In silico design of EGFR^{L858R/T790M/C797S} inhibitors via 3D-QSAR, molecular docking, ADMET properties and molecular dynamics simulations

Abstract

The development of L858R/T790M/C797S mutations in EGFR is one of the main reasons for the emergence of resistance after third-generation treatment of non-small cell lung cancer (NSCLC). Therefore, the development of 4th generation drugs needs urgent attention. To overcome resistance, in silico drug discovery and Design approaches were employed on a library of 29 novel 9-heterocyclyl substituted 9H-purines derivatives with EGFR^{L858R/T790M/C797S} inhibition for anticancer activity against NSCLC. The COMSIA/EHA model (Q² = 0.584, R² = 0.816, and $R_{pred}^2$ = 0.73) showed a stable and reliable predictive ability of NSCLC activity, which was tested by several validation methods. Molecular docking studies reveal crucial interactions with EGFR^{L858R/T790M/C797S} inhibition for NSCLC activity. Based on theoretical methods, we designed 10 new compounds with good activity potential, which were tested using ADMET properties. Next, the molecular docking results were examined by molecular dynamics simulations to verify the stability of hydrogen bonding interactions with important residues such as MET790, MET793 and SER797, which are essential for the design of 4th generation EGFR Inhibitors to combat drug-resistant NSCLC.

Non-small cell lung cancer; 3D-QSAR; EGFR^{L858R/T790M/C797S} inhibition; ADMET; Molecular docking; Molecular dynamics simulation

1. Introduction

Lung cancer is responsible for the highest cancer mortality in men and women worldwide [1]. According to World Health Organization (WHO) estimates in 2020 [2], the number of lung cancer cases is expected to exceed 2.21 million and cause 1.80 million deaths worldwide. Thus, the cure rate for lung cancer is less than 20%, indicating that we need to strengthen our efforts to fight lung cancer. Non-small cell lung cancer (NSCLC) accounts for 85% of lung cancers [3]. The Epidermal Growth Factor Receptor (EGFR) is a receptor tyrosine kinase of the ErbB family, which plays a key role in the transmission of cell signalling that leads to proliferation, differentiation, apoptosis and migration [4], [5], [6]. The discovery of EGFR tyrosine kinase inhibitors (EGFR-TKIs) represents an important milestone in NSCLC treatment procedures [7]. However, mutations in the EGFR gene have had a considerable impact on these procedures, resulting in three generations of EGFR-TKIs.

First-generation EGFR-TKIs (Gefitinib and Erlotinib) have been shown to be effective in treating NSCLC in patients with the L858R mutation in EGFR [8], [9], [10]. However, the appearance of the T790M mutation in the EGFRT gene has led to the therapeutic failure of first-generation inhibitors [11]. To overcome the resistance to the T790M mutation, second-generation EGFR-TKIs, namely Afatinib and Dacomitinib, have been developed [12]. However, these inhibitors cause serious side effects [13]. To overcome the mutations and observed toxicity, the development of third-generation EGFR-TKIs (rociletinib, osimertinib) avoided these problems with acceptable side effects and improved action against the T790M mutation [14]. Despite the significant therapeutic success of third-generation EGFR-TKIs, the development of new resistance occurring on target is the C797S mutation, rendering osimertinib clinically ineffective [15]. Therefore, the discovery of new drugs to overcome the L858R/T790M/C797S mutants of the EGFR represents an urgent therapeutic need for the treatment of NSCLC.

In recent years, the use of molecular modeling techniques has yielded very impressive results in the new drug discovery process [16], [17], [18]. For this purpose, three-dimensional quantitative structure–activity relationship (3D-QSAR), molecular docking, pharmacokinetic parameters (ADMET) and Molecular dynamic (MD) simulation have been performed for the design of new molecule able to overcome the resistance of EGFR^{L858R/T790M/C797S} mutation. In this study, 29 EGFR^{L858R/T790M/C797S} inhibitors were studied using two methods, 3D-QSAR and molecular docking to identify key structural factors affecting inhibitory activity. In addition, it helps us to obtain recommendations for the design of new drug candidates from the ZINC database. To test their drug-like potential, all designed compound was evaluated by calculating the ADMET properties. Furthermore, Molecular dynamics simulations were performed for 100 ns to estimate the ligand stability within the protein under normal physiological conditions.

2. Material and methods

2.1. Data set collection and biological activities

In this work, a series of 29 novel 9-heterocyclyl substituted 9H-purines derivatives (Fig. 1) as EGFR-TKIs of the L858R/T790M/C797S mutants was collected from the recent literature [19]. Thus, the observed activity values (IC50) were converted to their negative logarithm pIC50 (−logIC50) values to build the 3D-QSAR models. To build construct the 3D-QSAR models (Table 1), the dataset containing 29 compounds was randomly divided into a training set containing 21 compounds to generate the 3D-QSAR models (Table 1), while the remaining 8 compounds 5 represented by the symbol “*” superscript were used as a test set to verify the performance and accuracy of the established models.

Figure 1.

Structures of 9-heterocyclyl substituted 9H-purines derivatives.

Table 1.

Observed and predicted anticancer activity against NSCLC of 9-heterocyclyl substituted 9H-purines derivatives.

Compounds	R1	R2	IC50 (μM)	pIC50 (obs)	pIC50 (pred)	Residual
1	H		0.11	6.958	6.726	0.231
2	F		0.58	6.236	6.124	0.112
3*	H		0.012	7.921	7.607	0.313
4	F		0.027	7.568	7.48	0.087
5*	F		0.12	6.921	7.222	−0.301
6	F		0.28	6.553	6.709	−0.156
7	H		0.53	6.276	6.919	−0.643
8	F		1.22	5.914	6.332	−0.418
9*	H		0.065	7.187	6.875	0.311
10	F		0.3	6.523	6.319	0.204
11*	F		0.038	7.421	6.98	0.441
12	F		0.12	6.921	6.784	0.136
13	H		0.17	6.769	6.221	0.547
14	H		0.26	6.585	7.068	−0.483
15*	H		0.017	7.769	7.696	0.072
16	H		0.0016	8.796	8.058	0.738
17	F		0.38	6.421	6.534	−0.113
18	F		1.52	5.818	5.638	0.179
19	F		1.81	5.742	5.722	0.019
20*	F		0.0075	8.125	7.892	0.232
21	H		0.00088	9.055	8.919	0.135
22	F		0.35	6.456	7.092	−0.635
23	F		0.0025	8.602	8.345	0.256
24	F		0.0018	8.745	8.029	0.715
25	F		0.0086	8.065	8.759	−0.694
26*	F		0.066	7.181	6.959	0.221
27	F		0.01	8	8.163	−0.163
28	F		0.0075	8.125	8.179	−0.054
29*	F		0.05	7.301	7.023	0.278

2.2. Structural minimization and alignment

Molecular alignment is one of the most important parameters of 3D-QSAR studies, as this step is a determining factor of the successful model [20]. All ligands were constructed with the sketch module and optimized under the Tripos force field, using Gasteiger-Hückel charges [21]. The convergence criterion of the Powell gradient algorithm was set to 0.005 kcal/(mol Å) and a maximum iteration coefficient of 10,000 to obtain a stable conformation [22]. All optimized molecular structures were aligned to the most active compound 21 of the series, using the ALIGN DATABASE method available in SYBYL-X 2.0 software.

2.3. 3D-QSAR studies

Ligand-guided alignment was used to derive the CoMSIA models using SYBYL-X 2.1. The aligned inhibitors of the training set were placed inside a 3D cubic lattice with a grid spacing of 2 Å in all Cartesians directions. The CoMSIA descriptors, which includes five force fields (steric, electrostatic, hydrophobic, hydrogen bond donors and acceptors), were calculated using a sp3 hybridized carbon probe atom with a van der Waals radius of 1.52 Å and a net charge of +1.0. An energy cut-off value was set at 30 kcal/mol. The minimum column filtering and attenuation factor were set to 2.0 and 0.3 kcal/mol, respectively.

2.3.1. Partial least squares analysis

The partial least squares (PLS) analysis was used to build the CoMSIA models [23]. First, the leave-one-out (LOO) cross-validation method was applied to obtain the optimal number of components (ONC) with the highest cross-validation coefficient of determination (Q²) to validate this step. Second, after determining the ONC, CoMSIA models were built based on statistical parameters such as the coefficient of determination (R²), the standard error of estimate (SEE), the F-value (Fischer test) and the contributions of each field. In order to test the predictive ability and stability of the models, several validation strategies were also applied.

2.4. 3D-QSAR models validation

2.4.1. Y-randomization

The Y randomization technique is a widely used approach to ensure the robustness of a 3D-QSAR model [24]. In this test, only the values of the dependent variable (biological activity) are randomly shuffled several times, while the values of the independent variable (descriptors) remain unchanged. These random data are then used to generate new 3D-QSAR models, the $Q_{yrand}^2$ and $R_{yrand}^2$ values of the random model are then tested against the Eriksson and Wold criteria [25] represented by the following rule:

• •$Q_{yrand}^2$ < 0.2 and $R_{yrand}^2$ < 0.2, no chance correlation
• •any $Q_{yrand}^2$ and 0.2 < $R_{yrand}^2$ < 0.3, negligible chance correlation
• •any $Q_{yrand}^2$and 0.3 < $R_{yrand}^2$ < 0.4, tolerable chance correlation
• •any $Q_{yrand}^2$and $R_{yrand}^2$ > 0.4, recognized chance correlation.

An additional criterion based on the parameter called ${}^{\mathrm{c}}R_r^2$ which must be greater than 0.5, was calculated according to the following equation:

$${}^{\mathrm{c}}R_r^2=R\times \sqrt{(R^2-R_r^2)}$$ (1)

2.4.2. External validation of the CoMSIA model

To verify the predictive capabilities of the proposed models, the biological activities of the test set of 8 compounds were predicted. The predictive ability of the models is measured by the external validation coefficient of determination ($R_{pred}^2$) calculated by the following formula:

$$R_{pred}^2=(SD-PRESS)/SD$$ (2)

The value of SD is the sum of the squared deviations between the biological activity of the compounds in the test set and the mean activity of the training set, and the predictive sum of squares (PRESS) is the sum of the squared deviations between the predicted and actual activities of the test set compounds. To test the reliability of 3D-QSAR models in predicting the activity of new compounds, external validation statistical criteria was introduced by Golbraikh and Tropsha [26] described by the following equations:

$${\mathrm{r}}_0^2=1-\frac{\sum {({\mathrm{Y}}_{\mathrm{test}\left(\mathrm{pred}\right)}-\mathrm{k}{\mathrm{Y}}_{\mathrm{test}\left(\mathrm{pred}\right)})}^2}{\sum {({\mathrm{Y}}_{\mathrm{test}\left(\mathrm{pred}\right)}-{\bar{\mathrm{Y}}}_{\mathrm{test}\left(\mathrm{pred}\right)})}^2}$$ (3)

$${\mathrm{r}}_0^{\prime 2}=1-\frac{\sum {({\mathrm{Y}}_{\mathrm{test}}-\mathrm{k}{\mathrm{Y}}_{\mathrm{test}})}^2}{\sum {({\mathrm{Y}}_{\mathrm{test}}-{\bar{\mathrm{Y}}}_{\mathrm{test}})}^2}$$ (4)

$$\mathrm{K}=\frac{\sum ({\mathrm{Y}}_{\mathrm{test}}-{\mathrm{Y}}_{\mathrm{test}(\mathrm{pred})})}{\sum {({\mathrm{Y}}_{\mathrm{test}(\mathrm{pred})})}^2}$$ (5)

$${\mathrm{K}}^{\prime }=\frac{\sum ({\mathrm{Y}}_{\mathrm{test}}-{\mathrm{Y}}_{\mathrm{test}(\mathrm{pred})})}{\sum {({\mathrm{Y}}_{\mathrm{test}})}^2}$$ (6)

Where ${\mathrm{r}}_0^2$ and ${\mathrm{r}}_0^{\prime 2}$ indicate the squared correlation coefficients through origin between predicted and observed activity, as well as observed and predicted activity of the test set, respectively. k and k′ are the slopes for regression through origin.

Furthermore, according to Roy's statistical criteria [25] to validate the predictive ability of the model, it is important to calculate the difference between the values of r², $r_0^2$ and $r_0^{\prime 2}$ according the following equation:

$$r_m^2=r^2(1-\sqrt{(r^2-r_0^2)}$$ (7)

$$r_m^{\prime 2}=r^2(1-\sqrt{(r^2-r_0^{\prime 2})}$$ (8)

2.4.3. Applicability domain

The applicability domain (AD) [27] has been applied to determine the region of chemical space of a 3D QSAR model, which can be used to reliably predict new compounds. In this study, one of the simple approaches to determine the DA is the leverage approach. The leverage values (h) were calculated for each molecule for a graphical detection of outliers in the Williams diagram by the following relationship:

$$h_i=(x_i^t{(X{X}^t)}^{-1}{x}_i)\text{with }(\mathrm{i}=1,2,\dots \mathrm{n})$$ (9)

In this equation: xi: the descriptor vector of the tested compound, X: the matrix n*(k − 1) where n is the number of compounds and k the number of descriptors in the training set and the exponent t refers to the matrix/vector transposition. This diagram validates the reliability of the 3D-QSAR model if the value of h is lower than the critical value of the leverage (h*) (h* = 2.5 (k + 1)/n) [28], [29], where n is the number of compounds and k is the number of descriptors in the training set. The AD of the 3D-QSAR model is located to the left of the vertical line of h* = 0.47.

2.5. Molecular docking investigation

Molecular docking is a reliable computational method, which can be performed to analyze the interactions between potential drugs and the active site of the target protein, in order to better understand the key structural requirements of a geometric model as a function of binding energy. In the present work, we performed the molecular docking study of two compounds 4 and 2 with the same R1 and R2 substitution, but with different R and S enantiomers, with compound 4 having higher activity than compound 2, in the active site of EGFR^{L858R/T790M/C797S} (PDB code 6lud), which is downloaded from the RCSB protein database [30]. Firstly, Discovery Studio software [31] was used to prepare the protein by removing all water molecules, ligands and non-protein parts. AutoDock Tools software version 1.5.6 was used to analyze ligand-protein interactions [32]. The 3D grid was created by the AUTOGRID algorithm to measure the energies of ligand-protein interactions [33]. The grid maps were generated using 60 Å in all Cartesian directions, with grid-spacing 0.375 Å, the grid center coordinates are approximately (−53.330 Å, −5.149 Å, and −17.579 Å) for the location of the ligand in the complex. The 2D and 3D visualizations and the analysis of the docked conformations according to the established interactions were performed using Discovery Studio [31].

2.5.1. Docking validation protocol

The molecular docking simulation was verified by re-docking the crystallized ligand (drug Osimertinib) into the protein (6lud.pdb). Therefore, the crystallised ligand was extracted from the protein and docked to the same protein. The lowest energy pose of the docked ligand and the crystallised ligand was superimposed, in order to calculate the root mean square deviation (RMSD). To validate the docking process, the RMSD must be less than 2 Å [34], [35], [36].

2.6. ADMET prediction studies

The development of computational technology in the pharmaceutical field has made it possible to identify new drug candidates, reducing the number of experimental tests and improving the success rate. With this in mind, the in silico study provides a pathway for ADMET (Absorption, Distribution, Metabolism, Excretion and Toxicity) pharmacokinetic parameters [37], [38], with drug absorption in the human gut, drug penetration into the central nervous system and the blood-brain barrier, metabolism is the chemical biotransformation of a drug by the body, excretion is the elimination of a drug by the body and the toxicity levels of a drug.

2.7. Molecular dynamic simulation studies

Molecular dynamics (MD) simulations were carried out using the NAMD (Nanoscale Molecular Dynamics) program [39]. We used the CHARMM-GUI server to generate the NAMD input files [40]. The CHARMM36 force field was applied to proteins and small molecules. Protein-ligand systems were solvated with the TIP3P water model in a 10 Å cubic box around and neutralised using the KCl salt at the 0.15 M ionic concentration [41]. The energy of all systems was minimised for 10,000 steps using the steepest descent method, and equilibrated under constant atom number, volume and temperature (NVT) at 310 K for 10 ns in an ensemble. Finally, the system was subjected to 100 ns of unconstrained MD simulations in a constant atomic number, pressure and temperature (NPT) ensemble with temperature (310 K) and pressure (1 atm) [42]. The analyses of the MD trajectories were used to generate the root mean square deviation (RMSD), root mean square fluctuation (RMSF) using Visual Molecular Dynamics (VMD) software [43] to check the stability of the systems.

3. Results and discussion

3.1. Molecular minimization and alignment

The compound-based molecular alignment method was performed using the available alignment rule to build a powerful 3D-QSAR model. The most active compound 20 of the series was chosen as the model molecule to align the data set and to visualise the CoMSIA model contour maps. Fig. 2 illustrates the minimisation and alignment of all 3D structures of the data sets to the common core based on compound 21.

Figure 2.

Alignment of all compounds.

3.2. CoMSIA studies

The statistical parameters of the CoMSIA models are presented in Table 2. For the CoMSIA analysis, the combination of the five molecular fields was used to develop the different CoMSIA models. However, the hydrogen bond donor field is zero for the molecules in the series. The results indicate that the best combination was the electrostatic field, the hydrophobic field and the hydrogen bond acceptor field (EHA), with The contribution rates were 24.7%, 11.7% and 63.6%, respectively. In the CoMSIA/EHA model, the Q² was 0.571 with two as the optimal number of components, The R² was 0.852 with a reliable SEE of 0.434 and the F-test value is 51.812, highest predictive value was obtained $R_{pred}^2$ = 0.715 for the external validation of the test set.

Table 2.

The PLS statistical results.

	Q2	R2	RMSE	F	ONC	$R_{pred}^2$	Fractions
							Ster	Elec	Hyd	Don	Acc
CoMSIA/SEA	0.584	0.844	0.445	48.687	2	0.603	0.121	0.216		-	0.663
CoMSIA/SEH	0.386	0.816	0.497	25.198	3	0.408	0.249	0.432	0.319	-
CoMSIA/SHA	0.579	0.849	0.437	50.773	3	0.676	0.170		0.132	-	0.698
CoMSIA/EHA	0.571	0.852	0.434	51.812	2	0.715		0.247	0.117	-	0.636
CoMSIA/SEHA	0.567	0.846	0.442	69.893	2	0.680	0.112	0.198	0.108	-	0.582

Overall, the proposed model is considered a reliable predictive model, if the Q² and R² values are greater than 0.5 and 0.6, respectively, with a prediction value $R_{pred}^2$ of the new compound activity greater than 0.6. Thus, the CoMSIA/EHA model indicates statistical significance and good predictive quality, which was confirmed by the predictive ability of the external validation. For more precision concerning the stability and prediction of the CoMSIA/EHA model, several validation methods such as the Y-randomization test, the Tropsha and Golbraikh criteria and Roy criteria were used. Table 3 presents the results of the Y-randomization test of the CoMSIA/EHA model.

Table 3.

Statistical parameters of the COMSIA/EHA model after several Y-Randomization tests.

Iteration	COMSIA/EHA
	$Q_{yrand}^2$	$R_{yrand}^2$	${}^{\mathrm{c}}R_r^2$
1	−0.22	0.340	0.66
2	−0.239	0.388	0.628
3	−0.176	0.330	0.666
4	0.060	0.360	0.647
5	−0.143	0.404	0.617

In the results in Table 3, five random mixtures of the Y vector were performed, according to the criteria of the Y-randomization test by Eq. (1) described above. The values of $Q_{yrand}^2$, $R_{yrand}^2$ and ${}^{\mathrm{c}}R_r^2$ indicate that the random correlation in the training set is tolerable. This revealed that the results obtained from the original CoMSIA/EHA model were not due to chance correlation. The results of the external validation test calculated by Eq. (3)-(8) with the Tropcha and Roy criteria for the CoMSIA/EHA model are listed in Table 4.

Table 4.

Statistical data from the external validation of CoMSIA/EHA model.

Parameter	Validation Criteria	CoMSIA/EHA
Q2	Q2 > 0.5	0.571
r2	r2 > 0.6	0.715
$|r_0^2-r_0^{\prime 2}|$	$|r_0^2-r_0^{\prime 2}|<0.3$	0.02
k	0.85 < k < 1.15	1.026
$\frac{r^2-r_0^2}{r^2}$	$\frac{r^2-r_0^2}{r^2}<0.1$	0.02
K'	0.85 < k′ < 1.15	0.973
$\frac{r^2-r_0^{\prime 2}}{r^2}$	$\frac{r^2-r_0^2}{r^2}<0.1$	0.084
$r_m^2$	$r_m^2>0.5$	0.628
$r_m^{\prime 2}$	$r_m^{\prime 2}>0.5$	0.581

The results in Table 4 reveal that the CoMSIA/EHA model is in perfect agreement with the Tropsha and Golbraikh as well as roy criteria. The CoMSIA/EHA model passed all validation tests, showing a better accuracy in predicting the activity of new compounds. Therefore, to determine the applicability domain of this model, we used William's graph presented in Fig. 3.

Figure 3.

The applicability domain of the CoMSIA/EHA model.

The DA of the CoMSIA/EHA model was assessed by a leverage analysis represented by a Williams diagram (Fig. 3). In the Williams diagrams, the results indicate that all leverage values of the training and test sets were below the critical leverage value (h* = 0.47), except for one outlier of compound 29, which was above the critical leverage, this compound belongs to the test set. The test set of the CoMSIA/EHA model was accurately predicted because there were no outliers for the training set. Therefore, we can reliably predict the anticancer activity of new compounds using this model. Thus, CoMSIA/EHA contour maps were used to analyse the structural requirements for the design of new active compounds.

3.3. CoMSIA model visualization

The CoMSIA/EHA model was used to visualise the three-dimensional equipotentiality map, using the compound 21 with the highest activity as a template. Fig. 4 shows the electrostatic, hydrophobic and hydrogen bond acceptor fields contour maps of the CoMSIA model.

Figure 4.

(a) Electrostatic, (b) hydrophobic and (c) Hydrogen bond acceptor for CoMSIA contour map analysis of compound 21.

In the CoMSIA electrostatic contour maps (Fig. 4(a)), we observed a red contour near the pyrrolidine and piperidine of the R₂ indicating an electronegative substitution can improve the activity. This is consistent with the fact that all compounds with electronegative substituents at the R₂ position show higher activity. Therefore, the presence of electronegative groups in R₂ substitutions could have better activity.

In the CoMSIA hydrophobic contour map (Fig. 4(b)), we observed a gray contour near the pyrrolidine and piperidine of the R₂ substitution, indicating that hydrophilic substitution is required in this region. This is consistent with the fact that compound 1-6 with tert-butyl formate as the hydrophilic group has a higher activity than compound 6-12, respectively. Thus, the presence of a hydrophilic group bound to pyrrolidines or piperidines can increase the biological activity.

In the hydrogen bond acceptor contour map (Fig. 4(c)), the red contours near the R₂ substitution indicate that the substitution of the hydrogen bond acceptor in this position is unfavourable. Thus, the presence of hydrogen bond acceptor groups decreased the biological activity, which is due to the nature of the receptor in this region which can be a hydrogen bond acceptor. This can be explained by the fact that compounds 24 and 28 with hydrogen bond donor groups show better activities, and with the fact that compounds with hydrogen substitution in the R₁ group are more active than compounds with fluorine substitution in the same position. However, a small magenta contour near the nitrogen atom of the pyrrolidine or piperidine substitution R₂ is also observed indicating that the hydrogen bond acceptor atoms are favourable in this position only. This can be explained by the fact that compound 21 shows the best activities of the series. Generally, hydrogen bond acceptor groups are unfavorable to inhibitory activity. From Table 2, we notice that the hydrogen bond acceptor field plays a key role in predicting anticancer activity. In the case of the CoMSIA/EHA model, the hydrogen bond acceptor field explains 63.6% of the variance, which explains why hydrogen bond acceptor groups are essential for the inhibition of this process.

3.4. Molecular docking analysis

The molecular docking study was conducted to obtain information on key structural requirements and to analyze the established interaction with EGFR^{L858R/T790M/C797S} protein. Fig. 5 presents the interaction modes obtained by molecular docking for compounds 2 (IC₅₀ = 0.58) and 3 (IC₅₀ = 0.012).

Figure 5.

3D and 2D presentation of the interactions of compounds 3 and 2 in the binding sites of EGFR^{L858R/T790M/C797S} protein. (a) Compound 3 (binding energy −9.52 kcal/mol). (b) Compound 2: (binding energy −9.96 kcal/mol).

Molecular docking interaction of compound 3 with EGFR^{L858R/T790M/C797S}, as observed in Fig. 5 (a), showed three hydrogen bonding interactions with MET790, MET793 and SER797 at distances of 3.07 Å, 1.65 Å and 2.52 Å, respectively. In addition, Pi-sulfur interactions are observed between the two purine rings with the amino acid MET792 (4.12 Å, 5.38 Å). However, compound 2 with EGFR^{L858R/T790M/C797S} (Fig. 5 (b)) forms only two hydrogen bonding interactions with the amino acid with MET790 (2.19 Å) and MET793 (1.79 Å) and one halogen bonding interaction with the amino acid ASP855 (2.69 Å), as well as Two Pi-sulfur interactions are observed between the two purine rings with the amino acid MET790 (4.35 Å, 5.86 Å).

Compounds 3 and 2 formed two hydrogen bond and pi-sulfur interactions with the same residues and position for both ligands. However, the more active compound 3 formed an additional hydrogen bonding interaction with the most important mutated residue, Ser797, in the EGFR binding region, which enhanced the inhibitory activity. Interestingly, the presence of these hydrogen bonds strengthened the binding of the compounds to the protein and allowed the compounds to have strong inhibitory activity.

According to Table 1, the fluorine substituent in the R₁ group decreased the inhibitory activity compared to the hydrogen substituent, this could be due to the formation of a halogen interaction with the fluorine substituent that blocks the ligand in the active site, thus decreasing the stability of the ligand-receptor complex. Furthermore, the CoMSIA/EHA hydrogen bond acceptor field contour map shows that ligands with hydrogen bond acceptor groups in are unfavourable for anticancer activity. This observation is fully consistent with the 3D docking results which clearly show that the EGFR^{L858R/T790M/C797S} protein is a hydrogen bond acceptor at this position, validating our hypothesis in the CoMSIA contour map section.

3.4.1. Molecular docking validation

In order to validate the ability of docking algorithms to predict the conformation of the ligand (Osimertinib) bound to the EGFR^{L858R/T790M/C797S} proteins, a self-docking of the crystal ligands was performed to test the accuracy of the docking procedure. Fig. 6 shows the superimposed view between the conformation of the docked ligand and native ligand, with an RMSD value of 0.869 Å less than 2 Å. To visualise the quality of the docking poses in the protein, a visual inspection of the interactions between the crystallized and docked ligand was performed (Fig. 7).

Figure 6.

Re-docking pose with RMSD value of 0.869 Å (Original = Green, Docked = blue).

Figure 7.

(a) 2D visualization showing the interactions of the docked ligand. (b) 2D visualization showing the interactions of the native ligand.

The results of the visual inspection show that we obtained the same interaction modes as in the case of the experimental interaction of the drug Osimertinib, observed in Fig. 7(a, b), this indicates a high reliability of the docking protocols to produce the binding mode of the novel EGFR^{L858R/T790M/C797S} inhibitors.

3.5. Drug molecular design

The main objective of this study is to design new EGFR protein inhibitors to overcome the L858R/T790M/C797S mutations. In this regard, new drug candidates are designed using recommendations from 3D-QSAR and molecular docking analysis (Fig. 8) on the structural characteristics of compound 21. In this study, ten 9H-purine derivatives (N₁-N₁₀) were designed to improve anticancer activity, the proposed substitutions were taken from the ZINC database. Thus, these newly designed compounds were aligned using compound 21 as a template. The previously established CoMSIA/EHA model and molecular docking with EGFR^{L858R/T790M/C797S} protein were used to predict the activity of these new compounds. The structure of the newly designed molecules, the predicted pIC50 and IC50 values, the calculation of the leverage threshold h* as well as the molecular docking interactions are presented in Table 5.

Figure 8.

Summary of Structural requirements based on CoMSIA/EHA contour map and molecular docking analyses.

Table 5.

The leverage threshold h* and predicted IC₅₀ based on CoMSIA/EHA model and Molecular Docking interactions for the newly designed compounds.

Compounds	Structures of newly designed molecules	PIC50 (pred)	IC50 (pred)	leverage threshold h*	Interactions with EGFRL858R/T790M/C797S protein
					Binding affinity (kcal/mol)	Number of Hydrogen bonds	Amino acid residues with hydrogen bonding (distances)
N1		7.5432	0.0286	0.8065	−10.3	2	MET793 (1.81 Å, 2.54 Å)
N2		9.1590	0.0007	0.3717	−10.55	4	SER797 (2.35 Å, 2.73 Å) MET793 (1.71 Å) MET790 (2.57 Å)
N3		8.8819	0.0013	0.3486	−9.17	4	ASP800 (2.92 Å) SER797 (2.27 Å) MET793 (1.81 Å) LEU718 (2.19 Å)
N4		7.0516	0.0887	0.4616	−9.95	2	GLU804 (2.31 Å) SER797 (2.41 Å)
N5		8.6883	0.002	0.4160	−9.3	3	SER797 (2.22Å) MET793 (2.04 Å, 2.71 Å)
N6		7.0456	0.09	0.2360	−10.69	2	MET790 (2.73 Å) MET793 (1.78 Å)
N7		7.2338	0.058	0.4589	−10.57	2	SER797 (2.35 Å) Met793 (1.91 Å)
N8		8.0158	0.0096	0.1332	−9.3	3	SER797 (2.38Å) MET793 (2.05 Å, 2.53 Å)
N9		7.5243	0.0299	0.2335	−9.2	3	SER797 (1.96 Å) MET793 (2.31 Å, 2.49 Å)
N10		7.3045	0.0496	0.6577	−10.38	5	SER797 (2.24 Å, 2.90 Å) MET793 (2.21 Å, 2.16 Å) MET790 (2.37 Å)

The first step of a reliable prediction is to verify that all predicted new molecules belong to the application domain of the proposed model. The CoMSIA/EHA model has a critical leverage value h* = 0.47, all the designed novel compounds have a threshold leverage value h* lower than the critical leverage value, except for compounds N1 and N10 which could belong to another chemical family, this result shows that N2-N9 compounds were reliably predicted anticancer activity. Using the CoMSIA/EHA model, the compound T2 also showed better activity than all compounds in the data set. Furthermore, we carried out molecular docking for ten new compounds designed with the EGFR^{L858R/T790M/C797S} protein. The results indicate that almost all compounds formed hydrogen bonding interactions with the amino acids at positions 797 and 790 of the protein, both of which play an important role in the therapeutic failure of the anticancer drug. However, the inhibitory activity of compounds N1 and N10 cannot be reliably predicted, but the molecular docking results are encouraging and may lead to a new inhibitor family. Overall, all compounds show a good level of inhibitory activity and can theoretically overcome the problem of drug resistance in lung cancer.

3.6. ADMET prediction

To verify that the designed compounds can become drugs, we use the ADMET pharmacokinetic parameters by the online tool pkCSM [44]. The ADMET parameters of the newly designed compounds are listed in Table 6.

Table 6.

In silico ADMET prediction of newly designed inhibitors.

Compounds	Absorption	Distribution			Metabolism							Excretion	Toxicity
	Intestinal absorption (human)	VDss (human)	BBB permeability	CNS permeability	Substrate		Inhib itor					Total Clearance	AMES toxicity
					CYP
					2D6	3A4	1A2	2C19	2C9	2D6	3A4
	Numeric (% Absorbed)	Numeric (Log L/kg)	Numeric (Log BB)	Numeric (Log PS)	Categorical (Yes/No)							Numeric (Log ml/min/kg)	Categorical (Yes/No)
N1	90.063	0.556	−1.608	−2.823	No	No	No	Yes	Yes	No	Yes	0.758	No
N2	87.61	0.623	−1.43	−2.525	No	Yes	No	Yes	Yes	No	Yes	0.229	No
N3	67.671	0.432	−1.492	−3.755	No	No	No	No	Yes	No	Yes	0.485	Yes
N4	78.808	0.585	−1.709	−2.826	No	No	No	Yes	Yes	No	Yes	0.221	Yes
N5	82.792	−0.197	−1.252	−2.146	No	Yes	Yes	Yes	Yes	No	Yes	0.302	Yes
N6	97.823	−0.039	−1.485	−3.391	No	Yes	No	Yes	Yes	No	Yes	0.638	No
N7	78.764	0.036	−1.775	−3.944	No	No	No	No	Yes	No	Yes	0.149	Yes
N8	81.253	0.058	−1.917	−4.339	No	No	No	No	Yes	No	Yes	0.089	No
N9	90.08	0.063	−1.697	−4.004	No	No	No	No	Yes	No	Yes	0.328	No
N10	94.626	0.079	−1.489	−3.844	No	No	No	Yes	Yes	No	Yes	0.5	No

An absorption value of less than 30% is considered poor intestinal absorption, the ten compounds designed showed a value between (67.671% and 97.823%) indicating good intestinal absorption, volume of distribution (VDss) is considered low if logVDss < −0.15 and high if logVDss > 0.45, central nervous system (CNS) and blood-brain barrier (BBB) permeability standard values (>−2 to <−3 LogPS and >0.3 to <−1 Log BB), respectively, for a given compound a LogBB < −1 corresponds to poor distribution to the brain, while LogBB >0.3 are likely to cross the BBB and LogPS >3, to cross the BBB and LogPS > −2 considered to penetrate the CNS, while LogPS < −3 are difficult to move into the CNS [45]. Thus, the compounds N1, N2, N6 and N8-N10 have an excellent potential to cross the barriers.

The enzymatic metabolism refers to the chemical biotransformation of drugs in the human body, which plays a crucial role in the metabolic stability of drugs in the body [46]. The cytochrome P450 enzymes (CYP1A2, CYP3A4, CYP2C19, CYP2D6 and CYP2C9) found in the liver are the main enzymes of drug metabolism, being responsible for the biotransformation of more than 90% of drugs. Inhibition of these metabolising enzymes can increase the concentration of the active drug in the body. In this study, CYP3A4 was the main human enzyme responsible for the metabolism of the third-generation drug for treating NSCLC [47], [48], [49]. The results show that all newly designed compounds appear to be CYP3A4 inhibitors, but only compounds N2, N5 and N6 appear to be CYP3A4 substrates. All newly designed compounds showed a low total clearance value, which means accumulation and persistence of the drugs in the body. Finally, compounds N1, N2, N6 and N8-N10 showed negative toxicity. Overall, the newly designed compounds N1, N2, N6 and N8-N10 exhibit good pharmacokinetic properties. The results of this study may represent excellent drug candidates for the treatment of NSCLC to overcome the L858R/T790M/C797S mutations in EGFR-TKIs. Fig. 9 shows the 2D visualization of the molecular docking results of the best predicted compounds, which were then subjected to a molecular dynamics study

Figure 9.

2D interactions of the best ligands N1, N2, N6 and N8-N10 in the binding sites of EGFR^{L858R/T790M/C797S} protein.

3.7. Molecular dynamic simulation results

After performing molecular docking studies and ADMET properties of the predicted compounds, MD simulation of the best candidates N1, N2, N6 and N8-N10 was carried out, with RMSD and RMSF parameters to analyse the stability of the target protein. The RMSD and RMSF plots of the EGFR^{L858R/T790M/C797S} complex by Osimertinib and with the best predicted compounds are shown in Fig. 10 (a) and (b), respectively.

Figure 10.

a) The RMSD values of the EGFR^{L858R/T790M/C797S} protein in complex with Osimertinib and six best ligands at 100 ns, and b) The RMSF values of the EGFR^{L858R/T790M/C797S} protein residues in complex with Osimertinib and six best.

According to the RMSD plot analysis, all systems experienced a rapid increase in RMSD values from 0.53 Å to 1.2 Å within 40 ns, which is largely due to the initial kinetic shock experienced by all systems in the molecular dynamics studies. Subsequently, the RMSD values of all systems fluctuated within a similar distance range of 1.2 Å and 1.5 Å, indicating that all systems reached stability and equilibrium. The most stable complexation was the experimental crystallographic structure of the EGFR^{L858R/T790M/C797S} protein in combination with the drug Osimertinib (6lud) obtained from the RCSB protein database with an RMSD value of 1.082 Å. Osimertinib serves as the reference ligand in this study. From the RMSD plot, it can be seen that all the proposed ligands have a similar trajectory to the reference ligand with a slight difference, the RMSD values of ligands N1, N2, N6, N8, N9 and N10 complexed with EGFR^{L858R/T790M/C797S} were 1.231 Å, 1.137 Å, 1.173 Å, 1.169 Å, 1.159 Å and 1.251 Å, respectively. All ligands had a value of about 1.3 Å and the N2 ligand had the best stability. A study by Beura et al. [50] showed that an RMSD value of less than 3 Å is an indicator of the conformational stability of protein-ligand complexes. Therefore, the docking results of all predicted ligand complexations with the EGFR^{L858R/T790M/C797S} show higher conformational stability of ligand-receptor complexes.

The RMSF trajectories assess the effect of ligand binding on the flexibility of the protein during the MD simulation, which shows the stability of the protein per residue during the simulation. The higher the RMSF value, the more flexible the residue, while the lower the RMSF value, the more stable the residue. In general, the majority of the residuals have the same RMSF values, with higher fluctuations in the different ranges, such as LYS754 (1.35 Å), LEU782 (1.14 Å), GLY874 (1.20 Å), SER921 (1.12 Å) and GLU1005 (1.11 Å), these residues are not involved as they are located in the inactive regions of the EGFR^{L858R/T790M/C797S} protein. However, crucial residues in the active site such as LEU718, ASP800, GLU804, MET790, MET793, SER797 and ARG858 show smaller RMSF fluctuations of less than 0.4 Å, which could be related to the generation of more hydrogen bonding interactions for greater stability of the ligands with the EGFR^{L858R/T790M/C797S} protein. These results confirm the RMSD analysis that all predicted candidates with the EGFR^{L858R/T790M/C797S} protein show greater conformational stability.

4. Conclusion

In summary, 3D-QSAR, docking, ADMET and MD simulation methods were performed for the design of new drug candidates capable of overcoming drug resistance in third generation NSCLC. In the 3D-QSAR study, the best selected model (CoMSIA/EHA) has high stability and predictive ability, which were assessed using external validation, Y-randomisation test and applicability domain. The CoMSIA/SEHA contour map analyses provided a better understanding of the relationship between structure and activity, which was a progression in guiding the design of new potent compounds. The molecular docking analysis shows the importance of hydrogen bonds established with key residues, which also confirmed the importance of residues such as MET790, MET793 and SER797 for the active site of EGFR protein. Based on the precise recommendation provided by 3D-QSAR and molecular docking analysis, 10 new compounds with considerable activity were designed using the virtual zinc base. The ADMET properties were used to select the best pharmacokinetic profile of the proposed compounds. Finally, MD simulations verified the accuracy of the molecular docking results in terms of reliability and stability, in which essential residues formed hydrogens between the proposed compounds and the EGFR protein. The newly designed compounds N1, N2, N6 and N8-N10 could be good drug candidates to overcome resistance to third generation drugs against NSCLC.

Declarations

Author contribution statement

Hanine Hadni: Conceived and designed the experiments; Analyzed and interpreted the data; Wrote the paper.

Menana Elhallaoui: Contributed reagents, materials, analysis tools or data; Wrote the paper.

Funding statement

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability statement

Data will be made available on request.

Declaration of interests statement

The authors declare no conflict of interest.

Additional information

No additional information is available for this paper.