Pith. sign in

REVIEW 5 major objections 6 minor 88 references

A molecule representation built purely from functional groups matches or beats graph and sequence models on 33 property benchmarks while staying chemically interpretable.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-04 18:44 UTC pith:DE3LKFLQ

load-bearing objection A plausible functional-group representation with internally consistent ablations, but the headline SOTA claims rest on quoted baselines and a potentially leaky PubChem-mined vocabulary. the 5 major comments →

arxiv 2509.09619 v1 pith:DE3LKFLQ submitted 2025-09-11 cs.LG

Functional Groups are All you Need for Chemically Interpretable Molecular Property Prediction

Roshan Balaji , Joe Bobby , Nirav Pravinbhai Bhatt This is my paper
classification cs.LG
keywords molecular property predictionfunctional groupschemical interpretabilityautoencoder latent embeddingSMILES substructure miningsequential pattern miningToxAlertsscaffold split
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that the chemistry concept of a functional group — the recurring substructure that gives a molecule its reactivity and properties — is sufficient, on its own, for competitive molecular property prediction. The authors encode each molecule as a multi-hot bit vector over two functional-group vocabularies: 2,672 curated patterns from the ToxAlerts database and 30,000 patterns mined from PubChem SMILES strings by sequential pattern mining. A tied-weight autoencoder compresses these sparse vectors into a latent embedding, which is concatenated with 211 RDKit 2D descriptors and fed to a small feedforward predictor. Across 33 benchmark datasets spanning physiology, biophysics, physical chemistry, quantum mechanics, bioactivity, pharmacokinetics, and peptide cleavage, the combined FGR representation claims state-of-the-art or near-state-of-the-art accuracy under scaffold splits. Because every input feature is a named chemical group, the model's attributions can be checked against known structure-activity relationships, and the paper reports that they match.

Core claim

The central claim is that a molecular representation built only from functional-group presence — no explicit atom graph, no sequence model, no 3D geometry — can reach state-of-the-art property prediction while remaining chemically transparent. Specifically, the paper argues that combining expert-curated functional groups (FG) with automatically mined ones (MFG) is better than either alone; that the autoencoder-compressed latent space retains enough information for downstream MLPs to outperform previous SOTA in several categories (e.g., beating KANO on five of eight MoleculeNet classification tasks, improving physical-chemistry regression by about 8.66%, and beating DMPNN on all six peptide/b

What carries the argument

The load-bearing object is the functional-group vocabulary, used to build multi-hot input vectors. Two vocabularies are constructed: FG (2,672 curated SMARTS patterns from ToxAlerts, each tied to a known toxicological mechanism) and MFG (up to 30,000 frequent SMILES substrings mined from PubChem by sequential pattern mining with a frequency threshold of eta = 500). Each molecule becomes a binary vector; a tied-weight autoencoder with focal-loss reconstruction and an uncorrelated-bottleneck penalty compresses the vector into a latent embedding, optionally concatenated with 211 L2-normalized RDKit descriptors; property prediction is done by a feedforward MLP trained end-to-end with reconstruct

Load-bearing premise

The method assumes that SMILES substrings correspond to the same chemical substructure across different molecules, so a pattern mined from one molecule stays chemically meaningful in another; the SOTA comparisons also assume the quoted baselines were measured under the same scaffold-split and tuning protocol.

What would settle it

A reader could build a control with the identical architecture but a shuffled or random substructure vocabulary: if the accuracy gains persist, the functional-group chemistry is not the cause. More directly, one could evaluate FGR on a curated set of structural isomers and tautomers — if accuracy collapses there while graph models hold, the representation's blind spot is confirmed.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the FGR claim holds, molecular property prediction does not need graph or sequence encoders: a compact functional-group vocabulary plus an MLP suffices for state-of-the-art or near-SOTA results on many benchmarks.
  • Because the input features are named chemical groups, model attributions yield directly testable hypotheses for medicinal chemists, such as which substructures drive BACE1 inhibition or E. coli activity.
  • The fixed-size representation scales to arbitrarily long molecules such as peptides, where graph message passing struggles.
  • The framework performs well on low-data physical-chemistry datasets even without pre-training, suggesting a data-efficient route to property prediction.
  • The paper itself notes limitations: bit clashes from overlapping FG/MFG substructures, inability to distinguish structural isomers, and weaker performance on imbalanced datasets (HIV, MUV) and on quantum-mechanics tasks requiring 3D geometry.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural stress test is a control experiment where the MFG vocabulary is replaced with random SMILES substrings; if accuracy holds, the chemical semantics are not load-bearing, which would reframe the contribution as a factorization trick rather than a chemistry-driven representation.
  • Because FG and MFG bits overlap, a deduplication or hierarchical grouping of substructures could reduce redundancy and possibly improve both accuracy and interpretability — an extension the paper leaves open.
  • The SMILES-substring assumption implies the framework may be brittle for tautomeric or regiochemical variations; evaluating on isomer and tautomer test sets would clarify where the representation fails.
  • Since the latent space is human-legible, FGR could be used as a probe to compare whether larger graph models arrive at the same functional-group attributions, effectively turning FGR into an interpretive lens for opaque models.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes FGR, a functional-group-based molecular representation for property prediction. Molecules are encoded as multi-hot vectors over two vocabularies: 2,672 curated functional groups from ToxAlerts (FG) and up to 30,000 mined substrings from PubChem SMILES obtained by sequential pattern mining (MFG). These vectors are compressed by a tied-weight autoencoder with an uncorrelated-bottleneck penalty and focal reconstruction loss; the latent vector is optionally concatenated with 211 RDKit 2D descriptors and passed to a feedforward head. Results are reported on 33 datasets spanning MoleculeNet, MolMapNet, peptide-cleavage, and antibiotic benchmarks, with scaffold splits. The paper also presents alignment/uniformity analyses and feature-attribution studies claiming that important groups match established chemistry. The abstract claims state-of-the-art performance on all 33 benchmarks and says the framework leverages pre-training on unlabeled molecules.

Significance. If the claims were established, the contribution would be significant: a simple, parameter-efficient, chemically interpretable representation competitive with graph, sequence, and contrastive models would be practically valuable, and the explicit attempt to validate attributions against literature chemistry is a strength. The paper provides a code repository, comprehensive benchmark coverage, and multiple attribution algorithms. However, the headline SOTA claim is not currently supported by the evidence as presented: the tables show several clear losses to KANO, the baselines come from heterogeneous original publications rather than a unified protocol, and the PubChem-mined vocabulary may overlap with benchmark test molecules. These issues are fixable within the manuscript's scope, but they affect the central claim.

major comments (5)
  1. [Abstract; Tables 1–2] The abstract's claim of state-of-the-art performance on 33 benchmarks is contradicted by the paper's own tables. KANO outperforms FGR on BACE, MUV, HIV, Lipophilicity, qm8, and qm9 (e.g., Table 1: BACE 93.1 vs 89.3, HIV 85.1 vs 78.3; Table 2: qm8 MAE 0.0123 vs 0.0297, qm9 0.00320 vs 0.00547). Section 5 itself describes the framework as performing 'at par and sometimes better' than SOTA. Please either correct the claims to 'competitive on many tasks and best on several' or provide statistical evidence that supports the stronger SOTA statement.
  2. [§4.5.1, §4.5.2; Tables 1–5] All FGR results are reported under scaffold splits, but the baseline numbers are quoted from earlier papers without being re-run under the same split, metric, and tuning protocol. The original MolMapNet and KekuleScope evaluations, for instance, use different splitting conventions, and Tables 3–4 report only means with no uncertainty. Reported margins such as Tox21 +0.4 AUC, SIDER +2.6 AUC, and ESOL RMSE −0.05 are plausibly within cross-protocol noise. Please re-run all baselines under the identical protocol or clearly state that the comparisons are not controlled.
  3. [§4.2.2] The MFG vocabulary is mined from the full PubChem corpus (>114M molecules) with no documented exclusion of benchmark molecules. Several benchmarks used here, including MUV, HIV, Tox21, and ToxCast, are derived from PubChem or NCI bioassay records. If test-set structures contributed to the pattern-mining vocabulary, the representation can be said to have been constructed with knowledge of the test distribution, creating a transductive leak. Please quantify the overlap between benchmark SMILES and the PubChem sample used for mining, exclude overlapping structures from vocabulary construction, or ablate the effect of such overlap.
  4. [Abstract vs. §5] The abstract states that FGR 'leverages pre-training on a large dataset of unlabeled molecules,' but §5 lists pre-training the autoencoder on unlabeled molecules as future work, and §4.4 only notes the loss can optionally be combined with unlabeled data. If the reported experiments do not use pre-training, the abstract overstates the method; if they do, the conclusion is inaccurate. Please clarify and align the abstract, method description, and conclusions.
  5. [§4.2.2] The key assumption that identical SMILES substrings correspond to identical substructures across molecules is not validated. SMILES is non-unique, and ring closures, bond-order conventions, and tautomers can express the same substructure differently. Because the whole MFG representation depends on this assumption, please test the mined patterns against canonical SMILES and common tautomer/ring variants, and report the fraction of mined vocabulary affected by string-level artifacts.
minor comments (6)
  1. [§4.5.1 vs §2.2.1] The text says five independent runs in §4.5.1 but three independent runs in §2.2.1 and §2.3. Please make the number of runs consistent.
  2. [Algorithm 1] Algorithm 1 is not a sufficiently precise specification of the sequential pattern mining procedure. The update rule for V, the criterion for replacing substrings, and the termination condition are underspecified. Provide a complete pseudo-code description or cite an exact implementation/package.
  3. [Tables 3–4] These tables report only means, without standard deviations, for FGR and baselines. Please add uncertainty estimates or confidence intervals so that comparisons can be interpreted statistically.
  4. [§2.4.2] The uniformity analysis is described qualitatively from density plots. Report a numerical uniformity metric (e.g., divergence from a uniform angular distribution) to support claims about which representation is 'more uniform.'
  5. [§3] The phrase 'sp3-hybridized carbon atoms' is used as if it were a functional group in the vocabulary. Please clarify how this feature is encoded in the FG/MFG vocabularies, since it is an atom-hybridization property rather than a substructure group.
  6. [Eq. (3)] The uncorrelated-bottleneck loss is written as a sum over p×p entries without explicitly indexing the off-diagonal terms of Cov(z_G). Please define the covariance matrix and sum to be over i≠j.

Circularity Check

0 steps flagged

No circular derivation: FGR's property predictions are externally evaluated; the main concerns are benchmark-protocol and data-leak risks, not circularity.

full rationale

The paper is an empirical representation-learning study, not a derivation whose outputs coincide with its inputs by construction. The FGR pipeline fixes a functional-group vocabulary a priori from ToxAlerts and PubChem, encodes molecules as multi-hot vectors, compresses them with an autoencoder, and trains a feedforward network on benchmark labels (Section 4). No equation in Section 4 defines a reported test metric as an algebraic function of a fitted parameter in a way that would force a particular score; the benchmark numbers are computed on held-out scaffold-split test sets and compared with external baselines. There are no load-bearing self-citations: the references to GNN, KANO, MolMapNet, etc., are independent external works, and the paper does not import a uniqueness theorem from its own authors. The interpretability claims are validated against the independent chemistry literature, not by appealing to the framework's own assumptions. The genuine weaknesses of the paper are correctness/leak issues rather than circularity: (i) baseline values in Tables 1-5 are quoted from earlier papers under possibly different splits and tuning protocols, so the SOTA margins are not established under a single controlled comparison; (ii) the MFG vocabulary is mined from the full PubChem corpus with no documented exclusion of benchmark molecules, creating a possible transductive leak (Section 4.2.2); and (iii) the Abstract claims the framework 'leverages pre-training on a large dataset of unlabeled molecules,' while Section 5 says 'Future work can explore the effect of pre-training the autoencoder on a large dataset of unlabeled molecules,' a direct inconsistency. These are important evaluation and reporting concerns, but none of them makes the central prediction reduce by definition to the model's inputs or to a self-citation chain. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The representation depends on hand-tuned vocabularies and loss weights, and on the explicit domain assumption that SMILES substrings equal functional groups. No new physical entities are introduced; the only invented object is the MFG substring vocabulary, whose external validity is not established.

free parameters (4)
  • SPM frequency threshold eta = 500
    Chosen by hand in Section 4.2.2; controls which SMILES substrings enter the MFG vocabulary.
  • SPM maximum vocabulary size MVS = 30000
    Chosen by hand in Section 4.2.2; caps the MFG vocabulary and sets the input dimensionality.
  • Autoencoder latent dimension l = not reported in main text, deferred to SI S8-S10
    Core bottleneck hyperparameter of the representation, tuned by cross-validation.
  • Loss weights alpha, beta and focal gamma/alpha_t = set via cross-validation
    Balances reconstruction, uncorrelated-bottleneck and prediction losses in Sections 4.3 and 4.4.
axioms (4)
  • domain assumption SMILES substrings correspond to consistent chemical substructures across molecules
    Section 4.2.2 states 'The method assumes that the same SMILES sub-strings will represent sub-structures that appear across different molecules.' This is load-bearing for MFG.
  • domain assumption Scaffold split makes test molecules structurally novel and comparable across baselines
    Section 4.5.1 describes the scaffold split, but baseline numbers in Tables 1 to 5 are quoted from prior papers without re-running under the same protocol.
  • domain assumption ToxAlerts SMARTS patterns are valid, endpoint-relevant functional groups
    Section 4.2.1 uses curated toxicology alerts as the FG vocabulary for all tasks, including non-toxicity endpoints such as solubility and quantum properties.
  • domain assumption A linear encoder on multi-hot functional group vectors retains enough information for property prediction
    Equations (1) and (2) define a single-layer encoder with an uncorrelated bottleneck; representation power relies on the keys and descriptor concatenation.
invented entities (1)
  • MFG vocabulary (mined functional group substrings) no independent evidence
    purpose: Second hot-encoding channel for molecules, combined with curated FG to form the FGR input
    MFG patterns are SMILES substrings, not chemically validated functional groups. There is no falsifiable handle outside benchmark performance, and the vocabulary is built from PubChem, which overlaps benchmark molecules.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Functional Groups are All you Need for Chemically Interpretable Molecular Property Prediction." pith.science (2026). https://pith.science/paper/DE3LKFLQ

@misc{pith2026250909619,
  author       = {Pith},
  title        = {Pith review of: Functional Groups are All you Need for Chemically Interpretable Molecular Property Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DE3LKFLQ}},
  note         = {Machine review of arXiv:2509.09619}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Molecular property prediction using deep learning (DL) models has accelerated drug and materials discovery, but the resulting DL models often lack interpretability, hindering their adoption by chemists. This work proposes developing molecule representations using the concept of Functional Groups (FG) in chemistry. We introduce the Functional Group Representation (FGR) framework, a novel approach to encoding molecules based on their fundamental chemical substructures. Our method integrates two types of functional groups: those curated from established chemical knowledge (FG), and those mined from a large molecular corpus using sequential pattern mining (MFG). The resulting FGR framework encodes molecules into a lower-dimensional latent space by leveraging pre-training on a large dataset of unlabeled molecules. Furthermore, the proposed framework allows the inclusion of 2D structure-based descriptors of molecules. We demonstrate that the FGR framework achieves state-of-the-art performance on a diverse range of 33 benchmark datasets spanning physical chemistry, biophysics, quantum mechanics, biological activity, and pharmacokinetics while enabling chemical interpretability. Crucially, the model's representations are intrinsically aligned with established chemical principles, allowing chemists to directly link predicted properties to specific functional groups and facilitating novel insights into structure-property relationships. Our work presents a significant step toward developing high-performing, chemically interpretable DL models for molecular discovery.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

88 extracted references · 11 canonical work pages

  1. [1]

    & Nicolaou, C

    Shen, J. & Nicolaou, C. A. Molecular property prediction: recent trends in the era of artificial intel- ligence.Drug Discovery Today: Technologies32-33, 29–36 (2019). URL https://www.sciencedirect. com/science/article/pii/S1740674920300032

    2019

  2. [2]

    Walters, W. P. & Barzilay, R. Applications of Deep Learning in Molecule Generation and Molecular Property Prediction.Accounts of Chemical Research54, 263–270 (2021). URL https://doi.org/10. 1021/acs.accounts.0c00699. Publisher: American Chemical Society

    2021

  3. [3]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O. & Dahl, G. E. Neural message passing for Quantum chemistry (2017)

    2017

  4. [4]

    URL https://doi.org/10.1021/acs.jcim

    Yang, K.et al.Analyzing Learned Molecular Representations for Property Prediction.Journal of Chemical Information and Modeling59, 3370–3388 (2019). URL https://doi.org/10.1021/acs.jcim. 9b00237. Publisher: American Chemical Society

    doi:10.1021/acs.jcim 2019

  5. [5]

    URL https://www.nature.com/articles/ s42256-021-00438-4

    Fang, X.et al.Geometry-enhanced molecular representation learning for property predic- tion.Nature Machine Intelligence4, 127–134 (2022). URL https://www.nature.com/articles/ s42256-021-00438-4. Publisher: Nature Publishing Group

    2022

  6. [6]

    & Hahn, M

    Rogers, D. & Hahn, M. Extended-Connectivity Fingerprints.Journal of Chemical Information and Modeling50, 742–754 (2010). URL https://doi.org/10.1021/ci100050t. Publisher: American Chemical Society

    doi:10.1021/ci100050t 2010

  7. [7]

    L., Leland, B

    Durant, J. L., Leland, B. A., Henry, D. R. & Nourse, J. G. Reoptimization of MDL Keys for Use in Drug Discovery.Journal of Chemical Information and Computer Sciences42, 1273–1280 (2002). URL https://doi.org/10.1021/ci010132r. Publisher: American Chemical Society

    doi:10.1021/ci010132r 2002

  8. [8]

    & Bender, A

    Cort´ es-Ciriano, I. & Bender, A. KekuleScope: prediction of cancer cell line sensitivity and com- pound potency using convolutional neural networks trained on compound images.Journal of Cheminformatics11, 41 (2019). URL https://doi.org/10.1186/s13321-019-0364-5

    doi:10.1186/s13321-019-0364-5 2019

  9. [9]

    Shen, W. X.et al.Out-of-the-box deep learning prediction of pharmaceutical properties by broadly learned knowledge-based molecular representations.Nature Machine Intelligence3, 334–343 (2021). URL https://www.nature.com/articles/s42256-021-00301-6. Publisher: Nature Publishing Group

    2021

  10. [10]

    SMILES, a chemical language and information system

    Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules.Journal of Chemical Information and Computer Sciences28, 31–36 (1988). URL https://doi.org/10.1021/ci00057a005. Publisher: American Chemical Society

    doi:10.1021/ci00057a005 1988

  11. [11]

    & Aspuru-Guzik, A

    Krenn, M., H¨ ase, F., Nigam, A., Friederich, P. & Aspuru-Guzik, A. Self-referencing embedded strings (SELFIES): A 100% robust molecular string representation.Machine Learning: Science and Technology1, 045024 (2020). URL https://dx.doi.org/10.1088/2632-2153/aba947. Publisher: IOP Publishing

    doi:10.1088/2632-2153/aba947 2020

  12. [12]

    & Bjerrum, E

    Irwin, R., Dimitriadis, S., He, J. & Bjerrum, E. J. Chemformer: a pre-trained transformer for computational chemistry.Machine Learning: Science and Technology3, 015022 (2022). URL https: //dx.doi.org/10.1088/2632-2153/ac3ffb. Publisher: IOP Publishing

    doi:10.1088/2632-2153/ac3ffb 2022

  13. [13]

    & Do˘ gan, T

    Y¨ uksel, A., Ulusoy, E., ¨Unl¨ u, A. & Do˘ gan, T. SELFormer: molecular representation learning via SELFIES language models.Machine Learning: Science and Technology4, 025035 (2023). URL https://dx.doi.org/10.1088/2632-2153/acdb30. Publisher: IOP Publishing

    doi:10.1088/2632-2153/acdb30 2023

  14. [14]

    B., Hodas, N

    Goh, G. B., Hodas, N. O., Siegel, C. & Vishnu, A. SMILES2Vec: An Interpretable General-Purpose Deep Neural Network for Predicting Chemical Properties (2018). URL http://arxiv.org/abs/1712. 02034. ArXiv:1712.02034

    Pith/arXiv arXiv 2018

  15. [15]

    URL https://www.nature.com/articles/ s42256-022-00580-7

    Ross, J.et al.Large-scale chemical language representations capture molecular structure and prop- erties.Nature Machine Intelligence4, 1256–1264 (2022). URL https://www.nature.com/articles/ s42256-022-00580-7. Publisher: Nature Publishing Group

    2022

  16. [16]

    Kipf, T. N. & Welling, M. Semi-Supervised Classification with Graph Convolutional Networks (2017). URL https://openreview.net/forum?id=SJU4ayYgl

    2017

  17. [17]

    & Jegelka, S

    Xu, K., Hu, W., Leskovec, J. & Jegelka, S. How Powerful are Graph Neural Networks? (2018). URL https://openreview.net/forum?id=ryGs6iA5Km. 21

    2018

  18. [18]

    URL https://www.ijcai.org/proceedings/2020/392

    Song, Y.et al.Communicative Representation Learning on Attributed Molecular Graphs (2020). URL https://www.ijcai.org/proceedings/2020/392. ISSN: 1045-0823

    2020

  19. [19]

    Liu, S., Demirel, M. F. & Liang, Y. N-gram graph: simple unsupervised representation for graphs, with applications to molecules (2019)

    2019

  20. [20]

    & Lee, C.-K.Motif-based graph self-supervised learning for molecular property prediction, NIPS ’21, 15870–15882 (Curran Associates Inc., Red Hook, NY, USA, 2024)

    Zhang, Z., Liu, Q., Wang, H., Lu, C. & Lee, C.-K.Motif-based graph self-supervised learning for molecular property prediction, NIPS ’21, 15870–15882 (Curran Associates Inc., Red Hook, NY, USA, 2024)

    2024

  21. [21]

    Rong, Y.et al.Self-supervised graph transformer on large-scale molecular data (2020)

    2020

  22. [22]

    & Barati Farimani, A

    Wang, Y., Wang, J., Cao, Z. & Barati Farimani, A. Molecular contrastive learning of representations via graph neural networks.Nature Machine Intelligence4, 279–287 (2022). URL https://www. nature.com/articles/s42256-022-00447-x. Publisher: Nature Publishing Group

    2022

  23. [23]

    URL https://www.nature.com/articles/ s42256-023-00654-0

    Fang, Y.et al.Knowledge graph-enhanced molecular contrastive learning with functional prompt.Nature Machine Intelligence5, 542–553 (2023). URL https://www.nature.com/articles/ s42256-023-00654-0. Publisher: Nature Publishing Group

    2023

  24. [24]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature Machine Intelligence1, 206–215 (2019)

    Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature Machine Intelligence1, 206–215 (2019). URL https://www. nature.com/articles/s42256-019-0048-x. Publisher: Nature Publishing Group

    2019

  25. [25]

    URL https://doi.org/10.1093/nar/gkv951

    Kim, S.et al.PubChem Substance and Compound databases.Nucleic Acids Research44, D1202– D1213 (2016). URL https://doi.org/10.1093/nar/gkv951

    doi:10.1093/nar/gkv951 2016

  26. [26]

    & Veeramachaneni, K

    Zytek, A., Arnaldo, I., Liu, D., Berti-Equille, L. & Veeramachaneni, K. The Need for Interpretable Features: Motivation and Taxonomy.SIGKDD Explor. Newsl.24, 1–13 (2022). URL https://dl. acm.org/doi/10.1145/3544903.3544905

    arXiv 2022

  27. [27]

    A., Poda, G

    Sushko, I., Salmina, E., Potemkin, V. A., Poda, G. & Tetko, I. V. ToxAlerts: A Web Server of Structural Alerts for Toxic Chemicals and Compounds with Potential Adverse Reactions.Journal of Chemical Information and Modeling52, 2310–2316 (2012). URL https://doi.org/10.1021/ci300245q. Publisher: American Chemical Society

    doi:10.1021/ci300245q 2012

  28. [28]

    Pre-training Molecular Graph Representation with 3D Geometry(2021)

    Liu, S.et al. Pre-training Molecular Graph Representation with 3D Geometry(2021). URL https: //openreview.net/forum?id=xQUe1pOKPam

    2021

  29. [29]

    Maaten, L. v. d. & Hinton, G. Visualizing Data using t-SNE.Journal of Machine Learning Research 9, 2579–2605 (2008). URL http://jmlr.org/papers/v9/vandermaaten08a.html

    2008

  30. [30]

    Davies, D. L. & Bouldin, D. W. A Cluster Separation Measure.IEEE Transactions on Pattern Analy- sis and Machine IntelligenceP AMI-1, 224–227 (1979). URL https://ieeexplore.ieee.org/document/ 4766909. Conference Name: IEEE Transactions on Pattern Analysis and Machine Intelligence

    1979

  31. [31]

    Aromatic interactions at the ligand-protein interface: Implications for the development of docking scoring functions.Chemical Biology & Drug Design91, 380–390 (2018)

    Brylinski, M. Aromatic interactions at the ligand-protein interface: Implications for the development of docking scoring functions.Chemical Biology & Drug Design91, 380–390 (2018)

    2018

  32. [32]

    C.et al.3D Molecular Descriptors Important for Clinical Success.Journal of Chemical Information and Modeling53, 327–342 (2013)

    Kombo, D. C.et al.3D Molecular Descriptors Important for Clinical Success.Journal of Chemical Information and Modeling53, 327–342 (2013). URL https://doi.org/10.1021/ci300445e. Publisher: American Chemical Society

    doi:10.1021/ci300445e 2013

  33. [33]

    URL https://doi.org/10.1021/ acs.jmedchem.4c00652

    Guan, Q.et al.Triazoles in Medicinal Chemistry: Physicochemical Properties, Bioisosterism, and Application.Journal of Medicinal Chemistry67, 7788–7824 (2024). URL https://doi.org/10.1021/ acs.jmedchem.4c00652. Publisher: American Chemical Society

    2024

  34. [34]

    L.et al.The effect of halogenation on blood-brain barrier permeability of a novel peptide drug.Peptides20, 1229–1238 (1999)

    Gentry, C. L.et al.The effect of halogenation on blood-brain barrier permeability of a novel peptide drug.Peptides20, 1229–1238 (1999)

    1999

  35. [35]

    Schultz, T. W. & Yarbrough, J. W. Trends in structure-toxicity relationships for carbonyl-containing alpha,beta-unsaturated compounds.SAR and QSAR in environmental research15, 139–146 (2004)

    2004

  36. [36]

    Domalaon, R., Zhanel, G. G. & Schweizer, F. Short Antimicrobial Peptides and Peptide Scaffolds as Promising Antibacterial Agents.Current Topics in Medicinal Chemistry16, 1217–1230 (2016)

    2016

  37. [37]

    A., Ismail, N

    Mohamed, E. A., Ismail, N. S. M., Hagras, M. & Refaat, H. Medicinal attributes of pyridine scaffold as anticancer targeting agents.Future Journal of Pharmaceutical Sciences7, 24 (2021). URL https://doi.org/10.1186/s43094-020-00165-4

    doi:10.1186/s43094-020-00165-4 2021

  38. [38]

    R., Schauperl, M

    Loeffler, J. R., Schauperl, M. & Liedl, K. R. Hydration of Aromatic Heterocycles as an Adversary 22 ofπ-Stacking.Journal of Chemical Information and Modeling59, 4209–4219 (2019). URL https: //www.ncbi.nlm.nih.gov/pmc/articles/PMC7032848/

    2019

  39. [39]

    P.et al.Flavonoid and Chalcone Scaffolds as Inhibitors of BACE1: Recent Updates

    Narayanan, A. P.et al.Flavonoid and Chalcone Scaffolds as Inhibitors of BACE1: Recent Updates. Combinatorial Chemistry & High Throughput Screening27, 1243–1256 (2024)

    2024

  40. [40]

    Mureddu, L. G. & Vuister, G. W. Fragment-Based Drug Discovery by NMR. Where Are the Successes and Where can It Be Improved?Frontiers in Molecular Biosciences9(2022). URL https://www. frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2022.834453/full. Publisher: Frontiers

    arXiv 2022

  41. [41]

    Mar ´ ın, I. D. G.et al.New compounds from heterocyclic amines scaffold with multitarget inhibitory activity on Aβaggregation, AChE, and BACE1 in the Alzheimer disease.PLOS ONE17, e0269129 (2022). URL https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0269129. Publisher: Public Library of Science

    work page doi:10.1371/journal.pone.0269129 2022

  42. [42]

    URL https://doi.org/10.1021/jm3011405

    Yonezawa, S.et al.Conformational Restriction Approach toβ-Secretase (BACE1) Inhibitors: Effect of a Cyclopropane Ring To Induce an Alternative Binding Mode.Journal of Medicinal Chemistry55, 8838–8858 (2012). URL https://doi.org/10.1021/jm3011405. Publisher: American Chemical Society

    work page doi:10.1021/jm3011405 2012

  43. [43]

    Ghosh, A. K. & Osswald, H. L. BACE1 (β-secretase) inhibitors for the treatment of Alzheimer’s disease.Chemical Society Reviews43, 6765–6813 (2014). URL https://pubs.rsc.org/en/content/ articlelanding/2014/cs/c3cs60460h. Publisher: The Royal Society of Chemistry

    2014

  44. [44]

    URL https: //www.sciencedirect.com/science/article/pii/S0960894X06001636

    Kimura, T.et al.Design and synthesis of potentβ-secretase (BACE1) inhibitors with P1’ carboxylic acid bioisosteres.Bioorganic & Medicinal Chemistry Letters16, 2380–2386 (2006). URL https: //www.sciencedirect.com/science/article/pii/S0960894X06001636

    2006

  45. [45]

    URL https://www.sciencedirect.com/science/ article/pii/S0223523418304446

    Ghobadian, R.et al.Novel tetrahydrocarbazole benzyl pyridine hybrids as potent and selective butryl cholinesterase inhibitors with neuroprotective andβ-secretase inhibition activities.European Journal of Medicinal Chemistry155, 49–60 (2018). URL https://www.sciencedirect.com/science/ article/pii/S0223523418304446

    2018

  46. [46]

    Rosa, L. C. S., Argolo, C. O., Nascimento, C. M. C. & Pimentel, A. S. Identifying Substructures That Facilitate Compounds to Penetrate the Blood–Brain Barrier via Passive Transport Using Machine Learning Explainer Models.ACS Chemical Neuroscience15, 2144–2159 (2024). URL https://doi. org/10.1021/acschemneuro.3c00840. Publisher: American Chemical Society

    work page doi:10.1021/acschemneuro.3c00840 2024

  47. [47]

    Mikitsh, J. L. & Chacko, A.-M. Pathways for Small Molecule Delivery to the Central Nervous System Across the Blood-Brain Barrier.Perspectives in Medicinal Chemistry6, 11–24 (2014). URL https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4064947/

    2014

  48. [48]

    & Stiefl, N

    Beckers, M., Fechner, N. & Stiefl, N. 25 Years of Small-Molecule Optimization at Novartis: A Retro- spective Analysis of Chemical Series Evolution.Journal of Chemical Information and Modeling62, 6002–6021 (2022). URL https://doi.org/10.1021/acs.jcim.2c00785. Publisher: American Chemical Society

    work page doi:10.1021/acs.jcim.2c00785 2022

  49. [49]

    URL https://doi.org/10.1186/ s13068-023-02401-4

    Sen, O.et al.Escherichia coli displays a conserved membrane proteomic response to a range of alcohols.Biotechnology for Biofuels and Bioproducts16, 147 (2023). URL https://doi.org/10.1186/ s13068-023-02401-4

    2023

  50. [50]

    Hussein, A

    M. Hussein, A. H., A. El-Adasy, A.-B., M. El-Saghier, A., Olish, M. & H. Abdelmonsef, A. Synthesis, characterization, in silico molecular docking, and antibacterial activities of some new nitrogen- heterocyclic analogues based on a p -phenolic unit.RSC Advances12, 12607–12621 (2022). URL https://pubs.rsc.org/en/content/articlelanding/2022/ra/d2ra01794f. P...

    2022

  51. [51]

    URL https://www.frontiersin.org/journals/microbiology/articles/10.3389/ fmicb.2021.738896/full

    Zhang, G.et al.A Natural Antimicrobial Agent: Analysis of Antibacterial Effect and Mechanism of Compound Phenolic Acid on Escherichia coli Based on Tandem Mass Tag Proteomics.Frontiers in Microbiology12(2021). URL https://www.frontiersin.org/journals/microbiology/articles/10.3389/ fmicb.2021.738896/full. Publisher: Frontiers

    arXiv 2021

  52. [52]

    B.et al.Recent Advances in Pyridine Scaffold: Focus on Chemistry, Syn- thesis, and Antibacterial Activities.BioMed Research International2023, 9967591 (2023)

    Islam, M. B.et al.Recent Advances in Pyridine Scaffold: Focus on Chemistry, Syn- thesis, and Antibacterial Activities.BioMed Research International2023, 9967591 (2023). URL https://onlinelibrary.wiley.com/doi/abs/10.1155/2023/9967591. eprint: 23 https://onlinelibrary.wiley.com/doi/pdf/10.1155/2023/9967591

    work page doi:10.1155/2023/9967591 2023

  53. [53]

    P., Biabani, M

    Kock, I., Maskey, R. P., Biabani, M. A. F., Helmke, E. & Laatsch, H. 1-Hydroxy-1-norresistomycin and Resistoflavin Methyl Ether: New Antibiotics from Marine-derived Streptomycetes†,††.The Jour- nal of Antibiotics58, 530–534 (2005). URL https://www.nature.com/articles/ja200573. Publisher: Nature Publishing Group

    2005

  54. [54]

    C., Johnson, J

    Pawlowski, A. C., Johnson, J. W. & Wright, G. D. Evolving medicinal chemistry strategies in antibiotic discovery.Current Opinion in Biotechnology42, 108–117 (2016). URL https://www. sciencedirect.com/science/article/pii/S0958166916301148

    2016

  55. [55]

    Hinton, G. E. & Salakhutdinov, R. R. Reducing the Dimensionality of Data with Neural Net- works.Science313, 504–507 (2006). URL https://www.science.org/doi/10.1126/science.1127647. Publisher: American Association for the Advancement of Science

    doi:10.1126/science.1127647 2006

  56. [56]

    RDKit: A software suite for cheminformatics, computational chemistry, and predictive modeling.Greg Landrum8, 31 (2013)

    Landrum, G. RDKit: A software suite for cheminformatics, computational chemistry, and predictive modeling.Greg Landrum8, 31 (2013)

    2013

  57. [57]

    URL https://www

    Daylight Theory: SMARTS - A Language for Describing Molecular Patterns. URL https://www. daylight.com/dayhtml/doc/theory/theory.smarts.html

  58. [58]

    & Sun, J

    Huang, K., Xiao, C., Hoang, T., Glass, L. & Sun, J. CASTER: Predicting Drug Interactions with Chemical Substructure Representation.Proceedings of the AAAI Conference on Artificial Intelli- gence34, 702–709 (2020). URL https://ojs.aaai.org/index.php/AAAI/article/view/5412. Number: 01

    2020

  59. [59]

    & Doll´ ar, P

    Lin, T.-Y., Goyal, P., Girshick, R., He, K. & Doll´ ar, P. Focal Loss for Dense Object Detection (2017). URL https://www.computer.org/csdl/proceedings-article/iccv/2017/1032c999/ 12OmNApu5iv. ISSN: 2380-7504

    2017

  60. [60]

    Bemis, G. W. & Murcko, M. A. The Properties of Known Drugs. 1. Molecular Frameworks.Journal of Medicinal Chemistry39, 2887–2893 (1996). URL https://doi.org/10.1021/jm9602928. Publisher: American Chemical Society

    doi:10.1021/jm9602928 1996

  61. [61]

    & Hinton, G

    Sutskever, I., Martens, J., Dahl, G. & Hinton, G. On the importance of initialization and momentum in deep learning (2013). URL https://proceedings.mlr.press/v28/sutskever13.html. ISSN: 1938-7228

    2013

  62. [62]

    & Neyshabur, B

    Foret, P., Kleiner, A., Mobahi, H. & Neyshabur, B. Sharpness-aware Minimization for Efficiently Improving Generalization (2020). URL https://openreview.net/forum?id=6Tm1mposlrM

    2020

  63. [63]

    Bradley, A. P. The use of the area under the ROC curve in the evaluation of machine learning algo- rithms.Pattern Recognition30, 1145–1159 (1997). URL https://www.sciencedirect.com/science/ article/pii/S0031320396001422

    1997

  64. [64]

    F., Teixeira, A

    Martins, I. F., Teixeira, A. L., Pinheiro, L. & Falcao, A. O. A Bayesian Approach to in Silico Blood- Brain Barrier Penetration Modeling.Journal of Chemical Information and Modeling52, 1686–1697 (2012). URL https://doi.org/10.1021/ci300124c. Publisher: American Chemical Society

    doi:10.1021/ci300124c 2012

  65. [65]

    URL https://tripod.nih.gov/tox21/challenge/

    Tox21 Data Challenge 2014. URL https://tripod.nih.gov/tox21/challenge/

    2014

  66. [66]

    M.et al.ToxCast Chemical Landscape: Paving the Road to 21st Century Toxicol- ogy.Chemical Research in Toxicology29, 1225–1251 (2016)

    Richard, A. M.et al.ToxCast Chemical Landscape: Paving the Road to 21st Century Toxicol- ogy.Chemical Research in Toxicology29, 1225–1251 (2016). URL https://doi.org/10.1021/acs. chemrestox.6b00135. Publisher: American Chemical Society

    doi:10.1021/acs 2016

  67. [67]

    Kuhn, M., Letunic, I., Jensen, L. J. & Bork, P. The SIDER database of drugs and side effects. Nucleic Acids Research44, D1075–D1079 (2016). URL https://doi.org/10.1093/nar/gkv1075

    work page doi:10.1093/nar/gkv1075 2016

  68. [68]

    M., Madhukar, N

    Gayvert, K. M., Madhukar, N. S. & Elemento, O. A Data-Driven Approach to Predicting Successes and Failures of Clinical Trials.Cell Chemical Biology23, 1294–1301 (2016). URL https://www. sciencedirect.com/science/article/pii/S2451945616302914

    2016

  69. [69]

    & Denny, R

    Subramanian, G., Ramsundar, B., Pande, V. & Denny, R. A. Computational Modeling of Beta- Secretase 1 (BACE-1) Inhibitors Using Ligand Based Approaches.Journal of Chemical Information and Modeling56, 1936–1949 (2016). URL https://doi.org/10.1021/acs.jcim.6b00290. Publisher: American Chemical Society

    work page doi:10.1021/acs.jcim.6b00290 1936

  70. [70]

    Rohrer, S. G. & Baumann, K. Maximum Unbiased Validation (MUV) Data Sets for Virtual Screening Based on PubChem Bioactivity Data.Journal of Chemical Information and Modeling49, 169–184 (2009). URL https://doi.org/10.1021/ci8002649. Publisher: American Chemical Society. 24

    doi:10.1021/ci8002649 2009

  71. [71]

    URL https://wiki.nci.nih.gov/display/ NCIDTPdata/AIDS+Antiviral+Screen+Data

    AIDS Antiviral Screen Data - NCI DTP Data - NCI Wiki. URL https://wiki.nci.nih.gov/display/ NCIDTPdata/AIDS+Antiviral+Screen+Data

  72. [72]

    Delaney, J. S. ESOL: Estimating Aqueous Solubility Directly from Molecular Structure.Journal of Chemical Information and Computer Sciences44, 1000–1005 (2004). URL https://doi.org/10. 1021/ci034243x. Publisher: American Chemical Society

    2004

  73. [73]

    Mobley, D. L. & Guthrie, J. P. FreeSolv: a database of experimental and calculated hydration free energies, with input files.Journal of Computer-Aided Molecular Design28, 711–720 (2014). URL https://doi.org/10.1007/s10822-014-9747-x

    doi:10.1007/s10822-014-9747-x 2014

  74. [74]

    URL https://doi.org/10.1093/nar/gkr777

    Gaulton, A.et al.ChEMBL: a large-scale bioactivity database for drug discovery.Nucleic Acids Research40, D1100–D1107 (2012). URL https://doi.org/10.1093/nar/gkr777

    doi:10.1093/nar/gkr777 2012

  75. [75]

    Blum, L. C. & Reymond, J.-L. 970 Million Druglike Small Molecules for Virtual Screening in the Chemical Universe Database GDB-13.Journal of the American Chemical Society131, 8732–8733 (2009). URL https://doi.org/10.1021/ja902302h. Publisher: American Chemical Society

    work page doi:10.1021/ja902302h 2009

  76. [76]

    & von Lilienfeld, O

    Ramakrishnan, R., Hartmann, M., Tapavicza, E. & von Lilienfeld, O. A. Electronic spectra from TDDFT and machine learning in chemical space.The Journal of Chemical Physics143, 084111 (2015). URL https://doi.org/10.1063/1.4928757

    doi:10.1063/1.4928757 2015

  77. [77]

    O., Rupp, M

    Ramakrishnan, R., Dral, P. O., Rupp, M. & von Lilienfeld, O. A. Quantum chemistry structures and properties of 134 kilo molecules.Scientific Data1, 140022 (2014). URL https://www.nature. com/articles/sdata201422. Publisher: Nature Publishing Group

    2014

  78. [78]

    & Schmidt, F

    Wenzel, J., Matter, H. & Schmidt, F. Predictive Multitask Deep Neural Network Models for ADME- Tox Properties: Learning from Large Data Sets.Journal of Chemical Information and Modeling59, 1253–1268 (2019). URL https://doi.org/10.1021/acs.jcim.8b00785. Publisher: American Chemical Society

    work page doi:10.1021/acs.jcim.8b00785 2019

  79. [79]

    URL https://doi.org/10.1021/acs.jmedchem.9b00959

    Xiong, Z.et al.Pushing the Boundaries of Molecular Representation for Drug Discovery with the Graph Attention Mechanism.Journal of Medicinal Chemistry63, 8749–8760 (2020). URL https://doi.org/10.1021/acs.jmedchem.9b00959. Publisher: American Chemical Society

    doi:10.1021/acs.jmedchem.9b00959 2020

  80. [80]

    & Pei, J

    Li, X., Xu, Y., Lai, L. & Pei, J. Prediction of Human Cytochrome P450 Inhibition Using a Multitask Deep Autoencoder Neural Network.Molecular Pharmaceutics15, 4336–4345 (2018). URL https: //doi.org/10.1021/acs.molpharmaceut.8b00110. Publisher: American Chemical Society

    work page doi:10.1021/acs.molpharmaceut.8b00110 2018

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.