INTRODUCTION
Vector-borne diseases are one of the important health problems in most tropical countries. The environmental conditions in a region define the vectors, their prevalence, natural reservoirs, as well as their interaction with differ- ent hosts. The imposition of the man in such a niche car- ries risks of contagion in variable degree. In the American continent, they are especially important due to the dynam- ic development process that is taking place in the region, which implies profound ecological changes, which are determinants for the emergence and spread of endemic diseases in the form of epidemic outbreaks. According to the World Health Organization, vector-borne diseases represent 17% of the estimated global burden of infec- tious diseases1.
At present, the Zika virus epidemic has become no-
torious and has been associated with the appearance of microcephaly in neonates, and Guillain-Barre syndrome.
According to scientific consensus, the infection due to this virus and the neurological disorders associated with the infection, constitute a public health emergency of inter- national concern2. The Zika virus (a Flavivirus) is trans- mitted through various routes that include sexual contact, perinatal transmission, blood transfusions and the bite of infected mosquitoes of the genus Aedes, mainly Ae. ae- gypti. This mosquito is also agent of other diseases, like dengue, chikungunya and yellow fever3. Recently, there has been a wide geographic spread of this vector, with reports of Zika fever from 61 countries and territories; 47 of them belongs to the American continent.
Since, there is no specific treatment for Zika virus, the emerging scenario highlights that the eco-friendly and ef- fective control measures for mosquito vectors is of crucial
Larvicidal activity prediction against Aedes aegypti mosquito using computational tools
Yudith Cañizares-Carmenate
1, Mirelys Hernandez-Morfa
1, Francisco Torrens
2,
Gloria Castellano
3&
Juan A. Castillo-Garit
1, 41Unit of Computer-Aided Molecular “Biosilico” Discovery and Bioinformatic Research (CAMD-BIR) Unit, Facultad de Química-Farmacia, Universidad Central “Marta Abreu”; 2Institut Universitari de Ciència Molecular, Universitat de València, Edifici d’Instituts de Paterna, València, Spain; 3Departamento de Ciencias Experimentales y Matemáticas, Facultad de Veterinaria y Ciencias Experimentales, Universidad Católica de Valencia “San Vicente Mártir”, Guillem de Castro; 4Unidad de Toxicología Experimental, Universidad de Ciencias Médicas de Villa Clara, Santa Clara, Villa Clara, Cuba
ABSTRACT
Background & objectives: Aedes aegypti is an important vector for transmission of dengue, yellow fever, chikun- gunya, arthritis, and Zika fever. According to the World Health Organization, it is estimated that Ae. aegypti causes 50 million infections and 25,000 deaths per year. Use of larvicidal agents is one of the recommendations of health organizations to control mosquito populations and limit their distribution. The aim of present study was to deduce a mathematical model to predict the larvicidal action of chemical compounds, based on their structure.
Methods: A series of different compounds with experimental evidence of larvicidal activity were selected to develop a predictive model, using multiple linear regression and a genetic algorithm for the selection of variables, implemented in the QSARINS software. The model was assessed and validated using the OECDs principles.
Results: The best model showed good value for the determination coefficient (R2 = 0.752), and others parameters were appropriate for fitting (s = 0.278 and RMSEtr= 0.261). The validation results confirmed that the model has good robustness (Q2LOO = 0.682) and stability (R2-Q2LOO = 0.070) with low correlation between the descriptors (KXX = 0.241), an excellent predictive power (R2ext = 0.834) and was product of a non-random correlation (R2Y-scr = 0.100).
Interpretation & conclusion: The present model shows better parameters than the models reported earlier in the literature, using the same dataset, indicating that the proposed computational tools are more efficient in identifying novel larvicidal compounds against Ae. aegypti.
Key words Aedes aegypti; larvicide; QSARINS software; vector-borne disease; Zika
importance4. It is possible to reduce the vector density shares, by adopting measures of environmental health in the community or by performing chemical control of vectors and/or with the technical guarantee of adequate entomological surveillance. The use of conventional in- secticides, like DDT and malathion in the past decades demonstrated a successful move; however, this success was short lived due to development of resistance against them in many mosquito strains, ecological imbalance and undesirable effects on non-target organisms. This has necessitated the continued effort to search and develop environmentally safe, biodegradable and low-cost lar- vicides/adulticides for killing larvae/adult mosquitoes, respectively5.
Several approaches have been employed to find new and better larvicidal candidates; including predictive models like quantitative structure-activity relationship (QSAR)6-8. Our research group has performed several QSAR studies applied to the prediction of different physi- cochemical, chemical, pharmacokinetical, toxicological as well as biological properties of different compounds9-17. However, it is of great importance to develop a model that can predict larvicidal activity. The main aim of the pres- ent study was to propose a new and better mathematical model for prediction of larvicidal activity of structurally diverse chemical compounds against Ae. aegypti mosqui- toes using QSAR INSubria software (QSARINS)18. This approach offers an alternative to conventional methods used to design new molecular entities with larvicidal properties by identifying the relationship between the chemical structure and its activity.
MATERIAL & METHODS Compounds and their optimization/modeling
A series of 55 compounds (41 compounds as train- ing set to fit the model and 14 for the external predic- tion), derived from natural products, with experimental larvicidal activity was used for the evaluation as used and proposed by Scotti et al6. The LC50 (50% lethal concentra- tion) values were also taken from the same study, wherein it was calculated on the basis of larvicidal assays. All compounds were optimized by a semi-empirical method, Austin Model 1 (AM1)19, and implemented in molecular orbital PACkage (MOPAC) to achieve minimum energy conformation. The chemical structures of the compounds are shown in Fig. 1.
Molecular descriptors
The modelled/optimized structures were used as the initial structures to calculate the molecular descriptors
via the DRAGON Software version 5.520. This software includes the 0D–3D families and generates 3224 descrip- tors for every molecule. Therefore, to compute only those variables were selected, which were significant for the study (excluding the constant or near constant variables).
However, the number of obtained variables were large;
hence, a genetic algorithm (implemented in QSARINS software) was used for variable selection.
QSAR-MLR model development
For the development of the multiple linear regression (MLR) model QSARINS software version 2.2 was used, which has been developed in the QSAR Research Unit, University of Insubria, Italy. It allows obtaining MLR models by ordinary least squares (OLS) method21. The model was developed based on criteria such as, the fitting (highest R2), robustness (greatest Q2LOO), stability (low- est R2–Q2LOO), and low correlation of descriptors (low- est KXX), amongst others, so that it has a small difference between fitting, cross-validation and external validation parameters.
Validation of model
According to the chemometric approach, the obtained model must be carefully checked and thoroughly validat-
Fig. 1: Molecular structure of the selected compounds6
ed. The QSARINS offers a number of tools to verify that the model meets the standards, set by the Organization for Economic Cooperation and Development (OECD) for the development, validation, acceptance and use of QSAR models in order to increase the confidence in the reliabil- ity of data predicted by it. Based on these principles, the QSAR-MLR model must meet the following criteria: (i) a defined end point; (ii) an unambiguous algorithm; (iii) a defined domain of applicability; (iv) appropriate mea- sures of goodness-of-fit, robustness and predictivity; and (v) a mechanistic interpretation, if possible22.
To comply with principle (i), the LC50 of compounds against the mosquito larvae was used as endpoint. For compliance with principle (ii), the molecular descriptors were calculated with the DRAGON software, and the QSARINS was used to develop the MLR model by OLS method and genetic algorithm for variable selection. This algorithm allows relating the descriptors of the chemical structure and their larvicidal activity.
The definition of the model applicability domain (principle iii), is considered a crucial problem in the che- mometric studies. Only predictions for compounds that fall within the domain of applicability may be considered reliable, but not the extrapolations of the models, even if they show robustness and significance. In order to define the applicability domain of the model, the Williams graph/
plot was used; so that reliable predictions of the model have leverage values lower than the critical leverage, ranging between ±3 standard deviations23. Compounds that were outside these ranges were considered outliers. In addition, the Insubria graph (included in QSARINS) was used, which is useful to assess the position of molecules, the reliability of the predictions of compounds lacking experimental response, and to compare them to the pre- dictions of the external prediction set (14 compounds).
In order to evaluate the measures of goodness-of-fit, robustness and predictivity of the model (principle 4), several exercises were adopted. First, the coefficient of determination, R2 was used to evaluate the fitness of the model, and a modified form R2adj was used to evaluate the efficacy of the model by adding a new descriptor. How- ever, these criteria do not provide information about the predictivity of the model or if it is the product of a casual correlation. Considering these aspects, an internal vali- dation was considered in order to check how robust the model is when small changes occur in the data. For inter- nal validation, the cross-validation techniques were used.
Initially, leave-one-out (LOO) exercises were done which involved the process of removing a single compound, and creating and validating the model against the individu- al molecules for the entire set. Later, a leave-many-out
(LMO) technique was used, which allows studying the be- havior of the model when a larger number of compounds are excluded. In order to demonstrate that the model is not the result of a casual correlation, the randomization (Y-Scrambling) procedure was applied24. In this process, the response variable are randomly located, so that there is no correlation with the descriptors and, as a result, the model performance should decay dramatically.
In order to prove predictivity of the model, an external validation was performed. For this, the obtained model equation was evaluated with the external prediction set, which were not used in the model calculation (they are only used to check the predictability of the model).
RESULTS QSAR-MLR modeling results
Large numbers of models were obtained while ex- ploring the best combinations of molecular descriptors that show a high correlation with the response variable (pLC50). Therefore, an analysis of the model’s parameters was done, taking into account the principle of parsimony.
Based on this analysis, a QSAR-MLR model was devel- oped to evaluate the larvicidal activity against Ae. aegypti (so that it is able to describe the maximum information with the least number of descriptors). The equation and the statistical parameters of the best model are as follows:
pLC50=1.516 (±0.330)-0.224 (±0.049)×X2v+0.220 (±0.057)×
GATS5p +26.243 (±10.848)×R3p++0.540 (±0.054)×AlogP
R2 = 0.752; R2adj = 0.724; R2-R2adj = 0.028; LOF = 0.105;
KXX= 0.241; Delta K= 0.050; RMSEtr= 0.261; MAEtr= 0.206; RSStr= 2.784; CCCtr= 0.858; s = 0.278; F = 27.258 Where, R2 is the coefficient of determination, R2adj is adjusted R2, s is standard error of estimate, F is variance ratio, and LOF is Friedman lack of fit25-26. The parameter Kxx is the correlation among descriptors27, delta K is the difference of the correlation between the descriptors (Kx) and the descriptors plus the responses (Kx). Moreover, RMSEtr is root-mean-square error in fitting (for the train- ing set), MAEtr is mean absolute error in fitting (calculated on the training set), RSStr is the residual sum of squares in the fitting and CCCtr is the concordance correlation coef- ficient calculated over the training set28-30.
Validation results according to OECD standards
The leverage (h) and standardized residual approach, described in the literature as Williams Plot (Fig. 2), used for applicability domain study shows the graph for the training and prediction sets. As shown in the figure, most
compounds are within the applicability domain of the model.
The Insubria plot enables an easy visual inspection of the diagonal values of the leverage values (HAT) against the predicted responses for each compound and is shown in the Fig. 3.
The results of the internal validation demonstrated the fitness and stability of the model. The most impor- tant techniques for internal validation are LOO, LMO and Y-scrambling. According to the results, it was clear that the internal predictions were good showing values of Q2LOO= 0.682, RMSEcv = 0.295 and MAEcv = 0.236. The Fig. 4 shows the values predicted by LOO vs experimen-
tal LC50 values for the training and test sets. The model’s performance for internal validation by LMO was Q2LMO= 0.663, although it is important to take into account that this technique gives better result with larger databases. Figure 5 shows Q2LMO vs KXY values (correlation between the descriptors and larvicidal activity).
The results of the Y-scrambling procedure indicate that the proposed model is well founded, and not just the result of chance correlation. The values of R2 and Q2 of every iteration are shown in Fig. 6 (Yellow and red circles, correspondingly), and their averages were R2Y-scr = 0.100 and Q2Y-scr = 0.177 (Light blue and dark blue, respectively).
The predictivity of the model was tested by using
Fig. 2: Williams plot—Leverage values vs Standardized residuals.
Fig. 3: Insubria graph—Leverage values vs Predicted data.
Fig. 4: Scatter plot of experimental pLC50 vs Predicted by LOO Q2 experiment.
Fig. 5: Plot of Q2 values of LMO validations compared with the Q2 value of the original model.
Fig. 6: Plot of Y-scrambled model compared with the original model.
Fig. 7: Scatter plot of experimental pLC50 vs Predicted values of pLC50 by model equation.
the external prediction set of 14 compounds. Its parameters were R2ext = 0.834, RMSEex = 0.272, MAEext = 0.209, PRESSext = 1.039, Q2-F1 = 0.718, Q2-F2 = 0.709, Q2-F3 = 0.729, CCCext = 0.869, r2m_aver = 0.763, r2m_
delta = 0.008; where, R2ext is the external determination coefficient31; RMSEext is the root-mean-square error in external prediction; MAEext is the mean absolute er- ror in external prediction; PRESSext is the predictive residual sum of squares (external validation); Q2-F132, Q2-F233, Q2-F334 are the explained variances in external prediction; CCCext is the concordance correlation coef- ficient28-30 and r2m-aver and r2m-delta are the Roy criteria:
Average and delta, respectively. The predictions of com- pounds in the external set is shown in Fig. 7 (Blue circles) and Table 1.
Table 1. LC50 predicted by model equation Compounds Exp.
endpointa Predicted
by model HAT i/ib
(h*=0.366) SPRc
1 2.790 2.938 0.153 0.580
2 2.360 2.753 0.116 1.504
3* 2.310 2.284 0.154 –0.103
4 2.040 2.228 0.167 0.739
5 3.470 3.585 0.141 0.445
6 3.320 3.234 0.079 –0.324
7* 3.640 3.656 0.078 0.058
8 3.590 3.807 0.330 0.955
9 3.490 3.517 0.088 0.101
10 3.660 3.972 0.173 1.235
11* 3.430 3.396 0.142 –0.132
12 3.090 3.133 0.058 0.158
13 3.320 3.128 0.050 –0.709
14* 3.660 3.708 0.148 0.188
15 3.850 3.600 0.227 –1.024
16 3.490 3.421 0.063 –0.257
17* 3.460 3.889 0.141 1.666
18 3.160 3.277 0.079 0.437
19 3.720 3.361 0.107 –1.365
20 2.650 3.037 0.052 1.431
21 2.160 2.267 0.342 0.476
22 2.290 1.915 0.250 –1.557
23* 2.040 1.863 0.276 –0.749
24 2.330 2.964 0.051 2.341
25 3.280 2.991 0.042 –1.064
26 3.040 3.015 0.034 –0.092
27 2.400 2.143 0.199 –1.033
28 2.660 2.642 0.182 –0.070
29* 3.400 3.326 0.056 –0.273
30 3.350 3.172 0.062 –0.660
31* 2.690 3.189 0.703 3.293
32 3.390 3.342 0.156 –0.189
33 2.840 2.842 0.243 0.008
34 3.280 3.187 0.046 –0.343
35 2.050 2.694 0.057 2.384
36 2.410 2.501 0.138 0.352
37 2.710 2.606 0.077 –0.390
38* 2.590 3.071 0.067 1.791
39 2.480 2.554 0.091 0.279
40* 2.850 2.947 0.136 0.377
41 3.510 3.198 0.159 –1.225
42 3.240 3.180 0.052 –0.221
43 2.440 2.606 0.077 0.621
44 2.880 2.424 0.133 –1.760
45* 2.470 2.906 0.107 1.659
46 3.420 3.354 0.080 –0.247
47 3.550 3.257 0.047 –1.079
48* 3.000 3.283 0.066 1.052
49 2.280 2.583 0.161 1.190
50 3.700 3.361 0.097 –1.282
51 3.110 3.125 0.101 0.058
52* 2.950 3.102 0.259 0.634
53 3.080 3.324 0.171 0.965
54 3.640 3.283 0.066 –1.330
55* 2.470 2.646 0.098 0.666
*Compounds used in the test set; aExperimental endpoint expressed as pLC50mol/l;
bLeverage values (diagonal elements of the HAT matrix HAT i/ib); cStandardized residuals predicted (SPR) from the model; h*–Critical leverage value.
DISCUSSION
Several predictive QSAR model based studies have been employed to find new compounds with po- tential larvicidal activity; however, the development of newer models demand high accuracy, robustness and ap- plicability domain. The model developed in this study showed a R2 value of 0.752, indicating an appropriate fit for modeling larvicidal activity against Ae. aegypti.
This means that the model explains more than 75% of the experimental variance. There was no overfitting in the model representing a good fit with a minimum number of descriptors, due to the low value of the LOF parameter (0.105) and a R2adj (0.724). The LOF is used to penalize the addition of descriptors in the model equation and should be as low as possible; and the value of R2adj should be very similar to the R2 value, which were fulfilled by the model developed. The correlation between the descriptors of the model was low, since the KXX value was small (0.241), indicating that there was no redundant information in the selected descriptors. In addition, the correlation between the descriptors and the response variable was appropriate, in accordance with the Delta K parameter (0.05), with a small error in the calculations of training and parameters estimation (RMSEtr = 0.261; MAEtr = 0.206; s = 0.278).
From the scatter plot of the predicted response against experimental, it was evident that the compounds tend to be located next to the diagonal line. The compounds with extreme predictions were assessed as potential outliers.
The applicability domain was assessed by use of two approaches: William's Plot and Insubria graph. As evident from the Williams Plot, there was only one compound, i.e. 31 with a leverage value (h = 0.703) greater than the critical leverage value (h* = 0.366), with standard devia- tion value of 3.293, which is out of limits, indicating it as a possible outlier. Similar to the results obtained in Wil- liam’s graph, the compound 31 was also excluded in the Insubria graph, as it was behaving as an outlier.
In order to evaluate the robustness of the model, several internal validation strategies were carried out.
The parameters of the cross-validation LOO technique showed a well-known criterion: the explained variance in the prediction by LOO (Q2LOO), which iteratively excludes a compound of the dataset. The model’s performance for internal validation by LMO was similar to LOO (Q2LMO= 0.663), although it is important to take into account that this technique gives better result with larger databases.
It may be noted that Q2LMO values were similar to each other and comparable to KXY values, corroborating the good fit and stability of the model. The values of R2 and Q2 of every iteration, and their averages (R2Y-scr and Q2Y-scr)
showed that the model was good, since these parameters were lower with regard to the values of the model (R2Y-scr = 0.100 and Q2Y-scr = – 0.177). As obvious in the Figure 6, the red circles (Q2LMO for each iteration) were around the blue circle (R2 of the model). In the Y-scrambling experiment, the absence of change in correlation for the model was checked. The expected behavior is that the values of R2Y-scr and Q2Y-scr (for each iteration) should differ appreciatively from R2 and Q2 values of the model. As seen in Fig. 6, the values of R2 and Q2 of the model were far from the values obtained for those parameters in the Y-scrambling experi- ment, confirming the validity of the model.
In order to assess the model’s ability to predict new compounds, an external validation was also performed.
The procedure was carried out by applying the model equation, developed with the training set, to the external prediction set. Based on the results with the external set (R2ext = 0.834, RMSEext = 0.272,), it was inferred that the model has a good predictive power.
To study the structure-activity relationship of po- tential larvicidal agents, Scotti et al6 investigated the molecular interactions of 55 selected compounds (also used in this study) with the aid of chemometric methods which included principal components analysis (PCA), consensus PCA (CPCA), and partial least squares (PLS) regression. Hence, a direct comparison was carried out between their best model and the developed model. Their best model obtained with PLS exhibited R2= 0.714, while the present model showed a better value of R2= 0.752.
Similarly, in the LOO cross-validation, the present model showed slightly better results (Q2LOO = 0.682) with respect to Q2LOO = 0.679. Additionally, the results of the exter- nal validation (R2ext=0.834) were much better than their reports (R2ext = 0.623). Thus, it can be concluded that the derived model has an improved performance with respect to the Scotti et al6 model and could be a useful tool for the prediction of the larvicidal activity against Ae. aegypti.
CONCLUSION
In the present study, a QSAR-MLR model was developed by using molecular descriptors calculated us- ing the Dragon software, which adequately predicts the larvicidal activity against Ae. aegypti mosquitoes. The model developed with QSARINS software fulfilled all regulatory principles established by the OECD; the ro- bustness of the model was tested through internal valida- tion techniques (LOO, LMO, and Y-scrambling), and its predictability was determined with an external prediction set (external validation). The obtained model showed bet-
ter parameters than a model previously reported in the literature, using the same datasets, concluding that the proposed computational tools are efficient in identifica- tion of novel larvicidal compounds.
Conflict of interest
The authors declared no conflicts of interest.
ACKNOWLEDGEMENTS
The authors are grateful to Prof. Paola Gramatica and her group of the University of Insubria for providing the QSARINS software for this study. This work was sup- ported by Generalitat Valenciana (Project No. PROME- TEO/2016/094) and Valencia Catholic University Saint Vincent Martyr (Projects Nos. PRUCV/2015/617 and 2017).
REFERENCES
1. Factsheet: Vector-borne diseases. Geneva: World Health Orga- nization 2014. Factsheet No. 387. Available from: http://www.
who.int/campaigns/world-health-day/2014/fact-sheets/en/ (Ac- cessed on May 4, 2017).
2. Situation Report: Zika virus, microcephaly, Guillain-Barre syn- drome. Geneva: World Health Organization 2016. Available from: http://apps.who.int/iris/bitstream/10665/250724/1/zikas- itrep3Nov16-eng.pdf (Accessed on May 4, 2017).
3. Musso D, Nilles EJ, Cao-Lormeau V-M. Rapid spread of emerg- ing Zika virus in the Pacific area. Clin Microbiol Infect 2014;
20(10): O595–6.
4. Benelli G. Spread of Zika virus: The key role of mosquito vector control. Asian Pac J Trop Biomed 2016; 6(6): 468–71.
5. Magadula JJ, Innocent E, Otieno JN. Mosquito larvicidal and cytotoxic activities of three Annona species and isolation of ac- tive principles. J Med Plant Res 2009; 3(9): 674–80.
6. Scotti L, Tullius-Scotti M, Barros-Silva V, Lima-Santos SR, Cavalcanti SCH, Mendonça-Junior FJB. Chemometric studies on potential larvicidal compounds against Aedes Aegypti. Med Chem 2014; 10(2): 201–10.
7. Garza-Robledo AA, Martinez-Perales JF, Rodriguez-Castro VA, Quiroz-Martinez H. Effectiveness of spinosad and temephos for the control of mosquito larvae at a tire dump in Allende, Nuevo Leon, Mexico. J Am Mosq Control Assoc 2011; 27: 404–7.
8. Wang Z, Kim JR, Wang M, Shu S, Ahn YJ. Larvicidal activity of Cnidium monnieri fruit coumarins and structurally related com- pounds against insecticide-susceptible and insecticide-resistant Culex pipiens pallens and Aedes aegypti. Pest Manag Sci 2012;
68: 1041–7.
9. Marrero-Ponce Y, Martínez-Albelo E, Casañola-Martín G, Cas- tillo-Garit J, Echevería-Díaz Y, Zaldivar V, et al. Bond-based linear indices of the non-stochastic and stochastic edge-adjacen- cy matrix. 1. Theory and modeling of ChemPhys properties of organic molecules. Mol Div 2010; 14(4): 731–53.
10. Castillo-Garit JA, Marrero-Ponce Y, Torrens F, García- Domenech R, Rodríguez-Borges JE. Applications of bond- based 3D-chiral quadratic indices in QSAR studies related
to central chirality codification. QSAR Comb Sci 2009; 28:
1465–77.
11. Le-Thi-Thu H, Canizares-Carmenate Y, Marrero-Ponce Y, Tor- rens F, Castillo-Garit JA. Prediction of Caco-2 cell permeability using bilinear indices and multiple linear regression. Lett Drug Des Discov 2016; 13(2): 161–9.
12. Castillo-Garit JA, Casañola-Martin GM, Le-Thi-Thu H, Pham- The H, Barigye SJ. A simple method to predict blood-brain bar- rier permeability of drug-like compounds using classification trees. Med Chem 2017; doi: 10.2174/15734064136661702091 24302. [Epub ahead of print].
13. Brito-Sánchez Y, Castillo-Garit JA, Le-Thi-Thu H, González- Madariaga Y, Torrens F, Marrero-Ponce Y, et al. Comparative study to predict toxic modes of action of phenols from molecu- lar structures. SAR QSAR Environ Res 2013; 24(3): 235–51.
14. Castillo-Garit JA, Del Toro-Cortés O, Kouznetsov VV, Puen- tes CO, Romero Bohórquez AR, Vega MC, et al. Identification in silico and in vitro of novel trypanosomicidal drug-like com- pounds. Chem Biol Drug Des 2012; 80(1): 38–45.
15. Castillo-Garit JA, del Toro-Cortés O, Vega MC, Rolón M, Ro- jas de Arias A, Casañola-Martin GM, et al. Bond-based bilinear indices for computational discovery of novel trypanosomicidal drug-like compounds through virtual screening. Eur J Med Chem 2015; 96: 238–44.
16. Cañizares-Carmenate Y, Mena-Ulecia K, Perera-Sardiña Y, Tor- rens F, Castillo-Garit JA. An approach to identify new antihy- pertensive agents using Thermolysin as model: In silico study based on QSARINS and docking. Arab J Chem 2016. Available from: http://dx.doi.org/10.1016/j.arabjc.2016.10.003.
17. Castillo-Garit JA, Marrero-Ponce Y, Barigye SJ, Medina-Mar- rero R, Bernal MG, Vega JMGdl, et al. In silico antibacterial ac- tivity modeling based on the TOMOCOMD-CARDD approach.
J Braz Chem Soc 2015; 26: 1218–26.
18. Gramatica P, Chirico N, Papa E, Cassani S, Kovarich S. QSA- RINS, software for QSAR MLR model development and vali- dation: QSAR Research Unit in in Environmental Chemistry and Ecotoxicology. Varese, Italy: University of Insubria 2013.
19. Schmidt MW, Baldridge KK, Boatz JA, ST E, Gordon MS, Jen- sen JH, et al. General atomic and molecular electronic structure system. J Comput Chem 1993; 14: 1167–89.
20. Talete srl. DRAGON for Windows (Software for molecular de- scriptor calculations); v5.5; edn 2007.
21. Gramatica P, Chirico N, Papa E, Cassani S, Kovarich S. QSA- RINS: A new software for the development, analysis and valida- tion of QSAR MLR models. J Comput Chem (Software News and Updates) 2013; 34(24): 2121–32.
22. Worth AP, Bassan A, de Bruijn J, Gallegos Saliner A, Netzeva T, Patlewicz G, et al. The role of the European Chemicals Bureau in promoting the regulatory use of (Q)SAR methods. SAR QSAR Environ Res 2007; 18(1-2): 111–25.
23. Atkinson AC. Plots, transformations, and regression: An intro- duction to graphical methods of diagnostic regression analysis.
Oxford: Clarendon Press 1985; xi: 282.
24. Rücker C, Rücker G, Meringer M. y-Randomization and its variants in QSPR/QSAR. J Chem Inf Model 2007; 47(6): 2345–
25. Friedman JH. Multivariate adaptive regression splines. Ann Stat 57.
1991; 19(1): 1–67.
26. Friedman JH. Rejoinder: Multivariate adaptive regression splines. Ann Stat 1991; 19(1): 123–41.
27. Todeschini R, Consonni V, Maiocchi A. The K correlation in- dex: Theory development and its application in chemometrics.
Chemom Intell Lab Syst 1999; 46(1): 13–29.
28. Lin LI. A concordance correlation coefficient to evaluate repro- ducibility. Biometrics 1989; 45(1): 255–68.
29. Chirico N, Gramatica P. Real external predictivity of QSAR Models: How to evaluate it? Comparison of different validation criteria and proposal of using the concordance correlation coef- ficient. J Chem Inf Model 2011; 51(9): 2320.
30. Chirico N, Gramatica P. Real External Predictivity of QSAR Models. Part 2. New intercomparable thresholds for different validation criteria and the need for scatter plot inspection. J Chem Inf Model 2012; 52(8): 2044.
31. Golbraikh A, Tropsha A. Beware of q2! J Mol Graph Model 2002; 20(4): 269–76.
32. Shi LM, Fang H, Tong W, Wu J, Perkins R, Blair RM, et al.
QSAR models using a large diverse set of estrogens. J Chem Inf Comput Sci 2001; 41(1): 186–95.
33. Schüürmann G, Ebert R, Chen J, Wang B, Kühne R. External validation and prediction employing the predictive squared cor- relation coefficient-test set activity mean vs training set activity mean. J Chem Inf Model 2008; 48(11): 2140–5.
34. Consonni V, Ballabio D, Todeschini R. Comments on the defini- tion of the Q2 parameter for QSAR validation. J Chem Inf Model 2009; 49(7): 1669–78.
35. Ojha PK, Mitra I, Das RN, Roy K. Further exploring rm2 metrics for validation of QSPR models. Chemometr Intell Lab 2011;
107(1): 194–205.
Correspondence to: Dr Juan Alberto Castillo-Garit, Unidad de Toxicología Experimental, Universidad de Ciencias Médicas de Villa Clara, Santa Clara, 50200, Villa Clara, Cuba.
E-mail: [email protected]
Received: 3 March 2017 Accepted in revised form: 12 May 2017
