Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Phenotypic Profile-Informed Generation of Drug-Like Molecules via Dual-Channel Variational Autoencoders

T0 review · 4 major / 5 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 SmilesGEN generates drug-like molecules from gene-expression profiles by modeling drug-induced phenotypic changes as a latent-space subtraction, and reports Tanimoto similarities up to 0.98 with approved drugs.

desk verdict A plausible dual-VAE idea undercut by an unspecified pretraining corpus and max-only similarity metrics. read the letter →

arxiv 2506.02051 v1 pith:AFQZBWE3 submitted 2025-06-01 q-bio.BM cs.AIcs.LG

Hui Liu , Shiye Tian , Xuejun Liu This is my paper · ORCID
classification q-bio.BMcs.AIcs.LG
keywords denovodrugdesignvariationalautoencodergeneexpressionprofilesSMILESphenotypicdiscoveryTanimotosimilaritytranscriptionalresponse
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 SmilesGEN, a dual-channel variational autoencoder that generates drug-like molecules from gene-expression profiles. Its central claim is that explicitly modeling the perturbative effect of a molecule on cellular transcription — as a subtraction in a shared latent space — lets the model produce molecules that are more valid, unique, novel, and structurally similar to known ligands than existing phenotype-based generators. If true, this would make transcriptional signatures a more direct handle for de novo drug design, especially for diseases where no validated target exists. The paper reports that its best generated molecule for lung adenocarcinoma reaches a Tanimoto similarity of 0.98 against Alectinib, an approved drug.

What carries the argument

The load-bearing mechanism is the latent-space subtraction in Eq. (4), $Z_y = (\mu_x - \mu_s) + (\sigma_x + \sigma_s)\epsilon$, where $\mu_x,\sigma_x$ are the parameters of the post-treatment profile's encoding, $\mu_s,\sigma_s$ are the molecule's encoding parameters, and $\epsilon \sim \mathcal{N}(0,I)$. This constructs a random variable that should reconstruct the pre-treatment profile, forcing the molecule's latent code to represent the perturbation that the drug induces. The dual-channel architecture couples a pretrained GRU-based SMILES variational autoencoder (SmilesNet) with a feedforward expression-profile autoencoder (ProfileNet); the full objective sums the SMILES reconstruction loss, the post-treatment profile reconstruction loss, and the pre-treatment reconstruction loss from the subtracted latent variable.

What would settle it

Retrain SmilesNet after excluding every evaluation ligand (the known ligands in Table 1 and approved drugs in Table 2, including Alectinib) from the pretraining corpus, then regenerate the molecules; if the top Tanimoto similarities fall substantially, the high scores are retrieval of memorized structures rather than phenotype-conditioned generation.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the net effect of a drug perturbation on a cell can be encoded as the difference between the latent representation of the post-treatment expression profile and the latent representation of the drug molecule; SmilesGEN enforces this alignment by reconstructing the pre-treatment profile from the subtracted latent variable. Once trained, feeding a desired expression profile through the profile encoder and into the SMILES decoder generates molecules whose structures resemble known ligands of the relevant targets. Across ten target genes, the highest Tanimoto similarity with known ligands is consistently above 0.5, and in the therapeutic-generation task the best molecule for LUAD matches Alectinib at 0.98 Tanimoto similarity.

Load-bearing premise

The evaluation assumes the reference ligands and approved drugs were not present in the corpus used to pretrain SmilesNet, because the paper never describes the pretraining set or excludes evaluation compounds from it.

Editorial extensions

If this is right

  • If the central claim holds, expression profiles alone — without a defined molecular target — can be used to generate hit-like molecules, easing phenotypic drug discovery for poorly characterized diseases.
  • The latent subtraction gives an interpretable handle on drug-induced transcriptional change, so the model can in principle be steered by reversing a disease profile (as the paper does by inverting patient profiles) to seek molecules that push cells back toward a healthy state.
  • Scaffold-based optimization becomes possible: the paper shows benzene rings and short-chain ligands can be expanded into molecules with higher structural similarity to known ligands while keeping or improving drug-likeness (QED).
  • The reported 0.98 Tanimoto similarity to Alectinib suggests that, within the limits of the evaluation, the model can reproduce structural features of an approved drug from a purely transcriptional reversal signal.

Reading between the lines

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

  • Beyond the paper: the high Tanimoto scores could partly reflect memorization if evaluation ligands were present in SmilesNet's pretraining corpus; the paper never describes that corpus or any exclusion of evaluation compounds, so a held-out retraining test is needed to confirm true phenotype-conditioned generation.
  • Beyond the paper: the linear-additivity assumption in Eq. (4) is a strong simplification; an ablation that replaces the subtraction with a learned nonlinear perturbation mapping could indicate whether the gains come from this exact mechanism or from the dual-channel architecture more broadly.
  • Beyond the paper: the framework should transfer to other perturbation assays and to in-vitro validation; testing whether generated molecules actually reverse the target expression profile in a cell line would be the natural next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SmilesGEN, a dual-channel variational autoencoder that jointly models SMILES strings and gene expression profiles in a shared latent space. The key idea is to explicitly model the drug-induced perturbation by requiring that the latent code of a post-treatment expression profile, after subtraction of the molecule encoding, reconstructs the pre-treatment profile; at inference, the SmilesNet decoder is conditioned on an encoded expression profile to generate drug-like molecules. The authors report experiments on L1000 genetic perturbation profiles and TCGA cancer profiles, claiming that SmilesGEN outperforms existing phenotype-based generative models in validity, uniqueness, novelty, and Tanimoto similarity to known ligands, including a Tanimoto of 0.98 to the approved lung-cancer drug Alectinib. The paper includes the proposed architecture, training objective, experimental comparisons, and qualitative examples of scaffold optimization and therapeutic molecule generation.

Significance. The central idea of explicitly modeling the perturbative effect of a molecule on the cellular state in a shared latent space is a reasonable and potentially useful extension over methods that condition only on post-treatment expression profiles. If the empirical claims were supported, SmilesGEN could offer a practical tool for phenotype-informed drug design. The authors provide a GitHub link and state that source code and datasets are available, which is a positive aspect for reproducibility. However, the current evaluation has serious methodological weaknesses: the pretraining corpus for the molecule VAE is unspecified, the evaluation reports only maximum Tanimoto coefficients, and the subtraction formula in Eq. (4) appears statistically incorrect. These issues undermine the load-bearing claim of superior phenotype-conditioned generation, so the paper needs substantial revision before the results can be considered reliable.

major comments (4)
  1. [Section 4.3 and Table 1] The evaluation reports only the highest Tanimoto coefficient over 50 generated molecules per target. This is not a robust statistic: a single outlier or a single memorized molecule can dominate, and the claim that SmilesGEN 'consistently achieved significantly higher Tanimoto coefficients' is not supported by any significance test or distributional summary. The authors should report mean, median, and standard deviation of Tanimoto coefficients across the generated set, and perform a statistical comparison (e.g., permutation test or confidence intervals) against each baseline. Without such analysis, the superiority claim is not established.
  2. [Section 3.1 and 4.1] The pretraining corpus for SmilesNet is described only as 'large-scale SMILES descriptors'; its composition is never specified, and the paper does not state whether the evaluation ligands (DTC inhibitors/agonists and DrugBank approved drugs) were excluded from that corpus. A VAE trained to reconstruct SMILES strings can output near-exact copies of molecules in its training set when the latent code is close to the encoding of that molecule. The reported Tanimoto of 1.00 for TP53 and 0.98 for Alectinib strongly suggests retrieval of memorized structures rather than phenotype-conditioned generation. The authors must identify the pretraining dataset, confirm that evaluation compounds were not present, and ideally demonstrate that generated molecules are novel with respect to the pretraining set. This is a load-bearing issue for the central claim.
  3. [Eq. (4) and Section 3.4] The formula Zy = (μx − μs) + (σx + σs)ε is not consistent with the stated goal of modeling Zy as the difference of Zx and Zs. If Zx and Zs are independent Gaussian random variables with standard deviations σx and σs, the standard deviation of Zx − Zs is sqrt(σx^2 + σs^2), not σx + σs. The paper uses the sum of standard deviations, which overestimates the variance and has no clear probabilistic justification. Furthermore, the KL divergence term in Eq. (5), DKL(qθ(Zy|X)||pψ(Zy)), is undefined because qθ is the encoder for X, not for Zy; Zy is a deterministic function of Zx and Zs. The derivation of the objective needs to be corrected or carefully justified.
  4. [Algorithm 1] The training procedure is contradictory: line 6 states 'Freeze SmilesNet encoder parameter φ', but line 10 of the same algorithm says 'Update parameters θ, ψ, and φ'. Since φ is the SmilesNet encoder parameter, updating it contradicts the freeze instruction. This inconsistency makes the training protocol ambiguous and affects reproducibility. The authors should clarify which parameters are updated in each stage and ensure the pseudocode matches the text.
minor comments (5)
  1. [Throughout] There are several typos and inconsistent notations: 'SimlesGEN Framework' in Section 3.1, 'SmileGEN' in Section 3.5, 'PICK3CA' in Figure 2, and variable spacing in 'GxV AEs' (e.g., 'GxV AEs' vs 'GxVAEs'). These should be corrected for clarity.
  2. [Section 4.3] The statement about QED distribution, '90% known ligands displayed peak QED values below 0.55', is not directly supported by the referenced supplementary Figure S2, which is not included in the submission. Please ensure the figure is available and describe the statistical summary precisely.
  3. [Section 4.5] The use of disease-specific reversal profiles by multiplying averaged expression values by -1 is a heuristic that is not validated or discussed as a limitation. A brief justification or a sensitivity analysis would strengthen this section.
  4. [Section 3.4] The notation for the standard deviations σx and σs is ambiguous: it is not clarified whether these are the standard deviations of the approximate posteriors or the variances. In standard VAE formulations, the encoder outputs log-variance; explicitly defining these quantities would improve reproducibility.
  5. [References] Some references have incomplete information (e.g., [Tong et al., 2023] is a bioRxiv preprint without a DOI and [Yamanaka et al., 2023] lists page ranges but no volume). Please verify all references for completeness.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the VAE losses are explicitly supervised objectives, and the only self-citation is a non-load-bearing assumption reference.

full rationale

SmilesGEN's claimed derivation chain is a supervised conditional-VAE construction. Equation (2) trains SmilesNet to reconstruct a SMILES string, Eq. (3) trains ProfileNet to reconstruct post-treatment profiles, and Eq. (5) trains the decoder to reconstruct pre-treatment profiles from the latent subtraction defined in Eq. (4). Because the paper says ProfileNet is 'imposed' to reconstruct pre-treatment profiles, the restoration is an explicitly supervised training objective rather than a predicted discovery. The generation step at inference, feeding an encoded expression profile into the SmilesNet decoder, is likewise the trained conditional decoder; no claim is made that the model predicts a quantity outside its fitted objectives. The only self-citation, [Huang and Liu, 2024] for linear additivity, is accompanied by an external citation (Hetzel et al., 2022) and is introduced as an explicit assumption ('we assume'), so it is not load-bearing for the central empirical claim. The strongest concern is that the undisclosed SmilesNet pretraining corpus may contain the evaluation ligands (DTC ligands and DrugBank approved drugs), which could inflate Tanimoto scores via memorization; however, that is a benchmark-integrity risk, not a step that reduces by the paper's own equations. Without evidence that the reference compounds are in the pretraining set, this cannot be scored as circularity under the stated rules. No load-bearing circular step is exhibited.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical entities, particles, forces, or dimensions are introduced; the 'dual-channel VAE' is an architectural construct, not an invented entity. The central claims rest on the linear additivity assumption, decoder transfer, evaluation premises, and data disjointness.

free parameters (3)
  • Perturbation subtraction coefficient = 1 (implicit)
    Eq. (4) assumes the difference between post- and pre-treatment latent means equals the molecule latent with unit coefficient; no scaling is learned, so the linear additivity is imposed without validation.
  • Variance combination in Eq. (4) = σx+σs
    The paper uses the sum of standard deviations for the reparameterization of Zy, rather than sqrt(σx^2+σs^2), which would be the variance of the difference of independent Gaussian variables; this is an ad hoc choice.
  • Reversal factor for disease profiles = -1
    In Section 4.5, patient expression profiles are averaged and multiplied by -1 to create desired profiles, an ad hoc heuristic without biological justification.
assumptions (5)
  • ad hoc to paper Linear additivity of drug effects in learned latent space
    Eq. (4) assumes Zy = (μx - μs) + (σx+σs)ε reconstructs the pre-treatment profile; the unit coefficient and the sum of standard deviations are asserted without derivation or validation.
  • domain assumption Decoder transfer across latent shift
    The ProfileNet decoder pψ, trained to reconstruct post-treatment profiles from Zx, is assumed to also reconstruct pre-treatment profiles from Zy; this requires the decoder to generalize linearly across the latent space.
  • domain assumption Ground truth validity of known ligands as targets
    The evaluation assumes molecules with high Tanimoto similarity to known inhibitors/agonists of a knocked-down gene will phenotypically mimic that perturbation; structural similarity to known ligands is used as a proxy for activity.
  • domain assumption Disjointness of pretraining and evaluation molecules
    The paper does not state whether the SMILES corpus used to pretrain SmilesNet contains the DTC ligands or approved drugs used in evaluation; the empirical claims depend on this disjointness.
  • ad hoc to paper Reversal of expression profiles direction
    Section 4.5 assumes that multiplying averaged patient profiles by -1 yields a profile whose induction would be therapeutic; this is asserted without evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phenotypic Profile-Informed Generation of Drug-Like Molecules via Dual-Channel Variational Autoencoders." pith.science (2026). https://pith.science/paper/AFQZBWE3

@misc{pith2026250602051,
  author       = {Pith},
  title        = {Pith review of: Phenotypic Profile-Informed Generation of Drug-Like Molecules via Dual-Channel Variational Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFQZBWE3}},
  note         = {Machine review of arXiv:2506.02051}
}
read the original abstract

The de novo generation of drug-like molecules capable of inducing desirable phenotypic changes is receiving increasing attention. However, previous methods predominantly rely on expression profiles to guide molecule generation, but overlook the perturbative effect of the molecules on cellular contexts. To overcome this limitation, we propose SmilesGEN, a novel generative model based on variational autoencoder (VAE) architecture to generate molecules with potential therapeutic effects. SmilesGEN integrates a pre-trained drug VAE (SmilesNet) with an expression profile VAE (ProfileNet), jointly modeling the interplay between drug perturbations and transcriptional responses in a common latent space. Specifically, ProfileNet is imposed to reconstruct pre-treatment expression profiles when eliminating drug-induced perturbations in the latent space, while SmilesNet is informed by desired expression profiles to generate drug-like molecules. Our empirical experiments demonstrate that SmilesGEN outperforms current state-of-the-art models in generating molecules with higher degree of validity, uniqueness, novelty, as well as higher Tanimoto similarity to known ligands targeting the relevant proteins. Moreover, we evaluate SmilesGEN for scaffold-based molecule optimization and generation of therapeutic agents, and confirmed its superior performance in generating molecules with higher similarity to approved drugs. SmilesGEN establishes a robust framework that leverages gene signatures to generate drug-like molecules that hold promising potential to induce desirable cellular phenotypic changes.

Figures

Figures reproduced from arXiv: 2506.02051 by the authors.

Figure 1
Figure 1. SmilesGEN framework comprises two interacting variational autoencoders, SmilesNet and ProfileNet, which work jointly to extract [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Performance comparison in terms of validity, uniqueness and novelty of generated molecules by SmilesGEN and four competing [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Performance evaluation for scaffold-based molecule optimization. (a) Optimized molecules based on benzene ring toward specific [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Distribution of QED values of generated molecules using [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Valid Property-Enhanced Contrastive Learning for Targeted Optimization & Resampling for Novel Drug Design

    cs.LG 2025-08 conditional novelty 5.0 of 10

    VECTOR+ combines contrastive learning and Gaussian mixture sampling to generate novel, synthetically plausible inhibitors from low-data datasets, with improved docking scores over known compounds.

Reference graph

Works this paper leans on

39 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    Trends in kinase drug discovery: targets, indications and inhibitor design.Nature Reviews Drug Discovery, 20(11):839–861,

    [Attwoodet al., 2021 ] Misty M Attwood, Doriano Fabbro, Aleksandr V Sokolov, Stefan Knapp, and Helgi B Schi¨oth. Trends in kinase drug discovery: targets, indications and inhibitor design.Nature Reviews Drug Discovery, 20(11):839–861,

    work page 2021

  2. [3]

    Application of generative autoencoder in de novo molecular design.Molecular informatics, 37(1- 2):1700123,

    [Blaschkeet al., 2018 ] Thomas Blaschke, Marcus Olive- crona, Ola Engkvist, J ¨urgen Bajorath, and Hongming Chen. Application of generative autoencoder in de novo molecular design.Molecular informatics, 37(1- 2):1700123,

    work page 2018

  3. [7]

    Attention-based generative models for de novo molecular design.Chemical Science, 12(24):8362–8372,

    [Dollaret al., 2021 ] Orion Dollar, Nisarg Joshi, David AC Beck, and Jim Pfaendtner. Attention-based generative models for de novo molecular design.Chemical Science, 12(24):8362–8372,

    work page 2021

  4. [8]

    Automatic chemical design using a data-driven continuous representation of molecules.ACS central science, 4(2):268–276,

    [G´omez-Bombarelliet al., 2018 ] Rafael G ´omez- Bombarelli, Jennifer N Wei, David Duvenaud, Jos´e Miguel Hern ´andez-Lobato, Benjam ´ın S ´anchez- Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D Hirzel, Ryan P Adams, and Al ´an Aspuru- Guzik. Automatic chemical design using a data-driven continuous representation of molecules.ACS centr...

    work page 2018

  5. [9]

    Goodfellow, Jean Pouget- Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C

    [Goodfellowet al., 2014 ] Ian J. Goodfellow, Jean Pouget- Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks.Communications of the ACM, 63:139 – 144,

    work page 2014

  6. [10]

    Pre- dicting cellular responses to novel drug perturbations at a single-cell resolution.Advances in Neural Information Processing Systems, 35:26711–26722,

    [Hetzelet al., 2022 ] Leon Hetzel, Simon Boehm, Niki Kil- bertus, Stephan G ¨unnemann, Fabian Theis, et al. Pre- dicting cellular responses to novel drug perturbations at a single-cell resolution.Advances in Neural Information Processing Systems, 35:26711–26722,

    work page 2022

  7. [14]

    Auto-Encoding Variational Bayes.arXiv e- prints, page arXiv:1312.6114, December

    [Kingma and Welling, 2013] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes.arXiv e- prints, page arXiv:1312.6114, December

    arXiv 2013

  8. [16]

    Learning deep generative models of graphs.arXiv preprint arXiv:1803.03324,

    [Liet al., 2018 ] Yujia Li, Oriol Vinyals, Chris Dyer, Razvan Pascanu, and Peter Battaglia. Learning deep generative models of graphs.arXiv preprint arXiv:1803.03324,

    arXiv 2018

Show all 39 references
  1. [17]

    Re- cent advances in therapeutic strategies for triple-negative breast cancer.Journal of hematology & oncology, 15(1):121,

    [Liet al., 2022 ] Yun Li, Huajun Zhang, Yulia Merkher, Lin Chen, Na Liu, Sergey Leonov, and Yongheng Chen. Re- cent advances in therapeutic strategies for triple-negative breast cancer.Journal of hematology & oncology, 15(1):121,

    2022

  2. [18]

    Use of artificial intelligence to enhance phenotypic drug discovery.Drug Discovery Today, 26(4):887–901,

    [Malandraki-Miller and Riley, 2021] Sophia Malandraki- Miller and Paul R Riley. Use of artificial intelligence to enhance phenotypic drug discovery.Drug Discovery Today, 26(4):887–901,

    2021

  3. [19]

    The emerging role of mass spectrometry-based proteomics in drug discovery.Nature Reviews Drug Discovery, 21(9):637–654,

    [Meissneret al., 2022 ] Felix Meissner, Jennifer Geddes- McAlister, Matthias Mann, and Marcus Bantscheff. The emerging role of mass spectrometry-based proteomics in drug discovery.Nature Reviews Drug Discovery, 21(9):637–654,

    2022

  4. [20]

    De novo generation of hit-like molecules from gene expression signatures using artificial intelligence

    [M´endez-Lucioet al., 2020 ] Oscar M ´endez-Lucio, Benoit Baillif, Djork-Arn ´e Clevert, David Rouqui ´e, and Joerg Wichard. De novo generation of hit-like molecules from gene expression signatures using artificial intelligence. Nature communications, 11(1):10,

    2020

  5. [21]

    De novo design of bioac- tive small molecules by artificial intelligence.Molecular informatics, 37(1-2):1700153,

    [Merket al., 2018 ] Daniel Merk, Lukas Friedrich, Francesca Grisoni, and Gisbert Schneider. De novo design of bioac- tive small molecules by artificial intelligence.Molecular informatics, 37(1-2):1700153,

    2018

  6. [22]

    Op- portunities and challenges in phenotypic drug discovery: an industry perspective.Nature reviews Drug discovery, 16(8):531–543,

    [Moffatet al., 2017 ] John G Moffat, Fabien Vincent, Jonathan A Lee, J ¨org Eder, and Marco Prunotto. Op- portunities and challenges in phenotypic drug discovery: an industry perspective.Nature reviews Drug discovery, 16(8):531–543,

    2017

  7. [24]

    Fame: fragment-based conditional molecular gen- eration for phenotypic drug discovery

    [Phamet al., 2022 ] Thai-Hoang Pham, Lei Xie, and Ping Zhang. Fame: fragment-based conditional molecular gen- eration for phenotypic drug discovery. InProceedings of the 2022 SIAM International Conference on Data Mining (SDM), pages 720–728. SIAM,

    2022

  8. [25]

    Deep reinforcement learning for de novo drug design.Science advances, 4(7):eaap7885,

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

    2018

  9. [26]

    off-target

    [Sadri, 2023] Arash Sadri. Is target-based drug discovery efficient? discovery and “off-target” mechanisms of all drugs.Journal of medicinal chemistry, 66(18):12651– 12677,

    2023

  10. [27]

    Generating focused molecule libraries for drug discovery with recurrent neural networks.ACS central science, 4(1):120–131,

    [Segleret al., 2018 ] Marwin HS Segler, Thierry Kogej, Christian Tyrchan, and Mark P Waller. Generating focused molecule libraries for drug discovery with recurrent neural networks.ACS central science, 4(1):120–131,

    2018

  11. [28]

    Molecular generation for desired transcriptome changes with adversarial autoencoders.Frontiers in Pharmacol- ogy, 11:269,

    [Shayakhmetovet al., 2020 ] Rim Shayakhmetov, Maksim Kuznetsov, Alexander Zhebrak, Artur Kadurin, Sergey Nikolenko, Alexander Aliper, and Daniil Polykovskiy. Molecular generation for desired transcriptome changes with adversarial autoencoders.Frontiers in Pharmacol- ogy, 11:269,

    2020

  12. [29]

    Corsello, and et al

    [Subramanianet al., 2017 ] Aravind Subramanian, Rajiv Narayan, Steven M. Corsello, and et al. A next generation connectivity map: L1000 platform and the first 1,000,000 profiles.Cell, 171(6):1437–1452.e17,

    2017

  13. [30]

    How were new medicines discovered?Nature reviews Drug discovery, 10(7):507–519,

    [Swinney and Anthony, 2011] David C Swinney and Jason Anthony. How were new medicines discovered?Nature reviews Drug discovery, 10(7):507–519,

    2011

  14. [32]

    Transi- gen: Deep representation learning of chemical-induced transcriptional profile.bioRxiv, pages 2023–11,

    [Tonget al., 2023 ] Xiaochu Tong, Ning Qu, Xiangtai Kong, Shengkun Ni, Kun Wang, Lehan Zhang, Yiming Wen, Sulin Zhang, Xutong Li, and Mingyue Zheng. Transi- gen: Deep representation learning of chemical-induced transcriptional profile.bioRxiv, pages 2023–11,

    2023

  15. [33]

    Phenotypic drug discovery: recent successes, lessons learned and new directions.Nature Reviews Drug Discovery, 21(12):899–914,

    [Vincentet al., 2022 ] Fabien Vincent, Arsenio Nueda, Jonathan Lee, Monica Schenone, Marco Prunotto, and Mark Mercola. Phenotypic drug discovery: recent successes, lessons learned and new directions.Nature Reviews Drug Discovery, 21(12):899–914,

    2022

  16. [34]

    A transformer-based gener- ative model for de novo molecular design.arXiv preprint arXiv:2210.08749,

    [Wanget al., 2022 ] Wenlu Wang, Ye Wang, Honggang Zhao, and Simone Sciabola. A transformer-based gener- ative model for de novo molecular design.arXiv preprint arXiv:2210.08749,

    2022 arXiv

  17. [35]

    The cancer genome atlas pan- cancer analysis project.Nature Genetics, 45(10):1113– 1120,

    [Weinsteinet al., 2013 ] John N Weinstein, Eric A Collison, Gordon B Mills, Kenna R Shaw, Brad A Ozenberger, Kyle Ellrott, Ilya Shmulevich, Chris Sander, Joshua M Stuart, and Houtan Noushmehr. The cancer genome atlas pan- cancer analysis project.Nature Genetics, 45(10):1113– 1120,

    2013

  18. [36]

    Drugbank: a comprehensive resource for in silico drug discovery and exploration.Nucleic acids research, 34(suppl 1):D668– D672,

    [Wishartet al., 2006 ] David S Wishart, Craig Knox, An Chi Guo, Savita Shrivastava, Murtaza Hassanali, Paul Stothard, Zhan Chang, and Jennifer Woolsey. Drugbank: a comprehensive resource for in silico drug discovery and exploration.Nucleic acids research, 34(suppl 1):D668– D672,

    2006

  19. [38]

    De novo molecule design using molecular generative mod- els constrained by ligand–protein interactions.Journal of chemical information and modeling, 62(14):3291–3306,

    [Zhang and Chen, 2022] Jie Zhang and Hongming Chen. De novo molecule design using molecular generative mod- els constrained by ligand–protein interactions.Journal of chemical information and modeling, 62(14):3291–3306,

    2022

  20. [39]

    Deep learning enables rapid identification of potent ddr1 kinase inhibitors.Nature biotechnology, 37(9):1038–1040, 2019

    [Zhavoronkovet al., 2019 ] Alex Zhavoronkov, Yan A Iva- nenkov, Alex Aliper, Mark S Veselov, Vladimir A Al- adinskiy, Anastasiya V Aladinskaya, Victor A Terentiev, Daniil A Polykovskiy, Maksim D Kuznetsov, Arip Asadu- laev, et al. Deep learning enables rapid identification of ...

    2019

  21. [2006]

    De novo drug design based on patient gene expression profiles via deep learning.Molecular Informatics, 42(8- 9):2300064,

    [Yamanakaet al., 2023 ] Chikashige Yamanaka, Shunya Uki, Kazuma Kaitoh, Michio Iwata, and Yoshihiro Yamanishi. De novo drug design based on patient gene expression profiles via deep learning.Molecular Informatics, 42(8- 9):2300064,

    2023

  22. [2011]

    Drug target com- mons: a community effort to build a consensus knowledge base for drug-target interactions.Cell chemical biology, 25(2):224–229,

    [Tanget al., 2018 ] Jing Tang, Balaguru Ravikumar, Zaid Alam, Anni Rebane, Markus V ¨ah¨a-Koskela, Gopal Ped- dinti, Arjan J van Adrichem, Janica Wakkinen, Alok Jaiswal, Ella Karjalainen, et al. Drug target com- mons: a community effort to build a consensus knowledge base for ...

    2018

  23. [2013]

    Gxvaes: Two joint vaes generate hit molecules from gene expression profiles

    [Li and Yamanishi, 2024] Chen Li and Yoshihiro Yamanishi. Gxvaes: Two joint vaes generate hit molecules from gene expression profiles. InProceedings of the AAAI Confer- ence on Artificial Intelligence, volume 38, pages 13455– 13463,

    2024

  24. [2014]

    Gex2sgen: designing drug-like molecules from desired gene expression signa- tures.Journal of Chemical Information and Modeling, 63(7):1882–1893,

    [Daset al., 2023 ] Dibyajyoti Das, Broto Chakrabarty, Raj- gopal Srinivasan, and Arijit Roy. Gex2sgen: designing drug-like molecules from desired gene expression signa- tures.Journal of Chemical Information and Modeling, 63(7):1882–1893,

    2023

  25. [2017]

    De novo generation of multi-target compounds using deep generative chemistry.Nature Communications, 15(1):3636,

    [Munsonet al., 2024 ] Brenton P Munson, Michael Chen, Audrey Bogosian, Jason F Kreisberg, Katherine Licon, Ruben Abagyan, Brent M Kuenzi, and Trey Ideker. De novo generation of multi-target compounds using deep generative chemistry.Nature Communications, 15(1):3636,

    2024

  26. [2018]

    Empirical evalua- tion of gated recurrent neural networks on sequence mod- eling.ArXiv, abs/1412.3555,

    [Chunget al., 2014 ] Junyoung Chung, C ¸ aglar G ¨ulc ¸ehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evalua- tion of gated recurrent neural networks on sequence mod- eling.ArXiv, abs/1412.3555,

    2014 arXiv

  27. [2020]

    Triomphe: transcriptome-based inference and generation of molecules with desired phenotypes by machine learning.Journal of Chemical Information and Modeling, 61(9):4303–4320,

    [Kaitoh and Yamanishi, 2021] Kazuma Kaitoh and Yoshi- hiro Yamanishi. Triomphe: transcriptome-based inference and generation of molecules with desired phenotypes by machine learning.Journal of Chemical Information and Modeling, 61(9):4303–4320,

    2021

  28. [2021]

    The future of phenotypic drug discovery.Cell chemical biology, 28(3):424–430,

    [Berg, 2021] Ellen L Berg. The future of phenotypic drug discovery.Cell chemical biology, 28(3):424–430,

    2021

  29. [2022]

    Predict- ing single-cell cellular responses to perturbations using cycle consistency learning.Bioinformatics, 40(Supple- ment 1):i462–i470,

    [Huang and Liu, 2024] Wei Huang and Hui Liu. Predict- ing single-cell cellular responses to perturbations using cycle consistency learning.Bioinformatics, 40(Supple- ment 1):i462–i470,

    2024

  30. [2023]

    Molgan: An implicit generative model for small molecular graphs.arXiv preprint arXiv:1805.11973,

    [De Cao and Kipf, 2018] Nicola De Cao and Thomas Kipf. Molgan: An implicit generative model for small molecular graphs.arXiv preprint arXiv:1805.11973,

    2018 arXiv

  31. [2024]

    Hierarchical generation of molecular graphs us- ing structural motifs

    [Jinet al., 2020 ] Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Hierarchical generation of molecular graphs us- ing structural motifs. InInternational conference on ma- chine learning, pages 4839–4848. PMLR,

    2020

Pith tools

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