Pith. sign in

REVIEW 5 major objections 5 minor 56 references

ScaffoldGPT: A Scaffold-based GPT Model for Drug Optimization

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

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

Pith's one-line read ScaffoldGPT beats 8 baselines in drug optimization on two targets.

desk verdict ScaffoldGPT is a competent engineering paper with a useful decoding trick, but its headline benchmark results rest on an ambiguous train/test split and a docking surrogate that is both the reward and the metric. read the letter →

arxiv 2502.06891 v3 pith:WMXSFMJ4 submitted 2025-02-09 q-bio.BM cs.CLcs.LG

Xuefeng Liu , Songhao Jiang , Ian Foster , Jinbo Xu , Rick Stevens This is my paper · ORCID
classification q-bio.BMcs.CLcs.LG
keywords drugoptimizationgenerativepretrainedtransformermolecularscaffoldsreinforcementlearningpolicycontrollabledecodingdockingscoreSMILES
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 ScaffoldGPT, a generative pretrained transformer that optimizes an existing drug by modifying its SMILES string while keeping its molecular scaffold intact. The central claim is that a three-stage pipeline—incremental pretraining on scaffold–molecule pairs, advantage-alignment policy optimization (APO) finetuning, and a token-level Top-N decoding strategy that ranks candidates by reward rather than likelihood—produces the highest average normalized reward on two drug-optimization benchmarks, a COVID protease (3CLPro) and a cancer protein (RTCB). On both targets ScaffoldGPT surpasses the strongest baseline by a small margin (0.675 vs 0.666 and 0.682 vs 0.673), while raising docking, druglikeness, synthesizability, and solubility scores. If these results hold, drug optimization can be treated as a sequence-generation problem where structural preservation and multi-objective improvement are obtained in one framework.

What carries the argument

The central object is the ScaffoldGPT generator itself: a GPT-2-like Transformer with a Byte-Pair-Encoding tokenizer, pretrained in two phases (first on plain molecules, then on scaffold-to-molecule pairs) and then finetuned with APO. The load-bearing mechanism is the three-stage pipeline. The reward function in Eq. (7) normalizes five critics—docking (from a surrogate model), druglikeness, synthesizability, solubility, and Tanimoto similarity—into a single scalar that APO optimizes. The Top-N decoding strategy (Eq. 13) is the distinctive new piece: at each token position it keeps the n candidate continuations with the highest best-of-N reward, rather than the highest likelihood, which lets the model steer generation toward higher multi-objective reward while staying close to the scaffold.

What would settle it

Re-score the 1,280 evaluated molecules per benchmark with a standard physics-based docking program and compute the Spearman rank correlation with the surrogate docking scores; a correlation below roughly 0.5 on either target would indicate the surrogate is unreliable and the reward signal used in training may be misaligned with true binding.

Watch

Extended reading notes

Core claim

ScaffoldGPT's central discovery is that combining two-phase incremental pretraining with APO finetuning and reward-guided Top-N decoding yields a drug-optimization model that outperforms eight baselines on both a viral and a cancer target. Table 1 reports the highest average normalized reward on 3CLPro (0.675) and RTCB (0.682), with the best baselines scoring 0.666 and 0.673 respectively, while the Tanimoto similarity to the original drug remains around 0.745–0.747. Ablation studies show that each of the three stages contributes: two-phase incremental pretraining improves validity over one-phase training; APO finetuning pushes properties up; and Top-N decoding alone beats Top-K/Top-P/Top-PK sampling, selecting a slightly different token path that yields higher reward without breaking the scaffold.

Load-bearing premise

The results assume the surrogate docking model from Appendix A.4 ranks compounds by true binding affinity, since the same surrogate supplies both the reward and the evaluation and is never checked against real docking on the generated molecules.

Editorial extensions

If this is right

  • If the reported results hold, drug optimization can be viewed as a three-stage generative pipeline: pretrain on molecular structure, finetune with multi-objective rewards, then steer decoding by reward rather than likelihood.
  • The same pretrained ScaffoldGPT can be redirected to a new target by swapping the critics in the reward function (Eq. 7), so the framework is not tied to 3CLPro or RTCB.
  • Because Top-N consistently outperforms likelihood-based sampling in the ablations, reward-guided token selection can be layered onto existing pretrained molecule generators without retraining.

Reading between the lines

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

  • Top-N decoding is not specific to the five properties used here: any reward critic, including predicted toxicity, ADMET, or experimentally measured activity, could be plugged into Eq. (7), so the strategy may transfer to other sequence-optimization settings.
  • The paper hints at a fragment-based third pretraining phase (Eq. 6); an editor's extension would be to test whether that phase improves scaffold diversity or generation quality further.
  • The most direct test of the paper's central claim is to re-score the generated molecules with a standard physics-based docking program or an experimental assay; the paper's reliance on the same surrogate for both training and evaluation leaves this validation open.
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

5 major / 5 minor

Summary. The paper proposes ScaffoldGPT, a three-stage GPT-based framework for scaffold-constrained drug optimization. Stage 1 is a two-phase incremental pretraining on ZINC data, first learning SMILES token sequences and then learning scaffold-to-molecule mappings. Stage 2 applies Advantage-aligned Policy Optimization (APO) with a multi-critic reward that combines normalized docking, druglikeness, synthesizability, solubility, and Tanimoto similarity scores. Stage 3 introduces Top-N, a token-level decoding strategy that selects next tokens according to best-of-N reward rather than likelihood. The method is evaluated on 3CLPro (COVID) and RTCB (cancer) benchmarks against REINVENT4 variants, DrugImprover, Molsearch, MIMOSA, and DrugEx v3, with Table 1 reporting the highest average normalized reward for ScaffoldGPT on both targets and ablations demonstrating the contribution of each stage.

Significance. If the evaluation is sound, the work would be a useful contribution to drug optimization: the three-stage pipeline is clearly described, the incremental pretraining and token-level reward-guided decoding are novel design choices, and the appendices provide unusually detailed hyperparameter settings that aid reproducibility. However, the central claim of benchmark outperformance rests on three evaluation-design issues: the possible overlap between the APO fine-tuning set and the evaluation set, the use of the same surrogate docking model as both the training reward and the evaluation metric, and the fact that the primary metric is exactly the training reward for ScaffoldGPT but not for the baselines. These issues are load-bearing for the abstract's claim and must be resolved before the results can be interpreted as evidence of general drug-optimization capability.

major comments (5)
  1. [Appendix A.7] The sentence 'For APO finetuning and experimentation, we have two different set of 1280 molecules' is ambiguous: it could mean two disjoint sets, or it could mean the same set of 1,280 molecules is used both for APO fine-tuning and for the Table 1 evaluation. If the latter, the evaluation is in-sample: the policy has been reward-optimized on the exact scaffolds it is later asked to optimize, whereas the REINVENT4 baselines receive only dataset-level training and no per-scaffold APO adaptation. This would make the headline Avg Norm Reward gains potentially attributable to specialization to the evaluation inputs rather than to a general capability. The authors must state explicitly whether the APO fine-tuning set and the evaluation set are disjoint, describe how the split was made, and report results on the held-out portion.
  2. [Appendix A.4 / Section 5.1] The docking score used in the reward function (Eq. 7) is the surrogate model of Vasan et al. (2023), described in Appendix A.4, and the same surrogate is used as the evaluation metric 'Docking Score' in Section 5.1 and Table 1. The paper reports the surrogate's validation R² (0.842 for 3CLPro, 0.73 for RTCB) but provides no validation of the surrogate on the generated molecules against an actual docking program such as AutoDock Vina or OEDOCK. Without such external validation, the docking improvements in Table 1 (e.g., -9.343 vs -8.698 for 3CLPro) and the reward signals that drive APO and Top-N may reflect surrogate bias rather than genuine binding affinity. A concrete test would be to run a standard docking tool on a random sample of generated molecules and report the correlation between the surrogate scores and the docking scores.
  3. [Section 5.1 / Eq. (7), (12), (13)] The primary evaluation metric, Average Normalized Reward, is exactly the reward function Rnorm in Eq. (7) that is used as the training signal for APO (Eq. 12) and for Top-N token selection (Eq. 13). Thus ScaffoldGPT is directly optimized to maximize the evaluation metric, while the baselines (e.g., REINVENT4 variants, Molsearch, MIMOSA, DrugEx v3) are not trained with this composite reward. This makes the comparison in Table 1 biased in favor of ScaffoldGPT: the reported gains may reflect the advantage of having the evaluation metric as the training objective rather than superior drug-optimization ability. To support the claim of outperformance, the authors should either fine-tune the baselines using the same Rnorm reward, or report additional independent metrics not used in training (for example, validated docking and external property predictors).
  4. [Section 5.1 / Table 1] The paper states that 'Average Norm Reward is the average of the normalized values ... across all valid molecules,' meaning invalid SMILES are excluded from the primary metric. ScaffoldGPT's validity is substantially lower than that of the baselines (0.944±0.094 on 3CLPro and 0.826±0.100 on RTCB, versus 0.995–1.000 for most baselines). Excluding invalid molecules can reward a model that generates few valid but high-scoring molecules, and it makes the comparison unfair to baselines with higher validity. The authors should report the primary metric over all generated molecules, assigning zero reward to invalid SMILES, or at least demonstrate that the conclusions in Table 1 are robust to this choice.
  5. [Table 1] No statistical significance tests are reported for the differences between methods. Several ScaffoldGPT rows have large standard deviations (e.g., Avg Norm Reward 0.675±0.031 for 3CLPro and validity 0.944±0.094), so it is unclear whether the reported improvements over the best baselines (e.g., 0.675 vs 0.666 for 3CLPro) are statistically significant. The authors should provide significance tests or confidence intervals over the five runs, and also describe how the 1,280 evaluation molecules were sampled and whether the same scaffolds appear in the training set.
minor comments (5)
  1. [Section 5.3 / Figure 3] The ablation of one-phase versus two-phase incremental training is presented as a small unnumbered table inside Figure 3, but the text refers to it as 'Table 3'; either renumber the table or change the cross-reference.
  2. [Section 4, Eq. (7) area] The sentence 'we adopt the define of reward function from Liu et al. (2024c)' contains a grammatical error; 'define' should be 'definition'.
  3. [Table 1 caption] The caption says 'eight baselines' but the table lists eleven entries including 'Original' and six REINVENT4 variants along with DrugImprover, Molsearch, MIMOSA, and DrugEx v3; correct the baseline count.
  4. [Abstract / Section 5.2] The abstract's claim that ScaffoldGPT 'outperforms the competing baselines' is broader than what Table 1 shows: on individual metrics such as similarity, ScaffoldGPT is below Molsearch and MIMOSA. The claim should be qualified to refer to the average normalized reward, the designated primary objective.
  5. [Figures 1 and 2] Figure 1(a) and Figure 2 show a three-phase incremental pretraining including fragments, while Section 4 describes only two phases with the fragment phase presented as an optional extension. Clarify whether the reported results use the two-phase or three-phase setup.

Circularity Check

2 steps flagged · score 7.0 of 10

Main benchmark gains reduce by construction: Table 1's primary metric is exactly the reward optimized by APO and Top-N, and the docking column is produced by the same surrogate used as the docking critic.

  1. fitted input called prediction [Section 4 Stage 2 Eq. (7); Section 5.1 Critics and evaluation metric; Appendix A.2]
    "Rnorm (Y ) := Rnorm (Y |S) = λ · Norm (CTanimoto (S, Y)) + |C|−1 i=0 λ · Norm (Ci(Y )) (7) ... Average Norm Reward is the average of the normalized values of the docking score, druglikeness, synthesizability, solubility, and similarity across all valid molecules. This is the most important metric. ... The top five epochs with the highest historical average normalized reward (as detailed in Section 5.1) are selected. From these five epochs, the epoch with the highest product of validity and average normalized reward is chosen as the final model for generation."

    Eq. (7) defines Rnorm as a weighted sum of normalized Tanimoto similarity and normalized druglikeness, synthesizability, solubility, and docking critics; Section 5.1 defines the headline metric as the average of exactly these five normalized components. Stage 2's APO gradient (Eq. 12) uses RAP = Rc(Y) − Rc(S) with Rc computed from BON over Rnorm, and Stage 3's Top-N (Eq. 13) selects tokens by BON/Rnorm. Thus the Avg Norm Reward reported in Table 1 is the very function Stages 2 and 3 maximize. Appendix A.2 further selects the final checkpoint by the highest historical average normalized reward.

  2. other [Section 5.1 Critics and evaluation metric; Appendix A.4; Eq. (7) CDocking]
    "Docking Score assesses the drug’s potential to bind and inhibit the target site. To enable efficient computation, we employ a docking surrogate model (See Appendix A.4) to output this score. ... The validation r2 values are 0.842 for 3CLPro and 0.73 for the RTCB dataset."

    The same surrogate is CDocking inside the reward (Eq. 7), so it contributes to the APO reward signal and to Top-N token selection, and it is also the model that produces the Docking column of Table 1. Docking improvements are therefore scores from the exact function used to guide generation, not an independent docking measurement. Because no true docking is run on the generated molecules, the docking result is the optimized surrogate value and cannot independently confirm better binding.

full rationale

The paper's central derivation chain is: pretrain on ZINC, APO-finetune with reward Rnorm, decode with Top-N using Rnorm, then evaluate with average normalized reward (the same Rnorm) and the docking surrogate (the same CDocking). The headline benchmark thus reduces substantially to the optimization objective. This is not a case of harmless self-citation: the self-citations to DrugImprover and ERP supply APO, the dataset, and the reward schema, but the circularity is the objective/evaluation identity, which the paper states in its own equations. The docking surrogate from Vasan et al. is shared by reward and evaluation, making the docking column partly optimized rather than independently validated. Appendix A.7's data-split sentence is ambiguous: 'For APO finetuning and experimentation, we have two different set of 1280 molecules that were selected from each of the RTCB and 3CLPro datasets...' If this means the same 1,280 molecules per target are used for APO fine-tuning and for Table 1 evaluation, the evaluation is in-sample and compounds the reward/evaluation identity; if it means two disjoint sets, the paper does not make that split explicit or report held-out results. The ablation studies of Top-N and incremental pretraining are internally consistent, but they too use the same reward metric, so they do not provide an external anchor. Score 7 reflects that the main comparative claims are substantially evaluations of the fitted objective rather than independent predictions.

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

The method rests on several untested premises: the surrogate docking model is taken as ground truth for both training and evaluation; the Tanimoto similarity is assumed to preserve beneficial properties; and equal weights in the reward are chosen by hand. The evaluation further assumes the test molecules are disjoint from the APO fine-tuning set, which Appendix A.7 contradicts.

free parameters (6)
  • Reward weights (λ) = 0.2 for each of 5 critics
    Hand-chosen equal weights in Eq. (7) and Table 8; define the primary metric and the APO objective.
  • Min/max normalization bounds for docking = [-10, 10]
    Used for min-max normalization of the docking critic (Table 8); chosen by hand.
  • Top-N candidate count (N) = 1
    Number of reward-ranked next tokens selected during decoding (Table 7).
  • K candidate set size = 16
    Maximum number of next-token candidates in Top-N (Table 7).
  • Best-of-N rollout count in APO = 8 (searched over 4, 6, 8)
    Number of sampled molecules per prompt in BON reward estimation (Eq. 8 and Table 8).
  • Docking score selection range for the 1280 molecules = -14 to -6
    Molecules are selected within this docking range (Appendix A.7), which affects benchmark difficulty.
assumptions (5)
  • domain assumption Tanimoto similarity to the original molecule is a valid measure of preserving beneficial chemical properties.
    Used in the reward function (Eq. 7) and as an evaluation metric. The paper cites Bender and Glen (2004) but does not validate this proxy for the specific drug optimization task.
  • domain assumption The surrogate docking model (BERT-like, from Vasan et al. 2023) accurately predicts true binding affinity for generated molecules.
    Used as the docking critic in the reward (Eq. 7) and as the docking score in evaluation (Section 5.1). Validation R² is 0.842 (3CLPro) and 0.73 (RTCB), and no check against real docking is provided.
  • ad hoc to paper The five critics with equal weights capture the relevant drug quality and define a fair primary metric.
    The primary metric Avg Norm Reward is the average of five normalized scores with hand-chosen weights (0.2 each in Table 8); no evidence supports this weighting.
  • domain assumption The 1,280 molecules used for evaluation are not used to fit the APO policy.
    Appendix A.7 states the same 1,280 molecules are used 'for APO finetuning and experimentation', which would violate this premise. The paper describes no held-out split.
  • domain assumption The original molecule's normalized reward is on the same scale as the generated molecules' reward.
    Appendix A.7 computes the original's reward with four components weighted 0.25 each, while generated molecules use five components weighted 0.2 each, so the comparison is not scale-invariant.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ScaffoldGPT: A Scaffold-based GPT Model for Drug Optimization." pith.science (2026). https://pith.science/paper/WMXSFMJ4

@misc{pith2026250206891,
  author       = {Pith},
  title        = {Pith review of: ScaffoldGPT: A Scaffold-based GPT Model for Drug Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WMXSFMJ4}},
  note         = {Machine review of arXiv:2502.06891}
}
read the original abstract

Drug optimization has become increasingly crucial in light of fast-mutating virus strains and drug-resistant cancer cells. Nevertheless, it remains challenging as it necessitates retaining the beneficial properties of the original drug while simultaneously enhancing desired attributes beyond its scope. In this work, we aim to tackle this challenge by introducing ScaffoldGPT, a novel Generative Pretrained Transformer (GPT) designed for drug optimization based on molecular scaffolds. Our work comprises three key components: (1) A three-stage drug optimization approach that integrates pretraining, finetuning, and decoding optimization. (2) A novel two-phase incremental pre-training strategy for scaffold-based drug optimization. (3) A token-level decoding optimization strategy, Top-N, that enabling controlled, reward-guided generation using the pretrained or finetuned GPT. We demonstrate via a comprehensive evaluation on COVID and cancer benchmarks that ScaffoldGPT outperforms the competing baselines in drug optimization benchmarks, while excelling in preserving original functional scaffold and enhancing desired properties.

Figures

Figures reproduced from arXiv: 2502.06891 by the authors.

Figure 1
Figure 1. Overview of ScaffoldGPT. (a) ScaffoldGPT decomposes molecular knowl￾edge top-down into fragments and then reconstructs it bottom-up through linking. (b) In Stage 1, it proposes an incremental pretraining strategy. (c) In Stage 2, it undergoes reinforcement learning-based post-training. (d) In Stage 3, it introduces a token-level con￾trollable decoding optimization approach. ScaffoldGPT is the first to introduce a th… view at source ↗
Figure 2
Figure 2. Two-phase incremental training of ScaffoldGPT. The first phase concentrates on recognizing the molecule, while the second phase builds connections between its scaffold and the original molecule. This process can be extended into a three-phase pretraining framework by establishing connections among the molecule’s fragments, scaffold, and orig￾inal molecule. we propose a novel framework for pre-training a Transformer … view at source ↗
Figure 3
Figure 3. Ablation study of one-phase vs two-phases. Evalua￾tions are based on five random seeds. Effectiveness of (two-phase) incremental training. In two-phase incremental pretraining, the intuition behind the first phase lies in training on critical keywords as knowledge pieces, reinforcing the memory of these key terms, particularly in longer sequences. We conducted an ablation study comparing our novel two-phase incremen… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 41 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Mol GPT : Molecular generation using a transformer-decoder model

    Viraj Bagal, Rishal Aggarwal, PK Vinod, and U Deva Priyakumar. Mol GPT : Molecular generation using a transformer-decoder model. Journal of Chemical Information and Modeling, 62 0 (9): 0 2064--2076, 2021

    work page 2021

  3. [3]

    Molecular similarity: a key technique in molecular informatics

    Andreas Bender and Robert C Glen. Molecular similarity: a key technique in molecular informatics. Organic & biomolecular chemistry, 2 0 (22): 0 3204--3218, 2004

    work page 2004

  4. [4]

    Paccmannrl: De novo generation of hit-like anticancer molecules from transcriptomic data via reinforcement learning

    Jannis Born, Matteo Manica, Ali Oskooei, Joris Cadow, Greta Markert, and Mar \' a Rodr \' guez Mart \' nez. Paccmannrl: De novo generation of hit-like anticancer molecules from transcriptomic data via reinforcement learning. Iscience, 24 0 (4), 2021

    work page 2021

  5. [5]

    PPL-MCTS : Constrained textual generation through discriminator-guided MCTS decoding

    Antoine Chaffin, Vincent Claveau, and Ewa Kijak. PPL-MCTS : Constrained textual generation through discriminator-guided MCTS decoding. arXiv preprint arXiv:2109.13582, 2021

    arXiv 2021

  6. [6]

    The cost of new drug discovery and development

    Michael Dickson and Jean Paul Gagnon. The cost of new drug discovery and development. Discovery medicine, 4 0 (22): 0 172--179, 2009

    work page 2009

  7. [7]

    Mermaid: an open source automated hit-to-lead method based on deep reinforcement learning

    Daiki Erikawa, Nobuaki Yasuo, and Masakazu Sekijima. Mermaid: an open source automated hit-to-lead method based on deep reinforcement learning. Journal of Cheminformatics, 13: 0 1--10, 2021

    work page 2021

  8. [8]

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

    Peter Ertl and Ansgar Schuffenhauer. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of cheminformatics, 1: 0 1--11, 2009

    2009

Show all 56 references
  1. [9]

    https://www.alcf.anl.gov/polaris, last accessed on 10-2-2023

    Argonne Leadership Computing Facility. https://www.alcf.anl.gov/polaris, last accessed on 10-2-2023

    2023

  2. [10]

    Neural scaling of deep chemical models

    Nathan C Frey, Ryan Soklaski, Simon Axelrod, Siddharth Samsi, Rafael Gomez-Bombarelli, Connor W Coley, and Vijay Gadepally. Neural scaling of deep chemical models. Nature Machine Intelligence, 5 0 (11): 0 1297--1305, 2023

    2023

  3. [11]

    Mimosa: Multi-constraint molecule sampling for molecule optimization

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

    2021

  4. [12]

    A new algorithm for data compression

    Philip Gage. A new algorithm for data compression. The C Users Journal, 12 0 (2): 0 23--38, 1994

    1994

  5. [13]

    Scaling laws for reward model overoptimization

    Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. In International Conference on Machine Learning, pages 10835--10866. PMLR, 2023

    2023

  6. [14]

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

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

    2017 arXiv

  7. [15]

    Covid-19 vaccines and variants of concern: A review

    Ikbel Hadj Hassine. Covid-19 vaccines and variants of concern: A review. Reviews in medical virology, 32 0 (4): 0 e2313, 2022

    2022

  8. [16]

    Molecular optimization by capturing chemist’s intuition using deep neural networks

    Jiazhen He, Huifang You, Emil Sandstr \"o m, Eva Nittinger, Esben Jannik Bjerrum, Christian Tyrchan, Werngard Czechtizky, and Ola Engkvist. Molecular optimization by capturing chemist’s intuition using deep neural networks. Journal of cheminformatics, 13 0 (1): 0 1--17, 2021

    2021

  9. [17]

    Transformer-based molecular optimization beyond matched molecular pairs

    Jiazhen He, Eva Nittinger, Christian Tyrchan, Werngard Czechtizky, Atanas Patronov, Esben Jannik Bjerrum, and Ola Engkvist. Transformer-based molecular optimization beyond matched molecular pairs. Journal of cheminformatics, 14 0 (1): 0 18, 2022

    2022

  10. [18]

    Structure modification in chemical databases

    Peter W Kenny and Jens Sadowski. Structure modification in chemical databases. Chemoinformatics in drug discovery, pages 271--285, 2005

    2005

  11. [19]

    Self-referencing embedded strings ( SELFIES ): A 100\ Machine Learning: Science and Technology, 1 0 (4): 0 045024, 2020

    Mario Krenn, Florian H \"a se, AkshatKumar Nigam, Pascal Friederich, and Alan Aspuru-Guzik. Self-referencing embedded strings ( SELFIES ): A 100\ Machine Learning: Science and Technology, 1 0 (4): 0 045024, 2020

    2020

  12. [20]

    R D kit: Open-source cheminformatics software

    Greg Landrum et al. R D kit: Open-source cheminformatics software. https://www.rdkit.org. Accessed Oct 2023

    2023

  13. [21]

    Machine translation decoding beyond beam search

    R \'e mi Leblond, Jean-Baptiste Alayrac, Laurent Sifre, Miruna Pislar, Jean-Baptiste Lespiau, Ioannis Antonoglou, Karen Simonyan, and Oriol Vinyals. Machine translation decoding beyond beam search. arXiv preprint arXiv:2104.05336, 2021

    2021 arXiv

  14. [22]

    Druggpt: A gpt-based strategy for designing potential ligands targeting specific proteins

    Yuesen Li, Chengyi Gao, Xin Song, Xiangyu Wang, Yungang Xu, and Suxia Han. Druggpt: A gpt-based strategy for designing potential ligands targeting specific proteins. bioRxiv, pages 2023--06, 2023

    2023

  15. [23]

    Drugimprover: Utilizing reinforcement learning for multi-objective alignment in drug optimization

    Xuefeng Liu, Songhao Jiang, Archit Vasan, Alexander Brace, Ozan Gokdemir, Thomas Brettin, Fangfang Xia, Ian Foster, and Rick Stevens. Drugimprover: Utilizing reinforcement learning for multi-objective alignment in drug optimization. In NeurIPS 2023 Workshop on New Frontiers of...

    2023

  16. [24]

    Entropy-reinforced planning with large language models for de novo drug discovery

    Xuefeng Liu, Chih-Chan Tien, Peng Ding, Songhao Jiang, and Stevens Rick. Entropy-reinforced planning with large language models for de novo drug discovery. ICML, 2024 a

    2024

  17. [25]

    Drugex v3: scaffold-constrained drug design with graph transformer-based reinforcement learning

    Xuhan Liu, Kai Ye, Herman WT van Vlijmen, Adriaan P IJzerman, and Gerard JP van Westen. Drugex v3: scaffold-constrained drug design with graph transformer-based reinforcement learning. Journal of Cheminformatics, 15 0 (1): 0 24, 2023 b

    2023

  18. [26]

    Sora: A review on background, technology, limitations, and opportunities of large vision models

    Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024 b

    2024 arXiv

  19. [27]

    Reinvent 4: Modern ai--driven generative molecule design

    Hannes H Loeffler, Jiazhen He, Alessandro Tibo, Jon Paul Janet, Alexey Voronov, Lewis H Mervin, and Ola Engkvist. Reinvent 4: Modern ai--driven generative molecule design. Journal of Cheminformatics, 16 0 (1): 0 20, 2024

    2024

  20. [28]

    The different mechanisms of cancer drug resistance: a brief review

    Behzad Mansoori, Ali Mohammadi, Sadaf Davudian, Solmaz Shirjang, and Behzad Baradaran. The different mechanisms of cancer drug resistance: a brief review. Advanced pharmaceutical bulletin, 7 0 (3): 0 339, 2017

    2017

  21. [29]

    LLM drug discovery challenge: A contest as a feasibility study on the utilization of large language models in medicinal chemistry

    Kusuri Murakumo, Naruki Yoshikawa, Kentaro Rikimaru, Shogo Nakamura, Kairi Furui, Takamasa Suzuki, Hiroyuki Yamasaki, Yuki Nishigaya, Yuzo Takagi, and Masahito Ohue. LLM drug discovery challenge: A contest as a feasibility study on the utilization of large language models in m...

    2023

  22. [30]

    Exploring deep recurrent models with reinforcement learning for molecule design

    Daniel Neil, Marwin Segler, Laura Guasch, Mohamed Ahmed, Dean Plumbley, Matthew Sellwood, and Nathan Brown. Exploring deep recurrent models with reinforcement learning for molecule design. In ICLR, 2018

    2018

  23. [31]

    An empirical evaluation of github copilot's code suggestions

    Nhan Nguyen and Sarah Nadi. An empirical evaluation of github copilot's code suggestions. In Proceedings of the 19th International Conference on Mining Software Repositories, pages 1--5, 2022

    2022

  24. [32]

    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 0 (1): 0 1--14, 2017

    2017

  25. [33]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 0 2...

    2022

  26. [34]

    Deep reinforcement learning for de novo drug design

    Mariya Popova, Olexandr Isayev, and Alexander Tropsha. Deep reinforcement learning for de novo drug design. Science advances, 4 0 (7): 0 eaap7885, 2018

    2018

  27. [35]

    Drug repurposing: Progress, challenges and recommendations

    Sudeep Pushpakom, Francesco Iorio, Patrick A Eyers, K Jane Escott, Shirley Hopper, Andrew Wells, Andrew Doig, Tim Guilliams, Joanna Latimer, Christine McNamee, Alan Norris, Philippe Sanseau, David Cavalla, and Munir Pirmohamed. Drug repurposing: Progress, challenges and recomm...

    2019

  28. [36]

    Markov decision processes: Discrete stochastic dynamic programming

    Martin L Puterman. Markov decision processes: Discrete stochastic dynamic programming. John Wiley & Sons, 2014

    2014

  29. [37]

    C5 T 5: Controllable generation of organic molecules with transformers

    Daniel Rothchild, Alex Tamkin, Julie Yu, Ujval Misra, and Joseph Gonzalez. C5 T 5: Controllable generation of organic molecules with transformers. arXiv preprint arXiv:2108.10307, 2021

    2021 arXiv

  30. [38]

    To beam or not to beam: That is a question of cooperation for language GANs

    Thomas Scialom, Paul-Alexis Dray, Jacopo Staiano, Sylvain Lamprier, and Benjamin Piwowarski. To beam or not to beam: That is a question of cooperation for language GANs . Advances in Neural Information Processing Systems, 34: 0 26585--26597, 2021

    2021

  31. [39]

    Neural machine translation of rare words with subword units

    Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015

    2015 arXiv

  32. [40]

    Deep reinforcement learning for multiparameter optimization in de novo drug design

    Niclas St hl, 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 0 (7): 0 3166--3176, 2019

    2019

  33. [41]

    Molsearch: search-based multi-objective molecular generation and property optimization

    Mengying Sun, Jing Xing, Han Meng, Huijun Wang, Bin Chen, and Jiayu Zhou. Molsearch: search-based multi-objective molecular generation and property optimization. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, pages 4724--4732, 2022

    2022

  34. [42]

    Reinforcement learning for systems pharmacology-oriented and personalized drug design

    Ryan K Tan, Yang Liu, and Lei Xie. Reinforcement learning for systems pharmacology-oriented and personalized drug design. Expert Opinion on Drug Discovery, 17 0 (8): 0 849--863, 2022 a

    2022

  35. [43]

    Drlinker: Deep reinforcement learning for optimization in fragment linking design

    Youhai Tan, Lingxue Dai, Weifeng Huang, Yinfeng Guo, Shuangjia Zheng, Jinping Lei, Hongming Chen, and Yuedong Yang. Drlinker: Deep reinforcement learning for optimization in fragment linking design. Journal of Chemical Information and Modeling, 62 0 (23): 0 5907--5917, 2022 b

    2022

  36. [44]

    Matched molecular pair analysis in short: algorithms, applications and limitations

    Christian Tyrchan and Emma Evertsson. Matched molecular pair analysis in short: algorithms, applications and limitations. Computational and structural biotechnology journal, 15: 0 86--90, 2017

    2017

  37. [45]

    Benchmarking language-based docking models

    Archit Vasan, Rick Stevens, Arvind Ramanathan, and Vishwanath Venkatram. Benchmarking language-based docking models. 2023

    2023

  38. [46]

    A reinforcement learning approach for protein--ligand binding pose prediction

    Chenran Wang, Yang Chen, Yuan Zhang, Keqiao Li, Menghan Lin, Feng Pan, Wei Wu, and Jinfeng Zhang. A reinforcement learning approach for protein--ligand binding pose prediction. BMC bioinformatics, 23 0 (1): 0 1--18, 2022 a

    2022

  39. [47]

    A transformer-based generative model for de novo molecular design

    Wenlu Wang, Ye Wang, Honggang Zhao, and Simone Sciabola. A transformer-based generative model for de novo molecular design. arXiv preprint arXiv:2210.08749, 2022 b

    2022 arXiv

  40. [48]

    Smiles, a chemical language and information system

    David Weininger. Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules. Journal of chemical information and computer sciences, 28 0 (1): 0 31--36, 1988

    1988

  41. [49]

    A brief overview of chatgpt: The history, status quo and potential future development

    Tianyu Wu, Shizhu He, Jingping Liu, Siqi Sun, Kang Liu, Qing-Long Han, and Yang Tang. A brief overview of chatgpt: The history, status quo and potential future development. IEEE/CAA Journal of Automatica Sinica, 10 0 (5): 0 1122--1136, 2023

    2023

  42. [50]

    Covid-19 pathophysiology: A review

    Koichi Yuki, Miho Fujiogi, and Sophia Koutsogiannaki. Covid-19 pathophysiology: A review. Clinical immunology, 215: 0 108427, 2020

    2020

  43. [51]

    Planning with large language models for code generation

    Shun Zhang, Zhenfang Chen, Yikang Shen, Mingyu Ding, Joshua B Tenenbaum, and Chuang Gan. Planning with large language models for code generation. arXiv preprint arXiv:2303.05510, 2023 a

    2023 arXiv

  44. [52]

    Universal approach to de novo drug design for target proteins using deep reinforcement learning

    Yunjiang Zhang, Shuyuan Li, Miaojuan Xing, Qing Yuan, Hong He, and Shaorui Sun. Universal approach to de novo drug design for target proteins using deep reinforcement learning. ACS omega, 8 0 (6): 0 5464--5474, 2023 b

    2023

  45. [53]

    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 0 (1): 0 10752, 2019

    2019

  46. [54]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  47. [55]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  48. [56]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.