2,4-disubstituted 6-fluoroquinolines as potent antiplasmodial agents: QSAR, homology modeling, molecular docking and ADMET studies

Abstract

Objective

This work was designed to study 2,4-disubstituted 6-fluoroquinolines as antiplasmodial agents by using in silico techniques, to aid in the design of novel analogs with high potency against malaria and high inhibition of Plasmodium falciparum translation elongation factor 2 (PfeEF2), a novel drug target.

Methods

Quantitative structure-activity relationships (QSAR) of 2,4-disubstituted 6-fluoroquinolines were studied with the genetic function approximation technique in Material Studio software. The 3D structure of PfeEF2 was modeled in the SWISS-MODEL workspace through homology modeling. A molecular docking study of the modeled PfeEF2 and 2,4-disubstituted 6-fluoroquinolines was conducted with Autodock Vina in Pyrx software. Furthermore, the in silico pharmacokinetic properties of selected compounds were investigated.

Results

A robust, reliable and predictive QSAR model was developed that related the chemical structures of 2,4-disubstituted 6-fluoroquinolines to their antiplasmodium activities. The model had an internal squared correlation coefficient R² of 0.921, adjusted squared correlation coefficient R²_adj of 0.878, leave-one-out cross-validation coefficient Q²_cv of 0.801 and predictive squared correlation coefficient R²_pred of 0.901. The antiplasmodium activity of 6-fluoroquinolines was found to depend on the n5Ring, GGI9, TDB7u, TDB8u and RDF75i physicochemical properties: n5Ring, TDB8u and RDF75i were positively associated, whereas GGI9 and TDB7u were negatively associated, with the antiplasmodium activity of the compounds. Stable complexes formed between the compounds and modeled PfeEF2, with binding affinity ranging from −8.200 to −10.700 kcal/mol. Compounds 5, 11, 16, 22 and 24 had better binding affinities than quinoline-4-carboxamide (DDD107498), as well as good pharmacokinetic properties, and therefore may be better inhibitors of this novel target.

Conclusion

QSAR and docking studies provided insight into designing novel 2,4-disubstituted 6-fluoroquinolines with high antiplasmodial activity and good structural properties for inhibiting a novel antimalarial drug target.

Introduction

Malaria poses a great danger to public health. This disease is caused by the Plasmodium parasite, which is transmitted between humans by female Anopheles mosquitoes. Five species of the parasite are pathogenic to humans, among which Plasmodium falciparum and Plasmodium vivax are the most threatening.^1,2 Malaria caused an estimated 619,000 deaths in 2021, of which 76% (∼470,000) were in children younger than 5 years,³ amounting to a child dying from the disease nearly every minute. The highest malaria burden is in sub-Saharan Africa, which had an estimated 234 million cases and 593,000 deaths in 2021. Nigeria has 26.6% and 31.3% of the global malaria cases and deaths, respectively.³ Prevention of malaria through vector control is challenging, because of increasing mosquito resistance to the most commonly used insecticides (pyrethroids) in insecticide-treated nets, and increasing spread of an urban-adapted mosquito species (Anopheles stephensi).³

Treatment through chemotherapy is challenging because of rising resistance to many antimalarial drugs and artemisinin-based combination therapy, a recommended therapy for P. falciparum (the most deadly parasite).³ Therefore, new potent antiplasmodial drugs with novel mechanisms of action must be developed.⁴ Quinoline-4-carboxamide (DDD107498) was discovered as a potent antimalarial compound active against multiple life-cycle stages of the parasite. The molecular target of this compound is translation elongation factor 2 (eEF2). However, the interaction of DDD107498 with the target (PfeEF2) is not well understood.⁴ Hochegger et al. (2019) synthesized new analogs of quinoline-4-carboxamide to improve the antiplasmodial activity (in vitro and in vivo) and to understand the structure-activity-relationships of these analogs.⁴ In silico techniques are computer-aided modeling methods used in screening chemical databases to identify novel drug candidate.⁵ This work was aimed at conducting a quantitative structure-activity relationships (QSAR) study of quinoline-4-carboxamide analogs (6-fluoroquinolines) to build a model relating the antiplasmodium activity to the physicochemical properties of 6-fluoroquinolines, to increase understanding of the structure-activity-relationships of the compounds. We additionally performed a molecular docking study of 6-fluoroquinolines with homology modeled P. falciparum translation elongation factor 2 (PfeEF2) as the molecular target, to understand the modes of interaction of the compounds with the potential target. This information may be used to design better inhibitors of the novel target as antimalarial drug candidates.

Materials and Methods

Data collection

A dataset of 28 compounds of 6-fluoroquinoline derivatives and their in vitro activities against the chloroquine-sensitive strain NF54 of P. falciparum was obtained from the literature.⁴ The antiplasmodial activities of the 6-fluoroquinolines were obtained as IC₅₀ (nM) and converted to pIC₅₀ {−logIC₅₀ (M)} to normalize the distribution of the values for QSAR building.⁵ The structures and names of the 6-fluoroquinolines and their respective activities (pIC₅₀) are presented in Table 1.

Table 1.

Antiplasmodial activity and leverage of the 6-fluoroquinolines.

C/N	Compound	Experimental pIC50	Theoretical pIC50	Residual	Leverage
1a		5.132	5.403	−0.271	0.720
2		5.057	5.254	−0.197	0.281
3		5.223	5.111	0.111	0.190
4a		5.335	4.649	0.687	0.419
5		5.252	4.887	0.365	0.205
6a		4.827	5.808	−0.980	0.679
7		4.424	4.370	0.053	0.511
8a		4.194	4.014	0.179	0.586
9		4.996	5.360	−0.364	0.120
10		5.712	5.782	−0.070	0.107
11a		9.678	9.294	0.384	0.786
12		5.507	5.978	−0.471	0.292
13		8.301	8.275	0.026	0.528
14		8.959	8.720	0.239	0.336
15a		7.699	7.176	0.523	0.563
16		7.745	7.858	−0.114	0.658
17		5.147	5.504	−0.357	0.137
18a		6.004	5.676	0.329	0.432
19a		6.638	6.181	0.457	0.815
20		5.983	7.138	−1.155	0.365
21		5.123	5.017	0.106	0.410
22		5.019	4.620	0.399	0.462
23		8.398	8.244	0.154	0.282
24		6.215	5.570	0.645	0.367
25		6.585	6.505	0.080	0.125
26		6.276	6.489	−0.214	0.356
27a		7.056	7.172	−0.117	1.000
28		8.699	7.936	0.763	0.267

^aTest set, C/N = compound number.

Generation of molecular descriptors and pretreatment

The molecular structures of the 6-fluoroquinolines (Table 1) were drawn with Chemdraw version 12.0.2 software, and their equilibrium geometries were obtained in Spartan 14 software by optimization with the parametric semi-empirical (PM6) quantum mechanics method.^6,7 The molecular descriptors of the optimized 6-fluoroquinolines were generated with PaDEL-Descriptor software version 2.20.⁸ After redundant and highly correlated descriptors were removed, normalization with Eq. (1) was performed to give each descriptor an equal chance of appearing in the model.⁹ These steps were achieved with Drug Theoretical and Cheminformatics Laboratory (DTC Lab) pretreatment and normalization software and the following equation:

$$X_{ni}=\frac{X_i-X_{\mathit{min}}}{X_{\mathit{max}}-X_{\mathit{min}}}$$ (1)

where X_ni and X_i are the normalized and unnormalized descriptor values for molecule i for a particular descriptor, and X_min and X_max are the minimum and maximum values for the descriptor.

Model building and validation

The Kennard-Stone algorithm in DTC Lab Data Division software was used to divide the data into two sets.¹⁰ This algorithm selects training set compounds from the data set by first selecting two compounds separate from each, other on the basis of Euclidean distance, and including them in the training set. Sequentially, the algorithm removes compounds from the dataset and includes them in the training set to maximize the Euclidean distance between the x-vectors of the already selected compound and the remaining compounds in the dataset. This process is repeated until the specified number of training set compounds is selected.¹¹ This algorithm has the advantages of selecting training set compounds that are uniformly distributed along the data space and test set compounds that fall within the measured space.¹⁰

A total of 70% of the data (training set) was used in building the model, and 30% of the data (test set) was used to validate the model. The activities (pIC₅₀) of the training set compounds were used as the dependent variable and their descriptors served as independent variables in regression analysis to build the model with the genetic function approximation (GFA) technique in Material Studio software version 8.0.¹² GFA uses a genetic algorithm to identify the best model among possible QSAR models. It automatically selects group of descriptors at random, according to the user-specified number, and uses them to build regression models, then assesses the models with the Friedman function (LOF), a measure of model fitness expressed in Eq. (2).¹³ Many models are built on the basis of the user-specified number of generations, and the best model is the one with the lowest LOF score¹⁴ calculated as follows:

$$LOF=\frac{SEE}{{\left(1-\frac{c+dp}{M}\right)}^2}$$ (2)

where p is the total number of descriptors in the model, c is the number of terms in the model, M is the number of compounds in the training set, d is the user-defined smoothing parameter, and SEE is the standard error of estimation, which is the same as the standard deviation of the model, defined as;

$$SEE=\sqrt{\frac{{\left(Y_{\exp -Y_{prd}}\right)}^2}{N-P-1}}$$ (3)

The built model was validated with the squared correlation coefficient R²; adjusted correlation coefficient R²_adj; cross-validation coefficient Q²_cv; and external validation coefficient R²_pred, defined by Eqs. (4), (5), (6), (7), respectively.¹⁵

$$R^2=\frac{{\left\{\sum \left(Y_{exp}-{ӯ}_{exp}\right)\left(Y_{prd}-{ӯ}_{prd}\right)\right\}}^2}{{\sum {\left(Y_{exp}-{ӯ}_{exp}\right)}^2\left(Y_{prd}-{ӯ}_{prd}\right)}^2}$$ (4)

$$R_{adj}^2=\frac{\left(n-1\right)\left(R^2-p\right)}{N-p-1}$$ (5)

$$Q_{cv}^2=1-\frac{{\sum \left(Y_{prd}-Y_{\mathit{exp}}\right)}^2}{{\sum \left(Y_{\mathit{exp}}-{ӯ}_{prd}\right)}^2}$$ (6)

$$R_{pred}^2=1-\frac{\sum {\left(Y_{prd}-Y_{\mathit{exp}}\right)}^2}{{\left(Y_{\mathit{exp}}-{\mathrm{ӯ}}_{prd}\right)}^2}$$ (7)

Y_exp and Y_prd are the experimental and predicted activity of the training set compounds, respectively, in Eqs. (4), (6), and the experimental and predicted activity of the test set compounds in Eq. (7). ӯ_prd is the mean experimental activity of the training set compounds. N and p in Eqs. (3), (5) are the number of molecules in the training set and number of descriptors in the model, respectively.

To validate the reliability of the developed model, we computed the randomization parameters R²_r, Q²_r and cR²_p, which were robust and not obtained by chance. The training set was used to generate random multi-linear regression models through random shuffling of the activity of the compounds (dependent variables) while keeping their descriptors (independent variables) stable. R²_r and Q²_r were computed as the average of the squared correlation coefficient and cross-validation coefficient of the random models. The coefficient of determination cR²_p was computed with Eq. (8):¹⁶

$${{cR}_p^2=R^{2}x\left(R^2-R_r^2\right)}^2$$ (8)

where R² is the squared correlation coefficient for the nonrandomized model, and R²_r is the average of the squared correlation coefficients of the random models.

Descriptor analyses

Inter-correlation among the descriptors in the built model was verified by correlation analysis of the descriptors. The variance inflation factor (VIF) for each descriptor was computed with Eq. (9) to further confirm their inter-correlation:¹⁷

$${\text{VIF}}_{\mathrm{i}}=\frac{1}{1-R_{ij}^2}$$ (9)

where VIF_i is the variance inflation factor for a descriptor i in the model, and R²_ij is the correlation coefficient of the multiple regression between descriptor i and the remaining j descriptors in the model.

The mean effect (ME) for each descriptor in the model was computed with Eq. (10) to evaluate the relative influence of the descriptors in the model:¹⁸

$${\text{ME}}_{\mathrm{j}}=\frac{{\beta }_j{\sum }_{i=1}^{i=n}{d}_{ij}}{{{\sum }_j^m\beta }_j{\sum }_i^n{d}_{ij}}$$ (10)

where ME_j is the mean effect for descriptor j in a model, d_ij is the value of descriptor j in the descriptor matrix for each molecule in the training set, β_j is the coefficient of descriptor j in the model, m is the number of descriptors in the model, and n is the number of molecules in the training set.

Applicability domain of the model

The applicability domain, described by Williams's plot of the built model, was generated with the leverage method (Eq. (11)) to identify outliers and influential compounds in the dataset.¹⁹

$${h=X\left(X^{T}X\right)}^{-1}{X}^T$$ (11)

where X is the descriptor matrix, and X^T is the transpose matrix of X. The leverages of the compounds are the diagonal of the matrix h. The warning leverage h^∗ is the maximum value above which a compound is considered to be influential, and is expressed as:

$$h^{\ast }=\frac{3\left(p+1\right)}{n}$$ (12)

where n is the number of compounds in the training set, and p is the number of descriptors in the model.

Homology modeling

The crystal structure of PfeEF2 has not been elucidated. Therefore, we used comparative modeling to build the 3D structure of PfeEF2 and subsequently performed a docking study. The protein sequence of eEF2 for P. falciparum (isolate NF54), obtained from UniProtKB (http://www.uniprot.org) accession code W7JNW7, comprised 832 amino acids.²⁰ The code was submitted to the SWISS-MODEL workspace (https://www.swissmodel.expasy.org) to search for evolutionarily related structures matching the target sequence.²¹ Suitable target-template alignments were identified with the Basic Local Alignment Search Tool (BLAST) and hidden Markov models (HMMs) with an HMM-HMM–based lightning-fast iterative sequence search (HHblits).^22,23 The 3D structure of PfeEF2 was built with the highest ranked target-template alignment in ProMod3 version 3.2.1.²⁴ The built structure was assessed with the qualitative model energy analysis (QMEAN) scoring function and global model quality estimation.^24,25

Molecular docking

The modeled PfeEF2 was saved as a PDB file and used as receptor for docking with 6-fluoroquinolines as ligands. The ligands were prepared by saving their structures as PDB files from Spartan software, and the receptor was prepared by removal of co-crystallized ligands and hetero-atoms in Discovery Studio software.²⁶ Both the ligands and the receptor were converted to PDBQT files and docked with Autodock Vina in Pyrx software with a grid box dimension of 75.7994 Å × 100.6790 Å × 118.096 Å, and centers of 75.9456, 38.8172 and −1.2721 (X, Y and Z coordinates, respectively), to cover the complete surface of the protein.^27,28 The interactions in docked structures were visualized with Discovery Studio Visualizer.²⁹

In silico drug-likeness and ADMET prediction

In silico prediction of drug-likeness, and adsorption, distribution, metabolism, excretion and toxicity (ADMET) of compounds with excellent binding affinity toward PfeEF2 was conducted with the SwissADME and ADMETlab 2.0 online platforms.^30,31 These predictions are essential for evaluating the potential of a potent molecule to have effective pharmacokinetics and toxicity.³²

Results

The best QSAR model developed on the basis of GFA that related the chemical structures of the 6-fluoroquinolines to their antiplasmodial activities is presented below:

pIC₅₀ = 2.428645705n5Ring − 4.587432351GGI9 − 7.127111879TDB7u + 7.951073945TDB8u + 0.078997316RDF75i − 8.731499618

Discussion

GFA was deployed to generate QSAR models relating the physicochemical properties of 6-fluoroquinolines with substitutions at ring positions 2 and 4 to their antiplasmodial activities. The model that best predicted the antiplasmodial activity of the compounds is reported herein. The model surpassed all validation parameters (Table 2) for good prediction, as indicated by the low residual (difference between experimental and theoretical activity) values of the compounds (Table 1). The linearity of the plots of the model's predicted and experimental activity (Figure 1a), and the difference in R² and Q²_cv of <0.3, further supported the model's predictive ability.²⁶ The R² values for the random models generated were all below the minimum value (0.6) for an acceptable model, thus indicating that the main model was not a product of chance. This finding was further confirmed by the average values of R²_r, Q²_r and cR²_p (Table 3).^15,16 Hence, the built model was considered robust, reliable and stable. The distribution of the compounds on opposite sides of the line 0 standardized residual (Figure 1b) indicated the absence of systematic error in model building.

Table 2.

Validation parameters of the model.

Parameter	Threshold	Model value	RemarkR
R2	R2 > 0.6	0.912	Passed5
R2adj	R2adj > 0.5	0.878	Passed5
F	Large	27.001	Passed5
Q2cv	Q2cv > 0.5	0.801	Passed5
|R2 − Q2cv|	|R2 − Q2cv| < 0.3	0.111	Passed33
SEE	Low	0.505	Passed34
R2pred	R2pred > 0.5	0.901	Passed5
r2	r2 > 0.6	0.908	Passed35
r02		0.908
r′02		0.898
|r02 − −r'02|	|r02 − −r'02| < 0.3	0.010	Passed35
K	0.85 < k < 1.15	1.022	Passed35
(r2 − −r02)/r2	(r2 − −r02)/r2 < 0.1	0.000	Passed35

^RNote: r² and r₀² are the squared correlation coefficients of the plot of the experimental versus predicted pIC₅₀ of the test set compounds with and without intercepts, respectively, and k is the gradient of the plot with intercept. r'₀² is the reverse of r₀².

Figure 1.

(a) Plot of predicted versus experimental pIC₅₀ of the compounds. (b) Plot of standardized residual activity against experimental pIC₅₀ of the compounds. (c) Plot of standardized residual activity against leverage of the compounds (Williams plot).

Table 3.

Results for random models.

Model	R2	Q2
Original	0.912	0.802
Random 1	0.197	−0.779
Random 2	0.212	−1.077
Random 3	0.199	−1.108
Random 4	0.282	−0.515
Random 5	0.271	−0.388
Random 6	0.104	−1.029
Random 7	0.250	−1.413
Random 8	0.399	−0.031
Random 9	0.359	−0.213
Random 10	0.510	−0.046
Random model parameters
Average R2:	0.336
Average Q2:	−0.527
Average cRp2:	0.767

The five descriptors best relating the structural features of the studied compounds to their antiplasmodial activities, as demonstrated in the model, were ring count (n5Ring), topological charge index (GGI9), 3D topological distance based autocorrelation (TDB7u and TDB8u) and 3D radial distribution function (RDF75i) descriptors. Pearson's correlation analysis was performed on the descriptors to verify their inter-correlation; further verification was performed with evaluation of VIF. The test results (Table 4) indicated no significant inter-correlation among the descriptors, on the basis of VIF values below 10. Therefore, the combination of descriptors significantly related the antiplasmodial activities of the studied compounds to their structures. The mean effects (Table 4) of the descriptors indicated their relative strength in influencing the pIC₅₀ of the 6-fluoroquinolines.

Table 4.

Pearson's correlation, VIF and ME of the descriptors in the model.

Descriptors	Inter-correlations					VIF	ME
	n5Ring	GGI9	TDB7u	TDB8u	RDF75i
n5Ring	1.000	−0.329	0.351	0.186	−0.121	1.309	0.034
GGI9	−0.329	1.000	−0.430	−0.281	0.651	2.107	−0.090
TDB7u	0.351	−0.430	1.000	0.862	−0.566	5.036	−3.119
TDB8u	0.186	−0.281	0.862	1.000	−0.545	4.523	3.973
RDF75i	−0.121	0.651	−0.566	−0.545	1.000	2.414	0.202

The first descriptor in the model was n5Ring, defined as the 5-membered ring count. This descriptor relates the present of 5-membered rings in the structures of 6-fluoroquinolines to their pIC₅₀. The positive coefficient of this descriptor in the model indicated that the presence of 5-membered rings in the structures of 6-fluoroquinolines positively contributes to antiplasmodial activity. The second descriptor in the model was GGI9, defined as the 9-ordered raw topological charge index descriptor. This descriptor estimates the charge transfers between pairs of atoms that are nine bonds apart.³⁶ The negative coefficient for this descriptor in the model indicated that the presence of two atoms that are nine bonds apart and have a high electronegativity difference contributes negatively to 6-fluoroquinoline antiplasmodial activity. The third and fourth descriptors in the model, TDB7u and TDB8u, describe topological distance, on the basis of autocorrelation of lag 7 (TDB7u) and 8 (TDB8u), all unweighted. This class of descriptors was calculated on the basis of the average Euclidean distance between all atoms located at a given topological distance (distance between two atoms in molecular graph representation).³⁷ For the TDB7u and TDB8u descriptors, the topological distances considered were those between two atoms seven and eight bonds apart, respectively. All atoms were treated equally, because they were unweighted. An increase in TDB7u was associated with a decrease in antiplasmodial activity of 6-fluoroquinolines, because of its negative coefficient in the model. In contrast, an increase in TDB8u was associated with an increase in antiplasmodial activity, because of its positive coefficient in the model. The fifth descriptor in the model, RDF75i, is the radial distribution function 075/weighted by the first ionization potential.³⁸ The descriptor measures the first ionization potential of an atom or group of atoms at a radius 4.5 Å from the geometrical center of the molecules. The positive coefficient for this descriptor in the model indicated that an increase in its value was associated with increased antiplasmodial activity of 6-fluoroquinolines.

The domain of applicability for the model was presented by a Williams plot (Figure 1c), as the area within ±3 standardized residuals, and leverage 0–0.950. The applicability domain showed no outliers, and all compounds were within ±3 standardized residuals. However, the leverage of one influential compound (compound 27) exceeded 0.950, the warning leverage for the model. The structure of this influential compound slightly differs from that of the other compounds in the dataset (Table 1) and therefore should not be considered the template when the model is used to design novel 6-fluoroquinolines.

DDD107498, the template for designing the dataset used in this work, has been reported to inhibit PfeEF2 as a molecular target.⁴ PfeEF2 is crucial for protein synthesis and is responsible for the GTP-dependent ribosomal translocation along mRNA; therefore, its discovery may open new avenues for antimalarial drug discovery.³⁹ However, its crystal structure is not available, and its interaction with DDD107498 is unclear. Therefore, we designed a homology model for use in our docking study. The crystal structure of the closet template was searched with the protein sequence of eEF2 in P. falciparum (isolate NF54). ADP-ribosylated ribosomal translocase from Saccharomyces cerevisiae (PDB: 1U2R) was identified as the closest template, with 61.300% identity, 0.480 similarity, 1.000 coverage and 2.600 Å resolution, as determined by X-ray crystallography.⁴⁰ Figure 2a shows the single sequence alignment of modeled PfeEF2 (Figure 2c) with the 1U2R (Figure 2b) amino acid sequence; amino acids of the 1U2R that aligned with the model are indicated in bold. The modeled PfeEF2 had a global model quality estimation score of 0.790 and QMEAN score of 0.770. The former scoring function estimates the accuracy of the modeled structure, whereas the latter assesses the quality of the model.^24,25 The closer the values of the scoring functions are to 1, the better the built model. Therefore, the modeled PfeEF2 was considered good and reliable.

Figure 2.

(a) Single sequence alignment of modeled PfeEF2 (model-01) with 1U2R (1u2r.1.A) amino acid sequences. (b) 3D structure of 1U2R. (c) 3D structure modeled PfeEF2.

The plot in Figure 3a, shows the modeled PfeEF2 local quality estimate. The plot indicated a good local quality estimate, because most of the residue scores were close to 1, and the average was 0.770. Figure 3b compares the structure of the modeled PfeEF2 with the non-redundant aligned PDB structures, on the basis of a plot of normalized QMEAN scores (Z-scores) against protein sizes (residues). The plot indicated that the score of the structure of modeled PfeEF2 (red star) was within that of experimentally determined structures, on the basis of the number of residues. The model had a Z-score of −1.200, thus indicating good agreement with an experimental structure of similar size.⁴¹ Figure 3c shows the Ramachandran plot of modeled PfeEF2 for all non-glycine/proline residues. This plot provided insight into the backbone dihedral angles of amino acid residues in PfeEF2 against energetically favored regions of dihedrals of protein residues in general. In the plot, the green contour indicates the favored regions; 95.410% of residues were Ramachandran favored, and the MolProbity score was 1.88.

Figure 3.

(a) Local quality estimate of the residue graph. (b) Comparison of the modeled PfeEF2 structure with a non-redundant set of PDB structures. (c) Ramachandran plot of the modeled PfeEF2 for all non-glycine/proline residues.

All 28 compounds in our dataset were docked with the modeled pfeEF2, and their binding affinities ranged from −8.200 to −10.700 kcal/mol, thereby indicating strong interaction of the compounds with the amino acids of PfeEF2. DDD107498, which was experimentally suggested to form a stable complex with PfeEF2, was docked and found to have a binding affinity of −9.900 kcal/mol.³⁹ This binding affinity was greater than that of five compounds in the dataset (Table 5), thus indicating that the compounds formed more stable complexes with the target than DDD107498, and consequently may be better inhibitors. Compound 5 had the best binding affinity (−10.700 kcal/mol), possibly because of its better interaction with the target. Figure 4 shows the structures of the interactions of compound 5 with PfeEF2. The interactions involved the following: (1) Three conventional hydrogen bonds: one of bond length 2.480 Å from a hydrogen on THR323 to fluorine in the quinoline moiety of the compound; one of bond length 2.190 Å from the nitrogen of the quinoline moiety of the compound to oxygen in TRP317; and one of bond length 2.06 Å from the nitrogen of the carboxamide group of the compound to oxygen in the ASP322 residue. (2) A carbon hydrogen bond of bond length 3.139 Å from the methyl carbon of the 2-{4-[(morpholin-4-yl)methyl]phenyl} substituent of the compound to oxygen in the GLY294 residue. (3) A halogen interaction between the carbon of the ASP322 residue and fluorine in the quinoline moiety. (4) Three electrostatic interactions: one of π-cation type between the amino group of ARG221 and the π-orbital of the N-benzyl substituent in the carboxamide group, and two of π-anion type from oxygen in the ASP322 residue to two rings of the quinoline moiety. (5) Six hydrophobic interactions: one of π-sigma type from carbon in the THR218 residue to the phenyl ring of the quinoline moiety; two of π-π type between the π-orbital in the PHE79 and HIS81 residues, and the π-orbital of the N-benzyl substituent in the carboxamide group; one of alkyl–alkyl type between the alkyl carbon in LEU316 and the morpholin carbon of the 2-{4-[(morpholin-4-yl)methyl]phenyl} substituent; and two π-alkyl type between PRO319 and the two rings of the quinoline moiety.

Table 5.

Docking results between modeled pfeEF2 and selected 6-fluoroquinolines.

Ligand	Binding affinity (kcal/mol)	Hydrogen bonding	Hydrophobic interaction	Electrostatic interaction
		Amino acid (bond length, Å)	Amino acidbond type	Amino acidbond type
5	−10.700	THR323 (2.48)a, TRP317 (2.19)a, ASP322 (2.06)a, GLY294 (3.13)b	THR218c, PHE79d, HIS81d, LEU316e, PRO319f, PRO319f	ASP322g, ARG221h, ASP322i, ASP322i
11	−10.200	SER107 (2.62)a, SER107 (2.04)a, HIS527 (2.97)a, THR771 (2.98)a, GLY780 (1.79)a, GLU143 (3.71)b	TYR473c, PHE783d, SER106j, TYR473f, ARG775f, MET471f	GLU525g
16	−10.000	SER474 (2.79)a, THR771 (2.84)a, GLY780 (1.98)a	TYR473c, LEU507e, ARG775f, MET471f	GLU525g
22	−10.200	LYS455 (1.80)a, TYR473 (2.46)a, SER474 (2.38)a, GLU525 (2.85)a, PHE783 (3.66)b, PRO784 (3.28)b	TYR473c, LEU526e, LEU526e, HIS102f, TYR473f, ARG775f, LEU526f	ALA782g, ASP104i, GLU525i
24	−10.100	ARG114 (2.65)a, GLN781 (2.73)a, TYR186 (2.23)a, GLY142 (3.53)b, GLU143 (3.25)b, ASP188 (3.57)b	TYR186d	THR185g, ASP728i, ASP728i
Q4C	−9.900	SER107 (2.10)a, SER107 (2.07)a, ARG775 (3.33)b, GLY780 (3.24)b, ASP104 (3.47)b	TYR473d, TYR473d, CYS523f, GLY524j, ARG775f, CYS523f	PHE105g

Q4C = Quinoline-4-carboxamide. ^aConventional hydrogen bond. ^bCarbon hydrogen bond. ^cπ- sigma. ^dπ-π. ^eAlkyl-alkyl. ^fπ-alkyl. ^gHalogen. ^hπ-cation. ⁱπ-anion. ^jAmide-π.

Figure 4.

3D and 2D structures of compound 5 and PfeEF2 interactions.

No correlation was observed between the binding affinities of the docked compounds and their antiplasmodial activity against a chloroquine-sensitive strain (PfNF54), thus indicating the possibility of multiple targets for the compounds. However, the results provide insight into the binding nature of PfeEF2 and may be used in designing good inhibitors.

Table 6 presents the predicted ADMET and drug-likeness properties of the five compounds with the best binding affinity. The partition coefficient (logP ≤ 5), molecular weight (<500 g/mol), H-bond acceptors (≤10) and H-bond donors (≤5), on the basis of Lipinski's rule, were predicted for the compounds.⁴² The results indicated that compounds 5 and 24 had no violations, whereas compounds 11, 16 and 22 each had one violation of Lipinski's rule (Table 6). The compounds satisfying Lipinski's rule (fewer than two violations) therefore were considered orally active.⁴² The water solubility (log mol/L) of the compounds indicated that compounds 5 and 11 were moderately soluble (−6 log mol/L < −4), and compounds 16, 22 and 24 were soluble (−4 log mol/L < −2). Therefore, the compounds are not expected to have poor oral absorption.⁴³ The polar surface areas (Ų) of the compounds were within the satisfactory range and hence were considered orally bioavailable. Similarly, zero pan assay interference compounds (PAINS) alerts were obtained for all compounds, thus indicating that they may serve as lead compounds.⁴⁴

Table 6.

In silico drug-likeness and ADMET properties of the most inhibitory compounds.

Drug-likeness	Compounds
	5	11	16	22	24
Partition coefficient (logP)	4.700	3.770	3.990	3.520	3.100
Molecular weight	469.550	577.690	563.710	547.660	491.060
Number of H-bond acceptors	5	8	8	7	7
Number of H-bond donors	1	1	1	1	1
Number of rotatable bonds	7	11	11	9	8
Lipinski's violations	0	1	1	1	0
Water solubility (log mol/L)	−5.193	−4.159	−3.440	−2.849	−2.563
Polar surface area (Å2)	54.460	87.240	70.170	78.010	60.940
PAINS alert	0	0	0	0	0
Absorption
Caco-2 permeability (log cm/s)	−4.935	−4.989	−5.197	−5.134	−4.899
Human intestinal absorption (%)	99.700	99.900	99.900	99.800	99.300
Distribution
Volume distribution (L/kg)	2.328	1.830	2.109	2.319	2.357
BBB penetration (log cm/s)	−0.188	−0.072	−0.038	−0.041	−0.010
Metabolism
CYP1A2 inhibitor	No	No	No	No	No
CYP2C19 inhibitor	No	Yes	Yes	Yes	No
CYP2C9 inhibitor	Yes	Yes	No	No	No
CYP2D6 inhibitor	Yes	Yes	Yes	Yes	Yes
CYP3A4 inhibitor	Yes	Yes	Yes	Yes	Yes
Excretion
Half-life	0.015	0.017	0.017	0.072	0.029
Toxicity
Ames toxicity	0.833	0.289	0.046	0.030	0.144
Rat oral acute toxicity	0.436	0.426	0.326	0.786	0.429
Carcinogenicity	0.092	0.388	0.088	0.125	0.088

The Caco-2 permeability (>−5.15 log cm/s) of all compounds except compound 16 (<−5.15 log cm/s), and the intestinal absorption (>30%) of all compounds, indicated good absorption potential. The volume distribution (0.04 < VD < 20 L/kg) of the compounds indicated good distribution characteristics, and the blood–brain barrier (logBBB cm/s > −1) of the compounds indicated their ability to cross the barrier and thus not cause any problems with the central nervous system.³² Metabolism of the compounds was predicted on the basis of their interaction with cytochrome P450 (CYP) (Table 6). All compounds were found to be inhibitors of CYP2D6 and CYP3A4, but not CYP1A2. Compounds 5 and 11 were found to be inhibitors of CYP2C9, whereas compounds 16, 22 and 24 were not. Compounds 11, 16 and 22 were found to be inhibitors of CYP2C19, whereas compounds 5 and 24 were not. Investigation of the interactions of molecules with CYP isoforms is key to understanding drug metabolism.⁶ The excretion of the compounds was verified by prediction of their half-lives in terms of the probability (from 0 to 1) of having a long half-life. The half-lives of the compounds (Table 6) indicated high clearance. The toxicity prediction is given as the probability (from 0 to 1) of being toxic. The results (Table 6) indicated that the compounds had low probabilities of being toxic, except for compounds 5 and 22, which had high probabilities of Ames and rat oral acute toxicity, respectively. Therefore, only compound 5 might be mutagenic, and compound 22 might have toxicity to mammals.³² The ADMET results indicated good pharmacokinetic properties; therefore, the studied compounds may be considered for drug development.

Conclusion

This work developed a robust and reliable QSAR model that relates the structures of 2,4-disubstituted 6-fluoroquinolines to their antiplasmodium activity, on the basis of GFA. The model had R², R²_adj, Q²_cv, and R²_pred values of 0.921, 0.878, 0.801 and 0.901, respectively. Our findings indicated that the n5Ring, GGI9, TDB7u, TDB8u and RDF75i descriptors were the physicochemical properties most strongly associated with 6-fluoroquinoline antiplasmodium activity. n5Ring, TDB8u and RDF75i were positively associated, whereas GGI9 and TDB7u were negatively associated, with the antiplasmodium activities of the compounds. A docking study indicated formation of stable complexes between the compounds and modeled PfeEF2, with binding affinities ranging from −8.200 to −10.700 kcal/mol. Compounds 5, 11, 16, 22 and 24 had better binding affinity than DDD107498 and good pharmacokinetic properties; therefore, these compounds may serve as better inhibitors of this novel target. Our findings may be used to design novel 2,4-disubstituted 6-fluoroquinolines with high antiplasmodial potency and good structural properties of inhibiting the novel antimalarial drug target PfeEF2.

Source of funding

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

Conflict of interest

The authors have no conflict of interest to declare.

Ethical approval

No ethical issues are reported.

Authors contributions

GAS, AWM and AU conceived and designed the study. GAS and AWM obtained the data. AWM drew and optimized the structures. AWM calculated descriptors, generated the model and performed statistical analysis. GAS conducted homology modeling and molecular docking. GAS, AWM and AU interpreted the results. MTI performed pharmacokinetic predictions. AMW wrote the paper. All authors have critically reviewed and approved the final draft and are responsible for the content and similarity index of the manuscript.