CoChat
Sign Up
Cited by 160Computational Drug Discovery MethodsTopological and Geometric Data AnalysisMachine Learning in Materials Science

Quantitative Toxicity Prediction Using Topology Based Multitask Deep Neural Networks

Open access
Wu, Kedi, Wei, Guo-Wei·2018·Journal of Chemical Information and Modeling
DOI 10.1021/acs.jcim.7b00558
Open this paper in CoChat — free accountOpen PDF

Abstract

The understanding of toxicity is of paramount importance to human health and environmental protection. Quantitative toxicity analysis has become a new standard in the field. This work introduces element specific persistent homology (ESPH), an algebraic topology approach, for quantitative toxicity prediction. ESPH retains crucial chemical information during the topological abstraction of geometric complexity and provides a representation of small molecules that cannot be obtained by any other method. To investigate the representability and predictive power of ESPH for small molecules, ancillary descriptors have also been developed based on physical models. Topological and physical descriptors are paired with advanced machine learning algorithms, such as the deep neural network (DNN), random forest (RF), and gradient boosting decision tree (GBDT), to facilitate their applications to quantitative toxicity predictions. A topology based multitask strategy is proposed to take the advantage of the availability of large data sets while dealing with small data sets. Four benchmark toxicity data sets that involve quantitative measurements are used to validate the proposed approaches. Extensive numerical studies indicate that the proposed topological learning methods are able to outperform the state-of-the-art methods in the literature for quantitative toxicity analysis. Our online server for computing element-specific topological descriptors (ESTDs) is available at http://weilab.math.msu.edu/TopTox/ .

AI snapshot

Written from the paper’s full text.

TL;DR

This paper introduces element-specific topological descriptors (ESTDs) from element-specific persistent homology and pairs them with multi-task deep neural networks (MT-DNN) for quantitative toxicity prediction. Across four benchmark toxicity datasets, the proposed topological learning methods are competitive with or outperform the state-of-the-art TEST software, especially when ESTDs are combined with physical descriptors and MT-DNN.

Key findings

  • On the fathead minnow LC50 set, MT-DNN with only ESTDs achieved R² = 0.738 and RMSE = 0.763, outperforming TEST consensus (R² = 0.728, RMSE = 0.768); the best result used all descriptors with consensus (R² = 0.789, RMSE = 0.677).
  • On the Daphnia magna LC50-DM set, MT-DNN with ESTDs achieved R² = 0.788 and RMSE = 0.805, compared with TEST consensus R² = 0.739 and RMSE = 0.911.
  • On the Tetrahymena pyriformis IGC50 set, consensus with all descriptors gave the best result (R² = 0.802, RMSE = 0.438), while TEST consensus gave R² = 0.764 and RMSE = 0.475.
  • On the oral rat LD50 set, consensus with all descriptors gave the best result (R² = 0.653, RMSE = 0.568), compared with TEST consensus R² = 0.626 and RMSE = 0.594; ESTD-based models had coverage 0.999 versus 0.997 for all descriptors.

Methods

  • ESTDs were computed from element-specific persistent homology using 25 element-specific networks (4 single-element and 21 two-element networks), yielding 28 ESTDs per network and 700 ESTDs total; auxiliary physical descriptors (330 components) were built from atomic charge, surface area, and electrostatic solvation free energy.
  • Four quantitative toxicity datasets were used: LC50 (823 molecules; 659 train/164 test), LC50-DM (353; 283/70), IGC50 (1792; 1434/358), and LD50 (7413 total, 7403 used; 5931/1482 or 5924/1479 after dropping As-containing molecules).
  • Models included random forest, gradient boosting decision tree, single-task DNN, and MT-DNN; results were averaged over 50 runs, with evaluation by R², RMSE, MAE, and prediction coverage.

Limitations

  • Stated: for the LD50 set, some molecules containing As were dropped due to force field failure; high-dimensional barcodes were not used because small-molecule barcodes are sparse; consensus models performed worse than MT-DNN on the small LC50-DM set, likely due to overfitting from many descriptors.
  • Stated: persistent homology filtration radius is not a chemical bond and topological connectivity is not a physical relationship.
  • Evident from the text: evaluation is limited to four benchmark datasets and comparisons to TEST; no prospective experimental validation is reported, and ESTDs slightly underperformed TEST consensus on the IGC50 set.

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

Integrity signals

Registry

Crossref · 10.1021/acs.jcim.7b00558

Cites

45 of 45 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 17 cited sources checked · 0 supported · 0 flagged

  • Findingown data
    § Abstract · p. 1

    The proposed topological learning methods are able to outperform the state-of-the-art methods in the literature for quantitative toxicity analysis.

    “Extensive numerical studies indicate that the proposed topological learning methods are able to outperform the state-of-the-art methods in the literature for quantitative toxicity analysis.”
    Quote in the paperSubstantial claim
  • Findingnone given
    § 1 Introduction · p. 4

    The proposed topological learning provides a very competitive description of relatively small drug-like molecules.

    “We demonstrate that the proposed topological learning provides a very competitive description of relatively small drug-like molecules.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 3.3 Daphnia magna LC 50 test set · p. 26

    The RMSE of Group contribution is the smallest among TEST methods for Daphnia magna LC50.

    “Though the RMSE of Group contribution is the smallest, its coverage is only 0.657 % which largely restricts this method's applicability.”
    Quote in the paperNot a substantial claim
  • Findingreasoning
    § 3.5 Oral rat LD 50 test set · p. 28

    None of the methods is able to provide a 100% coverage of the oral rat LD50 data set.

    “However, none of methods is able to provide a 100% coverage of this data set.”
    Quote in the paperNot a substantial claim
  • Findingown data
    § 3.5 Oral rat LD 50 test set · p. 29

    Our predictions are relatively stable for the oral rat LD50 set as R2s do not essentially fluctuate.

    “Meanwhile, our predictions are also relatively stable for this particular set as R 2 s do not essentially fluctuate.”
    Quote in the paperNot a substantial claim
  • May combine two claims. 29
  • Classified as probably not a substantial claim. 17
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data
    § 4.3 Alternative element specific networks for generating ESTDs · p. 32

    On the LC50-DM set, the accuracy of the new ESTDs is comparably lower (still higher than T.E.S.T consensus).

    “However on LC50-DM set, the accuracy is comparably lower (still higher than T.E.S.T consensus).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.2 The predictive power of ESTDs for toxicity · p. 31

    Models using only ESTDs achieve higher accuracy than the T.E.S.T. consensus method on the LC50 set and LC50DM set.

    “LC50 set and LC50DM set. Models using only ESTDs achieve higher accuracy than T.E.S.T. consensus method.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.3 Daphnia magna LC 50 test set · p. 26

    Five independent models in TEST software give significantly different predictions for Daphnia magna LC50, with RMSEs ranging from 0.810 to 1.190 log units.

    “Indeed, five independent models in TEST software give significantly different predictions, as indicated by RMSEs shown in Table 5 ranging from 0.810 to 1.190 log units.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.2 Feathead minnow LC 50 test set · p. 25

    Our MT-DNN model outperforms TEST consensus in both R2 and RMSE with only ESTDs as input for fathead minnow LC50.

    “As Table 4 indicates, our MT-DNN model outperforms TEST consensus both in terms of R 2 and RMSE with only ESTDs as input.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.1 The impact of descriptor selection and potential overfitting · p. 30

    When the number of descriptors is increased from 222, 254, 308, 411 to 1030, RMSE does not increase and R2 does not change much for the LC50 set.

    “When the number of descriptors is increased from 222, 254, 308, 411 to 1030, RMSE does not increase and R 2 does not change much.”
    Quote in the paperSubstantial claim

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

    Findingnone given
    § 3.5 Oral rat LD 50 test set · p. 29

    Two molecules in the test set that contain As element cannot be properly optimized for energy computation.

    “since 2 molecules in the test set that contains As element cannot be properly optimized for energy computation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3.3 Daphnia magna LC 50 test set · p. 27

    For Daphnia magna LC50 models using ESTDs, physical descriptors, and all descriptors, R2 improved from around 0.5 to 0.788, 0.705, and 0.726, respectively.

    “For models using ESTDs, physical descriptors and all descriptors, the R 2 has been improved from around 0.5 to 0.788, 0.705, and 0.726, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.4 Tetraphymena pyriformis IGC 50 test set · p. 27

    For IGC50, the R2 of different TEST methods fluctuates from 0.600 to 0.764 and TEST consensus prediction yields the best result for TEST software with R2 of 0.764.

    “As shown in Table 6, the R 2 of different TEST methods fluctuates from 0.600 to 0.764 and Test consensus prediction again yields the best result for TEST software with R 2 of 0.764.”
    Quote in the paperNot a substantial claim

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

    Findingreasoning
    § 3.5 Oral rat LD 50 test set · p. 29

    Our ESTDs have slightly higher coverage than physical descriptors (all combined descriptors) for the oral rat LD50 set.

    “It should also be noted that our ESTDs have slightly higher coverage than physical descriptors (all combined descriptors) since 2 molecules in the test set that contains As element cannot be properly optimized for energy computation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3.2 Feathead minnow LC 50 test set · p. 25

    The best result for fathead minnow LC50 is generated by the consensus method using all descriptors, with R2 of 0.789 and RMSE of 0.677 log(mol/L).

    “The best result is generated by consensus method using all descriptors, with R 2 of 0.789 and RMSE of 0.677 log(mol/L).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.3 Daphnia magna LC 50 test set · p. 26

    The Daphnia magna LC50 set is the smallest in set size, with 283 training molecules and 70 test molecules.

    “The Daphinia Magna LC50 set is the smallest in terms of set size, with 283 training molecules and 70 test molecules, respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.2 The predictive power of ESTDs for toxicity · p. 31

    On the LD50 set, the consensus result of ESTDs tops T.E.S.T. software in terms of both R2 and RMSE, and MT-DNN results outperform all non-consensus T.E.S.T methods.

    “LD50 set. Consensus result of ESTDs tops T.E.S.T. software in terms of both R 2 and RMSE and MT-DNN results outperform all non-consensus T.E.S.T methods.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.4 Tetraphymena pyriformis IGC 50 test set · p. 27

    Although ESTDs slightly underperform compared to physical descriptors for IGC50, its MT-DNN results defeat most TEST methods except the FDA method.

    “Although ESTDs slightly underperform compared to physical descriptors, its MT-DNN results are able to defeat most TEST methods except FDA method.”
    Quote in the paperSubstantial claim

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

    Findingnone given
    § 3.5 Oral rat LD 50 test set · p. 28

    The oral rat LD50 set contains the largest molecule pool with 7413 compounds.

    “The oral rat LD 5 0 set contains the largest molecule pool with 7413 compounds.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3.2 Feathead minnow LC 50 test set · p. 25

    TEST consensus gives the best prediction among TEST results for fathead minnow LC50, with a correlation coefficient of 0.728 and RMSE of 0.768 log(mol/L).

    “TEST consensus gives the best prediction 44 among TEST results, reporting a correlation coefficient of 0.728 and RMSE of 0.768 log(mol/L).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [44] Martin, T.User’s Guide for T.E.S.T. (version 4.2) (Toxicity Estimation Software Tool): A Program to Estimate Toxicity from Molecular Structure;USEPA, 2016.

      Not checked in the free run.

  • May combine two claims.

    Findingown data
    § 3.5 Oral rat LD 50 test set · p. 29

    Consensus method using all descriptors again yield the best results for all combinations, with optimal R2 of 0.653 and RMSE of 0.568 log(mol/kg) for the oral rat LD50 set.

    “Consensus method using all descriptors again yield the best results for all combinations, with optimal R 2 of 0.653 and RMSE of 0.568 log(mol/kg).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.5 Oral rat LD 50 test set · p. 28

    The TEST consensus result improves overall prediction accuracy of other TEST methods by about 10%, while other non-consensus methods all yield low R2 and high RMSE.

    “The TEST consensus result improves overall prediction accuracy of other TEST methods by about 10 %, however, other non-consensus methods all yield low R 2 and high RMSE.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3.1 An overview of data sets · p. 24

    For the first three datasets (LC50, LC50-DM, and IGC50), all molecules were properly included; for the LD50 set, some molecules containing element As were dropped out due to force field failure.

    “Note that for the first three datasets (i.e., LC 50 , LC 50 -DM and IGC 50 set), all molecules were properly included. However, for LD 50 set, some molecules involved element As were dropped out due to force field failure.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.5 Oral rat LD 50 test set · p. 29

    GBDT and MT-DNN with all descriptors yield the best (similar) results, giving slightly better results compared to TEST consensus.

    “In particular, GBDT and MT-DNN with all descriptors yield the best (similar) results, giving slightly better results compared to TEST consensus.”
    Quote in the paperSubstantial claim

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

    Findingnone given
    § 3.4 Tetraphymena pyriformis IGC 50 test set · p. 27

    The diversity of molecules in the IGC50 set is low and the coverage of TEST methods is relatively high compared to previous LC50 sets.

    “The diversity of molecules of in IGC 50 set is low and the coverage of TEST methods is relatively high compared to previous LC 50 sets.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3.4 Tetraphymena pyriformis IGC 50 test set · p. 27

    For IGC50, when all descriptors are used, predictions by GBDT and MT-DNN outperform TEST consensus, with R2 of 0.787 and RMSE of 0.455 log(mol/L).

    “When all descriptors are used, predictions by GBDT and MT-DNN outperform TEST consensus, with R 2 of 0.787 and RMSE of 0.455 log(mol/L).”
    Quote in the paperSubstantial claim

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

    Premisecitation
    § 1 Introduction · p. 4

    The integration of ESPH and machine learning gives rise to some of the most accurate predictions of protein-ligand binding affinities and mutation induced protein stability changes.

    “The integration of ESPH and machine learning gives rise to some of the most accurate predictions of protein-ligand binding affinities 27,28 and mutation induced protein stability changes. 26,28”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [27] source

      Not checked in the free run.

    • cites [28] source

      Not checked in the free run.

    • cites [26] Cang Z. X., 2017, Bioinformatics

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationnone given
    § Abstract · p. 1

    Element specific persistent homology (ESPH) retains crucial chemical information during the topological abstraction of geometric complexity and provides a representation of small molecules that cannot be obtained by any other method.

    “ESPH retains crucial chemical information during the topological abstraction of geometric complexity and provides a representation of small molecules that cannot be obtained by any other method.”
    Quote in the paperNot a substantial claim

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

    Premisenone given
    § 1 Introduction · p. 4

    ESPH has not been validated for its potential utility in small molecular characterization, analysis, and modeling.

    “However, ESPH has not been validated for its potential utility in small molecular characterization, analysis, and modeling.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Small molecules involve a large number of element types, are more diversified in their chemical compositions, are rich in structural variability, and are very sensitive to structural and compositional differences.

  • May combine two claims.

    Premisecitation
    § 1 Introduction · p. 2

    Multi-task learning strategies have brought new insights to bioinformatics because compounds from related assays may share features at various feature levels, which is helpful when the data set is small.

    “Indeed, MT learning strategies have brought new insights to bioinformatics since compounds from related assays may share features at various feature levels, which is extremely helpful if data set is small.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [9] Widmer C., 2012

      Not checked in the free run.

    • cites [10] source

      Not checked in the free run.

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

    Interpretationcitation
    § 3.3 Daphnia magna LC 50 test set · p. 26

    The usually low RMSE and MAE values of Group contribution might be typos.

    “Since Ref. 44 states that 'The consensus method achieved the best results in terms of both prediction accuracy and coverage', these usually low RMSE and MAE values might be typos.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [44] Martin, T.User’s Guide for T.E.S.T. (version 4.2) (Toxicity Estimation Software Tool): A Program to Estimate Toxicity from Molecular Structure;USEPA, 2016.

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given
    § 3 Results · p. 23

    Except for specifically mentioned cases, all results shown in the following tables are the average outputs of 50 numerical experiments.

    “Note that except for specifically mentioned, all our results shown in following tables are the average outputs of 50 numerical experiments.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given
    § 2.1.1 Persistent homology · p. 9

    The birth, death and persistence of a Betti number carry important chemical and/or biological information, which is the basis of the present method.

    “The birth, death and persistence of a Betti number carry important chemical and/or biological information, which is the basic of the present method.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given
    § 2.4.4 Network parameters and training · p. 20

    The ensemble results of different DNN models are essentially capable of reducing bias from individual predictions.

    “The underlying reason may be that the ensemble results of different DNN models is essentially capable of reducing bias from individual predictions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 4.1 The impact of descriptor selection and potential overfitting · p. 30

    Our models are essentially insensitive to the number of descriptors and thus there is little overfitting.

    “This behavior suggests that our models are essentially insensitive to the number of descriptors and thus there is little overfitting.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § 3.1 An overview of data sets · p. 24

    The TEST tool encounters a similar problem because results from two TEST models are unavailable and coverage from various TEST models is always smaller than one.

    “Apparently, the TEST tool encounters a similar problem since results from two TEST models are unavailable, and the coverage (fraction of molecules predicted) from various TEST models is always smaller than one.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning
    § 4.1 The impact of descriptor selection and potential overfitting · p. 30

    MT-DNN architecture can potentially take advantage over related tasks, which in turn reduces the potential overfitting on single dataset by the alternative training procedure.

    “MT-DNN architecture can also potentially take advantage over related tasks, which in turn reduces the potential overfitting on single dataset by the alternative training procedure.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 4.4 A potential improvement with consensus tools · p. 32

    It is likely that large number of descriptors may cause overfitting issues for most machine learning algorithms, and consequently generate large deviations, which eventually result in a large error of consensus method.

    “It is likely due to the fact that large number of descriptors may cause overfitting issues for most machine learning algorithms, and consequently generate large deviations, which eventually result in a large error of consensus method.”
    Quote in the paperSubstantial claim

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

    Premisecitation
    § 2.5 Evaluation criteria · p. 21

    Golbraikh et al. proposed a protocol to determine if a QSAR model has predictive power.

    “Golbraikh et al. 43 proposed a protocol to determine if a QSAR model has a predictive power.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [43] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). The quote only partly states this claim.

    Interpretationnone given
    § 3.5 Oral rat LD 50 test set · p. 29

    The inability to optimize the two As-containing molecules for energy computation is not an issue with our persistent homology computation.

    “However this is not an issue with our persistent homology computation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given
    § 2.1.3 ESTDs for small molecules · p. 14

    The birth position of a Betti-0 bar can signal the formation of hydrogen bonding, and the death position represents the disappearance of such effects, which reflects the strength of these effects.

    “The birth position can signal the formation of hydrogen bonding, and the death position represents the disappearance of such effects, which in turn reflects the strength of these effects.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 4.1 The impact of descriptor selection and potential overfitting · p. 31

    Our MT-DNN architecture is very robust against feature selection and can avoid overfitting.

    “Therefore our MT-DNN architecture is very robust against feature selection and can avoid overfitting.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation
    § 1 Introduction · p. 2

    Multi-task learning was proposed partially to deal with the data sparsity problem commonly encountered in QSAR applications.

    “Multi-task (MT) learning 8 was proposed partially to deal with data sparsity problem, which is commonly encountered in QSAR applications.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] Caruana, R.Learning to Learn;Springer, 1998; pp95–133. source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given
    § 3 Results · p. 23

    To eliminate randomness in neural network training, 50 models are built for each set of parameters and their average output is used as the final prediction.

    “Similarly, to eliminate randomness in neural network training, we build 50 models for each set of parameters and then use their average output as our final prediction.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § 3.3 Daphnia magna LC 50 test set · p. 27

    By learning related problems jointly and extracting shared information from different data sets, MT-DNN architecture can simultaneously perform multiple prediction tasks and enhances performances especially on small datasets.

    “This result suggests that by learning related problems jointly and extracting shared information from different data sets, MT-DNN architecture can simultaneously perform multiple prediction tasks and enhances performances especially on small datasets.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Different small molecules with different measured toxicity endpoints comprise distinct physical or chemical features, while descriptors such as the occurrence of certain chemical structure can result in similar toxicity property.
    • Rests on an assumption: A small data set with incomplete statistical distribution to establish an accurate predictive model can often be significantly benefited from relatively large data sets with more complete statistical distributions.
    • Rests on an assumption: Overfitting issue challenges traditional machine learning approaches especially when the number of samples is less than the number of descriptors.

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

    Interpretationnone given
    § 5 Conclusion · p. 34

    Our topological descriptors are much easier to construct than physical descriptors, which depend on physical models and force fields.

    “In fact, our topological descriptors are much easier to construct than physical descriptors, which depend on physical models and force fields.”
    Quote in the paperNot a substantial claim
41 of 41

  • The paper relies on this without support.

    Different quantitative toxicity endpoints are inherently correlated.

    “Additionally, the inherent correlation among different quantitative toxicity endpoints makes our topology based multitask strategy a viable approach to quantitative toxicity predictions.”

  • The paper relies on this without support.

    Small molecules involve a large number of element types, are more diversified in their chemical compositions, are rich in structural variability, and are very sensitive to structural and compositional differences.

    “In fact, unlike proteins, small molecules involve a large number of element types and are more diversified in their chemical compositions. They are also rich in structural variability in structures, including cis-trans distinctions and chiral and achiral stereoisomers. Small molecular properties are very sensitive to their structural and compositional differences.”

  • The paper relies on this without support.

    Toxicity experiments are usually very time consuming, cost intensive, and give rise to ethical concerns when it comes to animal tests.

    “Nevertheless, such experiments are usually very time consuming and cost intensive, and even give rise to ethical concerns when it comes to animal tests.”

  • The paper relies on this without support.

    Both random forest and gradient boosting decision tree are essentially insensitive to parameters and robust to redundant features.

    “Most importantly, both RF and GBDT are essentially insensitive to parameters and robust to redundant features.”

  • The paper relies on this without support.

    A typical shallow neural network is not able to construct hierarchical features and thus falls short in revealing more abstract properties, making it difficult to model complex non linear relationships.

    “However, it is not able to construct hierarchical features and thus falls short in revealing more abstract properties, which makes it difficult to model complex non linear relationships.”

  • The paper relies on this without support.

    Small molecules are geometrically relatively simple and their barcodes of high dimensions are usually very sparse.

    “However, small molecules are geometrically relatively simple and their barcodes of high dimensions are usually very sparse.”

  • The authors state this as an assumption.

    Different small molecules with different measured toxicity endpoints comprise distinct physical or chemical features, while descriptors such as the occurrence of certain chemical structure can result in similar toxicity property.

    “More concretely, it is reasonable to assume that different small molecules with different measured toxicity endpoints comprise distinct physical or chemical features, while descriptors such as the occurrence of certain chemical structure, can result in similar toxicity property.”

  • The paper relies on this without support.

    A small data set with incomplete statistical distribution to establish an accurate predictive model can often be significantly benefited from relatively large data sets with more complete statistical distributions.

    “Another advantage of MT learning is that a small data set with incomplete statistical distribution to establish an accurate predictive model can often be significantly benefited from relatively large data sets with more complete statistical distributions.”

  • The paper relies on this without support.

    Prediction accuracy can be increased by reducing the prediction coverage.

    “The prediction coverage, or fraction of chemical predicted, of corresponding methods is also taken into account since the prediction accuracy can be increased by reducing the prediction coverage.”

  • The paper relies on this without support.

    Overfitting issue challenges traditional machine learning approaches especially when the number of samples is less than the number of descriptors.

    “Indeed, overfitting issue challenges traditional machine learning approaches especially when the number of samples is less than the number of descriptors.”

  • The paper relies on this without support.

    The units of the four toxicity endpoints are of the same magnitude order, so no rescaling is needed.

    “Although the units are not exactly the same, it should be pointed out that no additional attempt was made to rescale the values since endpoints are of the same magnitude order.”

  • The paper relies on this without support.

    Experimental measurement of quantitative toxicity is typically expensive and time-consuming.

    “Experimental measurement of quantitative toxicity is extremely valuable, but is typically expensive and time consuming, in addition to potential ethic concerns.”

  • The paper relies on this without support.

    Theoretical prediction of quantitative toxicity has become a useful alternative in pharmacology and environmental science.

    “Theoretical prediction of quantitative toxicity has become a useful alternative in pharmacology and environmental science.”

  • The paper relies on this without support.

    The performances of toxicity prediction methods depend not only on the descriptors, but also on machine learning algorithms, which makes model evaluation difficult.

    “The performances of these methods depend not only on the descriptors, but also on machine learning algorithms, which makes the model evaluation a difficult task.”
14 of 14

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Wu et al. improved traditional molecular descriptors using ESPH and auxiliary descriptors, and achieved the highest degree of fitness and accuracy with RF, Gradient Boosting Decision Tree, single-task deep learning, and multi-task deep learning.

    “Wu et al. recently improved traditional molecular descriptors using element specific persistent homology (ESPH) and auxiliary descriptors, where ESPH includes topological information from intermolecular interactions and homology analysis on each component of molecules. On this basis, they performed RF, Gradient Boosting Decision Tree, single-task deep learning, multi-task deep learning, multi-task”

    From Machine Learning Based Toxicity Prediction: From Chemical Structural Description to Transcriptome Analysis (2018) · cites this paper as [118]

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.

  • DNN learning rates disagree

    Major

    The training description gives a learning rate of 0.0001, while Table 2 gives 0.001.

  • DNN layer counts and widths disagree

    Minor

    The training description specifies four 1,000-neuron layers, but Table 2 specifies seven layers, the last four having 100 neurons.

  • Coverage fraction is incorrectly written as a percentage

    Minor

    Coverage is reported as a fraction, and Table 5 gives group contribution coverage as 0.657. That fraction is 65.7%, not 0.657%.

    Reported 0.657 · recomputed 65.7

  • Filtration example reverses the connectivity rule

    Minor

    The pyridine example says atoms closer than the filtration value do not connect, whereas the stated Vietoris-Rips rule connects points within that radius.

  • Feature-selection threshold differs from the reported threshold

    Minor

    The methods list 1e-4 as one of four selected thresholds; Table 8 instead reports 1e-3 alongside the other three.

  • Stated Daphnia RMSE range excludes a table value

    Minor

    The discussion says the five independent TEST-model RMSEs start at 0.810, but Table 5 reports 0.803 for group contribution.

6 of 6

Cites

Flagged first, then by citations

Research works printed but not linked: 5

  1. Widmer C., 2012, Workshop on Unsupervised and Transfer Learning, 207
  2. Sutskever I., 2014, Adv. Neural Infor. Process. Syst, 3104
  3. Cang Z. X., 2017, Bioinformatics, 33, 3549
  4. Pedregosa F., 2011, J. Mach. Learn. Res., 12, 2825
  5. Nantasenamat C., 2009, EXCLI J., 8, 74

Cited by

Sorted by citations

Related

Authors
Wu, K., & Wei, G.-W.
Affiliations
Michigan State University (US)
Corresponding author
Wei, Guo-Wei
ORCID
Venue
Journal of Chemical Information and Modeling
Publisher
American Chemical Society (ACS)
Published
2018-01-31
Volume
58
Issue
2
Pages
520-531
ISSN
  • 1549-9596 (ISSN-L)
  • 1549-960X
DOI
10.1021/acs.jcim.7b00558
OpenAlex
W2773405722
Semantic Scholar
59586266612964c581cfdb92b58dea71ee5c7efa
Cited by
160
References
45
Citations by year
2026: 14 · 2025: 18 · 2024: 15 · 2023: 26 · 2022: 17 · 2021: 23 · 2020: 24 · 2019: 17 · 2018: 6
FWCI
12.94
Type
journal article
Language
en
Open access
Locations
  • Journal of Chemical Information and Modeling (publishedVersion)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • Division of Mathematical Sciences (DMS-1721024)
  • Division of Information and Intelligent Systems (IIS-1302285)
  • Center for Mathematical Molecular Biosciences, Michigan State University
Topics
  • Physical Sciences › Computer Science › Computational Theory and Mathematics › Computational Drug Discovery Methods
  • Physical Sciences › Computer Science › Computational Theory and Mathematics › Topological and Geometric Data Analysis
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
Keywords
algebraic topology, multi-task learning, topological indices, physical descriptors, random forest, gradient boosting decision tree, deep neural network
MeSH terms
Machine Learning, Algorithms, Neural Networks, Computer, Toxicity Tests

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 2026-10-10.

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

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