Modelling the cytotoxic activity of pyrazolo-triazole hybrids using descriptors calculated from the open source tool “PaDEL-descriptor”

Peer review under responsibility of Taibah University.

Abstract

In this study, we developed QSAR models for the anti-proliferative activity of pyrazolo-triazole hybrids [(1-benzyl-1H-1,2,3-triazol-4-yl)(1,3-diphenyl-1H-pyrazol-4-yl) methanone] on human brain cancer (U87MG), lung cancer (A549), prostate cancer (PC-3), and colon cancer (HT-29) cell lines. We employed K-means cluster analysis to split the data sets. Statistically robust models were generated [pIC₅₀ (U87MG): R = 0.873, Q² = 0.554, $R_{pred}^2=0.866;$ pIC₅₀ (A549): R = 0.879, Q² = 0.637, $R_{pred}^2=0.858;$ pIC₅₀ (PC3): R = 0.953; Q² = 0.850; $R_{pred}^2=0.796;$ pIC₅₀ (HT-29): R = 0.962, Q² = 0.891; $R_{pred}^2=0.707]\text{.}$ The reliability of these models was confirmed by acceptable validation parameters, and these models also satisfied the Golbraikh and Tropsha acceptable model criteria. The QSAR study highlighted the atomic feature and molecular descriptors, information content descriptors, and topological and constitutional descriptors that affect anti-cancer activity.

Keywords:

• Pyrazolo-triazole hybrids
• Cancer
• QSAR
• PaDEL-Descriptor
• S-MLR
• K-means clustering

1 Introduction

The conventional drug discovery and development cascade, from a theoretical concept to the market, consumes almost 15 years with large monetary expenses [1,2]. Computational methodologies have emerged as a crucial component of the drug discovery process. Many in silico tools have been developed to shorten the drug discovery cycle to minimize the expense and risk of failure [3,4]. Computer-aided drug design (CADD) is an important method that can help from hit identification to lead optimization and beyond [5,6]. Computer-aided drug design (CADD) including quantitative structure-activity relationships (QSAR), pharmacophore modelling, molecular field analysis, molecular docking and de novo drug design, which allow the development of predictive models for potential lead candidate discovery and optimization [7,8].

Quantitative structure-activity relationship (QSAR) modelling is used to find the relationship between an observed activity and the corresponding structural features of chemo types [9–15]. By predicting the biological activity of a novel chemical entity from the QSAR models, various molecular descriptors are calculated [16–21]. A large number of software programmes that are able to calculate such molecular descriptors are available commercially as well as freely [9]. “PaDEL-Descriptor” (http://www.yapcwsoft.com/dd/padeldescriptor/) is a premium QSAR modelling software that calculates 797 descriptors (663 1D, 2D descriptors, and 134 3D descriptors) and 10 types of fingerprints [9,10].

Previously, we developed many QSAR models for varieties of diseases, such as cancer and AIDS [12–17,19,21]. Now, our attention is focussed on a series of pyrazolo-triazole hybrids that exhibit promising anti-proliferative properties, as reported by Reddy and co-workers [22]. Thus far, no QSAR study on pyrazolo-triazole hybrids has been reported. Hence, we have attempted to develop some robust statistically significant QSAR models to predict the biological potency of those pyrazolo-triazole hybrids. The information generated from the QSAR could be effective to understanding the structure – activity relationships of the series under consideration and subsequently design a new anticancer lead compound.

2 Materials and methods

2.1 Datasets and biological activity

The antiproliferative activities of several (1-benzyl-1H-1,2,3-triazol-4-yl)(1,3-diphenyl-1H-pyrazol-4-yl) methanone derivatives (Table 1) on different human cancer cell lines [U87MG (brain), A549 (lung), PC3 (prostate), and HT-29 (colon)] were selected from Reddy et al. [22] to develop the chemometric tools. On the basis of the four different cell lines, four datasets were generated, and the negative logarithm of the IC₅₀ (μM) value was considered as the dependent parameter or the response variable for the QSAR models [18,20,23–25].

Table 1 Antiproliferative activities (IC₅₀-μM) of the (1-benzyl-1H-1,2,3-triazol-4-yl)(1,3-diphenyl-1H-pyrazol-4-yl)methanone derivatives (1–30).

No^a		R	IC50 (μM)
			U87MG	A549	PC-3	HT-29
1		F	6.45	7.07	>50	22.4
2		Cl	>50	>50	>50	>50
3		CF3	>50	>50	>50	>50
4		OCH3	8.87	7.21	4.21	5.71
5		3,4 OCH3	3.37	4.69	2.69	2.92
6		3,4,5 OCH3	7.85	5.56	6.81	4.63
7		F	4.76	5.92	17.87	14.48
8		Cl	9.43	8.5	16.04	15.92
9		CF3	24.76	>50	>50	>50
10		OCH3	3.76	2.13	5.87	2.54
11		3,4 OCH3	1.14	2.3	3.34	2.48
12		3,4,5 OCH3	5.3	3.8	4.4	3.09
13		F	8.91	4.98	>50	12.59
14		Cl	4.24	6.04	11.05	9.92
15		CF3	6.54	5.72	13.4	12.9
16		OCH3	3.3	5.3	6.3	1.78
17		3,4 OCH3	0.86	1.93	3.71	2.42
18		3,4,5 OCH3	3.95	2.42	5.62	3.5
19		F	15.5	13.7	9.7	14.7
20		Cl	5.89	11.21	>50	7.22
21		CF3	4.24	14.5	>50	29.11
22		OCH3	3.63	4.84	7.84	6.99
23		3,4 OCH3	0.97	2.24	3.52	1.78
24		3,4,5 OCH3	1.15	5.32	5.65	4.4
25		F	18.9	>50	>50	28.6
26		Cl	26.7	>50	>50	24.1
27		CF3	8.25	>50	>50	14.72
28		OCH3	2.47	5.5	>50	4.39
29		3,4 OCH3	3.61	3.47	2.32	2.6
30		3,4,5 OCH3	3.81	6.45	5.6	3.89

a Compound number.

2.2 Generation and selection of descriptors

The 2D structures of the pyrazolo-triazole hybrids were drawn using ChemDraw Ultra 8.0.3 [26]. Then, the PaDEL-Descriptors [9] were calculated for the compounds on the MS Windows platform. Whole constitutional, topological, electronic, thermodynamic, geometric, and autocorrelation descriptors were calculated by PaDEL software. These descriptors encoded different aspects of the molecular structures and were considered as independent parameters in the chemometric models.

2.3 Correlation analysis

Initially, a correlation analysis was carried out with the response variable and independent parameters of the compounds using STATISTICA 7 software [27]. The information less descriptors that showed no variation between structures and/or descriptors and those that did not cover the whole space were initially deleted from our study [20]. One of each inter-correlated pair of variables with a correlation coefficient greater than 0.60 was eliminated to decrease the redundancy.

2.4 Test set and training set selection

The splitting of datasets is very important in chemometric model development because of the influence of the chemical nature of the training set compounds on the predictive ability of a developed QSAR model. Non-hierarchical clustering techniques such as the K-means clustering technique are widely used to standardize a descriptor matrix consisting of different descriptors [28,29]. In the K-means clustering analysis, the separation and arrangement of different objects into several groups was performed according to their distance in multidimensional space. In this study, we used the STATISTICA 7 software [27] based K-means clustering analysis tool.

2.5 Stepwise multiple linear regression (S-MLR)

Stepwise multiple linear regression [18–20] methods built multistep models step by step. The following procedures were followed to build S-MLR models.

(i)	By altering the descriptor or variable combination, a previous step was repeated with the stepping criterion of F = 4 for inclusion and F = 3.99 for exclusion.

(ii)	When a specified maximum number of steps had been reached or/and when stepping is no longer possible with the given stepping criterion, the search was terminated.

2.6 Statistical methods

Several validation parameters were employed to verify the statistical quality of all of the models. The correlation coefficient (R), squared correlation coefficient (R²), adjusted R² ($R_A^2$), variance ratio (F) at specified degree of freedom (df) and standard error of the estimate (SEE) [19,28–30] were calculated using STATISTICA 7 software [27].

We also reported the p-value and t-value for various parameters in each model. The effects of the intercorrelation of the descriptors was further verified by the variant inflation factor (VIF) [18,19]. A VIF value greater than 10 indicates a defective QSAR model.$\text{VIF}=\frac{1}{1-r^2}$where r² is the multiple correlation coefficient of one descriptor's effect regressed over the remaining molecular descriptors.

2.7 Validation of the 2D-QSAR Models

The leave-one-out (LOO) cross-validation method, the predicted residual sum of squares (PRESS), and the standard deviation of error of prediction (SDEP) [28–30] were applied as internal validation tools to validate our developed 2D-QSAR models. The predictive powers of all of the developed QSAR models were justified by these internal cross-validation techniques.

To predict the anti-proliferative activities of the pyrazolo-triazole hybrids, it was also important to calculate the $R_{pred}^2$, $r_{m(test)}^2$, rm²_(overall), Δrm²_(overall), and other similar parameters. [29]. $R_{pred}^2$ was calculated by the equation$R_{pred}^2=1-\frac{\sum {(y_i-{\overline{y}}_i)}^2}{\sum {(y_i-y_{mean})}^2}$where y_i and ${\overline{y}}_i$ are the actual and predicted activities of the ‘ith’ molecule in the training set, respectively, and y_mean is the average activity of all molecules in the test set.

The matrices $r_m^2$ were calculated to understand both the internal and external predictive capacities,$r_m^2=r^2\times (1-\sqrt{r^2-r_0^2})$where r² is the squared correlation coefficient between the observed vs predicted response values of the test set compounds and $r_0^2$ represents the corresponding values for regression through the origin (observed vs. predicted).

The Y-randomization of the model (cR²p) was evaluated to ensure that the model is not based on chance for 100 scrambling iterations [19]. Additional parameters proposed by Golbraikh and Tropsha et al. [30] were also determined for the developed QSAR models to ensure robustness.

3 Results and discussion

3.1 Selection of training and test sets

All datasets were tested on the basis of the anti-cancer activity on human cancer cell lines [U87MG (brain cancer), A549 (lung cancer), PC3 (prostate cancer), and HT-29 (colon cancer)] to develop the QSAR models (Table 1). Each dataset (n = 28 in U87MG cell line; n = 24 in A549 cell line; n = 19 in PC3 cell line; and n = 27 in HT-29 cell line) of the pyrazolo-triazole hybrids (Fig. 1.) was divided on the basis of the clusters, as derived from the K-means clustering method in STATISTICA version 7 software [27].

Fig. 1 General structure of pyrazolo-triazole hybrids (1-benzyl-1H-1,2,3-triazol-4-yl)(1,3-diphenyl-1H-pyrazol-4-yl) methanone.

Four clusters (three clusters in the case of the PC3 cell line) were generated, and from each cluster, 25% of the compounds were taken at random to form the test set. It was also ensured that the property range of the test set compounds spans over the range of the training set compounds. Thus, the dataset was divided into training (n_Train = 21 in U87MG cell line; n_Train = 18 in A549 cell line; n_Train = 14 in PC3 cell line; n_Train = 20 in HT-29 cell line) and test (n_Test = 7 in U87MG cell line; n_Test = 6 in A549 cell line; n_Test = 5 in PC3 cell line; n_Test = 7 in HT-29 cell line) sets, as shown in Supplementary Table T1.

Four clusters (three clusters in the case of the PC3 cell line) were generated, and from each cluster, 25% of the compounds were taken at random to form the test set. It was also ensured that the property range of the test set compounds spans over the range of the training set compounds. Thus, the dataset was divided into training (n_Train = 21 in U87MG cell line; n_Train = 18 in A549 cell line; n_Train = 14 in PC3 cell line; n_Train = 20 in HT-29 cell line) and test (n_Test = 7 in U87MG cell line; n_Test = 6 in A549 cell line; n_Test = 5 in PC3 cell line; n_Test = 7 in HT-29 cell line) sets, as shown in Supplementary Table T1.

The QSAR models were generated based on the training set compounds, and the test set compounds were taken for validation purposes. The criterion for selecting the best model was based on the correlation coefficient. The recommended ratio between the number of predictor parameters and the number of data points was 1:5 for all of these models [19,28,31,32].

3.2 Stepwise multiple linear regression (S-MLR)

The correlation matrixes among descriptors used to develop regression-based QSAR models are provided in Supplementary Table T2. The best models that we found through the stepwise multiple linear regression analysis of all of these datasets are as follows.

The correlation matrixes among descriptors used to develop regression-based QSAR models are provided in Supplementary Table T2. The best models that we found through the stepwise multiple linear regression analysis of all of these datasets are as follows.(1) $pI{C}_{50}(\text{U}87\text{MG})=-28.596(\pm 11.939)+0.647(\pm 0.128)ATSC6e-4.288(\pm 0.976)VE1\_Dt+9.661(\pm 3.333)PetitjeanNumber+15.183(\pm 5.978)SpMin1\_Bhs\text{.}$(1) $n_{Train}=21;R=0.873;R^2=0.762;R_A^2=0.703;F(\mathrm{4,16})=12.806;\text{SEE}=0.212;Q^2=0.554;\text{PRESS}=1.353;\text{SDEP}=0.254;r{m}_{(LOO)}^2=0.510;\Delta r{m}_{(LOO)}^2=0.156;n_{Test}=7;R_{pred}^2=0.866;r_{m(test)}^2=0.723;Avg{r}_{m(test)}^2=0.765;\Delta r_{m(test)}^2=0.083;r{m}_{(overall)}^2=0.566;\Delta r{m}_{(overall)}^2=0.051;c{R}^{2}p=0.638\text{.}$(2) $pI{C}_{50}(\text{A}549)=3.033(\pm 0.580)-6.977(\pm 1.090)MATS8c+205.629(\pm 51.458)JGI10+0.050(\pm 0.017)VR3\_Dzp+0.187(\pm 0.074)VE3\_Dzv\text{.}$(2) $n_{Train}=18;R=0.879;R^2=0.772;R_A^2=0.703;F(\mathrm{4,13})=11.020;\text{SEE}=0.118;Q^2=0.637;\text{PRESS}=0.290;\text{SDEP}=0.127;r{m}_{(LOO)}^2=0.590;\Delta r{m}_{(LOO)}^2=0.120;n_{Test}=6;R_{pred}^2=0.858;r_{m(test)}^2=0.637;Avg{r}_{m(test)}^2=0.542;\Delta r_{m(test)}^2=0.188;r{m}_{(overall)}^2=0.697;\Delta r{m}_{(overall)}^2=0.219;c{R}^{2}p=0.687\text{.}$(3) $pI{C}_{50}(\text{PC}-3)=53.755(\pm 8.220)-3.823(\pm 0.428)AATS6e-295.415(\pm 96.956)ASP-7-8.421(\pm 2.892)SpMin6\_Bhi\text{.}$(3) $n_{Train}=14;R=0.953;R^2=0.908;R_A^2=0.880;F(\mathrm{3,10})=32.881;\text{SEE}=0.078;Q^2=0.850;\text{PRESS}=0.0993;\text{SDEP}=0.084;r{m}_{(LOO)}^2=0.822;\Delta r{m}_{(LOO)}^2=0.059;n_{Test}=5;R_{pred}^2=0.796;r_{m(test)}^2=0.657;Avg{r}_{m(test)}^2=0.553;\Delta r_{m(test)}^2=0.207;r{m}_{(overall)}^2=0.787;\Delta r{m}_{(overall)}^2=0.148;c{R}^{2}p=0.833\text{.}$and(4) $pI{C}_{50}(\text{HT}-29)=-17.371(\pm 8.018)-819.204(\pm 72.655)AATSC8c+6.953(\pm 0.798)VE1\_Dzv+9.997(\pm 4.207)SpMin3\_Bhs\text{.}$(4) $n_{Train}=20;R=0.962;R^2=0.926;R_A^2=0.912;F(\mathrm{3,16})=66.924;\text{SEE}=0.118;Q^2=0.891;\text{PRESS}=0.326,\text{SDEP}=0.128;r{m}_{(LOO)}^2=0.879;\Delta r{m}_{(LOO)}^2=0.060;n_{Test}=7;R_{pred}^2=0.707;r_{m(test)}^2=0.488;Avg{r}_{m(test)}^2=0.575;\Delta r_{m(test)}^2=0.174;r{m}_{(overall)}^2=0.769;\Delta r{m}_{(overall)}^2=0.062;c{R}^{2}p=0.827\text{.}$

Model (1) and Model (2) both explain 70.3% and predict 55.4% and 63.7% of the variances of the U87MG and A549 cell line anti-proliferative activities, respectively, whereas Model (3) explains 88% and predicts 85% of the variance of the PC-3 cell line inhibition. The highest explanation and predictive ability are shown by Model (4), which can explains 91% and predict 89% of the variances of the HT-29 cell line antiproliferative activity. The VIF, t-value and p-value for Models (1)–(4) are shown in Table 2.

Table 2 VIF, t-value and p-value for Models (1)–(4).

Parameter	t-value	p-value	VIF
Model (1)
Intercept	−2.395	0.029	–
ATSC6e	5.055	0.000	1.057
VE1_Dt	−4.391	0.000	1.390
PetitjeanNumber	2.898	0.010	1.115
SpMin1_Bhs	2.539	0.021	1.326
Model (2)
Intercept	5.227	0.000	–
MATS8c	−6.401	0.000	1.456
JGI10	3.996	0.001	1.422
VR3_Dzp	2.883	0.012	1.543
VE3_Dzv	2.507	0.026	1.630
Model (3)
Intercept	6.539	0.000	–
AATS6e	−8.937	0.000	1.999
ASP-7	−3.046	0.012	2.189
SpMin6_Bhi	−2.911	0.015	1.655
Model (4)
Intercept	−2.166	0.045	–
AATSC8c	−11.275	0.000	1.992
VE1_Dzv	8.717	0.000	1.571
SpMin3_Bhs	2.376	0.030	1.499

The differences between the R² and Q² values of all models are less than 0.30. All models showed rm²_(LOO) values greater than 0.50 and have good predictive ability, as indicated by $R_{pred}^2$ values higher than 0.50. These four models demonstrated Avg $r_{m(test)}^2$ and rm²_(overall) values greater than 0.50. The validation metric values are within the acceptable limits, confirming the acceptable overall fitting of these models. The values found for the cR²p (model randomization) of all models are much greater than the threshold value of 0.50, ensuring that the models are not the outcome of mere chance alone. The statistical results are listed in Table 3. Additionally, the four models satisfied the acceptability criteria suggested by Golbraikh and Tropsha (Table 4). The observed, calculated, residual, predicted residual, and LOO-predicted activities of the training set compounds of Model (1)–(4) are shown in Supplementary Table T3. The observed and predicted values of the test set compounds of Models (1)–(4) are provided in Supplementary Table T4. The observed (Obs) vs. LOO-predicted (Pred) activities of the training set compounds for Models (1)–(4) are plotted graphically in Fig. 2. It is very interesting that compounds 11, 17, 23 have a potent broad spectrum of activities against all of the tested cancer cell lines. To check the significance of the developed QSAR models, the predicted biological activities of these compounds were compared with the observed activities, suggesting good agreement (Supplementary Table T5).

Fig. 2 Graphical representation of Observed (Obs) versus LOO-predicted (Pred) activities of training set compounds for (a) Model (1), (b) Model (2), (c) Model (3), (d) Model (4).

Table 3 Important statistical parameters of Models (1)–(4).

Model	nTrain	R	$R_A^2$	SEE	Q^2	PRESS	rm^2(LOO)	nTest	$R_{pred}^2$	$r_{m(test)}^2$	rm^2(overall)	cR^2p
1	21	0.873	0.703	0.212	0.554	1.353	0.510	7	0.866	0.723	0.566	0.638
2	18	0.879	0.703	0.118	0.637	0.290	0.590	6	0.858	0.637	0.697	0.687
3	14	0.953	0.880	0.078	0.850	0.0993	0.822	5	0.796	0.657	0.787	0.833
4	20	0.962	0.912	0.118	0.891	0.326	0.879	7	0.707	0.488	0.769	0.827

Table 4 Golbraikh and Tropsha acceptable model criteria of Models (1)–(4).

Sl. No.	Parameter	Threshold value	Model (1)	Model (2)	Model (3)	Model (4)
(i)	Q^2	Q^2 > 0.5	0.55422	0.63712	0.85033	0.8912
(ii)	r^2	r^2 > 0.6	0.90963	0.94474	0.84756	0.82154
(iii)	$|r_0^2-r_{0^{\prime }}^2|$	$|r_0^2-r_{0^{\prime }}^2|<0.3$	0.04583	0.21287	0.2568	0.08512
(iv)	k	0.85 < k < 1.15	1.00128	0.99564	1.00085	1.01793
(v)	k′	0.85 < k′ < 1.15	0.99798	1.00391	0.99842	0.98127

(i) Q² is the cross-validated correlation coefficient, (ii) Parameter r² is the squared correlation coefficient between the observed vs predicted response values of the test set compounds, (iii) $r_0^2$ and $r_{0^{\prime }}^2$ represent the corresponding values for regression through the origin (observed vs. predicted and predicted vs. observed, respectively), (iv) The slopes of the regression lines through the origin are assigned by k and k′.

The differences between the R² and Q² values of all models are less than 0.30. All models showed rm²_(LOO) values greater than 0.50 and have good predictive ability, as indicated by $R_{pred}^2$ values higher than 0.50. These four models demonstrated Avg $r_{m(test)}^2$ and rm²_(overall) values greater than 0.50. The validation metric values are within the acceptable limits, confirming the acceptable overall fitting of these models. The values found for the cR²p (model randomization) of all models are much greater than the threshold value of 0.50, ensuring that the models are not the outcome of mere chance alone. The statistical results are listed in Table 3. Additionally, the four models satisfied the acceptability criteria suggested by Golbraikh and Tropsha (Table 4). The observed, calculated, residual, predicted residual, and LOO-predicted activities of the training set compounds of Model (1)–(4) are shown in Supplementary Table T3. The observed and predicted values of the test set compounds of Models (1)–(4) are provided in Supplementary Table T4. The observed (Obs) vs. LOO-predicted (Pred) activities of the training set compounds for Models (1)–(4) are plotted graphically in Fig. 2. It is very interesting that compounds 11, 17, 23 have a potent broad spectrum of activities against all of the tested cancer cell lines. To check the significance of the developed QSAR models, the predicted biological activities of these compounds were compared with the observed activities, suggesting good agreement (Supplementary Table T5).

3.3 Descriptor Contributions in the QSAR models

3.3.1 Model 1

ATSC6e is defined as the centred Broto-Moreau autocorrelation – lag 6/weighted by the Sanderson electronegativities [33]. The electronegative substituent, e.g., F or Cl, decreases whereas electron-donating group, e.g., methoxy, increases the value of the coefficient ATSC6e. It shows a positive impact on the biological activity. Compound 23 has the highest ATSC6e value and has methoxy substitutions on both the A and B ring. Thus, this compound is the most potent against the U87MG cell line.

VE1_Dt represents the cocoefficient sum of the last eigenvector from the detour matrix [34]. The negative coefficient of VE1_Dt implies that with the increase of these values, the glioblastoma cell inhibitory activity of the pyrazolo-triazole hybrids may be decreased. The presence of this descriptor in the QSAR model suggests that three methoxy substitutions in the A-ring are unfavourable (compounds 25–30), whereas di- or tri-methoxy substitutions are preferable in the B ring.

SpMin1_Bhs and PetitjeanNumber have positive contributions towards the anti-proliferative activity of the analogues on cancerous U87MG cell lines. SpMin1_Bhs demonstrates the smallest absolute eigenvalue of the Burden modified matrix – n 1/weighted by the relative I-state. Fluorine/fluorine-containing substituents, e.g., CF₃, decrease the values of the coefficient and so are detrimental to the activity. The PetitjeanNumber is defined as the shape of the cation. The value of the shape parameter PetitjeanNumber increases when the substituents are changed from F or Cl to CF₃ or –OCH₃ on the B ring, which increases the activity.

3.3.2 Model 2

MATS8c negatively contributed towards the antiproliferative activity on A549 cell lines. It represents the Moran autocorrelation lag 8/weighted by charges [34]. Compound 17 has the lowest MATS8c value, and it shows the highest biological potency on the lung cancer cell line.

JGI10 is the mean topological charge index of order 10 [34], and it demonstrated a positive impact on the lung cancer cell line inhibition. The mean topological charge index can be increased by adding high-electronegativity compounds to the A ring of (1-benzyl-1H-1,2,3-triazol-4-yl)(1,3-diphenyl-1H-pyrazol-4-yl)methanone derivatives (compound 7, 10, 13, 15, 16, 21).

VR3_Dzp is defined as the logarithmic Randic-like eigenvector-based index from the Barysz matrix/weighted by polarizabilities, whereas VE3_Dzv represents the logarithmic coefficient sum of the last eigenvector from the Barysz matrix/weighted by the van der Waals volumes [34]. These two descriptors positively contributed towards the A549 cell line inhibition, which means that compounds having good VR3_Dzp and VE3_Dzv values will show promising antiproliferative activity.

3.3.3 Model 3

AATS6e is defined as the average Broto-Moreau autocorrelation – lag 6/weighted by the Sanderson electronegativities [34]. The negative coefficient of AATS6e implies that with a decrease in the value of AATS6e, the prostate cancer cell inhibitory activity of the pyrazolo-triazole hybrids will increase. The highest and lowest AATS6e values are shown by compound 7 (containing an electron-withdrawing group) and compound 23 (bearing an electron-donating group), respectively. Thus, electronegativity on the B-ring of the pyrazolo-triazole hybrid is detrimental to the PC-3 cell line inhibition.

ASP-7 (ChiPath Descriptor Average simple path, order 7) and SpMin6_Bhi (smallest absolute eigenvalue of Burden modified matrix – n 6/weighted by relative first ionization potential) negatively contributed towards the inhibition of prostate cancer cell lines [33]. The potential anticancer activity shown by compounds 4 and 5 had a low ionization potential compared with that of others.

3.3.4 Model 4

AATSC8c shows a negative impact on the HT-29 cell line antiproliferative activity. It denotes the average centred Broto-Moreau autocorrelation – lag 8/weighted by charges [34]. Substitution on the B ring of pyrazolo-triazole hybrids with trifluoromethyl may lead to a less active molecule against colon cancer cell lines (compounds 15, 27).

VE1_Dzv is the coefficient sum of the last eigenvector from the Barysz matrix/weighted by van der Waals volumes [34]. It positively contributed towards the inhibition of colon cancer cell lines.

SpMin3_Bhs is defined as the smallest absolute eigenvalue of the Burden modified matrix – n 3/weighted by the relative I-state [34]. The positive contribution shown by SpMin3_Bhs indicates that the biological potency can be increased by increasing the SpMin3_Bhs value. Upon the addition of an electronegative atom or an electron-withdrawing group to ring B of these hybrid derivatives, there was a dramatic loss in the antiproliferative activity of the HT-29 cell lines (compounds 1, 7, 13, 19, and 25).

4 Conclusions

In conclusion, very important indications emerge from our QSAR study of pyrazolo-triazole hybrids (Fig. 3). The QSAR models suggest the importance of descriptors, such as ATSC6e, VE1_Dt, PetitjeanNumber, SpMin1_Bhs, MATS8c, VR3_Dzp, ASP-7 SpMin6_Bhi, AATSC8c, VE1_Dzv and SpMin3_Bhs. The QSAR study revealed that the tri-methoxy substitution on the A-ring was unfavourable for the antiproliferative activity (negative contribution of VE1_Dt), whereas the positive contribution of JGI10 suggested that the addition of electronegative atoms on the A-ring of pyrazolo-triazole hybrids leads to potent anti-cancer compounds. In the B-ring, trifluoromethyl and strong electronegative group substitutions were not suitable (negative contribution of AATSC8c and positive contribution of SpMin3_Bhs, respectively), but 3,4-OCH₃ substitution into the pyrazolo-triazole hybrid compound was preferable (positive contribution of ATSC6e) for anticancer activity. Thus, an electronegative substituent at the p-position in the A ring and a 3,4-OCH₃ substituent in the B ring make compounds 11, 17, and 23 potent broad spectrum anticancer compounds against all of the tested cell lines. This study also presents a rationale for choosing a hybrid moiety for anti-cancer drug design and development.

Fig. 3 Structural requirements of pyrazolo-triazole hybrids for anti-cancer activity.

Acknowledgements

S. Gayen thanks the University Grants Commission (UGC), New Delhi for awarding a UGC-Start up grant (No.F.30-106/2015-BSR). One of the authors (S.A. Amin) is grateful to the All India Council for Technical Education (AICTE) New Delhi for awarding a Post Graduate GPAT fellowship. We also express our thanks for the support from the Department of Pharmaceutical Sciences, Dr. Harisingh Gour University, Sagar, India.

Notes

Peer review under responsibility of Taibah University.

Appendix A

Supplementary data

Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.jtusci.2016.04.009.