QSAR without borders

In absolute terms, the improvement in prospective predictions from DNNs over other ML methods is less than notable.

“However, in absolute terms, the improvement is less than notable.”

The squared correlation coefficient (R2) of models generated with DNN was only 0.04 higher on average than those built with RF.

“Thus, the squared correlation coefficient (R2) of models generated with DNN was only 0.04 higher (on average) than those built with RF as shown in Figure 3.”

Comparison of DNNs to other ML methods like RF and XGBoost on standard industrial QSAR datasets shows a statistically significant improvement in prospective predictions.

“Comparison of DNNs to other ML methods like RF and XGBoost on standard industrial QSAR datasets shows a statistically significant improvement in prospective predictions as shown in studies conducted by some of the authors of this paper, and similar conclusions have been published elsewhere.63”

• cites [63] Lenselink EB, ten Dijke N, Bongers B, Papadatos G, van Vlijmen HWT, Kowalczyk W, IJzerman AP and van Westen GJP, J. Cheminform, 2017, 9, 45. 29086168 PMC5555960 sourceNot checked yet

  Not checked in the free run.

When trained on the same data sets and descriptors, DNN predictions are not different to those of other methods.

“When trained on the same data sets and descriptors, DNN predictions are not different to those of other methods.64”

• cites [64] Winkler DA and Le TC, Mol. Inform, 2017, 36, 1600118. 29115063Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

QSAR modeling was one of the first research fields that highlighted the importance of data curation, rigorous validation of developed models, and data reproducibility.

“Importantly, QSAR modeling was one of the first research fields that highlighted the importance of data curation,13 rigorous validation of developed models,14 and data reproducibility,15 that has recently become a significant concern to the scientific community.16”

• cites [13] Oprea T, Olah M, Ostopovici L, Rad R and Mracec M, in EuroQSAR 2002—Designing Drugs and Crop Protectants: Processes Problems and Solutions, eds. Ford M, Livingstone D, Dearden J and Van de Waterbeemd HH, Blackwell Publishing, New York, 2003, pp. 314– 315.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [14] Golbraikh A and Tropsha A, J. Mol. Graph. Model, 2002, 20, 269– 276. 11858635 sourceNot checked yet

  Not checked in the free run.

• cites [15] Fourches D, Muratov E and Tropsha A, J. Chem. Inf. Model, 2016, 56, 1243– 1252. 27280890 PMC5657146 sourceNot checked yet

  Not checked in the free run.

QSAR modeling is based upon the premise that structurally similar compounds exhibit similar biological effects, the similarity-property principle.

“QSAR modeling is based upon the premise that structurally similar compounds exhibit similar biological effects, often referred to as the similarity-property principle (SPP)”

• cites [1] Hansch C, Maloney P, Fujita T and Muir R, Nature, 1962, 194, 178– 180.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Focusing potency predictions around local regions of SAR continuity can often lead to QSAR models with high predictive power.

“Focusing potency predictions around local regions of SAR continuity can often lead to QSAR models with high predictive power.”

The existence of activity cliffs in compound data sets is a major factor limiting QSAR predictions, often much greater than intrinsic limitations of modeling.

“The existence of activity cliffs in compound data sets is a major factor limiting QSAR predictions, often much greater than intrinsic limitations of modeling.22”

• cites [22] Maggiora GM, J. Chem. Inf. Model, 2006, 46, 1535. 16859285 sourceNot checked yet

  Not checked in the free run.

In QSAR modeling the presence of SAR continuities and discontinuities in sets of active compounds is not mutually exclusive; continuous and discontinuous SARs coexist in many data sets.

“In QSAR modeling the presence of SAR continuities and discontinuities in sets of active compounds is not mutually exclusive. Rather, continuous and discontinuous SARs coexist in many data sets21 resulting in the presence of adjacent gently sloped and rugged regions in activity landscapes (Figure 2).”

• cites [21] Peltason L and Bajorath J, J. Med. Chem,, DOI: 10.1021/jm0705713. sourceNot checked yet

  Not checked in the free run.

If the size of the experimental error in the measured dependent variable approaches the magnitude of the variation across multiple molecules in the dataset, it becomes increasingly hard to generate meaningful models.

“If the size of the experimental error in the measured dependent variable approaches the magnitude of the variation across multiple molecules in the dataset, it becomes increasingly hard to generate meaningful models.”

Outside of the pharmaceutical industry, the availability of large volumes of published or otherwise public domain data in databases like ChEMBL and PubChem has transformed the field.

“Outside of the pharmaceutical industry, the availability of large volumes of published, or otherwise public domain data in databases like ChEMBL29 and PubChem 30 has transformed the field.”

• cites [29] Gaulton A, Bellis LJ, Bento AP, Chambers J, Davies M, Hersey A, Light Y, McGlinchey S, Michalovich D, Al-Lazikani B and Overington JP, Nucleic Acids Res., 2012, 40, D1100– 7. 21948594 PMC3245175 sourceNot checked yet

  Not checked in the free run.

• cites [30] Kim S, Thiessen PA, Bolton EE, Chen J, Fu G, Gindulyte A, Han L, He J, He S, Shoemaker BA, Wang J, Yu B, Zhang J and Bryant SH, Nucleic Acids Res., 2015, 44, D1202– 13. 26400175 PMC4702940 sourceNot checked yet

  Not checked in the free run.

Validation of QSAR models for properties of chemical mixtures is more complicated than for traditional QSAR.

“Validation of QSAR models for properties of chemical mixtures is more complicated.”

Metrics that measure the prevalence of activity cliffs in a dataset are good predictors of the modelability of that dataset.

“Metrics that measure the prevalence of activity cliffs in a dataset are good predictors of the modelability of that dataset.41”

• cites [41] Golbraikh A, Muratov E, Fourches D and Tropsha A, J. Chem. Inf. Model, 2014, 54, 1– 4. 24251851 PMC3984298 sourceNot checked yet

  Not checked in the free run.

The concept of modelability proposes that predictivity of QSAR models is limited by activity cliffs.

“Assuming this is not an issue, and considering activity and descriptors together, the relatively new concept of modelability41 proposes that predictivity of QSAR models is then limited by activity cliffs.”

• cites [41] Golbraikh A, Muratov E, Fourches D and Tropsha A, J. Chem. Inf. Model, 2014, 54, 1– 4. 24251851 PMC3984298 sourceNot checked yet

  Not checked in the free run.

In the pharmaceutical industry, unlike other areas where ML is applied, the data required to build models is limited, expensive, and resource-intensive.

“It is important to remember that in the pharmaceutical industry, unlike other areas where ML is applied, the data required to build models is limited, expensive, and resource-intensive.67”

• cites [67] Hochreiter S, Klambauer G and Rarey M, J. Chem. Inf. Model, 2018, 58, 1723– 1724. 30109927Not checked yet

  Not checked in the free run.

DNNs can model more than one activity at a time (multi-task models).

“They can straightforwardly model more than one activity at a time (multi-task models);68”

• cites [68] Ramsundar B, Liu B, Wu Z, Verras A, Tudor M, Sheridan RP and Pande V, J. Chem. Inf. Model,, DOI: 10.1021/acs.jcim.7b00146. 28692267 sourceNot checked yet

  Not checked in the free run.

A time-split test set better simulates the natural evolution of a typical drug discovery project.

“A better ways to simulate the natural evolution of a typical drug discovery project is to use a time-split test set,31 i.e., assigning compounds tested in later phases of the project to the test set.”

• cites [31] Sheridan RP, J. Chem. Inf. Model, 2013, 53, 783– 90. 23521722 sourceNot checked yet

  Not checked in the free run.

Failure of ToxCast and QSARWorld Bioavailability Challenge indicates critical importance of data curation.

“The failure of this92 and QSARWorld Bioavailability Challenge indicates critical importance of data curation.93”

• cites [93] Fourches D, Muratov E and Tropsha A, J. Chem. Inf. Model, 2010, 50, 1189– 204. 20572635 PMC2989419 sourceNot checked yet

  Not checked in the free run.

Tuning hyper-parameters using a validation set chosen from the same pool as the test set lets information leak, making predictions overly optimistic.

“Another practice in ML is to tune hyper-parameters using a validation set, where both the validation and test sets have been chosen from the same pool of compounds. In effect, this lets information about the test set to leak into the training set of the model, which makes predictions overly optimistic, and thus this practice should be avoided.”

Consensus QSAR models from CERAPP and CoMPARA predicted ER and AR potentials of 32,464 new chemicals.

“The resulting consensus QSAR models leveraged knowledge from the groups and were used to predict ER and AR potentials of 32,464 new chemicals.”

• cites [89] Mansouri K, Abdelaziz A, Rybacka A, Roncaglioni A, Tropsha A, Varnek A, Zakharov A, Worth A, Richard AM, Grulke CM, Trisciuzzi D, Fourches D, Horvath D, Benfenati E, Muratov E, Wedebye EB, Grisoni F, Mangiatordi GF, Incisivo GM, Hong H, Ng HW, Tetko IV, Balabin I, Kancherla J, Shen J, Burton J, Nicklaus M, Cassotti M, Nikolov NG, Nicolotti O, Andersson PL, Zang Q, Politi R, Beger RD, Todeschini R, Huang R, Farag S, Rosenberg SA, Slavov S, Hu X and Judson RS, Environ. Health Perspect, 2016, 124, 1023– 1033. 26908244 PMC4937869 sourceNot checked yet

  Not checked in the free run.

Toxicity alerts generally have lower prediction accuracies compared to QSAR models.

“However, toxicity alerts generally have lower prediction accuracies compared to QSAR models.99”

• cites [99] Alves VM, Muratov EN, Capuzzi SJ, Politi R, Low Y, Braga RC, Zakharov AV, Sedykh A, Mokshyna E, Farag S, Andrade CH, Kuz’min VE, Fourches D and Tropsha A, Green Chem., 2016, 18, 4348– 4360. 28503093 PMC5423727Not checked yet

  Not checked in the free run.

Incorporation of information about metabolic activation can improve toxicity QSAR models.

“Therefore, incorporation of information about metabolic activation can improve toxicity QSAR models.81”

• cites [81] Rybacka A, Rudén C, Tetko IV and Andersson PL, Chemosphere, 2015, 139, 372– 378. 26210185Not checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: Metabolites can cause toxicity even when the precursor has low toxicity.

Combination of in vitro and in silico predictions contributed better models for a number of in vivo endpoints.

“Indeed, it was reported that combination of in vitro and in silico predictions contributed better models for a number of in vivo endpoints.97”

• cites [97] Abdelaziz A, Sushko Y, Novotarskyi S, Körner R, Brandmaier S and Tetko IV, Comb. Chem. High Throughput Screen, 2015, 18, 420– 38. 25747436 sourceNot checked yet

  Not checked in the free run.

A recent study showed that multi-task modeling consistently improved the accuracy of models for prediction of 29 in-vivo endpoints using 87K chemical structures from RTECS.

“A recent study showed that multi-task modeling consistently improved the accuracy of models for prediction of 29 in-vivo endpoints using 87K chemical structures collected from the Registry of Toxic Effects of Chemical Substances (RTECS) database.98”

• cites [98] Sosnin S, Karlov D, Tetko IV and Fedorov MV, J. Chem. Inf. Model, 2019, 59, 1062– 1072. 30589269Not checked yet

  Not checked in the free run.

Recent progress in neural networks have accelerated development of robust and predictive QSAR models for in vitro toxicity.

“The results of the Tox21 Challenge indicated that recent progress in neural networks have accelerated development of robust and predictive QSAR models for in vitro toxicity.”

Tox21 Data Challenge winning team used a DNN to build multi-task models for 12 cellular stress assays.

“The winning team (as determined by the AUC metric) used a DNN to build multi-task models for these outcomes.86”

• cites [86] Mayr A, Klambauer G, Unterthiner T and Hochreiter S, Front. Environ. Sci, 2016, 3, 80.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Development of new experimental techniques such as RNA-Seq provides new types of data for in vitro assessment of toxicities that can be used for QSAR modeling.

“It should be emphasized that development of new experimental techniques such as deep-sequencing RNA-Seq,90 provides new types of data for in vitro assessment of toxicities that can also be used for QSAR modeling.91”

• cites [90] Wang Z, Gerstein M and Snyder M, Nat. Rev. Genet, 2009, 10, 57– 63. 19015660 PMC2949280 sourceNot checked yet

  Not checked in the free run.

• cites [91] Liu R, Yu X and Wallqvist A, J. Cheminform, 2015, 7, 4. 25717346 PMC4339691 sourceNot checked yet

  Not checked in the free run.

The majority of pharmaceutical agents interact with several, sometimes many, biological targets.

“Consequently, it is now generally accepted that the majority of pharmaceutical agents interact with several, sometimes many, biological targets.”

• cites [102] Roth BL, Sheffler DJ and Kroeze WK, Nat.Rev.Drug Discov, 2004, 3, 353– 359. 15060530Not checked yet

  Not checked in the free run.

In vitro assays in ToxCast were not predictive of complex in vivo endpoint.

“Since in vitro assays in ToxCast were not predictive of such complex endpoint,95”

• cites [95] Wetmore BA, Wambaugh JF, Ferguson SS, Sochaski MA, Rotroff DM, Freeman K, Clewell HJ, Dix DJ, Andersen ME, Houck KA, Allen B, Judson RS, Singh R, Kavlock RJ, Richard AM and Thomas RS, Toxicol. Sci, 2012, 125, 157– 174. 21948869 sourceNot checked yet

  Not checked in the free run.

It is still unclear from ML literature whether DNNs are distinctly better at QSAR tasks than standard methods, because exhaustive comparison has not been made.

“To summarize, it is still unclear from the ML literature whether DNNs are distinctly better at QSAR tasks than standard methods, because in most cases an exhaustive comparison has not been made.”

CNNs have provided more accurate predictions than descriptor-based DNNs in some cases.

“In some cases, CNN has provided more accurate predictions than descriptor based DNNs.71”

• cites [71] Coley CW, Barzilay R, Green WH, Jaakkola TS and Jensen KF, J. Chem. Inf. Model, 2017, 57, 1757– 1772. 28696688 sourceNot checked yet

  Not checked in the free run.

DNN methods require substantially larger datasets to fully capitalize on their advantages.

“DNN methods also require substantially larger datasets to fully capitalize on their advantages76”

• cites [76] Baskin II, Winkler D and Tetko IV, Expert Opin. Drug Discov, 2016, 11, 785– 795. 27295548 sourceNot checked yet

  Not checked in the free run.

This often generates beneficial therapeutic activities due to additive or synergistic pharmacological effects.

“This often generates beneficial therapeutic activities,102 due to additive or synergistic pharmacological effects.103”

• cites [102] Roth BL, Sheffler DJ and Kroeze WK, Nat.Rev.Drug Discov, 2004, 3, 353– 359. 15060530 sourceNot checked yet

  Not checked in the free run.

• cites [103] Lagunin A, Filimonov D and Poroikov V, Curr. Pharm. Des, 2010, 16, 1703– 1717. 20222853 sourceNot checked yet

  Not checked in the free run.

CNNs are better at predicting quantum chemical energies.

“For example, CNN is better at predicting quantum chemical energies.72”

• cites [72] Faber FA, Hutchison L, Huang B, Gilmer J, Schoenholz SS, Dahl GE, Vinyals O, Kearnes S, Riley PF and von Lilienfeld OA, J. Chem. Theory Comput, 2017, 13, 5255– 5264. 28926232 sourceNot checked yet

  Not checked in the free run.

DNNs provide the possibility of inverting the QSAR model (inverse QSAR).

“They provide the possibility of inverting the QSAR model (inverse QSAR), i.e. designing molecules directly from the model (so called generative models).73”

• cites [73] Merk D, Friedrich L, Grisoni F and Schneider G, Mol. Inform, 2018, 37, 1700153. PMC5838524 29319225 sourceNot checked yet

  Not checked in the free run.

A model built with an Associative Neural Network had similar prediction performance to the winning DNN in Tox21.

“Model built with an Associative Neural Network87 had similar prediction performance.”

• cites [87] Tetko IV, in Methods in molecular biology, Humana Press, Clifton, 2008, vol. 458, pp. 180– 197.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

These methods have generally lower prediction accuracy than ML approaches using traditional QSAR descriptors.

“However, these methods have generally lower prediction accuracy than ML approaches using traditional QSAR descriptors.88”

• cites [88] Wu Y and Wang G, Int. J. Mol. Sci, 2018, 19, 2358.Not checked yet

  Not checked in the free run.

New types of DNN have opened up new applications, allowing simpler molecular representations such as SMILES strings or chemical graphs to be used to generate useful toxicity models.

“The development of new types of DNN76 has opened up new applications, allowing simpler molecular representations, such as SMILES strings or chemical graphs to be used to generate useful toxicity (and other property) models.”

• cites [76] Baskin II, Winkler D and Tetko IV, Expert Opin. Drug Discov, 2016, 11, 785– 795. 27295548 sourceNot checked yet

  Not checked in the free run.

ToxCast Lowest Effect Level prediction challenge: highest prediction accuracy with lowest RMSE of 1.08 log units was achieved using a consensus prediction of Associative Neural Network models.

“The highest prediction accuracy with the lowest RMSE of 1.08 log units was achieved using a consensus prediction of Associative Neural Network87 models developed with several sets of descriptors.92”

• cites [92] Novotarskyi S, Abdelaziz A, Sushko Y, Körner R, Vogt J and Tetko IV, Chem. Res. Toxicol, 2016, 29, 768– 775. 27120770 PMC5413193 sourceNot checked yet

  Not checked in the free run.

Random-split validation makes predictions appear good because test and training sets cover about the same chemical space.

“Random-split validation (which is still a literature standard) makes predictions that appear to be good because the test and training sets cover about the same chemical space, a difficult constraint as predictions outside of the model AD are likely to be poor).”

When used separately, biological descriptors resulted in a model with lower statistical significance than the model based on chemical descriptors.

“Notably, when used separately, biological descriptors resulted in a model with lower statistical significance than the model based on chemical descriptors.”

• cites [128] Bologa CG, Ursu O, Halip L, Curpăn R and Oprea TI, Rev. Roum. du Chim, 2015, 60, 219– 226. PMC4557807 26346852Not checked yet

  Not checked in the free run.

Gonzalez-Diaz et al. developed PTML methods for simultaneous prediction of many target properties under several experimental conditions.

“More recently, Gonzalez-Diaz et al.109 developed the perturbation theory machine learning (PTML) methods that search for QSAR models capable of simultaneous prediction of many target properties under several experimental conditions.”

• cites [109] González-Díaz H, Arrasate S, Gómez-SanJuan A, Sotomayor N, Lete E, Besada-Porto L and Ruso JM, Curr. Top. Med. Chem, 2013, 13, 1713– 41. 23889050 sourceNot checked yet

  Not checked in the free run.

CGR technology can efficiently model optimal reaction conditions.

“In summary, the CGR technology can efficiently model optimal reaction conditions.”

Hybrid DNN models can rather accurately associate genes and small molecules to up- or down-regulation.

“These results suggested that ‘hybrid’ DNN models can rather accurately associate genes and small molecules to up- or down-regulation.”

GO terms and HTS results can be translated into input features for cheminformatics models.

“For instance, it was demonstrated that gene ontology (GO) terms125 and HTS results can be translated into input features for cheminformatics models.126”

• cites [125] The Gene Ontology Consortium, Nucleic Acids Res., 2017, 45, D331– D338. 27899567 PMC5210579 sourceNot checked yet

  Not checked in the free run.

• cites [126] Hsing M, Byler K and Cherkasov A, BMC Syst. Biol, 2008, 2, 80. 18796161 PMC2553323 sourceNot checked yet

  Not checked in the free run.

QSAR descriptors and GO terms can be combined within a unified QSAR model capable of predicting the effect of a given molecule on a particular gene.

“QSAR descriptors and GO terms can be combined within a unified QSAR model capable of predicting the effect of a given molecule on a particular gene.129”

• cites [129] Woo G, Fernandez M, Hsing M, Lack NA, Cavga AD and Cherkasov A, Bioinformatics, 2019, 10.1093/bioinformatics/btz645. 31504186 sourceNot checked yet

  Not checked in the free run.

General-purpose ML models in chemistry predict potential energies and other molecular properties accurately for a broad class of chemical systems.

“Recent breakthroughs in the development of ML methods in chemistry175 have produced general purpose models that predict potential energies and other molecular properties accurately for a broad class of chemical systems.”

• cites [175] Butler KT, Davies DW, Cartwright H, Isayev O and Walsh A, Nature, 2018, 559, 547– 555. 30046072 sourceNot checked yet

  Not checked in the free run.

PCM may provide good estimates of ligand-target affinity in a single model by combining data from multiple assays.

“Careful review of the published results of PCM modeling leads to the conclusion that it may provide good estimates of ligand-target affinity in a single model by combining data from multiple assays (Figure 4).”

In vivo rat oral toxicity was predicted by combining endpoints of 499 HTS assays with 548 circular Morgan descriptors.

“In a related work,128 in vivo rat oral toxicity was predicted by combining endpoints of 499 HTS assays (biological variables) with 548 circular Morgan descriptors (chemical variables).”

• cites [128] Bologa CG, Ursu O, Halip L, Curpăn R and Oprea TI, Rev. Roum. du Chim, 2015, 60, 219– 226. PMC4557807 26346852Not checked yet

  Not checked in the free run.

A recent study demonstrated significant advantages of PCM approach and inductive transfer of knowledge between targets over traditional methods.

“A recent study, has also demonstrated significant advantages of PCM approach and inductive transfer of knowledge between the targets over traditional methods.120”

• cites [120] Paricharak S, Cortés-Ciriano I, IJzerman AP, Malliavin TE and Bender A, J. Cheminform, 2015, 7, 15. 25926892 PMC4413554 sourceNot checked yet

  Not checked in the free run.

Ignoring the data curation step of the modeling workflow will significantly affect the quality of the training data and models derived from them.

“Ignoring the data curation step of the modeling workflow will significantly affect the quality of the training data and models derived from them.93”

• cites [93] Fourches D, Muratov E and Tropsha A, J. Chem. Inf. Model, 2010, 50, 1189– 204. 20572635 PMC2989419 sourceNot checked yet

  Not checked in the free run.

Reaction data from public databases is of mixed quality.

“Generally, reaction data from public databases is of mixed quality.”

• cites [139] Lin AI, Madzhidov TI, Klimchuk O, Nugmanov RI, Antipin IS and Varnek A, J. Chem. Inf. Model, 2016, 56, 2140– 2148. 27783508 sourceNot checked yet

  Not checked in the free run.

Binary ML models were trained to distinguish ATG genes from non-autophagy genes with cross-validated AUC values in the 0.95-0.99 range.

“These binary ML models were trained to distinguish ATG genes from non-autophagy genes (cross-validated AUC values were in the 0.95-0.99 range).”

• cites [133] Chen T and Guestrin C, in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD ‘16, ACM Press, New York, New York, USA, 2016, pp. 785– 794.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

ANI-1 predicted DFT energies of test set molecules with up to 10 heavy atoms with RMSE values below 0.57 kcal/mol.

“It predicted DFT energies of the test set molecules with up to 10 heavy atoms very well, with the resulting RMSE values below 0.57 kcal/mol.”

• cites [189] Smith JS, Isayev O and Roitberg AEAEE, Chem. Sci, 2017, 8, 3192– 3203. 28507695 PMC5414547 sourceNot checked yet

  Not checked in the free run.

The transfer learning-based ANI-1ccx method outperforms DFT on test cases, especially those where DFT fails to capture reaction thermochemistry.

“The computed reaction energy values demonstrated that the transfer learning-based ANI-1ccx method outperforms DFT on test cases, especially those where DFT fails to capture reaction thermochemistry.”

A SOAP-GAP model within a ML framework (GPR) achieved 99% accurate atomic-scale properties for Si surface reconstruction, stability of molecules, and protein ligands.

“Gaussian process regression (GPR) was used to generate very accurate Gaussian atomic potentials (GAP) and then to train a SOAP–GAP model within a ML framework (GPR) that achieved a 99% accurate atomic-scale properties for Si surface reconstruction, stability of molecules, and protein ligands.209”

• cites [209] Bartók AP, De S, Poelking C, Bernstein N, Kermode JR, Csányi G and Ceriotti M, Sci. Adv, 2017, 3, e1701816. 29242828 PMC5729016 sourceNot checked yet

  Not checked in the free run.

Hautier et al. used experimental data to create a probabilistic framework for ionic substitution capable of dealing with sparse spaces (quaternary configurations).

“For instance, Hautier et al.200 used experimental data to create a probabilistic framework for ionic substitution capable of dealing with sparse spaces (quaternary configurations).”

• cites [200] Hautier G, Fischer C, Ehrlacher V, Jain A and Ceder G, Inorg. Chem, 2011, 50, 656– 663. 21142147 sourceNot checked yet

  Not checked in the free run.

ANI-1ccx represents an attractive alternative to DFT and standard force fields for conformational searches, molecular dynamics, and calculation of reaction energies.

“Such methodology resulted in the development of the ANI-1ccx potential, which represents an attractive alternative to DFT and standard force fields for conformational searches, molecular dynamics, and the calculation of reaction energies.”

Sequence-to-sequence models applied to chemical reactions provided similar performance (ca 37% for top-1) to rule-based systems (35%).

“This methodology was applied to model reaction products and for retrosynthetic reaction route prediction, which provided similar performance (ca 37% for top-1) to rule-based systems (35%).140”

• cites [140] Liu B, Ramsundar B, Kawthekar P, Shi J, Gomes J, Luu Nguyen Q, Ho S, Sloane J, Wender P and Pande V, ACS Cent. Sci, 2017, 3, 1103– 1113. 29104927 PMC5658761 sourceNot checked yet

  Not checked in the free run.

Use of an advanced Transformer architecture boosted the accuracy of predictions to about 43%.

“A use of an advanced Transformer architecture, which was initially used for English-to-German translation, boosted the accuracy of predictions to about 43%.153”

• cites [153] Karpov P, Godin G and Tetko IV, Springer, Cham, 2019, pp. 817– 830.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

GTM can be used to predict properties of new reactions projected on the map.

“Unlike SOM and many other dimensionality reduction methods, GTM can be used to predict properties of new reactions projected on the map.”

De et al. used the smooth overlap of atomic positions (SOAPs) to introduce the "alchemical similarity" descriptor for comparing molecular and periodic structures.

“In a related study, De et al.184 used the smooth overlap of atomic positions (SOAPs) to introduce a very useful descriptor for comparing structures: the “alchemical similarity” for molecular and periodic structures.”

• cites [184] De S, Bartók AP, Csányi G and Ceriotti M, Phys. Chem. Chem. Phys, 2016, 18, 13754. 27101873 sourceNot checked yet

  Not checked in the free run.

Template-free methodology allows one to automatically enlarge the list of transformation rules as soon as new data are available.

“This allows one to automatically enlarge the list of transformation rules as soon as new data are available.”

QM7 is difficult to model by ML because of its relatively small size.

“This benchmark is difficult to model by ML because of its relatively small size.”

Currently, QSRR modeling is performed on large and diverse datasets that account for solvent effects and temperature for many types of chemical reactions using NN approaches.

“Currently, quantitative structure-reactivity relationship (QSRR) modeling is performed on large and diverse datasets that account for solvent effects and temperature for many types of chemical reactions using NN approaches.”

A 70% match with experimental conditions was found within the top-10 predictions of Gao et al.'s models.

“A 70% match with experimental conditions was found within the top-10 predictions.”

• cites [163] Gao H, Struble TJ, Coley CW, Wang Y, Green WH and Jensen KF, ACS Cent. Sci, 2018, 4, 1465– 1476. 30555898 PMC6276053 sourceNot checked yet

  Not checked in the free run.

The design of blue emitters for organic light-emitting diode devices was accomplished by virtual screening of half a million molecules.

“One such example is the design of blue emitters for organic light-emitting diode devices accomplished by virtual screening of half a million molecules .172”

• cites [172] Gómez-Bombarelli R, Aguilera-Iparraguirre J, Hirzel TD, Duvenaud D, Maclaurin D, Blood-Forsythe MA, Chae HS, Einzinger M, Ha D-G, Wu T, Markopoulos G, Jeon S, Kang H, Miyazaki H, Numata M, Kim S, Huang W, Hong SI, Baldo M, Adams RP and Aspuru-Guzik A, Nat. Mater, 2016, 15, 1120– 1127. 27500805 sourceNot checked yet

  Not checked in the free run.

Retrosynthesis predictions can be significantly improved by algorithms originally developed for very different purposes.

“This result indicates that retrosynthesis predictions can be significantly improved by algorithms originally developed for very different purposes.”

This virtual screening approach led to the successful discovery of three lead candidate compounds with state-of-the-art performance.

“This approach led the successful discovery of three lead candidate compounds with state-of-the-art performance,172 exemplifying the promise of closed loop discovery.”

• cites [172] Gómez-Bombarelli R, Aguilera-Iparraguirre J, Hirzel TD, Duvenaud D, Maclaurin D, Blood-Forsythe MA, Chae HS, Einzinger M, Ha D-G, Wu T, Markopoulos G, Jeon S, Kang H, Miyazaki H, Numata M, Kim S, Huang W, Hong SI, Baldo M, Adams RP and Aspuru-Guzik A, Nat. Mater, 2016, 15, 1120– 1127. 27500805 sourceNot checked yet

  Not checked in the free run.

Gao et al. reported NN-based models trained on ~10 million reactions from Reaxys that identify appropriate catalysts, solvents, reagents, and temperatures for specified reactions.

“Gao et al.163 reported NN-based models trained on ~10 million reactions from Reaxys that identify appropriate catalysts, solvents, reagents, and temperatures for a specified reactions.”

• cites [163] Gao H, Struble TJ, Coley CW, Wang Y, Green WH and Jensen KF, ACS Cent. Sci, 2018, 4, 1465– 1476. 30555898 PMC6276053 sourceNot checked yet

  Not checked in the free run.

Marcou et al. used CGR-derived fragment descriptors to train SVM, RF, and Naive Bayes classification models to predict optimal solvents and catalysts for the Michael reaction.

“For example, Marcou et al.162 used CGR-derived fragment descriptors to train SVM, RF, and Naive Bayes classification models to predict optimal solvents and catalysts for the Michael reaction.”

• cites [162] Marcou G, Aires de Sousa J, Latino DARS, de Luca A, Horvath D, Rietsch V and Varnek A, J. Chem. Inf. Model, 2015, 55, 239– 250. 25588070 sourceNot checked yet

  Not checked in the free run.

DOGS can usefully suggest a synthetic plan not only for the target molecule but also for its close analogs.

“Thus, DOGS can usefully suggest a synthetic plan not only for the target molecule but also for its close analogs”

PHOENICS employs Bayesian Neural Networks and a kernel density estimate approximation to balance exploration vs. exploitation.

“PHOENICS,168 for example, employs Bayesian Neural Networks and a kernel density estimate approximation to balance exploration vs. exploitation.”

• cites [168] Häse F, Roch LM, Kreisbeck C and Aspuru-Guzik A, ACS Cent. Sci, 2018, 4, 1134– 1145. 30276246 PMC6161047 sourceNot checked yet

  Not checked in the free run.

Lin et al. used the heuristic that similar reactions proceed under similar conditions to predict optimal reaction conditions.

“Lin et al.139 used the heuristic that similar reactions proceed under similar conditions to predict optimal reaction conditions.”

• cites [139] Lin AI, Madzhidov TI, Klimchuk O, Nugmanov RI, Antipin IS and Varnek A, J. Chem. Inf. Model, 2016, 56, 2140– 2148. 27783508 sourceNot checked yet

  Not checked in the free run.

This thinking has led to the revival of methods for developing closed-loop or autonomous approaches.

“This thinking has led to the revival of methods for developing closed-loop or autonomous approaches.”

This problem can be solved using Condensed Graphs of Reaction (CGRs) that combine the reactant and product information.

“This problem can be solved using Condensed Graphs of Reaction (CGRs) that combine the reactant and product information.”

DNA-encoded chemical libraries have made possible synthesis and testing of millions of compounds.

“More recently, DNA-econded chemical libraries have made possible synthesis and testing of millions of compounds165 and many big pharma companies have embraced this approach.”

• cites [165] Neri D and Lerner RA, Annu. Rev. Biochem, 2018, 87, 479– 502. 29328784 PMC6080696 sourceNot checked yet

  Not checked in the free run.

Models trained on 142,111 catalytic hydrogenation reactions demonstrated high accuracy (ca. 90%) for predicting optimal experimental conditions.

“Models trained on 142,111 catalytic hydrogenation reactions demonstrated high accuracy (ca. 90%) for predicting optimal experimental conditions.”

Fragment descriptors generated for CGRs were concatenated with solvent and temperature descriptors and used to train logk models for bimolecular nucleophilic substitution, bimolecular elimination, and different types of cycloaddition.

“Fragment descriptors generated for CGRs were concatenated with solvent and temperature descriptors and used to train logk models for bimolecular nucleophilic substitution,159 bimolecular elimination, and different types of cycloaddition.160”

• cites [159] Nugmanov RI, Madzhidov TI, Khaliullina GR, Baskin II, Antipin IS and Varnek AA, J. Struct. Chem, 2014, 55, 1026– 1032.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [160] Glavatskikh M, Madzhidov T, Horvath D, Nugmanov R, Gimadiev T, Malakhova D, Marcou G and Varnek A, Mol. Inform, 2019, 38, 1800077. 30134047 sourceNot checked yet

  Not checked in the free run.

DOGS applies 58 well-established chemical transformation rules to a set of 25144 readily available synthetic blocks from the Sigma-Aldrich catalog.

“This algorithm applies 58 well-established chemical transformation rules to a set of 25144 readily available synthetic blocks from the Sigma-Aldrich catalog.”

• cites [154] Hartenfeller M, Zettl H, Walter M, Rupp M, Reisen F, Proschak E, Weggen S, Stark H and Schneider G, PLoS Comput Biol, 2012, 8, e1002380. 22359493 PMC3280956 sourceNot checked yet

  Not checked in the free run.

Currently, efforts to derive generalized toxicity models based on ENMs characteristics have been based on datasets from single studies rather than integrated from the collective body of published data.

“Currently, efforts to derive generalized toxicity models based on ENMs characteristics have been based on datasets from single studies rather than integrated from the collective body of published data.237”

• cites [237] Ehrenberg MS, Friedman AE, Finkelstein JN, Oberdörster G and McGrath JL, Biomaterials, 2009, 30, 603– 610. 19012960 sourceNot checked yet

  Not checked in the free run.

The use of biomaterials as cell factories shows great promise.

“The use of biomaterials as cell factories263 shows great promise, and the large generated stem cell attachment, proliferation, and differentiation datasets were modelled by ML methods.”

• cites [263] Celiz AD, Smith JGW, Langer R, Anderson DG, Winkler DA, Barrett DA, Davies MC, Young LE, Denning C and Alexander MR, Nat. Mater, 2014, 13, 570– 579. 24845996Not checked yet

  Not checked in the free run.

Robotic devices can be used in a home setting, minimizing patient burden and inconvenience.

“Robotic devices are less sensitive to the skills and expertise of a human rater, can reduce inter- and intra-rater variability, can be used simultaneously for both assessment and rehabilitation, which can be done faster and more frequently, and can further be used in a home setting thus minimizing patient burden and inconvenience.276”

• cites [276] Lo AC, Guarino PD, Richards LG, Haselkorn JK, Wittenberg GF, Federman DG, Ringer RJ, Wagner TH, Krebs HI, Volpe BT, Bever CT, Bravata DM, Duncan PW, Corn BH, Maffucci AD, Nadeau SE, Conroy SS, Powell JM, Huang GD and Peduzzi P, N. Engl. J. Med, 2010, 362, 1772– 1783. 20400552 PMC5592692 sourceNot checked yet

  Not checked in the free run.

Materials characteristics measured with low accuracy and reproducibility will limit the predictivity of the QNAR models trained using them.

“Clearly, materials characteristics measured with low accuracy and reproducibility, will limit the predictivity of the QNAR models trained using them.”

Materials informatics has a clear similarity to the traditional workflow of QSAR modeling.

“To conclude this section, we highlight the clear similarity between materials informatics with the traditional workflow of QSAR modeling (see Figures 1 and 6).”

Previous studies have demonstrated the value of BNs for developing qualitative "toxicity/hazard" classification of ENMs based on using physicochemical and specialized descriptors.

“Previous studies have demonstrated the value of BNs for developing qualitative “toxicity/hazard” classification of ENMs based on using physicochemical and specialized descriptors .251”

• cites [251] Marvin HJP, Bouzembrak Y, Janssen EM, van der Zande M, Murphy F, Sheehan B, Mullins M and Bouwmeester H, Nanotoxicology, 2017, 11, 123– 133. 28044458 sourceNot checked yet

  Not checked in the free run.

Unlike bioinformatics, cell biology, and clinical medicine, there is a relative paucity of published examples of the application of QSAR or related ML-based methods to biomaterials, regenerative medicine, and stem cells studies.

“Unlike bioinformatics,253 cell biology, and clinical medicine,254 there is a relative paucity of published examples of the application of QSAR or related ML-based methods to biomaterials, regenerative medicine, and stem cells studies.”

• cites [253] Cheng C and Worzel WP, in Genetic Programming Theory and Practice XII, eds. Riolo R, Worzel WP and Kotanchek M, 2014, pp. 1– 15.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [254] Molina E, Uriarte E, Santana L, Matos M and Borges F, Curr. Bioinform, 2013, 8, 438– 451.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Several literature studies have identified causal relationships between the biological outcomes and important ENMs properties.

“Several literature studies have identified causal relationships between the biological outcomes and important ENMs properties.239”

• cites [239] Oh E, Liu R, Nel A, Gemill KB, Bilal M, Cohen Y and Medintz IL, Nat. Nanotechnol, 2016, 11, 479– 86. 26925827 sourceNot checked yet

  Not checked in the free run.

A feature selection algorithm based on artificial ant colonies was used to identify the subset of robotic metrics with the highest predictive power in order to minimize over-fitting.

“To minimize over-fitting, a feature selection algorithm based on artificial ant colonies, originally developed for QSAR applications, was used to identify the subset of robotic metrics that had the highest predictive power.280”

• cites [280] and SI and Agrafiotis D,, DOI: 10.1021/CI000336S. sourceNot checked yet

  Not checked in the free run.

Read-across techniques are increasingly used to estimate the properties of nanomaterials.

“To mitigate this limitation, read-across techniques are increasingly used to estimate the properties of nanomaterials.231”

• cites [231] Gajewicz A, Nanoscale, 2017, 9, 8435– 8448. 28604902 sourceNot checked yet

  Not checked in the free run.

Robotic technology can provide repeatable, reliable, and speedy assessment of continuous measures of impairment and its change during recovery.

“One way to minimize this measurement variability issue is to replace human raters with robotic technology that can provide repeatable, reliable and speedy assessment of continuous measures of impairment and its change during recovery.”

• cites [276] Lo AC, Guarino PD, Richards LG, Haselkorn JK, Wittenberg GF, Federman DG, Ringer RJ, Wagner TH, Krebs HI, Volpe BT, Bever CT, Bravata DM, Duncan PW, Corn BH, Maffucci AD, Nadeau SE, Conroy SS, Powell JM, Huang GD and Peduzzi P, N. Engl. J. Med, 2010, 362, 1772– 1783. 20400552 PMC5592692 sourceNot checked yet

  Not checked in the free run.

One of the biggest challenges in the field of biomaterials is generation of appropriate descriptors that capture relevant properties of these materials and can adequately represent their structure.

“As discussed above in the sections on (nano)materials informatics, one of the biggest challenges in the field of biomaterials is generation of appropriate descriptors that capture relevant properties of these materials and can adequately represent their structure, often poorly understood and characterized.19”

• cites [19] Fujita T and Winkler DA, J. Chem. Inf. Model, 2016, 56, 269– 274. 26754147Not checked yet

  Not checked in the free run.

Inter- and intra-sample variability in QNAR is a big issue that can dramatically affect the predictivity of a model.

“As is true for traditional QSAR, inter- and intra-sample variability in QNAR is a big issue that can dramatically affect the predictivity of a model.”

Materials informatics represents a prime example of a new discipline whose development was enabled and catalyzed by QSAR experience and approaches.

“Thus, materials informatics (and a closely related field of nanomaterials informatics described in the next section) represents a prime example of a new discipline, whose development was enabled and immensely catalyzed by the experience and approaches developed in QSAR.”

Models that predict toxicity of ENMs must account not only for the properties of the nanomaterials per se, but also for experimental conditions.

“Models that predict toxicity of ENMs must account not only for the properties of the nanomaterials per se, but also for experimental conditions (e.g., assay types, exposure concentrations, exposure period, organism and more).”

Evolutionary algorithms can also be used to generate closed loop, autonomous molecule and materials discovery systems.

“Evolutionary algorithms can also be used to generate closed loop, autonomous molecule and materials discovery system.”

• cites [174] Le TC and Winkler DA, Chem. Rev, 2016, 116, 6107– 6132. 27171499Not checked yet

  Not checked in the free run.

The relatively small size of the datasets currently available in the public domain leads to lower prediction accuracy and smaller applicability domains for QNAR models compared to those of QSAR models trained on large organic molecule data sets.

“Another challenge of QNAR modeling, similar to materials informatics is the relatively small size of the datasets currently available in the public domain. This leads to lower prediction accuracy and smaller applicability domains for QNAR models compared to those of QSAR models trained on large organic molecule data sets.”

The high structural diversity and complexity of nanomaterials typically lead to specific challenges, especially when it comes to the choice of molecular descriptors.

“However, the high structural diversity and complexity of nanomaterials typically lead to specific challenges,230 especially when it comes to the choice of molecular descriptors.”

• cites [230] Fourches D, Barnes J, Day NC, Bradley P, Reed JZ and Tropsha A, Chem. Res. Toxicol, 2010, 23, 171– 83. 20014752 PMC2850112 sourceNot checked yet

  Not checked in the free run.

COPD cluster 4, comprising young patients with few comorbidities, has the highest number of prescriptions for bronchodilators.

“Cluster 4 (young patients, few comorbidities) has the highest number of prescriptions for bronchodilators;”

• cites [283] Vazquez Guillamet R, Ursu O, Iwamoto G, Moseley PL and Oprea T, Health Informatics J., 2018, 24, 394– 409. 27856785 sourceNot checked yet

  Not checked in the free run.

For the COPD dataset, the optimal balance between the number of clusters and clustering overlap was found at a similarity threshold of 0.62.

“For this dataset, the optimal balance between the number of clusters and clustering overlap was found at similarity threshold 0.62.”

• cites [283] Vazquez Guillamet R, Ursu O, Iwamoto G, Moseley PL and Oprea T, Health Informatics J., 2018, 24, 394– 409. 27856785 sourceNot checked yet

  Not checked in the free run.

COPD cluster 6 is associated with predominantly cerebrovascular disease and younger patients under age 65.

“whereas cluster 6 is associated with predominantly cerebrovascular disease and younger (under 65) patients.”

• cites [283] Vazquez Guillamet R, Ursu O, Iwamoto G, Moseley PL and Oprea T, Health Informatics J., 2018, 24, 394– 409. 27856785 sourceNot checked yet

  Not checked in the free run.

The models had better prediction accuracy for the FM and MP scales, which are more closely related to motor function, than for the NIH and MR metrics.

“The models had better prediction accuracy for the FM and MP scales that are more closely related to motor function than the NIH and MR metrics.”

• cites [277] Hogan N and Krebs HI, in Progress in brain research, 2011, vol. 192, pp. 59– 68. 21763518 sourceNot checked yet

  Not checked in the free run.

• cites [278] Flash T and Hogan N, J. Neurosci, 1985, 5, 1688– 703. 4020415 PMC6565116 sourceNot checked yet

  Not checked in the free run.

COPD cluster 8 is associated with chronic kidney disease and type 2 diabetes in patients aged 40-65.

“cluster 8 is associated with chronic kidney disease (CKD) and type 2 diabetes in patients aged 40-65,”

• cites [283] Vazquez Guillamet R, Ursu O, Iwamoto G, Moseley PL and Oprea T, Health Informatics J., 2018, 24, 394– 409. 27856785 sourceNot checked yet

  Not checked in the free run.

Optimized robotic composites with as few as four features increased the effect size over a reference natural history trial by as much as 107% for the training set and 83% for the test set.

“Optimized robotic composites with as few as four features increased the effect size over a reference natural history trial282 by as much as 107% for the training and 83% for the test set.”

• cites [282] Kerr DM, Fulton RL, Lees KR and VISTA Collaborators, Stroke, 2012, 43, 1401– 1403. 22308254 sourceNot checked yet

  Not checked in the free run.

The above studies successfully adapted methods from computational chemistry and cheminformatics into in-depth analyses of health data.

“To summarize, the above studies successfully adapted methods from computational chemistry and cheminformatics into in-depth analyses of health data.”

The resulting models recapitulated the human-scored clinical scales with cross-validated R2 values of 0.73, 0.75, 0.63, and 0.60 for the FM, MP, NIH, and MR scales, respectively.

“The resulting models recapitulated the human scored clinical scales with a cross-validated R2 of 0.73, 0.75, 0.63, and 0.60 for the FM, MP, NIH and MR scales, respectively.”

• cites [277] Hogan N and Krebs HI, in Progress in brain research, 2011, vol. 192, pp. 59– 68. 21763518 sourceNot checked yet

  Not checked in the free run.

• cites [278] Flash T and Hogan N, J. Neurosci, 1985, 5, 1688– 703. 4020415 PMC6565116 sourceNot checked yet

  Not checked in the free run.

An increase of 83% in effect size would result in a 70% reduction in the number of patients required to achieve the typical 80% statistical power in a clinical trial.

“This result is highly significant as an increase of 83% in effect size would result in a 70% reduction in the number of patients required to achieve the typical 80% statistical power in a clinical trial.”

QSAR-like approaches are beginning to yield exciting results in research areas as diverse as quantum mechanics, materials and nanomaterials science, biomaterials, regenerative medicine, and health care.

“We have illustrated some of non-traditional applications in this review, demonstrating how QSAR-like approaches are beginning to yield exciting results in research areas as diverse as quantum mechanics, materials and nanomaterials science, biomaterials, regenerative medicine, and health care.”

The models showed lower but still useful predictive power for the external validation set of non-completers.

“The models also showed lower but still useful predictive power for the external validation set (non-completers).”

• cites [277] Hogan N and Krebs HI, in Progress in brain research, 2011, vol. 192, pp. 59– 68. 21763518 sourceNot checked yet

  Not checked in the free run.

• cites [278] Flash T and Hogan N, J. Neurosci, 1985, 5, 1688– 703. 4020415 PMC6565116 sourceNot checked yet

  Not checked in the free run.

COPD cluster 3 is associated with heart disease in patients over age 65.

“Cluster 3 is associated with heart disease (patients over age 65),”

• cites [283] Vazquez Guillamet R, Ursu O, Iwamoto G, Moseley PL and Oprea T, Health Informatics J., 2018, 24, 394– 409. 27856785 sourceNot checked yet

  Not checked in the free run.

The largest COPD cluster is characterized by a large proportion of patients over age 65 and depression.

“The largest cluster is characterized by a large proportion of patients over age 65 and depression;”

• cites [283] Vazquez Guillamet R, Ursu O, Iwamoto G, Moseley PL and Oprea T, Health Informatics J., 2018, 24, 394– 409. 27856785 sourceNot checked yet

  Not checked in the free run.

Many roadblocks and technical issues in statistical data modelling employed in different domains had already been addressed in the QSAR modeling literature, including papers on the impact of errors on QSAR analysis and the importance of data curation for stable and reproducible models.

“Examples include papers on the impact of the errors on QSAR analysis288 and the importance of data curation to achieve stable and reproducible models.93”

• cites [288] Young D, Martin D, Venkatapathy R, Harten P, Martin T, Venkatapathy R and Harten P, QSAR Comb. Sci, 2008, 27, 1337– 1345.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [93] Fourches D, Muratov E and Tropsha A, J. Chem. Inf. Model, 2010, 50, 1189– 204. 20572635 PMC2989419 sourceNot checked yet

  Not checked in the free run.

Using the sphere exclusion algorithm for clustering reduces the risk of bias because the method does not make a priori assumptions regarding numbers of clusters or similarity thresholds.

“Using the sphere exclusion algorithm for clustering reduces the risk of bias since the method does not make a priori assumptions regarding numbers of clusters or similarity thresholds.”

• cites [286] Taylor R, J. Chem. Inf. Model, 1995, 35, 59– 67.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The models were used to derive novel composite robotic endpoints with improved sensitivity and effect size compared to existing scales.

“Finally, the models were used to derive novel composite robotic endpoints with improved sensitivity (and effect size) compared to existing scales.”

• cites [277] Hogan N and Krebs HI, in Progress in brain research, 2011, vol. 192, pp. 59– 68. 21763518 sourceNot checked yet

  Not checked in the free run.

• cites [278] Flash T and Hogan N, J. Neurosci, 1985, 5, 1688– 703. 4020415 PMC6565116 sourceNot checked yet

  Not checked in the free run.