Modeling Kinase Inhibition Using Highly Confident Data Sets - Journal

Apr 30, 2018 - Here we analyzed kinase–inhibitor pairs with multiple activity points extracted from the ChEMBL database and identified the main sour...
0 downloads 0 Views 8MB Size
Article Cite This: J. Chem. Inf. Model. 2018, 58, 957−967

pubs.acs.org/jcim

Modeling Kinase Inhibition Using Highly Confident Data Sets Sorin Avram,* Alina Bora, Liliana Halip, and Ramona Curpăn* Department of Computational Chemistry, Institute of Chemistry Timişoara of Romanian Academy, 24 Mihai Viteazu Avenue, 300223-Timişoara, Romania

J. Chem. Inf. Model. 2018.58:957-967. Downloaded from pubs.acs.org by LINKOPING UNIV on 08/25/18. For personal use only.

S Supporting Information *

ABSTRACT: Protein kinases form a consistent class of promising drug targets, and several efforts have been made to predict the activities of small molecules against a representative part of the kinome. This study continues our previous work (Bora, A.; Avram, S.; Ciucanu, I.; Raica, M.; Avram, S. Predictive Models for Fast and Effective Profiling of Kinase Inhibitors. J. Chem. Inf. Model. 2016, 56, 895−905; www.chembioinf.ro) aiming to build and measure the performance of ligand-based kinase inhibitor prediction models. Here we analyzed kinase−inhibitor pairs with multiple activity points extracted from the ChEMBL database and identified the main sources of inconsistency. Our results indicate that lower IC50 values are usually less affected by errors and reflect more accurately the structure−activity relationship of the molecules against the target, ideally for quantitative structure− activity relationship studies. Further, we modeled the activities of 104 kinases using unbiased target-specific activity points. The performance of predictors built on extended connectivity fingerprints (ECFP4) and two-dimensional pharmacophore fingerprints (PFPs) are compared by means of tolerance intervals (TIs) (95%/95%) in virtual screening (VS) and classification tasks using external random (RandSets) and diversity-based (DivSets) test sets. We found that the two encodings perform superior to each other on different kinases in VS and that PFP models perform consistently better in classifying actives (higher sensitivity). Next, we combined the two encodings into a single one (PFPECFP) and demonstrated that especially in VS (as indicated by the exponential receiver operating curve enrichment metric (eROCE)), for the vast majority of kinases the model performance increased compared with the individual fingerprint models. These findings are highlighted in the more challenging DivSets compared with RandSets. The current paper explores the boundaries of inhibitor predictors for individual kinases to enhance VS and ultimately aid the discovery of novel compounds with desirable polypharmacology.



The human kinome comprises more than 500 kinases,15,16 and many are involved in a plethora of biological processes determinant for several pathological conditions, including cancers.17 For drug development, the design of selective compounds across several classes of kinases is a necessary and challenging research field.18 Several computational models have been developed to predict kinase ligands (in many cases inhibitors)14,19−21 at a nearly kinomewide level.14 Activity data (Ki, IC50, Kd, etc.) extracted from the ChEMBL database (ChEMBLdb),22,23 PubChem Bioassay,24 or Kinase Knowledgebase (KKB)25 have been used to train and test supervised machine learning models.10,14 Among these, random forest has been shown to generally outperform alternative methods

INTRODUCTION Learning algorithms have a wide spectrum of applications in various fields, including drug discovery.1−6 Numerous times machine learning has been successfully used to identify novel compounds that are active against a desired protein target.7 Powerful prediction models are able to promote valuable compounds in virtual screening (VS) of large chemical libraries8 and also to discriminate between active and inactive compounds in classification (CLS) tasks.9 In order to assess their usefulness, the prediction capabilities need to be estimated in real-life applications, e.g., in classification (i.e., the separation of actives from inactives by accurate class labeling) or in VS (i.e., the prioritization of actives by ranking a large chemical library according to a scoring function). In drug discovery, a series of prediction models (predictors) have been developed to tackle important targets such as kinases.10−14 © 2018 American Chemical Society

Received: December 20, 2017 Published: April 30, 2018 957

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling

inactives (IC50 > 10 μM). This was followed by the adjustment of the IC50 threshold of the actives (for each kinase set) on the basis of the distribution of the inhibitory values. The adjusted IC50 thresholds (see the Excel file in the Supporting Information) were computed using the package “hotspot”38,39 in the R statistical environment40 as previously done by Bora et al.14 Inconsistent Activity Analysis. In the initial set of APs, 1122 compound−kinase pairs for which multiple IC50 determinations are available, with values spread over 2 orders of magnitude, were considered inconsistent and retained for manual assay inspection. The extreme APs were analyzed on the basis of the information found in the assay “Description” field provided by ChEMBLdb. The activity information and compound structures were tracked down to the source articles. APs were tagged according to several characteristics (many related to the assay) that could affect the IC50 values: cellular (Cellular), tissue assays (Tissue), mutant kinase targets (Mutant), patent data (Patent), specification of ATP concentrations >100 μM (ATP), specification of reaction times >1 h (Time), curation errors that can be corrected (Error), and ambiguous activity information (Ambiguous). A last category, i.e., BioChemical, defined APs that were not included in the previous groups. Balanced Kinase Sets of Compounds. We obtained 33 037 compounds (with unique ChEMBL22,23 molecular identifiers) after the removal of inconsistent APs. Additionally, we used the PubChemKinIna data set,14 which comprises 38 957 compounds considered to be generally inactive against kinases on the basis of high-throughput screening results. All of the compounds were standardized using ChemAxon’s Standardizer41 (salt removal, largest disconnected fragment kept, functional group transformations and basic aromatization scheme used).14 From the 107 kinase sets, three sets contained ”, or “=”. Throughout this paper, we will refer to this data set as the initial set of APs. Further, APs were grouped according to unique compound−target pairs on the basis of molecule’s ChEMBL ID and the target’s UniProt ID. Pairs with multiple APs that were considered inconsistent were submitted to analysis (see the next section). Only corrected and biochemical assay outcomes were further used, and the geometric mean was computed to obtain a single IC50 value for each compound− kinase pair. The compounds in the kinase sets were split into two groups according to the inhibition activity: actives (IC50 ≤ 10 μM) and 958

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling number of actives were randomly selected from the smallest training cluster, i.e., the largest test set cluster. Thereby, the smallest (and more numerous) clusters, covering more diverse actives, were included in the test set, as exemplified for Rhoassociated protein kinase 1 (ROCK1) in Figures S1 and S2 in the Supporting Information. The 20% test sets, containing equally sized labels, were evaluated for CLS. In order to create real-life VS conditions,14 each test set was merged with the PubChemKinIna compounds tailored to avoid inactives used for training the corresponding model. Virtual screening conditions were assured by the large number of inactives encompassed in PubChemKinIna (∼39 000).14 Depending on the kinase set, the predictors were challenged to prioritize 10 to 200 actives for the vast majority of the kinases. Thereby, each model generated in the current study was evaluated separately on independent CLS and VS sets. Molecular Descriptors. Using ChemAxon’s Java API,41 we computed the following molecular encoding: ChemAxon’s molecular fingerprints (similarity search and clustering), ECFP4 (1024 bit and 512 bit length), and PFPs (standardized to the 210 bit format described by Bora et al.14). Additionally, PFP and ECFP4 encodings were joined into PFPECFP encodings of length 722 bits simply by appending to the 210 bit PFP vector the 512 bit version of ECFP. Learning Algorithm. Random forest models were computed on the basis of conditional inference trees using the function “cforest” with default parameters (ntrees = 500, mtry = square root of the total number of variables) from the package “party”45 available in the R statistical environment.40 The response classes were set as the majority class of each node, and class probabilities were computed from the conditional distribution of the response.45 Performance Intervals. Tolerance intervals (TIs) describe the spread of values within confidence limits computed for a certain percentile of the distribution. Here tolerance intervals were computed for each evaluation parameter on the basis of 50 samples (resulting from the resampling of the training and test sets) using the function “nptol.int” (Wilks nonparametric approach)46 in the package “tolerance” available in the R statistical environment.40 The one-sided tolerance interval for 95% of the data was computed at the 95% confidence level (95%/95% TI). Evaluation Parameters. A set of evaluation parameters (see Table 1) were computed using our in-house program ETCIv1.614 to describe the CLS (and discrimination) power of the models as well as their VS performance (the early enrichment in actives). Model performance in the CLS test sets (equal number of class samples) was assessed through sensitivity (Se, the fraction of actives correctly predicted), specificity (Sp, the fraction of inactives correctly predicted), accuracy (Acc, the fraction of correct predictions), and the area under the receiver operating characteristic (ROC) curve (AUC),47 and that in the VS test sets was assessed through the exponential ROC enrichment (eROCE)27,33,48 and the true positive rate (TPR) at 0.5% and 1% false positives (FPs).14,33,49 The eROCE parameter computes identical values (with the same meaning) as the Boltzmann enhanced discrimination of ROC (BEDROC)50 parameter in VS conditions but offers important advantages, i.e., it is easy to compute (it is a simple function of the FP rate) and exhibits increased robustness also in nontypical VS scenarios and wider application in comparing results from different studies and data sets.48

Table 1. Description of the Evaluation Parameters Used To Assess the Classification and Virtual Screening Capacities of the Predictors evaluation type classification

virtual screening

equationa

parameter sensitivity specificity accuracy

Se = TP/(TP + FN) Sp = TN/(TN + FP) Acc = (TP + TN)/(TP + FN + TN + FP)

area under the ROC curve

AUC = 1 −

exponential ROC curve enrichment (α = 20) true positive rate at x% false positives

eROCE =

1 TP + FN

1 TP + FN

TP + FN



FPR i

i=1

TP + FN



e−FPR iα

i=1

TPRx = TPR at x% FP, where x = [0.5%, 1%]

a

TP is the number of correctly predicted actives (true positives), TN is the number of correctly predicted inactives (true negatives), FP is the number of incorrectly identified actives (false positives), FN is the number of incorrectly identified inactives (false negatives), TPR is the true positive rate (i.e., the fraction of correctly predicted actives), and FPRi is the ratio of the number of mispredicted inactives to the total number of inactives when the ith active in the ranking list is retrieved.

describe the output of an extensive model evaluation study based on homogeneous activity data. The results of the current study comprise 77 633 investigated inhibitory APs implying 33 037 compounds and covering 107 kinases. We built and evaluated the VS and CLS performance of 31 200 random forest models (three types of molecular encodings) covering 104 kinase sets, which provided diversity-based test sets (DivSets) and random-based test sets (RandSets). The results are discussed in terms of tolerance intervals (i.e., ranges in which at least 95% of the evaluation values fall with a 95% level of confidence) computed for seven evaluation metrics describing the potential of the approaches. Inconsistencies in Activity Points. We found 1122 compound−kinase pairs (covering 3381 APs by 1407 assays in 778 documents) showing inconsistencies (multiple activity determinations for which the values spread over 2 orders of magnitude). We considered several possible causes that defined categories of APs (see Inconsistent Activity Analysis in Methods) and manually inspected the lowest and highest IC50 values as well as the corresponding assay properties in the published articles providing the APs. In Table 2 we report within each category the percentage of APs found as the highest and lowest IC50 values in the inconsistency sets. The highest IC50 values were encountered in activity determinations performed on tissues (whole blood in general, which are not flagged as tissue-based APs in the corresponding field in ChEMBL; Tissue 93%) and cells (also many times not correspondingly assigned; Cellular 77%) and APs with specified reaction times >1 h (Time 73%) or high ATP concentrations >100 μM (ATP 80%). Errors (Error) originating from activity value conversions affected the higher-IC50 APs in 90% of the cases. The lowest IC50 values are associated in ∼85% of the cases (out of 781) with target-specific biochemical assays (BioChemical). Activity determinations extracted from patents could not be verified because of limited access to the assay protocols and were separately categorized. In 73% of the cases studied herein, Patent APs were encountered within the lower IC50 values. These results suggest that the lower IC50 values are usually less affected by errors, avoid factors related to the cells or tissues, and reflect more accurately the structure−activity relationship (SAR) of the molecules against the target.



RESULTS AND DISCUSSION We begin by reporting our results of the inconsistent APs related to kinase inhibition, and then in the following sections we 959

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling

models to outperform each other equally on 34 kinases, but a superior TPR0.5 performance of ECFP-based models was encountered in 37 cases (compared with PFP-based models in 15 kinases). We found superior PFP-based models in classifying kinase inhibitors for 78 and 43 of the kinases in terms of Se and Acc, respectively. The results obtained on RandSets reveal more homogeneous results, with only a few superior TIs in the case of VS parameters as well as Sp for ECFP models and Se for PFP models. In terms of predicting inactives (Sp), there is no kinase for which PFP predictors outperformed ECFP predictors (Figure 1 and Table S1 in the Supporting Information). Considering the above findings, PFP and ECFP encodings seem to perform in a complementary manner: in VS evaluation, the two encodings show superior performance on different kinases, and in CLS, PFP tends to better classify actives while ECFP does better with inactives. These features were further speculated by joining the two fingerprints to obtain PFPECFP encodings (of length 722 bits). Performance of the PFPECFP Models. We found that PFPECFP models enhanced especially the VS on DivSets, as suggested by eROCE: the performance of PFPECFPs was superior to those of both individual encodings for 28 kinases and better than at least one of them for 62 kinases (Figure 1). In CLS, PFPECFP performed better than either PFP or ECFP for 71 kinases in classifying actives. Some improvement can be seen in the classification of inactives, but nonoverlapping TIs could not be detected for PFPECFP. In contrast to DivSets, the evaluation of RandSets revealed only a few cases of superior TIs among the three encodings used in modeling kinase inhibitors (see Table S1). For example, PFPECFPs outperformed the individual fingerprints for seven, nine, and seven kinases in terms of Sp, eROCE, and TPR0.5. DivSets versus RandSets. Commonly, in drug discovery a data set containing active and inactive compounds against a target is randomly split into training and test sets preserving proportional class samples.14,27,33 Here we compared this approach (i.e., RandSets) to DivSets, for which only inactives are randomly sampled while actives are kept constant, so that models learning from low-diversity sets of actives are challenged to predict highly diverse kinase inhibitors. In comparison with DivSets, RandSets evaluations resulted in consistently higher scores. Subtracting separately the lower and upper TI values of DivSets from the RandSets results, for all kinases, enables a direct comparison between the two types of splitting (Tables S2 and S3). In terms of the lower TI limits, in PFPECFP modeling the smallest difference is encountered for Sp (”, or “=”. Approximately 90% of these resulted from a single kinase panel study (i.e., Metz et al.52). According to the comparative review of four large kinase profiling panels by Sutherland et al.53 (which includes the Metz data set), the agreement for active compounds is only 37%. While we do not neglect the value of kinase panel assays, this low percentage, in addition to approximations regarding the conversion between the two activity types (Cheng−Prusoff equations),54 has persuaded us, at least for the purpose of the current study, to leave out Ki determinations. In the future, the revised IC50 APs assembled herein might help to identify and explain possible inconsistencies in the results of kinase profiling assays. Comparison among Molecular Encodings. The performances of the predictors built on different fingerprints were compared in terms of the TIs (95%/95%) of the evaluation parameters. Herein we considered that for a given kinase, model performance using a given fingerprint is superior to that with another fingerprint if its TI (computed for a particular metric) is higher and the two TIs do not overlap. The evaluation results are shown in Figures 1−8. Performance of PFP- and ECFP-Based Models. In VS testing on DivSets, eROCE showed PFP- and ECFP-based 960

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling

Figure 1. Stacked bar plots describing, for each of the 104 kinases, counts of superior TIs among three fingerprint models, i.e., PFP (green), ECFP (red), and PFPECFP (blue), in DivSets. For example, in the case of ABL1, in terms of eROCE, PFPECFP models are superior to both PFP and ECFP (count of 2), while PFP models perform superior to ECFP models (count 1); in the case of AKT1, eROCE TIs indicate that ECFP and PFPECFP models are superior to PFP models.

in ECFP models. Regardless of the fingerprint, the capacity to classify inactive compounds is highly correlated between DivSets and RandSets (Pearson coefficient values around 0.9), indicating that the prediction of inactive compounds might be only loosely dependent on the diversity of the actives in the data set. Our results suggest that in spite of the significant correlation in the lower TI values (between 0.669 for PFPECFP’s Se and 0.939 for ECFP’s Sp; see Table S4), DivSet evaluations are in general more severe compared with RandSets evaluations. This provides a way to fully exploit the data set by testing the boundaries of the modeling tools, which used on average 39 ± 9% of the total number of clusters for training. In DivSets, actives assigned for testing have Tanimoto values of 50% of the actives were retrieved before the first 0.5% of inactives in the ranking list (TPR0.5 > 0.5). Such high early enrichment was achieved only for 43 kinases in PFP modeling and 55 kinases in ECFP modeling. In terms of eROCE, scores >0.5 were found for 92, 87, and 68 kinases in PFPECFP-, PFP-, and ECFP-based modeling, respectively. In CLS tasks of PFP and PFPECFP models, we found 64 kinases for which both Se and Sp values are >0.5. In the case of ECFP, the same classification rates were reached for only 24 kinases. The discriminative power, also computed on the CLS 961

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling

Figure 2. Tolerance intervals and medians of exponential ROC enrichment (eROCE) values computed for PFP-, ECFP4-, and PFPECFP-based models.

Figure 3. Tolerance intervals (TIs) and medians of true positive rate at 0.5% FPs (TPR0.5) values computed for PFP-, ECFP4-, and PFPECFP-based models.

test sets, resulted in AUC ≥ 0.7 for 84, 80, and 73 kinases in PFPECFP-, PFP-, and ECFP-based models, respectively. The

AUC scores are generally high, suggesting that the probability scores computed by the models provide a good class separation. 962

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling

Figure 4. Tolerance intervals (TIs) and medians of true positive rate at 1% FPs (TPR1) values computed for PFP-, ECFP4-, and PFPECFP-based models.

Figure 5. Tolerance intervals (TIs) and medians of sensitivity (Se) values computed for PFP-, ECFP4-, and PFPECFP-based models.

Analysis of the upper TI limits can highlight models with poor performance. Regardless of the molecular encoding employed, in the case of RPS6KA3, CAMK2D, CDK5, CDK7, CLK1,

MAPK9, MAPK1, AXL, PTK2B, FYN, ZAP70, CSNK1D, and PLK1, the early enrichment in the top ∼0.5% of a chemical library could not exceed 50%. In the case of eROCE, upper TI 963

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling

Figure 6. Tolerance intervals (TIs) and medians of specificity (Sp) values computed for PFP-, ECFP4-, and PFPECFP-based models.

Figure 7. Tolerance intervals (TIs) and medians of accuracy (Acc) values computed for PFP-, ECFP4-, and PFPECFP-based models.

MAP3K8, and PLK1. One should bear in mind that models built for these kinases are still capable of successfully identifying kinase inhibitors, as suggested by the RandSets results (Figures 2−8).

limits below 0.5 were found only for CDK5 and PTK2B, independent of the fingerprint type. Both Se < 0.5 and Sp < 0.5 were found in the case of PRKCA, PRKCZ, RPS6KA3, CAMK2D, CDK5, CLK1, MAPK1, AXL, PTK2B, FYN, ZAP70, 964

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling

Figure 8. Tolerance intervals (TIs) and medians of area under the ROC curve (AUC) values computed for PFP-, ECFP4-, and PFPECFP-based models.

adding the features of the two fingerprints and widening the spectra of highly predictive kinase targets. Instead of reporting the average performance of the models, we have focused on estimating the lowest performance one would expect in predicting diverse kinase inhibitors (for each of the 104 kinases) with Tanimoto similarity of 0.7 were encountered for 98 kinases. Moreover, also in PFPECFP models, 90 kinases indicated concomitantly median Se and Sp > 0.8 and 96 kinases had AUC > 0.9. These results indicate a very good VS and CLS capacities of the PFPECFP models using target-specific clean kinase inhibitory APs.



CONCLUSIONS Bounded by the kinase sets studied herein, we have shown that differences in cellular (and tissular) and biochemical assays contribute to the majority of the extreme inhibitory values. In quantitative SAR (QSAR) modeling, ideally the independent variable should reflect as accurately as possible the activity of a series of compounds against the target of interest (without being affected by other assay properties), and therefore, we recommend the use of biochemical assay data for QSAR modeling but also for optimization and comparisons of docking-based methods. Activity cliff identification might also be affected on the same grounds. Furthermore, the results obtained herein suggest that taking the minimum activity value (if multiple values are available) can be in general a simple way to avoid biased compound−target interaction values. Of course, for drug discovery, successful prediction of cell-based kinase inhibitory determinations can be an even more valuable asset in hit or even lead identification and should be explored in future studies. Here we used 95%/95% tolerance intervals to estimate the potential of two-dimensional molecular fingerprint based-models. We found that each molecular encoding works better for some kinases compared with the others. Moreover, PFP-based models perform better on learning to predict (diverse) actives compared with ECFP-based models. We have shown that combining the two fingerprints into PFPECFP encodings enhances the models,



ASSOCIATED CONTENT

S Supporting Information *

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.jcim.7b00729. Figures and tables supporting the description of the results (PDF) Brief description of the balanced kinase sets (XLSX)



AUTHOR INFORMATION

Corresponding Authors

*E-mail: [email protected]; 5orin.4vram@ gmail.com. *E-mail: [email protected]; [email protected]. 965

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling ORCID

(13) Crisan, L.; Pacureanu, L.; Bora, A.; Avram, S.; Kurunczi, L.; Simon, Z. QSAR Study and Molecular Docking on Indirubin Inhibitors of Glycogen Synthase Kinase-3. Cent. Eur. J. Chem. 2013, 11, 63−77. (14) Bora, A.; Avram, S.; Ciucanu, I.; Raica, M.; Avram, S. Predictive Models for Fast and Effective Profiling of Kinase Inhibitors. J. Chem. Inf. Model. 2016, 56, 895−905. (15) Manning, G.; Whyte, D. B.; Martinez, R.; Hunter, T.; Sudarsanam, S. The Protein Kinase Complement of the Human Genome. Science 2002, 298, 1912−1934. (16) Miranda-Saavedra, D.; Barton, G. J. Classification and Functional Annotation of Eukaryotic Protein Kinases. Proteins: Struct., Funct., Genet. 2007, 68, 893−914. (17) Knight, Z. A.; Lin, H.; Shokat, K. M. Targeting the Cancer Kinome through Polypharmacology. Nat. Rev. Cancer 2010, 10, 130− 137. (18) Ekambaram, R.; Enkvist, E.; Vaasa, A.; Kasari, M.; Raidaru, G.; Knapp, S.; Uri, A. Selective Bisubstrate Inhibitors with Sub-Nanomolar Affinity for Protein Kinase Pim-1. ChemMedChem 2013, 8, 909−913. (19) Cao, D. S.; Zhou, G. H.; Liu, S.; Zhang, L. X.; Xu, Q. S.; He, M.; Liang, Y. Z. Large-Scale Prediction of Human Kinase-Inhibitor Interactions Using Protein Sequences and Molecular Topological Structures. Anal. Chim. Acta 2013, 792, 10−18. (20) Niijima, S.; Shiraishi, A.; Okuno, Y. Dissecting Kinase Profiling Data to Predict Activity and Understand Cross-Reactivity of Kinase Inhibitors. J. Chem. Inf. Model. 2012, 52, 901−912. (21) Schurer, S. C.; Muskal, S. M. Kinome-Wide Activity Modeling from Diverse Public High-Quality Data Sets. J. Chem. Inf. Model. 2013, 53, 27−38. (22) Bento, A. P.; Gaulton, A.; Hersey, A.; Bellis, L. J.; Chambers, J.; Davies, M.; Kruger, F. A.; Light, Y.; Mak, L.; McGlinchey, S.; Nowotka, M.; Papadatos, G.; Santos, R.; Overington, J. P. The ChEMBL Bioactivity Database: An Update. Nucleic Acids Res. 2014, 42, D1083− 1090. (23) Papadatos, G.; Gaulton, A.; Hersey, A.; Overington, J. P. Activity, Assay and Target Data Curation and Quality in the ChEMBL Database. J. Comput.-Aided Mol. Des. 2015, 29, 885−896. (24) Wang, Y.; Suzek, T.; Zhang, J.; Wang, J.; He, S.; Cheng, T.; Shoemaker, B. A.; Gindulyte, A.; Bryant, S. H. Pubchem Bioassay: 2014 Update. Nucleic Acids Res. 2014, 42, D1075−1082. (25) Sharma, R.; Schurer, S. C.; Muskal, S. M. High Quality, Small Molecule-Activity Datasets for Kinase Research. F1000Research 2016, 5, 1366. (26) Rogers, D.; Hahn, M. Extended-Connectivity Fingerprints. J. Chem. Inf. Model. 2010, 50, 742−754. (27) Avram, S.; Pacureanu, L. M.; Seclaman, E.; Bora, A.; Kurunczi, L. PLS-DA - Docking Optimized Combined Energetic Terms (PLSDADOCET) Protocol: A Brief Evaluation. J. Chem. Inf. Model. 2011, 51, 3169−3179. (28) Hert, J.; Willett, P.; Wilton, D. J.; Acklin, P.; Azzaoui, K.; Jacoby, E.; Schuffenhauer, A. Comparison of Topological Descriptors for Similarity-Based Virtual Screening Using Multiple Bioactive Reference Structures. Org. Biomol. Chem. 2004, 2, 3256−3266. (29) McGregor, M. J.; Muskal, S. M. Pharmacophore Fingerprinting. 1. Application to QSAR and Focused Library Design. J. Chem. Inf. Comput. Sci. 1999, 39, 569−574. (30) Reutlinger, M.; Koch, C. P.; Reker, D.; Todoroff, N.; Schneider, P.; Rodrigues, T.; Schneider, G. Chemically Advanced Template Search (CATS) for Scaffold-Hopping and Prospective Target Prediction for ’Orphan’ Molecules. Mol. Inf. 2013, 32, 133−138. (31) Chemaxon Documentation. https://docs.chemaxon.com (accessed August 2017). (32) Lang, T.; Flachsenberg, F.; von Luxburg, U.; Rarey, M. Feasibility of Active Machine Learning for Multiclass Compound Classification. J. Chem. Inf. Model. 2016, 56, 12−20. (33) Avram, S. I.; Pacureanu, L. M.; Bora, A.; Crisan, L.; Avram, S.; Kurunczi, L. ColBioS-FlavRC: A Collection of Bioselective Flavonoids and Related Compounds Filtered from High-Throughput Screening Outcomes. J. Chem. Inf. Model. 2014, 54, 2360−2370.

Sorin Avram: 0000-0002-4816-7446 Notes

The authors declare no competing financial interest. The balanced kinase sets used in this work and the PFPECFP models (implemented in Kinase Inhibition Predictor, KIP) are freely available at www.chembioinf.ro.



ACKNOWLEDGMENTS

This work was supported by a grant from the Romanian National Authority for Scientific Research and Innovation, CNCS− UEFISCDI (Project PN-II-RU-TE-2014-4-0422) and the Romanian Academy, Institute of Chemistry Timişoara (Project 1.2.4/2017). All of the authors are indebted to ChemAxon Ltd. for providing access to their software.



ABBREVIATIONS RandSets, random sets; DivSets, diversity sets; CLS, classification; VS, virtual screening; AP, activity point; TI, tolerance interval



REFERENCES

(1) Lavecchia, A. Machine-Learning Approaches in Drug Discovery: Methods and Applications. Drug Discovery Today 2015, 20, 318−331. (2) Wang, N. N.; Dong, J.; Deng, Y. H.; Zhu, M. F.; Wen, M.; Yao, Z. J.; Lu, A. P.; Wang, J. B.; Cao, D. S. ADME Properties Evaluation in Drug Discovery: Prediction of Caco-2 Cell Permeability Using a Combination of NSGA-II and Boosting. J. Chem. Inf. Model. 2016, 56, 763−773. (3) Wang, N.-N.; Deng, Z.-K.; Huang, C.; Dong, J.; Zhu, M.-F.; Yao, Z.J.; Chen, A. F.; Lu, A.-P.; Mi, Q.; Cao, D.-S. ADME Properties Evaluation in Drug Discovery: Prediction of Plasma Protein Binding Using NSGA-II Combining PLS and Consensus Modeling. Chemom. Intell. Lab. Syst. 2017, 170, 84−95. (4) Wang, N.-N.; Huang, C.; Dong, J.; Yao, Z.-J.; Zhu, M.-F.; Deng, Z.K.; Lv, B.; Lu, A.-P.; Chen, A. F.; Cao, D.-S. Predicting Human Intestinal Absorption with Modified Random Forest Approach: A Comprehensive Evaluation of Molecular Representation, Unbalanced Data, and Applicability Domain Issues. RSC Adv. 2017, 7, 19007−19018. (5) Deng, Y. H.; Wang, N. N.; Zou, Z. X.; Zhang, L.; Xu, K. P.; Chen, A. F.; Cao, D. S.; Tan, G. S. Multi-Target Screening and Experimental Validation of Natural Products from Selaginella Plants against Alzheimer’s Disease. Front. Pharmacol. 2017, 8, 539. (6) Merget, B.; Turk, S.; Eid, S.; Rippmann, F.; Fulle, S. Profiling Prediction of Kinase Inhibitors: Toward the Virtual Assay. J. Med. Chem. 2017, 60, 474−485. (7) Lima, A. N.; Philot, E. A.; Trossini, G. H.; Scott, L. P.; Maltarollo, V. G.; Honorio, K. M. Use of Machine Learning Approaches for Novel Drug Discovery. Expert Opin. Drug Discovery 2016, 11, 225−239. (8) Slynko, I.; Schmidtkunz, K.; Rumpf, T.; Klaeger, S.; Heinzlmeir, S.; Najar, A.; Metzger, E.; Kuster, B.; Schule, R.; Jung, M.; Sippl, W. Identification of Highly Potent Protein Kinase C-Related Kinase 1 Inhibitors by Virtual Screening, Binding Free Energy Rescoring, and in Vitro Testing. ChemMedChem 2016, 11, 2084−2094. (9) Hillisch, A.; Heinrich, N.; Wild, H. Computational Chemistry in the Pharmaceutical Industry: From Childhood to Adolescence. ChemMedChem 2015, 10, 1958−1962. (10) Ferrè, F.; Palmeri, A.; Helmer-Citterich, M. Computational Methods for Analysis and Inference of Kinase/Inhibitor Relationships. Front. Genet. 2014, 5, 196. (11) Avram, S.; Avram, S.; Crisan, L.; Pacureanu, L.; Kurunczi, L.; Bora, A. Self-Organizing Map Classification Model for the Prediction of MEK1 Inhibitors. Rev. Roum. Chim. 2015, 60, 167−173. (12) Crisan, L.; Bora, A.; Pacureanu, L.; Avram, S.; Kurunczi, L. PLS (Partial Least Square) Study for GSK-3 (Glycogen Synthase Kinase-3) Inhibition by Indirubin Derivatives. Rev. Chim. (Bucharest) 2012, 63, 481−488. 966

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967

Article

Journal of Chemical Information and Modeling (34) Gaulton, A.; Hersey, A.; Nowotka, M.; Bento, A. P.; Chambers, J.; Mendez, D.; Mutowo, P.; Atkinson, F.; Bellis, L. J.; Cibrian-Uhalte, E.; Davies, M.; Dedman, N.; Karlsson, A.; Magarinos, M. P.; Overington, J. P.; Papadatos, G.; Smit, I.; Leach, A. R. The ChEMBL Database in 2017. Nucleic Acids Res. 2017, 45, D945−D954. (35) Bain, J.; Plater, L.; Elliott, M.; Shpiro, N.; Hastie, C. J.; McLauchlan, H.; Klevernic, I.; Arthur, J. S.; Alessi, D. R.; Cohen, P. The Selectivity of Protein Kinase Inhibitors: A Further Update. Biochem. J. 2007, 408, 297−315. (36) UniProt Consortium. Reorganizing the Protein Space at the Universal Protein Resource (UniProt). Nucleic Acids Res. 2012, 40, D71−D75. (37) ChEMBL, version 22_1. https://www.ebi.ac.uk/chembl/ (accessed June 2017). (38) Darrouzet-Nardi, A.; Bowman, W. D. Hot Spots of Inorganic Nitrogen Availability in an Alpine-Subalpine Ecosystem, Colorado Front Range. Ecosystems 2011, 14, 848−863. (39) Anighoro, A.; Rastelli, G. Bear, a Molecular Docking Refinement and Rescoring Method. Comput. Mol. Biosci. 2013, 03, 27−31. (40) R Core Team. R: A Language and Environment for Statistical Computing, version 3.2.3; The R Foundation: Vienna, Austria, 2015; http://www.R-project.org/. (41) JChem Base was used for structure searching and chemical database access and management: JChem Base; ChemAxon: Budapest, Hungary, 2016; http://www.chemaxon.com. (42) Willett, P. Similarity-Based Virtual Screening Using 2D Fingerprints. Drug Discovery Today 2006, 11, 1046−1053. (43) Kuhn, M.; Wing, J.; Weston, S.; Williams, A.; Keefer, C.; Engelhardt, A.; Cooper, T.; Mayer, Z.; Kenkel, B.; the R Core Team; Benesty, M.; Lescarbeau, R.; Ziem, A.; Scrucca, L.; Tang, Y.; Candan, C.; Hunt, T. caret: Classification and Regression Training, version 6.0-41; http://CRAN.R-project.org/package=caret (accessed December 2017). (44) Marin, Y. E.; Seiberg, M.; Lin, C. B. Aldo-Keto Reductase 1c Subfamily Genes in Skin Are UV-Inducible: Possible Role in Keratinocytes Survival. Exp. Dermatol. 2009, 18, 611−618. (45) Strobl, C.; Hothorn, T.; Zeileis, A. Party On! A New, Conditional Variable-Importance Measure for Random Forests Available in the party Package. R J. 2009, 1, 14−17. (46) Wilks, S. S. Determination of Sample Sizes for Setting Tolerance Limits. Ann. Math. Stat. 1941, 12, 91−96. (47) Hanley, J. A.; McNeil, B. J. The Meaning and Use of the Area under a Receiver Operating Characteristic (ROC) Curve. Radiology 1982, 143, 29−36. (48) Avram, S. I.; Crisan, L.; Bora, A.; Pacureanu, L. M.; Avram, S.; Kurunczi, L. Retrospective Group Fusion Similarity Search Based on Eroce Evaluation Metric. Bioorg. Med. Chem. 2013, 21, 1268−1278. (49) Jain, A. N. Bias, Reporting, and Sharing: Computational Evaluations of Docking Methods. J. Comput.-Aided Mol. Des. 2008, 22, 201−212. (50) Truchon, J. F.; Bayly, C. I. Evaluating Virtual Screening Methods: Good and Bad Metrics for the ″Early Recognition″ Problem. J. Chem. Inf. Model. 2007, 47, 488−508. (51) Hu, Y.; Kunimoto, R.; Bajorath, J. Mapping of Inhibitors and Activity Data to the Human Kinome and Exploring Promiscuity from a Ligand and Target Perspective. Chem. Biol. Drug Des. 2017, 89, 834− 845. (52) Metz, J. T.; Johnson, E. F.; Soni, N. B.; Merta, P. J.; Kifle, L.; Hajduk, P. J. Navigating the Kinome. Nat. Chem. Biol. 2011, 7, 200−202. (53) Sutherland, J. J.; Gao, C.; Cahya, S.; Vieth, M. What General Conclusions Can We Draw from Kinase Profiling Data Sets? Biochim. Biophys. Acta, Proteins Proteomics 2013, 1834, 1425−1433. (54) Copeland, R. A. Comparing Relative Affinity. In Evaluation of Enzyme Inhibitors in Drug Discovery: A Guide for Medicinal Chemists and Pharmacologists; Copeland, R. A., Ed.; John Wiley & Sons: Hoboken, NJ, 2005; p 131.

967

DOI: 10.1021/acs.jcim.7b00729 J. Chem. Inf. Model. 2018, 58, 957−967