CoChat
Sign Up
Cited by 7Computational Drug Discovery MethodsMachine Learning in Materials ScienceGenerative Adversarial Networks and Image Synthesis

PCF-VAE: posterior collapse free variational autoencoder for de novo drug design

Open access
Bhadwal, Arun Singh, Kumari, Monika, Kumar, Anil·2025·Scientific Reports
DOI 10.1038/s41598-025-14285-5PMC PMC12488873
Open this paper in CoChat — free account

Abstract

Generating novel molecular structures with desired pharmacological and physicochemical properties is challenging due to the vast chemical space, complex optimization requirements, predictive limitations of models, and data scarcity. This study focuses on investigating the problem of posterior collapse in variational autoencoders, a deep learning technique used for de novo molecular design. Various generative variational autoencoders were employed to map molecule structures to a continuous latent space and vice versa, evaluating their performance as structure generators. Most state-of-the-art approaches suffer from posterior collapse, limiting the diversity of generated molecules. To address this challenge, a novel approach termed PCF-VAE was introduced to mitigate the issue of posterior collapse, reduce the complexity of SMILES representations, and enhance diversity in molecule generation. In comparison to state-of-the-art models, PCF-VAE has been evaluated and compared in the MOSES benchmark at different diversity levels. Depending on the diversity level, PCF-VAE has a validity of 98.01% at D = 1, 97.10% at D = 2, and 95.01% at D = 3. It is important to note that PCF-VAE effectively generates molecules with a 100% unique structure. Both intDiv and intDiv2 are measures of internal diversity; intDiv2 ranges from 85.87 to 86.33% and intDiv ranges from 85.87 to 89.01%. Additionally, at D = 1, D = 2, and D = 3, PCF-VAE generates 93.77%, 94.71%, and 95.01% novel molecules, respectively. The results indicate that this research provides valuable insights into the challenges of molecular generation and contributes to the design of novel molecules with desired properties.

AI snapshot

Written from the paper’s full text.

TL;DR

PCF-VAE is a conditional variational autoencoder for de novo drug design that replaces SMILES with GenSMILES and uses a dynamically weighted KL term to reduce posterior collapse. On the MOSES benchmark it reports high validity, 100% uniqueness, and controllable diversity, though it does not dominate every metric and can generate less chemically realistic molecules at high diversity.

Key findings

  • PCF-VAE reports validity of 0.9801 at D = 1, 0.9710 at D = 2, and 0.9501 at D = 3, with 100% uniqueness at both Unique@1k and Unique@10k across diversity levels.
  • It reports novelty of 0.9377, 0.9471, and 0.9501 at D = 1, D = 2, and D = 3, respectively; intDiv ranges from 0.8587 to 0.8901 and intDiv2 from 0.8527 to 0.8633.
  • The paper states PCF-VAE’s Frag metric surpasses all other SOTA methods, but it “does not consistently outperform other models across all metrics” and instead provides competitive overall performance.
  • During training on MOSES over 150 epochs, reconstruction accuracy rises from about 0.55 to nearly 0.80, and test validity rises from about 0.92 to nearly 0.98.

Methods

  • The model is a conditional VAE with GRU-based encoder and decoder, trained on GenSMILES representations that replace paired SMILES symbols with single-token, position-aware encodings for branches and rings.
  • Training uses the MOSES benchmark dataset: 1,936,962 molecular structures split into 1.6 million training, 176,000 test, and 176,000 scaffold molecules, derived from ZINC with molecular weight 250–350 Da, ≤7 rotatable bonds, XLogP ≤3.5, and medicinal chemistry filters.
  • Posterior collapse is addressed by modifying the ELBO with a dynamic weighting factor α on the KL divergence term, initially near zero and gradually increased; molecule generation uses latent samples with a controllable diversity parameter D_i = 1, 2, or 3.

Limitations

  • The paper states that PCF-VAE shows marginal variations in Scaffold metrics, particularly in TestSF, and struggles to maintain chemical realism in extreme diversity settings, occasionally generating unstable or synthetically infeasible molecules.
  • The paper also notes that PCF-VAE does not consistently outperform other models across all metrics, despite balancing diversity and competitive performance.
  • Evident from the text: evaluation is limited to the MOSES benchmark and reported metrics; no experimental synthesis or wet-lab validation of generated molecules is described.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41598-025-14285-5

Cites

37 of 37 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

5 of 8 cited sources checked · 1 supported · 0 flagged · accuracy not yet measured

  • Findingown data

    PCF-VAE intDiv ranges from 85.87 to 89.01%.

    “intDiv ranges from 85.87 to 89.01%”
    Quote in the paperSubstantial claim
  • Findingreasoning

    PCF-VAE generates a more informative latent space.

    “PCF-VAE proves successful in generating a more informative latent space”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Most state-of-the-art VAE-based approaches suffer from posterior collapse, limiting the diversity of generated molecules.
  • Findingown data

    PCF-VAE intDiv2 ranges from 85.87 to 86.33%.

    “intDiv2 ranges from 85.87 to 86.33%”
    Quote in the paperNot a substantial claim
  • Findingown data

    PCF-VAE has a validity of 98.01% at diversity level D = 1.

    “PCF-VAE has a validity of 98.01% at D = 1”
    Quote in the paperSubstantial claim
  • Findingnone given

    PCF-VAE generates molecules with a 100% unique structure.

    “PCF-VAE effectively generates molecules with a 100% unique structure.”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 15
  • May combine two claims. 20

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Several research studies propose intricate and expansive architectures to overcome the limitations of conditional VAEs.

    “To overcome these limitations, several research studies35,36 propose intricate and expansive architectures.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [35] Liao, Z., Xie, L., Mamitsuka, H. & Zhu, S. Sc2mol: A scaffold-based two-step molecule generator with variational autoencoder and transformer. Bioinformatics 39(1), 814 (2023). source

      The Introduction and §2.1 describe Sc2Mol as a two-step architecture combining a VAE with a transformer, so it is an example of a more elaborate molecule generator. Our copy does not say it was proposed specifically to overcome limitations of conditional VAEs; that motivation is unsupported. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [36] Skalic, M., Jiménez, J., Sabbadin, D. & De Fabritiis, G. Shape-based generative modeling for de novo drug design. J. Chem. Inform. Model. 59(3), 1205–1214 (2019). source

  • May combine two claims.

    Premisecitation

    Conditional VAE applicability in drug design is constrained by inherent drawbacks such as sub-optimal performance on standard SMILES representations, the posterior collapse problem, and inefficient exploration of latent space.

    “While the conditional VAE has been previously documented in the literature34, its applicability in drug design is constrained by inherent drawbacks such as sub-optimal performance on standard SMILES representations, the posterior collapse problem, and inefficient exploration of latent space.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [34] Lim, J., Ryu, S., Kim, J. W. & Kim, W. Y. Molecular generative model based on conditional variational autoencoder for de novo molecular design. J. Cheminform. 10(1), 1–9 (2018). source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The MOSES dataset comprises 1,936,962 molecular structures, divided into training, testing, and scaffold datasets containing 1.6 million, 176,000, and 176,000 molecules respectively.

    “The MOSES dataset comprises a total of 1,936,962 molecular structures, which are subsequently divided into training, testing, and scaffold datasets, containing 1.6 million, 176,000, and 176,000 molecules respectively.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [32] Polykovskiy, D. et al. Molecular sets (moses): A benchmarking platform for molecular generation models. Front. Pharmacol. 11, 565644 (2020). source

      The Dataset section reports three non-overlapping splits of 1,584,664 training, 176,075 test, and 176,226 scaffold-test molecules, supporting the rounded split sizes. Those counts sum to 1,936,965, three more than the claimed total of 1,936,962; the total is not stated in the available passage. (LLM, conf=0.90)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    To meet particular goals, optimization methods such as post-processing approaches (transfer learning and reinforcement learning) and Bayesian optimization are applied.

    “Furthermore, to meet particular goals, optimization methods are applied, including techniques such as post-processing approaches (e.g., transfer learning and reinforcement learning) and Bayesian optimization11.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [11] Gómez-Bombarelli, R. et al. Automatic chemical design using a data-driven continuous representation of molecules. ACS Central Sci. 4(2), 268–276 (2018). source

      The Introduction (page not available) describes optimizing molecules for desired properties and says gradient-based optimization can be combined with Bayesian inference. The Results and Discussion section, “Optimization of Molecules via Properties” (page not available), reports optimization using a Gaussian-process model. Our copy does not describe transfer learning or reinforcement learning as post-processing approaches, so the full list attributed to this paper is not supported. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisereasoning

    These intricate and expansive architectures attain a SOTA level of validity and diversity in generated molecules but require extended time for completion.

    “Although these architectures attain a SOTA level of validity and diversity in generated molecules, the intricate nature of the process leads to an extension in the time required for completion.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [35] Liao, Z., Xie, L., Mamitsuka, H. & Zhu, S. Sc2mol: A scaffold-based two-step molecule generator with variational autoencoder and transformer. Bioinformatics 39(1), 814 (2023). source

      Section 3.5.1, Table 2 reports 100% validity and 99.99% uniqueness for Sc2Mol on MOSES; Section 3.4 identifies uniqueness as a measure of molecular diversity. The cited results support high validity and diversity, but the passages do not establish that architectural complexity extends the time required for completion. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [36] Skalic, M., Jiménez, J., Sabbadin, D. & De Fabritiis, G. Shape-based generative modeling for de novo drug design. J. Chem. Inform. Model. 59(3), 1205–1214 (2019). source

  • May combine two claims.

    Findingnone given

    PCF-VAE demonstrates a reduction in the impact of posterior collapse, resulting in the generation of a greater variety of valid molecules on benchmark datasets.

    “PCF-VAE method demonstrates a reduction in the impact of posterior collapse, resulting in the generation of a greater variety of valid molecules on benchmark datasets.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Most state-of-the-art VAE-based approaches suffer from posterior collapse, limiting the diversity of generated molecules.

  • May combine two claims.

    Findingown data

    PCF-VAE achieved a validity score of up to 0.9801 and a novelty reaching 0.9501, with consistently high values for Frag and Scaf metrics.

    “Notably, PCF-VAE achieved a validity score of up to 0.9801, Novelty reaching 0.9501, and consistently high values for Fragmentation (Frag) and Scaffold (Scaf) metrics.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The MOSES benchmark and its metrics adequately capture the quality of generated molecules for drug design.

  • May combine two claims.

    Findingown data

    KL loss initially rises, reaching its maximum near the 20th epoch, subsequently declining consistently.

    “The KL loss initially rises, reaching its maximum near the 20th epoch, subsequently declining consistently.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    The generation of molecules by PCF-VAE near the specified target yields highly similar molecular structures.

    “It has been noted that the generation of molecules by PCF-VAE in the vicinity of the specified target yields highly similar molecular structures.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    PCF-VAE excels in generating molecules with high diversity and novelty but shows marginal variations in Scaffold metrics, particularly in the TestSF.

    “However, PCF-VAE excels in generating molecules with high diversity and novelty, it shows marginal variations in Scaffold metrics, particularly in the TestSF.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The MOSES benchmark and its metrics adequately capture the quality of generated molecules for drug design.

  • May combine two claims.

    Findingown data

    Molecular Weight and LogP distributions broaden and shift, while TPSA shows increased multimodality and divergence as diversity increases.

    “Molecular Weight and LogP distributions broaden and shift, while TPSA shows increased multimodality and divergence.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The property distributions of the actual data are the appropriate reference for structural realism.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Mutual Information varies across the epochs, exhibiting no distinct upward or downward trajectory.

    “Likewise, the mutual Information varies across the epochs, exhibiting no distinct upward or downward trajectory.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The test Validity curve started at approximately 0.92 and increased to near 0.98 by the 150th epoch.

    “Commencing from an initial value of approximately 0.92, it demonstrates a consistent increase, nearing 0.98 by the 150th epoch.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Reconstruction Loss exhibits a sharp drop in the early epochs, subsequently stabilizing as training progresses.

    “The Reconstruction Loss exhibits a sharp drop in the early epochs, subsequently stabilizing as training progresses.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In Table 2, PCF-VAE achieves a validity score of 0.9789 at diversity level 1, surpassing all other models.

    “Specifically, in Table 2, the PCF-VAE achieves a validity score of 0.9789 at diversity level 1, surpassing all other models.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The MOSES benchmark and its metrics adequately capture the quality of generated molecules for drug design.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    GenSMILES simplifies the complexity of SMILES strings.

    “This process simplifies the complexity of the SMILES strings”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Experimental studies have identified paired representations in SMILES as a primary contributor to invalid molecule generation.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    GenSMILES preserves the intrinsic semantic information of molecules.

    “GenSMILES has been specifically designed to preserve the intrinsic semantic information of the molecules.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    PCF-VAE reduces the complexity of SMILES representations.

    “reduce the complexity of SMILES representations”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Experimental studies have identified paired representations in SMILES as a primary contributor to invalid molecule generation.

  • May combine two claims.

    Interpretationreasoning

    GenSMILES enables the probabilistic model to understand long-term relationships inside SMILES strings while reducing their complexity.

    “This structural adaption enables the probabilistic model to understand long-term relationships inside the SMILES strings while simultaneously reducing their complexity.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Using GenSMILES reduces the complexity associated with SMILES representation.

    “By using GenSMILES, the complexity associated with SMILES representation is reduced.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The behavior of KL loss demonstrates the model's progressive alignment of the approximate posterior with the prior distributions, signifying enhanced latent space representation.

    “This behavior demonstrates the model’s progressive alignment of the approximate posterior with the prior distributions, signifying enhanced latent space representation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The enriched and diverse latent representation enhances the model's generative capabilities, empowering it to generate molecules that are valid, diverse, and distinctive.

    “The enriched and diverse latent representation enhances the model’s generative capabilities, empowering it to generate molecules that are valid, diverse, and distinctive qualities that hold significant importance in the generative process.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The MOSES benchmark and its metrics adequately capture the quality of generated molecules for drug design.
    • Rests on an assumption: High validity, uniqueness, diversity, and novelty are important qualities for de novo drug design.

  • May combine two claims.

    Interpretationnone given

    Adding reparameterized loss allows the model to find the best balance between reconstruction and KL divergence parts, which improves performance.

    “By adding reparameterized loss, the model can find the best balance between the reconstruction and KL divergence parts, which improves performance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The upward trend in test Validity validates the model's ability to generalize effectively, ensuring strong validity and robustness during training.

    “This upward trend validates the model’s ability to generalize effectively, ensuring strong validity and robustness during the training process.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Directly applying regularization techniques to datasets represented in SMILES format remains challenging due to structural and syntactic differences between SMILES strings and natural language sequences.

    “However, directly applying such techniques to datasets represented in the SMILES format remains challenging, owing to the structural and syntactic differences between SMILES strings and natural language sequences.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Experimental studies have identified paired representations in SMILES as a primary contributor to invalid molecule generation.

  • May combine two claims.

    Interpretationown data

    Higher diversity yields more novel molecules but with less resemblance to real molecular properties.

    “This trend reflects a trade-off: higher diversity yields more novel molecules but with less resemblance to real molecular properties.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The property distributions of the actual data are the appropriate reference for structural realism.

  • May combine two claims.

    Interpretationreasoning

    GenSMILES reduces dependency on symbol pairing and nested structural context, leading to higher validity in generative tasks and more interpretable molecular encodings.

    “These compact representations in GenSMILES reduce dependency on symbol pairing and nested structural context, leading to higher validity in generative tasks and more interpretable molecular encodings.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Experimental studies have identified paired representations in SMILES as a primary contributor to invalid molecule generation.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    GenSMILES replaces paired symbols with single-token, position-aware representations.

    “To address these issues, GenSMILES5 proposes an alternative representation that replaces paired symbols with single-token, position-aware representations.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [5] Bhadwal, A. S., Kumar, K. & Kumar, N. Gensmiles: An enhanced validity conscious representation for inverse design of molecules. Knowl.-Based Syst. 268, 110429 (2023). source

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    In the relevant literature, SMILES preprocessing, posterior collapse, and directed search have received insufficient consideration.

    “In the relevant literature discussed in Section 2, SMILES preprocessing, posterior collapse, and directed search have received insufficient consideration.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    PCF-VAE has reduced time complexity compared with a conditional VAE, attributed to its simplicity and resulting in fewer learnable parameters.

    “The model exhibits a reduced time complexity attributed to its simplicity, resulting in a less number of learnable parameters.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    PCF-VAE does not consistently outperform other models across all metrics but strikes a balance with diversity among the various metrics, providing competitive performance overall.

    “While the PCF-VAE does not consistently outperform other models across all metrics, it strikes a balance with diversity among the various metrics, providing competitive performance overall.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The MOSES benchmark and its metrics adequately capture the quality of generated molecules for drug design.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationown data

    The variability of Mutual Information indicates that the information encoded by the latent representation regarding the input data fluctuates dynamically throughout the training process.

    “This variability indicates that the information encoded by the latent representation regarding the input data fluctuates dynamically throughout the training process.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The metrics collectively demonstrate the model's efficacy in acquiring meaningful representations, reconstructing input data, and generalizing to novel samples.

    “These metrics collectively demonstrate the model’s efficacy in acquiring meaningful representations, reconstructing input data, and generalizing to novel samples.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Molecules generated in GenSMILES format can be smoothly translated back into valid SMILES strings using deterministic derivation rules that preserve valency and structural integrity.

    “These can be smoothly translated back into valid SMILES strings using a set of deterministic derivation rules that preserve valency and structural integrity, ensuring chemically valid outputs while improving syntactic robustness during generation.”
    Quote in the paperNot a substantial claim
34 of 34

  • The paper relies on this without support.

    Most state-of-the-art VAE-based approaches suffer from posterior collapse, limiting the diversity of generated molecules.

    “Most state-of-the-art approaches suffer from posterior collapse, limiting the diversity of generated molecules.”
    “A prevalent challenge encountered in most of the VAE based approaches is the occurrence of the posterior collapse problem, wherein the model fails to effectively utilize samples from the latent space.”

  • The paper relies on this without support.

    In the initial stages of training, RNN-based VAE architectures may encounter the posterior collapse effect.

    “However, in the initial stages of training, the RNN-based VAE architecture may encounter a phenomenon known as the posterior collapse effect.”

  • The paper relies on this without support.

    Experimental studies have identified paired representations in SMILES as a primary contributor to invalid molecule generation.

    “Experimental studies have identified these pair representations as a primary contributor to invalid molecule generation.”

  • The paper relies on this without support.

    The properties of molecules exhibit a strong correlation, such that any alteration in one property leads to equally significant changes in the other properties.

    “The properties of molecules exhibit a strong correlation, such that any alteration in one property leads to equally significant changes in the other properties.”

  • The paper relies on this without support.

    Generative models fail to effectively learn the molecular patterns in the SMILES representation of molecules, resulting in the production of undesired and invalid molecules.

    “Furthermore, the generative models fail to effectively learn the molecular patterns in the SMILES representation of molecules, resulting in the production of undesired and invalid molecules.”

  • The paper relies on this without support.

    Currently, unbounded generative neural networks comprise the majority of implementations.

    “Currently, as far as we know, unbounded generative neural networks comprise the majority of implementations.”

  • The paper relies on this without support.

    The MOSES benchmark and its metrics adequately capture the quality of generated molecules for drug design.

    “The PCF-VAE underwent assessment using the MOSES platform, as detailed by32.”

  • The paper relies on this without support.

    High validity, uniqueness, diversity, and novelty are important qualities for de novo drug design.

    “empowering it to generate molecules that are valid, diverse, and distinctive qualities that hold significant importance in the generative process.”

  • The paper relies on this without support.

    The property distributions of the actual data are the appropriate reference for structural realism.

    “The Fig. 5 compares the distributions of Molecular Weight, LogP, and TPSA properties between actual (red) and PCF-VAE-generated (blue) molecules at three diversity levels (D = 1, D = 2, D = 3).”
9 of 9

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Prose gives a different D = 1 validity score

    Major

    The stated score of 0.9789 differs from the 0.9801 reported for the same model and diversity level in Table 2.

  • Conclusion's novelty comparison conflicts with Table 2

    Major

    The conclusion says PCF-VAE outperformed NGram on measures including novelty. Table 2 reports NGram novelty of 0.9694, higher than PCF-VAE's highest reported novelty of 0.9501.

  • PCF-VAE does not consistently outperform VAE

    Major

    Table 3 reports a lower, better Test FCD for VAE (0.099) than for any PCF-VAE setting (0.2962, 0.3151, and 0.3233). Its Test SNN is also higher than all three PCF-VAE settings, contradicting consistent outperformance across Tables 2 and 3.

  • PCF-VAE does not have the highest validity in Table 2

    Major

    The claim that its D = 1 validity surpasses all other models conflicts with Table 2: both the Combinatorial and JTN-VAE models have validity of 1.0, above PCF-VAE's 0.9801.

  • FCD mean and covariance labels are reversed

    Minor

    One sentence assigns covariance to m_r and mean to C_r, while the next assigns mean to m and covariance to C for the corresponding generated distribution.

  • Abstract overstates the minimum IntDiv2 value

    Minor

    Table 2 reports IntDiv2 of 0.8527 (85.27%) for PCF-VAE at D = 1, below the abstract's stated minimum of 85.87%.

  • PCF-VAE does not surpass every model on Frag

    Minor

    Table 3 gives CharRNN a Test Frag score of 0.9998; the highest PCF-VAE Test Frag score is 0.9996 at D = 3.

  • Uniqueness is claimed at 100k but tabulated at 10k

    Minor

    The text says 100% uniqueness was attained at the 100k level, whereas Table 2 reports a Unique@10k metric, not Unique@100k.

8 of 8

Cites

Flagged first, then by citations

Research works printed but not linked: 4

  1. Jaques, N., Gu, S., Bahdanau, D., Hernández-Lobato, J.M., Turner, R.E., & Eck, D. Sequence tutor: Conservative fine-tuning of sequence generation models with kl-control. In: International Conference on Machine Learning, pp. 1645–1654 (2017). PMLR
  2. Zhang, M., Jiang, S., Cui, Z., Garnett, R., & Chen, Y. D-vae: A variational autoencoder for directed acyclic graphs. Adv. Neural Inform. Process. Syst. 32 (2019).
  3. Lim, J., Ryu, S., Kim, J. W. & Kim, W. Y. Molecular generative model based on conditional variational autoencoder for de novo molecular design. J. Cheminform. 10(1), 1–9 (2018). · DOI · Paper page
  4. Jin, W., Barzilay, R., & Jaakkola, T. Junction tree variational autoencoder for molecular graph generation. In: International Conference on Machine Learning, pp. 2323–2332 (2018). PMLR

Cited by

Sorted by citations

Authors
Bhadwal, A. S., Kumari, M., & Kumar, A.
Affiliations
  • Bennett University (IN)
  • University of Petroleum and Energy Studies (IN)
Corresponding author
Kumar, Anil
ORCID
Venue
Scientific Reports
Publisher
Springer Science and Business Media LLC
Published
2025-10-01
Volume
15
Issue
1
Pages
34152-34152
ISSN
2045-2322 (ISSN-L)
DOI
10.1038/s41598-025-14285-5
PMC
PMC12488873
OpenAlex
W4414706327
Cited by
7
References
32
Citations by year
2026: 6 · 2025: 1
FWCI
3.35
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • Europe PMC (PubMed Central) (submittedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • PubMed Central (submittedVersion, cc-by-nc-nd)
Topics
  • Physical Sciences › Computer Science › Computational Theory and Mathematics › Computational Drug Discovery Methods
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
Keywords
posterior collapse, variational autoencoder, de novo drug design, SMILES representation, molecular diversity, latent space
MeSH terms
Deep Learning, Autoencoder, Algorithms, Drug Design, Molecular Structure

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-21.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-21.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review