the-art structure-preserving molecular editing with far fewer parameters than large language models.

desk verdict Sensible graph-diffusion + RL editing system and a large new dataset, but the headline 74% margin sits on a self-built testbed that needs external confirmation. read the letter →

arxiv 2505.20131 v1 pith:ILM4QR7C submitted 2025-05-26 cs.LG q-bio.QM

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

classification cs.LGq-bio.QM

keywords moleculareditingdiscretegraphdiffusionreinforcementlearningstructurepreservationpropertyoptimizationinstructionfollowingbenchmarkdatasetdrugdiscovery

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

MolEditRL sets out to show that molecular editing is best done on the molecular graph itself, not on a text string. The paper builds a compact transformer that reads a source molecule, a target molecule, and a natural-language instruction together, injects bond connectivity directly into attention, and is trained in two stages: first a discrete diffusion pass that reconstructs targets from masked graphs, then a reinforcement-learning pass that rewards property changes while a KL term pulls the output back toward the pretrained structure-aware distribution. To evaluate this, the authors construct MolEdit-Instruct, three million source–target pairs spanning ten chemical properties and twenty single- and multi-property editing tasks. On that benchmark they report the highest accuracy at strict and relaxed similarity thresholds across all tasks, the lowest distributional distance, a 74% improvement in editing success rate over leading baselines, and a parameter count 98% smaller than the largest comparison model. The caring point is that precise, localized molecular edits may not need a very large language model at all.

What carries the argument

The workhorse is the Structure-Preserving Editing Network, a transformer with a graph-aware attention bias: at every layer the bias re-injects the source molecule's adjacency, at the first layer it injects the target adjacency, and at later layers it reuses the previous layer's attention, so the model's information flow is constrained by molecular topology rather than by token order alone. Around that network sits a two-stage training loop. Stage one is x0-parameterized discrete diffusion: a forward process masks atoms and bonds of the target graph, and the network learns to predict the clean graph conditioned on source graph and instruction, with an auxiliary cross-entropy term that keeps the instruction tokens aligned. Stage two is editing-aware reinforcement learning: the denoising trajectory is treated as an MDP, rewards from chemical property oracles arrive only at the final molecule, and a policy-gradient objective with x0-parameterization turns those rewards into a weighted cross-entropy loss over atoms and bonds, regularized by the KL divergence to the pretrained structure-aware prior. The KL term is the mechanism that prevents reward chasing from destroying the scaffold.

What would settle it

Run MolEditRL and the compared baselines on an independently constructed test set whose source molecules have low Tanimoto similarity to every MolEdit-Instruct training molecule, with property changes and instructions drawn from a different source; if the reported margin over the best baseline shrinks or disappears, the claim of general structure-preserving editing capability would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that a discrete graph diffusion model fine-tuned with a KL-regularized, editing-aware reinforcement-learning objective can edit molecules both accurately and structure-preservingly, outperforming larger sequence-based and instruction-tuned language models. MolEditRL encodes instruction tokens and source/target atom tokens into one sequence, adds learnable biases that carry bond-level connectivity into the attention scores, and pretrains the resulting network to reverse a masking process that corrupts atom and bond types of the target graph. Fine-tuning then frames each reverse-diffusion step as an action in a Markov decision process, scores the final molecule with deterministic chemical property oracles, normalizes rewards per batch, and optimizes a reward-weighted cross-entropy loss over atoms and bonds with an extra KL term toward the pretrained prior; gradients are taken at a fixed stride of denoising steps to keep training stable. Across the MolEdit-Instruct benchmark the paper reports the best editing accuracy at Tanimoto thresholds 0.65 and 0.15 for every single- and multi-property task, the best Fréchet ChemNet Distance, and a 74% relative gain in editing success rate while using about 2% of the parameters of the largest compared model.

Load-bearing premise

The load-bearing premise is that the paper's own MolEdit-Instruct benchmark, built from matched molecular pairs and template-generated prompts, fairly represents molecular editing, so that the advantage measured on it will transfer to editing tasks whose molecules and instructions were not generated by the same pipeline.

Editorial extensions

If this is right

Editing at a strict Tanimoto threshold of 0.65 is reported highest for MolEditRL on every single- and multi-property task, so if correct, instruction-following edits that stay close to the source scaffold are achievable without a large language model.
The same pretrained backbone adapts to three held-out properties by swapping in new reward oracles, so new optimization objectives can be added after deployment without retraining the diffusion model.
Because the model uses about 2% of the parameters of the largest compared model, practical molecular editing systems could run on much smaller infrastructure while maintaining or improving fidelity.
KL regularization against the pretrained denoiser keeps validity and similarity stable through RL fine-tuning, which is the property that makes the two-stage recipe usable for multi-objective tasks with three or four simultaneous constraints.

Reading between the lines

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

Editorial inference: because the benchmark pairs are generated by fragment-based matched molecular pairs, the reported gains may be concentrated in local edits such as substituent swaps; testing on edits that require ring-system changes would clarify whether the structure-preservation claim holds for larger topological rewrites.
Editorial inference: the recipe—graph-conditioned diffusion pretraining plus KL-regularized policy fine-tuning with deterministic oracle rewards—is portable to other discrete structured prediction tasks wherever a cheap oracle exists.
Editorial inference: a head-to-head ablation that keeps the RL objective identical but replaces the graph decoder with a SELFIES or SMILES decoder would isolate how much of the 74% gain comes from the graph representation itself rather than from the reward design.

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

Dear Colleague,

The thing to know: MolEditRL is a competent two-stage molecular editing system, and MolEdit-Instruct is a potentially useful dataset. The 74% improvement headline, though, is only as credible as the self-built benchmark it's measured on, and the paper doesn't give enough to check it.

What's actually new is modest but real. The structure-aware attention bias—source adjacency injected at all layers, target adjacency at the first layer, then propagated—is a simple and sensible way to bias discrete graph diffusion toward scaffold preservation. Combining that with KL-regularized x0-parameterized policy-gradient fine-tuning on an instruction-conditioned diffusion model is not in the cited prior work. The dataset itself is a genuine asset if it ships: 3M MMP/BRICS pairs, 10 properties, single- and multi-property prompts. That alone is worth something.

The methods section is coherent and the mechanisms check out. Using GDPO-style gradient estimation with a KL term to keep the fine-tuned policy near the pretrained prior is a sound way to balance reward and structure. I believe the system does what the authors say.

The soft spot is the evaluation, and it's load-bearing. The test pairs are generated by the same MMP/BRICS pipeline as training, and baselines are not adapted to those templates: GeLLM3O-P(6) is tested on ten properties after training on six; DrugAssist, BioT5, and REINVENT4 are essentially used as released. In that setting, the large margins in Tables 1–7 could reflect train/test distribution alignment more than general editing ability. Every number is a point estimate—no seeds, no error bars. The "unseen property" experiments still use pairs constructed the same way, and only MolEditRL is fine-tuned to those tasks. Also, the reward in Eq. 8 is the property oracle alone; structural preservation is carried by the KL prior, not by an explicit similarity term. That is a defensible design choice, but it means the structural-fidelity claim is softer than the name implies. Finally, the abstract says the dataset is on Hugging Face, but I see no URL or code link anywhere in the text.

I'd send this to review rather than desk-reject. The system is sensible, and the dataset could be a real contribution. But the referee needs to push on benchmark fairness, request multiple seeds, released code/data, an external test set, and properly adapted baselines before the SOTA claim can be taken at face value.

Best,

Referee Report

4 major / 6 minor

Summary. The paper proposes MolEditRL, a two-stage molecular editing framework. Stage one pretrains a RoBERTa-based transformer with a structure-aware attention bias over a discrete masked graph diffusion process, reconstructing target molecular graphs from source graphs plus natural language instructions. Stage two fine-tunes the diffusion policy with a KL-regularized reinforcement learning objective, using property rewards computed by RDKit and TDC oracles, with gradient estimation via x0-parameterization. To evaluate the method, the authors construct MolEdit-Instruct, a dataset of roughly 3 million source–target molecular pairs over 10 chemical properties and 20 single-property editing tasks plus multi-property combinations, generated by MMP/BRICS fragmentation with template prompts. Experiments report that MolEditRL outperforms BioT5, DrugAssist, GeLLM3O, MolGen, and REINVENT4 on editing accuracy at Tanimoto thresholds 0.65 and 0.15, achieves lower FCD, and does so with substantially fewer parameters, including a claimed 74% improvement in editing success rate.

Significance. If the comparative evaluation is valid, the paper makes a useful contribution: it explicitly models the discrete graph structure of molecules during editing, introduces a structure-preserving RL fine-tuning stage with a KL penalty to the pretrained prior, releases a large public dataset, and reports consistently strong numbers across single- and multi-property tasks. The use of standard oracles (RDKit, TDC), the relatively small model footprint, and the qualitative scaffold-preservation results are concrete strengths. However, the central SOTA claim rests on a self-constructed benchmark whose test distribution is generated by the same pipeline as the training data and on which the baselines are not adapted, and all reported results are single-run point estimates. These issues are load-bearing for the headline 'significantly outperforms' claim, so the conclusions are plausible but not yet established.

major comments (4)

[§4.1, Appendix C] The central SOTA claim relies entirely on MolEdit-Instruct, which is constructed by the authors from MMP/BRICS pairs and the template prompts in Table 4. Because the test set is sampled from the same pipeline and prompt templates as the training set, while the baselines (DrugAssist, BioT5, REINVENT4, MolGen, GeLLM3O) are used without template-specific adaptation—and GeLLM3O-P(6) was trained on six properties but evaluated on ten—the comparison conflates model capability with train/test distribution alignment. The reported 74% margin in the Abstract may therefore be a benchmark artifact. I would need an evaluation on at least one independently constructed benchmark (e.g., MolOpt-Instructions or MuMOInstruct) or an external validation set, plus baseline fine-tuning or adapter training on the MolEdit-Instruct templates, before the headline comparative claim is supported.
[Tables 1, 5, 6, 7; Figures 3–5] All quantitative results are single-run point estimates; no error bars, confidence intervals, or multiple seeds are reported anywhere in the main text or appendix. The phrases 'significantly outperforms' (Abstract, Section 1) and 'consistently achieves the highest' (Section 4.4) are used without a statistical test. Given the noise inherent in RL fine-tuning and stochastic diffusion sampling, the reported margins over DrugAssist and GeLLM3O cannot be assessed for robustness. Please report at least three seeds with standard deviations for all main metrics, and a paired significance test for the principal comparisons.
[§3.4, Eq. (8), Table 2] The reward function is underspecified. Eq. (8) defines r only as 'property improvements through chemical-rule-based oracles', and Table 2 lists only 'Reward Calculation (0, 0.2, 1.0)' with no mapping to the binary oracles. It is not clear whether r includes a Tanimoto-similarity term, how a direction/magnitude threshold for each property is applied, how multi-property rewards are aggregated, or how invalid molecules are penalized. Because the same oracles and thresholds appear to underlie the Accall evaluation metrics, the RL objective and the evaluation metric may be aligned by construction; please spell out the exact reward formula and confirm that the evaluation thresholds are not used to shape the reward during training.
[Eqs. (9)–(14)] The gradient derivation is difficult to follow and appears internally inconsistent as written. Eq. (9) places the KL penalty outside the reward expectation, but Eq. (10) includes both an advantage-weighted term and a negative KL derivative; Eq. (14) uses the uncentered reward r while Eq. (10) uses the normalized advantage A-hat. Since this objective is the core training mechanism, please state the exact surrogate loss used in the implementation, clarify the sign convention of the KL term, and explain how the x0-parameterization in Eqs. (12)–(13) is converted into the reward-weighted cross-entropy loss in Eq. (14).

minor comments (6)

[Abstract, Section 1] The '74% improvement' and '98% fewer parameters' claims need a concrete reference: which baseline and which metric define the 74%, and what is the parameter count of the comparison model?
[Figure 5(b) caption] The caption reads 'FCD at a fixed threshold of 0.15', but FCD is not thresholded by Tanimoto similarity; this appears to be a typo that should be corrected.
[Table 3] The explanation of the ± notation is confusing: 'positive values for property reduction and negative values for property increase' inverts the usual arrow convention (↑/↓). Please clarify the sign convention for each property direction.
[Section 4.2] The baseline description notes that GeLLM3O-P(6) was trained on six properties but is evaluated on ten. If the baseline cannot be fine-tuned to the full property set, this limitation should be stated in the table caption so readers can weigh the comparison fairly.
[Appendix E] The 'unseen property' experiments are not zero-shot: MolEditRL is fine-tuned with property-specific oracles for BBBP, HIA, and hERG. The text should be reworded to say 'adaptation to new reward oracles' rather than implying generalization without any task-specific training.
[Eq. (3)] The attention bias indexing assumes contiguous source and target atom positions; please clarify how padding, variable-length graphs, and graphs with different node orderings are handled in the implementation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MolEditRL's results are benchmark-specific but not definitionally derived from its inputs.

full rationale

The derivation chain is self-contained. The pretraining loss (Eq. 7) is a conditional masked-graph reconstruction objective over source molecule, instruction, and target graph; the RL objective (Eqs. 9–14) maximizes rewards computed by external RDKit/TDC oracles and is evaluated with independent Tanimoto-threshold accuracy and FCD metrics on a test distribution, not by fitting those metrics into the training loss. No parameter is fitted to the evaluation labels and then reported as a prediction; no equation defines a predicted quantity in terms of the target quantity. The benchmark MolEdit-Instruct is constructed by the authors, so the 74% margin and SOTA comparisons are specific to this self-built testbed and should be validated externally, but that is a dataset-fairness/generalization concern, not circularity. The paper contains no load-bearing self-citations: the cited baselines (DrugAssist, GeLLM3O, BioT5, REINVENT4) and method components (DDPO, GDPO, RoBERTa, RDKit, TDC) are external works. Therefore the central claim does not reduce by construction to its inputs.

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

The central claim rests on the fidelity of RDKit/TDC oracles, the x0-parameterization approximation, the KL-prior-as-structural-constraint assumption, and the representativeness of the MMP/BRICS dataset construction. No new physical entities are introduced. The model is a parameterized network trained on the authors' dataset, so the main unpaid inputs are the benchmark and the oracle approximations.

free parameters (5)

KL regularization coefficient beta = not reported
Balances reward maximization against the pretrained prior in Eq. 9; its value is chosen by hand and never stated in Appendix B.
Diffusion steps T and policy update stride t_s = T=2000, t_s=50
Manual choices that define the MDP and how often gradients are computed; results in Figure 4 depend on them.
Similarity thresholds tau = 0.65 and 0.15
Success is defined as property change plus Tanimoto similarity at these thresholds; they are chosen by the authors and strongly affect reported accuracy.
Dataset property-change thresholds = property-specific deltas (e.g., +/-0.050 DRD2 to +/-6.563 SA)
Pairs are retained only if property shifts are substantial and Tanimoto is in 0.650-0.982; these hand-set cutoffs determine task difficulty (Appendix C).
Frequency weighting lambda = 0.3
Sinusoidal frequency weighting in the pretraining loss, chosen by hand (Appendix B).

assumptions (5)

domain assumption RDKit and TDC property oracles are accurate proxies for real chemical and biological properties
The RL reward and all evaluation metrics rely on these computational oracles; for DRD2, GSK3beta, JNK3, BBBP, HIA, hERG the TDC oracles are model-based predictors, not experimental assays (Sections 3.4, 4.3, Appendix E).
domain assumption x0-parameterization approximation in Eq. 12 is accurate enough for policy-gradient estimation
The RL gradient replaces the full reverse transition with a point estimate of G0; the paper invokes Austin et al. without checking approximation error for molecular graphs (Section 3.4).
ad hoc to paper The KL penalty to the pretrained prior preserves structural similarity during RL fine-tuning
No Tanimoto or scaffold term appears in the reward; the paper assumes KL closeness to the pretrained model acts as the structural constraint (Eq. 9, Section 3.4).
domain assumption MMP/BRICS-based pair construction yields a representative distribution of realistic editing tasks
The dataset is built from ZINC/ChEMBL pairs and template prompts; representativeness for real lead-optimization edits is assumed (Section 4.1, Appendix C).
standard math Standard masked discrete diffusion forward process can model molecular graphs
Uses D3PM-style masking with beta(t) = (T-t+1)^(-1); standard but not formalized for graph validity (Section 3.3).

how reviews work

0 comments

Cite this review

Pith. "Pith review of MolEditRL: Structure-Preserving Molecular Editing via Discrete Diffusion and Reinforcement Learning." pith.science (2026). https://pith.science/paper/ILM4QR7C

@misc{pith2026250520131,
  author       = {Pith},
  title        = {Pith review of: MolEditRL: Structure-Preserving Molecular Editing via Discrete Diffusion and Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ILM4QR7C}},
  note         = {Machine review of arXiv:2505.20131}
}

read the original abstract

Molecular editing aims to modify a given molecule to optimize desired chemical properties while preserving structural similarity. However, current approaches typically rely on string-based or continuous representations, which fail to adequately capture the discrete, graph-structured nature of molecules, resulting in limited structural fidelity and poor controllability. In this paper, we propose MolEditRL, a molecular editing framework that explicitly integrates structural constraints with precise property optimization. Specifically, MolEditRL consists of two stages: (1) a discrete graph diffusion model pretrained to reconstruct target molecules conditioned on source structures and natural language instructions; (2) an editing-aware reinforcement learning fine-tuning stage that further enhances property alignment and structural preservation by explicitly optimizing editing decisions under graph constraints. For comprehensive evaluation, we construct MolEdit-Instruct, the largest and most property-rich molecular editing dataset, comprising 3 million diverse examples spanning single- and multi-property tasks across 10 chemical attributes. Experimental results demonstrate that MolEditRL significantly outperforms state-of-the-art methods in both property optimization accuracy and structural fidelity, achieving a 74\% improvement in editing success rate while using 98\% fewer parameters.

Figures

Figures reproduced from arXiv: 2505.20131 by the authors.

**Figure 1.** Performance, FCD, and parameter size comparison. Designing molecules with tailored properties is essential for drug discovery [25]. Unlike de novo molecular generation that creates molecules from scratch [33], molecular editing [14] focuses on precisely modifying existing molecules to optimize targeted properties while preserving known structure-activity relationships [11]. Traditional molecular editing approaches f… view at source ↗

**Figure 2.** Overview of MolEditRL. while selectively updating target representations. For tokens i and j at layer l, the raw attention score Aˆl i,j is computed as: Aˆl i,j = 1 √ dk (h l iWQ) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Performance by number of edited properties. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (10 more)

**Figure 4.** Figure 4: Impact of step size, fine-tuning strategy, and KL regularization. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Performance comparison under structural constraints. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Visualization of edits. Red highlights indicate structural changes from the source. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: More visualization of edits on 20 tasks. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

**Figure 8.** Figure 8: More visualization of edits on 20 tasks. [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]

**Figure 9.** Figure 9: More visualization of edits on 20 tasks. [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]

**Figure 10.** Figure 10: Qualitative comparison of molecular editing methods. [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]

**Figure 11.** Figure 11: Qualitative comparison of molecular editing methods. [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]

**Figure 12.** Figure 12: Qualitative comparison of molecular editing methods. [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]

**Figure 13.** Figure 13: Qualitative comparison of molecular editing methods. [PITH_FULL_IMAGE:figures/full_fig_p025_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 32 canonical work pages

[1]

Josep Arús-Pous, Simon Viet Johansson, Oleksii Prykhodko, Esben Jannik Bjerrum, Christian Tyrchan, Jean-Louis Reymond, Hongming Chen, and Ola Engkvist. 2019. Randomized SMILES strings improve the quality of molecular generative models.Journal of cheminformatics11 (2019), 1–13

work page 2019
[2]

Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg

work page
[3]

A Patrícia Bento, Anne Hersey, Eloy Félix, Greg Landrum, Anna Gaulton, Francis Atkinson, Louisa J Bellis, Marleen De Veij, and Andrew R Leach. 2020. An open source chemical structure curation pipeline using RDKit.Journal of Cheminformatics12 (2020), 1–16

work page 2020
[4]

Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. 2023. Training diffusion models with reinforcement learning.arXiv preprint arXiv:2305.13301(2023)

arXiv 2023
[5]

Ziqi Chen, Martin Renqiang Min, Srinivasan Parthasarathy, and Xia Ning. 2021. A deep generative model for molecule optimization via one fragment modification.Nature machine intelligence3, 12 (2021), 1040–1049

work page 2021
[6]

Andrew Dalke, Jerome Hert, and Christian Kramer. 2018. mmpdb: An open-source matched molecular pair platform for large multiproperty data sets.Journal of chemical information and modeling58, 5 (2018), 902–910

work page 2018
[7]

Nicola De Cao and Thomas Kipf. 2018. MolGAN: An implicit generative model for small molecular graphs.arXiv preprint arXiv:1805.11973(2018)

arXiv 2018
[8]

Vishal Dey, Xiao Hu, and Xia Ning. 2025. GeLLM3O: Generalizing Large Language Models for Multi-property Molecule Optimization.arXiv preprint arXiv:2502.13398(2025)

arXiv 2025

Show all 43 references

[9]

Yin Fang, Ningyu Zhang, Zhuo Chen, Lingbing Guo, Xiaohui Fan, and Huajun Chen. 2024. Domain-Agnostic Molecular Generation with Chemical Feedback. InThe Twelfth International Conference on Learning Representations

2024
[10]

Tianfan Fu, Cao Xiao, Xinhao Li, Lucas M Glass, and Jimeng Sun. 2021. Mimosa: Multi- constraint molecule sampling for molecule optimization. InProceedings of the AAAI Conference on Artificial Intelligence, V ol. 35. 125–133

2021
[11]

Corwin Hansch. 1969. Quantitative approach to biochemical structure-activity relationships. Accounts of chemical research2, 8 (1969), 232–239

1969
[12]

Jiazhen He, Huifang You, Emil Sandström, Eva Nittinger, Esben Jannik Bjerrum, Christian Tyrchan, Werngard Czechtizky, and Ola Engkvist. 2021. Molecular optimization by capturing chemist’s intuition using deep neural networks.Journal of cheminformatics13 (2021), 1–17

2021
[13]

Kexin Huang, Tianfan Fu, Wenhao Gao, Yue Zhao, Yusuf Roohani, Jure Leskovec, Con- nor W Coley, Cao Xiao, Jimeng Sun, and Marinka Zitnik. 2021. Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development.arXiv preprint arXiv:2102.09548(2021)

2021 arXiv
[14]

Chunngai Hui, Zhuo Wang, Shiping Wang, and Chunfa Xu. 2022. Molecular editing in natural product synthesis.Organic Chemistry Frontiers9, 5 (2022), 1451–1457

2022
[15]

Wengong Jin, Regina Barzilay, and Tommi Jaakkola. 2018. Junction tree variational autoencoder for molecular graph generation. InICML. PMLR, 2323–2332

2018
[16]

Wengong Jin, Regina Barzilay, and Tommi Jaakkola. 2020. Hierarchical generation of molecular graphs using structural motifs. InICML. PMLR, 4839–4848

2020
[17]

Mario Krenn, Florian Häse, AkshatKumar Nigam, Pascal Friederich, and Alan Aspuru-Guzik
[18]

Matt J Kusner, Brooks Paige, and José Miguel Hernández-Lobato. 2017. Grammar variational autoencoder. InInternational conference on machine learning. PMLR, 1945–1954

2017
[19]

Khiem Le and Nitesh V Chawla. 2024. Utilizing Large Language Models in an iterative paradigm with domain feedback for molecule optimization.arXiv preprint arXiv:2410.13147 (2024)

2024
[20]

Shengchao Liu, Weili Nie, Chengpeng Wang, Jiarui Lu, Zhuoran Qiao, Ling Liu, Jian Tang, Chaowei Xiao, and Animashree Anandkumar. 2023. Multi-modal molecule structure–text model for text-based retrieval and editing.Nature Machine Intelligence5, 12 (2023), 1447–1457

2023
[21]

Shengchao Liu, Jiongxiao Wang, Yijin Yang, Chengpeng Wang, Ling Liu, Hongyu Guo, and Chaowei Xiao. 2024. Conversational drug editing using retrieval and domain feedback. InThe Twelfth International Conference on Learning Representations

2024
[22]

Yinhan Liu. 2019. Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692364 (2019)

2019 arXiv
[23]

Yijing Liu, Chao Du, Tianyu Pang, Chongxuan Li, Min Lin, and Wei Chen. 2024. Graph diffusion policy optimization.NeurIPS37 (2024), 9585–9611

2024
[24]

Hannes H Loeffler, Jiazhen He, Alessandro Tibo, Jon Paul Janet, Alexey V oronov, Lewis H Mervin, and Ola Engkvist. 2024. Reinvent 4: Modern AI–driven generative molecule design. Journal of Cheminformatics16, 1 (2024), 20

2024
[25]

Chunhua Ma, Craig W Lindsley, Junbiao Chang, and Bin Yu. 2024. Rational molecular editing: a new paradigm in drug discovery. 11459–11466 pages

2024
[26]

Emmanuel Noutahi, Cristian Gabellini, Michael Craig, Jonathan SC Lim, and Prudencio Tossou
[27]

Marcus Olivecrona, Thomas Blaschke, Ola Engkvist, and Hongming Chen. 2017. Molecular de-novo design through deep reinforcement learning.Journal of cheminformatics9 (2017), 1–14

2017
[28]

Qizhi Pei, Wei Zhang, Jinhua Zhu, Kehan Wu, Kaiyuan Gao, Lijun Wu, Yingce Xia, and Rui Yan. 2023. Biot5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations.arXiv preprint arXiv:2310.07276(2023)

2023 arXiv
[29]

Mariya Popova, Olexandr Isayev, and Alexander Tropsha. 2018. Deep reinforcement learning for de novo drug design.Science advances4, 7 (2018), eaap7885

2018
[30]

Kristina Preuer, Philipp Renz, Thomas Unterthiner, Sepp Hochreiter, and Gunter Klambauer
[31]

Chence Shi, Minkai Xu, Zhaocheng Zhu, Weinan Zhang, Ming Zhang, and Jian Tang. 2020. Graphaf: a flow-based autoregressive model for molecular graph generation.arXiv preprint arXiv:2001.09382(2020)

2020 arXiv
[32]

Bhuvanesh Sridharan, Animesh Sinha, Jai Bardhan, Rohit Modee, Masahiro Ehara, and U Deva Priyakumar. 2024. Deep reinforcement learning in chemistry: A review.Journal of Computa- tional Chemistry45, 22 (2024), 1886–1898

2024
[33]

Mingyang Wang, Zhe Wang, Huiyong Sun, Jike Wang, Chao Shen, Gaoqi Weng, Xin Chai, Honglin Li, Dongsheng Cao, and Tingjun Hou. 2022. Deep learning approaches for de novo drug design: An overview.Current opinion in structural biology72 (2022), 135–144

2022
[34]

Zhenxing Wu, Odin Zhang, Xiaorui Wang, Li Fu, Huifeng Zhao, Jike Wang, Hongyan Du, Dejun Jiang, Yafeng Deng, Dongsheng Cao, et al. 2024. Leveraging language model for advanced multiproperty molecular optimization via prompt engineering.Nature Machine Intelligence (2024), 1–11. 11

2024
[35]

Yuran Xiang, Haiteng Zhao, Chang Ma, and Zhi-Hong Deng. 2024. Instruction-Based Molecular Graph Generation with Unified Text-Graph Diffusion Model.arXiv preprint arXiv:2408.09896 (2024)

2024 arXiv
[36]

Yida Xiong, Kun Li, Weiwei Liu, Jia Wu, Bo Du, Shirui Pan, and Wenbin Hu. 2024. Text- Guided Multi-Property Molecular Optimization with a Diffusion Language Model.arXiv preprint arXiv:2410.13597(2024)

2024
[37]

Geyan Ye, Xibao Cai, Houtim Lai, Xing Wang, Junhong Huang, Longyue Wang, Wei Liu, and Xiangxiang Zeng. 2025. Drugassist: A large language model for molecule optimization. Briefings in Bioinformatics26, 1 (2025), bbae693

2025
[38]

Jiaxuan You, Bowen Liu, Zhitao Ying, Vijay Pande, and Jure Leskovec. 2018. Graph con- volutional policy network for goal-directed molecular graph generation.Advances in neural information processing systems31 (2018)

2018
[39]

hydrogen bond acceptors,

Zhenpeng Zhou, Steven Kearnes, Li Li, Richard N Zare, and Patrick Riley. 2019. Optimization of molecules via deep reinforcement learning.Scientific reports9, 1 (2019), 10752. 12 Technical Appendices and Supplementary Material This appendix provides extended technical and exper...

2019
[2018]

Journal of chemical information and modeling58, 9 (2018), 1736–1741

Fréchet ChemNet distance: a metric for generative models for molecules in drug discovery. Journal of chemical information and modeling58, 9 (2018), 1736–1741

2018
[2020]

Self-referencing embedded strings (SELFIES): A 100% robust molecular string represen- tation.Machine Learning: Science and Technology1, 4 (2020), 045024. 10

2020
[2021]

Structured denoising diffusion models in discrete state-spaces.NeurIPS34 (2021), 17981–17993

2021
[2024]

Gotta be SAFE: a new framework for molecular design.Digital Discovery3, 4 (2024), 796–804

2024

Pith tools

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

[1] [1]

Josep Arús-Pous, Simon Viet Johansson, Oleksii Prykhodko, Esben Jannik Bjerrum, Christian Tyrchan, Jean-Louis Reymond, Hongming Chen, and Ola Engkvist. 2019. Randomized SMILES strings improve the quality of molecular generative models.Journal of cheminformatics11 (2019), 1–13

work page 2019

[2] [2]

Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg

work page

[3] [3]

A Patrícia Bento, Anne Hersey, Eloy Félix, Greg Landrum, Anna Gaulton, Francis Atkinson, Louisa J Bellis, Marleen De Veij, and Andrew R Leach. 2020. An open source chemical structure curation pipeline using RDKit.Journal of Cheminformatics12 (2020), 1–16

work page 2020

[4] [4]

Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. 2023. Training diffusion models with reinforcement learning.arXiv preprint arXiv:2305.13301(2023)

arXiv 2023

[5] [5]

Ziqi Chen, Martin Renqiang Min, Srinivasan Parthasarathy, and Xia Ning. 2021. A deep generative model for molecule optimization via one fragment modification.Nature machine intelligence3, 12 (2021), 1040–1049

work page 2021

[6] [6]

Andrew Dalke, Jerome Hert, and Christian Kramer. 2018. mmpdb: An open-source matched molecular pair platform for large multiproperty data sets.Journal of chemical information and modeling58, 5 (2018), 902–910

work page 2018

[7] [7]

Nicola De Cao and Thomas Kipf. 2018. MolGAN: An implicit generative model for small molecular graphs.arXiv preprint arXiv:1805.11973(2018)

arXiv 2018

[8] [8]

Vishal Dey, Xiao Hu, and Xia Ning. 2025. GeLLM3O: Generalizing Large Language Models for Multi-property Molecule Optimization.arXiv preprint arXiv:2502.13398(2025)

arXiv 2025