4-Aminoquinolines Active against Chloroquine-Resistant Plasmodium falciparum: Basis of Antiparasite Activity and Quantitative Structure-Activity Relationship Analyses

Abstract

Chloroquine (CQ) is a safe and economical 4-aminoquinoline (AQ) antimalarial. However, its value has been severely compromised by the increasing prevalence of CQ resistance. This study examined 108 AQs, including 68 newly synthesized compounds. Of these 108 AQs, 32 (30%) were active only against CQ-susceptible Plasmodium falciparum strains and 59 (55%) were active against both CQ-susceptible and CQ-resistant P. falciparum strains (50% inhibitory concentrations [IC₅₀s], ≤25 nM). All AQs active against both CQ-susceptible and CQ-resistant P. falciparum strains shared four structural features: (i) an AQ ring without alkyl substitution, (ii) a halogen at position 7 (Cl, Br, or I but not F), (iii) a protonatable nitrogen at position 1, and (iv) a second protonatable nitrogen at the end of the side chain distal from the point of attachment to the AQ ring via the nitrogen at position 4. For activity against CQ-resistant parasites, side chain lengths of ≤3 or ≥10 carbons were necessary but not sufficient; they were identified as essential factors by visual comparison of 2-dimensional (2-D) structures in relation to the antiparasite activities of the AQs and were confirmed by computer-based 3-D comparisons and differential contour plots of activity against P. falciparum. The advantage of the method reported here (refinement of quantitative structure-activity relationship [QSAR] descriptors by random assignment of compounds to multiple training and test sets) is that it retains QSAR descriptors according to their abilities to predict the activities of unknown test compounds rather than according to how well they fit the activities of the compounds in the training sets.

INTRODUCTION

The annual burden of malaria is overwhelming, with more than 300 to 500 million cases and 2 million deaths each year (3). In addition, its impact is exacerbated by the increasing prevalence of antimalarial resistance, which was the single most important factor in its recent worldwide resurgence (22, 41). The most important resistance is that of Plasmodium falciparum to 7-chloro-4-(4-diethylamino-1-methylbutylamino)-quinoline (chloroquine [CQ]), because P. falciparum is responsible for most morbidity and mortality, and especially for the deaths of children under the age of 5 years in sub-Saharan Africa (3). For these reasons, and because resistance to the artemisinins and to the combination of atovaquone and proguanil (Malarone) is now emerging (14, 20, 33), the development of antimalarials that are effective against drug-resistant parasites—and that are affordable for the people in the greatest need—is an urgent global health priority.

The established safety record of CQ, its ease of synthesis, and its low cost led us to examine the structure-activity relationships (SARs) responsible for the antiparasite activity of the 4-aminoquinolines (AQs) (10, 11, 26) with the goal of identifying the factors responsible for their activity against CQ-resistant parasites. The design and synthesis of AQ analogues during the past 5 to 10 years, in combination with biological testing against CQ-susceptible and -resistant P. falciparum strains, have also yielded molecular insights into the mechanisms of AQ action and resistance (27–30).

In this report, we present structure-activity analyses for 108 AQ analogues—including 68 newly synthesized compounds—to clarify the relationship between the chemical structures of the AQs and their antiparasite activities.

A main purpose of quantitative structure-activity relationship (QSAR) analyses is to make activity predictions for unknown compounds in order to guide the structure-based design of new analogues. Because we plan to extend the search for more-potent AQ analogues by using an iterative molecular design approach, useful quantitative models must have the ability to be readily extended to incorporate additional compounds. Therefore, we used computer-generated 3-dimensional (3-D) structures and chemical descriptors. This article examines 108 AQs using comparative molecular field analysis (CoMFA) (8) and comparative similarity index analysis (CoMSIA) (25), as well as the 2-D hologram QSAR (HQSAR) (32, 36) and UNITY molecular fingerprint (UFP) (6) paradigms. To eliminate bias from the selection of a single test set for the examination of the predictive ability of QSAR models, we randomly assigned 20 molecules to 20 different test sets and used the 88 excluded compounds as the associated training set in each case.

Partial least-squares (PLS) analysis reduces the complexity of a QSAR model by sorting the large number of descriptors into a small number of orthogonal principal components to facilitate least-squares fitting. However, these components still contain large numbers of descriptors. Inevitably, many of these descriptors are affected by the inherent noise of the method and the uncertainty of the biological data. Several methods to eliminate these uninformative descriptors have been described previously (43). These methods include the stability (mean/standard deviation [SD]) of PLS weights (2) or PLS coefficients (5) and retain only the most stable descriptors in the final QSAR models. In this study, descriptor normalization was used to prevent descriptors with large magnitudes or variances from swamping other descriptors. Ranking of the PLS contributions for each of the normalized descriptors identified those that best predicted the antiparasite activities (50% inhibitory concentrations [IC₅₀s]) of the unknown test AQs. In this way, the model was chosen by its ability to predict rather than to fit the data.

MATERIALS AND METHODS

Synthesis of AQ analogues.

AQ analogues were synthesized as described previously (9–12). Briefly, as an example, 7-chloro-AQs were synthesized by the condensation of m-chloroaniline with ethyl ethoxymethylene malonate to produce 4,7-dichloroquinoline. Nucleophilic substitution with different N,N-diethyldiaminoalkanes produced the desired analogues with side chains containing 2 to 12 carbons (9–12). Each AQ analogue was purified by column chromatography with an activated basic alumina or silica gel and was characterized by ¹H nuclear magnetic resonance (NMR) spectroscopy (¹⁹F NMR when applicable), mass spectroscopy, and C, H, and N elemental analyses, which agreed with the theoretical values (±0.4%). The purity of each compound was ≥95% with two high-performance liquid chromatography (HPLC) systems. The first consisted of a C₁₈ column (length, 250 mm; inner diameter, 4.6 mm; particle size, 5 μm) using isocratic elution with an acetonitrile-borate buffer (56:44) (pH 9.2) at 1 ml min⁻¹ and fluorescence detection based on excitation at 320 nm and emission at 380 nm. The second HPLC system consisted of a C₁₈ column with buffer A (0.1% trifluoroacetic acid [TFA]) and buffer B (0.1% TFA in 80% acetonitrile) using gradient elution from 10% buffer B to 70% buffer B at 1 ml min⁻¹ with UV detection at 215 nm. The structures and biological activities of the analogues used for QSAR are summarized in Table 1, and those of the AQ analogues that have not been reported previously are shown in detail in Table 2. Data for the 40 AQ analogues that have been reported previously are available in references 10, 11, and 31. Synthetic and analytical data for the remaining compounds will be published elsewhere.

Table 1.

Summary of the chemical structures and biological activities of 108 4-aminoquinoline analogues examined in quantitative structure-activity relationship analyses

Table 2.

Structures of 4-aminoquinoline analogues and their antiplasmodial activities against chloroquine-susceptible and -resistant Plasmodium falciparum strains^a

^aAntiplasmodial activities are expressed as IC₅₀s. The CQ-susceptible strain used for these determinations was Haiti 135, and the CQ-resistant strain used was Indochina I. The data presented in this table are for the 68 4-aminoquinoline (4-AQ) analogues that have not been reported previously. Data for the 40 AQ analogues reported previously are available in references 10, 11, and 32.

QSAR data set.

Computational QSAR studies examined 108 AQ analogues based on CQ. These AQs consisted of 72 7-Cl, 6 7-F, 7 7-Br, 6 7-I, 7 7-CF₃, 6 7-OCH₃, 1 6-Cl, 1 8-Cl, and 2 unsubstituted AQs. Side chains ranged from 2 to 12 carbons, and terminal side chain amine functions included 3 (CH₃)₂N, 48 (ethyl)₂N, 2 (propyl)₂N, 1 (butyl)₂N, 6 pyrrolidines, 28 piperidines, 4 azepanes, 3 azocanes, 3 primary amines, 8 secondary amines, and 2 unsymmetrical linear secondary amines. The in vitro inhibitory activities of these AQs against CQ-susceptible and CQ-resistant P. falciparum strains (Haiti 135 and Indochina I, respectively) were interpreted based on their IC₅₀s (AQs with IC₅₀s of ≤25 nM were considered active) (Table 2), which were converted to pIC₅₀ values (log 1/IC₅₀) and used as dependent variables in deriving 3-D QSAR models.

Malaria parasite strains.

The Plasmodium falciparum strains employed in these studies were a cloned CQ-susceptible parasite from Haiti (Haiti 135) (26, 39) and a cloned CQ-resistant parasite from Indochina (Indochina I) (7).

Antiplasmodial potency.

The in vitro antiplasmodial activities of AQ analogues, expressed as their effects on the growth of cloned susceptible and resistant strains of P. falciparum (IC₅₀s, given as nanomolar concentrations), were assessed by their abilities to inhibit [³H]hypoxanthine incorporation via the essential parasite purine salvage pathway (19, 29) as described previously (13). Briefly, synchronous cultures of early-ring-stage parasites (600 μl of a 2% red blood cell suspension with 0.2% parasitemia in 24-well tissue culture plates [Corning Cell Wells, catalog no. 25820; Corning, Corning, NY]) were grown in the in vitro culture system (40). Varying concentrations of the AQ analogues were added, and the parasites were grown for 48 h without supplemental hypoxanthine. [³H]hypoxanthine (0.5 μCi; Perkin-Elmer, Boston, MA) was then added, followed by an additional 16 to 18 h of incubation. Cell contents were harvested on 12-mm-diameter glass fiber filter disks (934 AH; Whatman, Clifton, NJ) using a Skatron (Sterling, VA) cell harvester, dried, placed in glass vials (Kimble, Toledo, OH) with 8 ml of liquid scintillation fluid (Cytoscint; ICN Radiochemicals, Costa Mesa, CA), and assayed in a liquid scintillation counter (Tri Carb 2100 TR; Packard, Meriden, CT). After initial screening to identify the relevant range of the dose-response curve, IC₅₀s were determined for each AQ (results are presented as mean IC₅₀s ± standard errors of the means [SEMs] [Table 2]).

Molecular modeling.

Molecular modeling studies were performed on an HP dual Xeon 3.2-GHz workstation operating Red Hat Enterprise Linux 4 using SYBYL, version 8.0 (Tripos, L.P., St. Louis, MO) and an Apple Macintosh (iMac Core 2 Duo 3.06-GHz OS X 10.5.6) using SYBYL, version 8.1. The conversational SMILES (Simplified Molecular Identification and Line Entry System) (1, 42) string for each AQ analogue was the input for the CONCORD (CONnectivity to CoORDinates) program (34, 35), and the default settings were used to provide the 3-D coordinates for each molecule. Partial atomic charges were calculated using the Gasteiger-Hückel method (16–18). The 3-D coordinates of each AQ analogue were imported into the SYBYL molecular modeling suite (24, 38) to create molecular databases containing the 108 AQ analogues. All molecules were modeled in the neutral form, since protonation of this highly congeneric series conferred no advantages.

QSAR analyses. (i) Molecular alignment.

Aminoquinolines are believed to act by binding to the growing faces of hematin crystals and thus preventing the further deposition of hematin. The subsequent accumulation of toxic free heme then kills the parasite. The quinoline moiety is thought to intercalate into the grooves on the growing face, with the side chains lying on the face of the hematin crystal (4). The basic side chain of the AQs is necessary to drive their accumulation in the acidic vacuole of the parasite. The molecular database was aligned separately by a root mean square (rms) fit to the common heavy atoms of the quinoline rings of CQ. SYBYL QSAR tables were then built from the molecular databases, with the compounds as rows and the target antiplasmodial biological data (IC₅₀s for CQ-susceptible and -resistant P. falciparum) and molecular descriptors as columns. The SYBYL QSAR module was used to define a 2-Å-spaced CoMFA region extending 4 Å beyond the aligned analogues, and this region grid was used for all 3-D QSAR analyses. The 108 AQ analogues were divided randomly (MATLAB random number seed, 98,517) into 20 independent training sets containing 88 compounds in order to derive the QSAR models and into 20 associated test sets containing the 20 excluded compounds in order to test the predictive power of the final QSAR models.

(ii) 3-D QSAR descriptors.

The predefined 3-D lattice region definition with 2-Å spacing was used for all CoMFA and CoMSIA descriptors in SYBYL (Tripos). CoMFA steric and electrostatic field energies were calculated at all intersections of the region grid using an sp³ hybridized carbon probe atom with a charge of +1 and a distance-dependent dielectric constant with the Tripos Force Field. The steric and electrostatic contributions were truncated to ±30 kcal mol⁻¹, and the electrostatic contributions were ignored at lattice intersections with maximal steric interactions. CoMSIA steric, electrostatic, hydrophobic, hydrogen bond donor, and hydrogen bond acceptor molecular similarity descriptors used the following default settings: probe atom radius, 1 Å; charge, +1; hydrophobicity, +1; attenuation factor, 0.3.

(iii) 2-D QSAR descriptors.

UNITY molecular fingerprints (UFPs) with a bit length of 992 were generated using the default settings in UNITY (Tripos). QSAR holograms with bin lengths of 97, 151, 199, 257, 307, 353, 3,500, and 10,000 and fragment lengths of 4 to 7 atoms, with nonhydrogen fragment atoms, bonds, and connections used to determine uniqueness, were generated using the HQSAR module of SYBYL (24, 38). Each HQSAR analysis of biological data for CQ-susceptible and CQ-resistant parasites produced the same estimate for the optimum hologram bin length (10,000), which was used for all subsequent analyses. Basicities (pK_as) for each of the nitrogen log P and log D values (at pH 5.5, 6.0, 6.5, 7.0, and 7.4) were calculated using the ACD/Labs Structure Design Suite (Advanced Chemistry Development, Toronto, Ontario, Canada).

QSAR descriptors were imported into MATLAB R2008a (MathWorks, Natick, MA), and QSAR algorithms were programmed in MATLAB with the Statistics Toolbox, version 7.0. Each principal component was required to increase the observed q² by ≥5% above the model with one less component in order to ensure that the models were not overfitting the noise in the data. This method provides a more conservative estimate than the number of components that correspond to the minima in the standard error.

(iv) Iterative QSAR descriptor optimization.

Because the 2-D and 3-D descriptors had markedly different magnitudes and variances, descriptors were normalized by centering (mean = 0) and scaling (SD = 1) using the z-score algorithm in MATLAB to prevent descriptors with large magnitudes or variances from swamping the other descriptors. The descriptors were ranked by the mean PLS contributions from the non-cross-validated analyses (n = 10), and the ranked descriptors were then added stepwise to the cross-validated QSAR models while monitoring the predictive q² (q²_prediction), q², and related standard errors of prediction (SDEP) from the non-cross-validated QSAR analyses. This approach allowed the most significant (most predictive) descriptors to be retained while eliminating those that contributed least to the final QSARs. The q² values (mean ± SEMs) (n = 10), numbers of components, and initial (SD ≠ 0) and final numbers of optimized descriptors from the PLS analyses are shown in Fig. 1, and the SDEP are shown in Fig. 2, for CQ-susceptible and CQ-resistant parasites.

Fig. 1.

Effects of descriptor optimization on q² for training compounds, numbers of components, and descriptors from cross-validated, partial least squares (PLS) analyses against P. falciparum.

Fig. 2.

Effects of descriptor optimization on standard errors of prediction (SDEP) for unknown test compounds from PLS analyses against P. falciparum.

(v) QSAR model validation.

The extrapolated predictive ability of each QSAR model was examined using the 20 randomized test sets of 20 compounds that had not been included in the training sets. The activities of these compounds were predicted from each final PLS analysis, and the SDEP and q²_prediction values were then calculated for each QSAR model. The q²_, SDEP, and q²_prediction (mean ± SEM) (n = 20) values of the test compounds for CQ-susceptible and CQ-resistant parasites are shown in Fig. 1 to 3, respectively, and are summarized in Table 3.

Fig. 3.

Effects of descriptor optimization on the predictive q² (q²_prediction) for unknown test compounds from PLS analyses against P. falciparum.

Table 3.

Summary of QSAR statistics with optimized descriptors

QSAR statistica	Value for:
	CQ-susceptible P. falciparum				CQ-resistant P. falciparum
	HQSAR	UFP	CoMFA	CoMSIA	HQSAR	UFP	CoMFA	CoMSIA
q2	0.49 ± 0.004	0.45 ± 0.004	0.59 ± 0.003	0.55 ± 0.006	0.68 ± 0.004	0.64 ± 0.004	0.54 ± 0.003	0.56 ± 0.004
SEP	0.32 ± 0.002	0.33 ± 0.003	0.27 ± 0.001	0.29 ± 0.003	0.35 ± 0.002	0.37 ± 0.001	0.42 ± 0.001	0.41 ± 0.001
PC	3	4	6	3	2	3	3	3
No. of descriptors	2,640	329	4,723	7,553	2,640	329	4,723	7,553
Optimized no. of descriptors	700	70	400	300	350	90	175	225
r2	0.75 ± 0.003	0.72 ± 0.004	0.84 ± 0.002	0.76 ± 0.003	0.76 ± 0.003	0.76 ± 0.003	0.64 ± 0.003	0.63 ± 0.004
SE	0.50 ± 0.003	0.55 ± 0.004	0.44 ± 0.002	0.49 ± 0.002	0.67 ± 0.004	0.68 ± 0.002	0.84 ± 0.004	0.85 ± 0.003
F	267.3 ± 4.7	229.8 ± 4.8	477.2 ± 8.4	281.3 ± 5.0	273.5 ± 3.3	280.6 ± 3.8	153.6 ± 2.1	149.5 ± 2.3
SDEP	0.33 ± 0.019	0.35 ± 0.024	0.30 ± 0.013	0.32 ± 0.019	0.35 ± 0.002	0.37 ± 0.013	0.38 ± 0.011	0.36 ± 0.015
q2prediction (n = 20)	0.47 ± 0.032	0.42 ± 0.030	0.54 ± 0.034	0.52 ± 0.031	0.60 ± 0.040	0.55 ± 0.037	0.50 ± 0.050	0.56 ± 0.040
Contribution (%)
Steric			61.3 ± 0.44	24.4 ± 0.15			63.5 ± 0.27	14.2 ± 0.08
Electrostatic			38.7 ± 0.25	13.6 ± 0.09			36.5 ± 0.12	12.8 ± 0.04
Hydrophobic				37.0 ± 0.21				37.0 ± 0.13
HB acceptor				6.3 ± 0.04				9.0 ± 0.02
HB donor				18.8 ± 0.05				27.0 ± 0.08

^aSEP, standard error of prediction; PC, number of principal components; HB, hydrogen bond.

RESULTS AND DISCUSSION

At a time when there is an unprecedented global effort to control and potentially eliminate malaria (21), antimalarial resistance remains the single most important obstacle to success. This is because there are few safe and affordable antimalarials that are effective against resistant parasites. It is also because the antimalarials that are currently key for treatment success (the artemisinins and the atovaquone-proguanil [Malarone] combination) (14, 20, 33) are now threatened by resistance, and previously promising antimalarials in development are threatened by toxicity and resistance (dapsone plus chlorproguanil [LapDap], dapsone plus artesunate [Dacart], and piperaquine) (23, 37).

As a result, increasing the number of AQs active against CQ-resistant P. falciparum, and thus expanding the depleted upstream pipeline for antimalarial development, is an overwhelmingly important step in the path to the control and elimination of malaria. This strategic argument has recently been strengthened by our phase 1 (safety and pharmacokinetic) studies of the lead compound (AQ-13) with human subjects. Those studies have shown that AQ-13 is economical (cost of goods, $0.03 per dose), is as safe as CQ (one of the safest drugs known) for human subjects, produces levels in blood and bioavailability similar to those of CQ, and has no identifiable cardiac, neurologic, or other toxicities at oral doses equivalent to 600, 1,200, 1,500, or 1,800 mg CQ base (31).

SARs against P. falciparum.

Based on the data reported here (Table 2) and in our previous studies (10, 11, 31), common features of AQs active against both CQ-susceptible and CQ-resistant P. falciparum are (i) an AQ ring without alkyl substitutions, (ii) a halogen at position 7 of the AQ ring (Cl, Br, or I but not F), (iii) a protonatable nitrogen at position 1 of the AQ ring, and (iv) a second protonatable nitrogen at the other end of the side chain from the point of attachment to the AQ ring via the 4-anilino-nitrogen. In contrast to these relatively stringent requirements for the AQ ring, there was considerable latitude in the side chains and terminal amines that permitted antiparasite activity. The results reported here demonstrate that AQs with side chains from 2 to 12 carbons and diverse terminal amines (e.g., diethylamine, piperidine, azepane, azocane) are active against CQ-susceptible P. falciparum at low nanomolar concentrations (Table 2) (10, 11, 31).

Because the 59 AQs active against CQ-resistant P. falciparum were also active against CQ-susceptible P. falciparum, they share the four structural features noted above. In addition, these 59 AQs have side chains either shorter or longer (≤3 or ≥ 10 carbons) than the intermediate-length isopentyl (5-carbon) side chain of CQ. During the past 5 to 10 years, these studies have had two goals: (i) to develop AQs with the potential to treat CQ-resistant P. falciparum infections in humans and (ii) to use 3-D QSAR modeling to identify the structural features essential for activity against CQ-resistant parasites. By comparing AQs active only against CQ-susceptible P. falciparum strains to AQs active against both CQ-susceptible and CQ-resistant P. falciparum strains, it should be possible to identify the structural features responsible for activity against resistant parasites—i.e., the structural features that prevent recognition and excretion (efflux) by the transporter responsible for CQ resistance in P. falciparum (PfCRT) (10, 11, 15). Based on compounds with similar AQ rings (with Cl, Br, or I at position 7) and terminal amines (N-diethylamine or piperidine), these results indicate that side chains of ≤3 or ≥10 carbons are necessary for activity against CQ-resistant P. falciparum. For instance, all 20 of the AQs active against CQ-susceptible P. falciparum that had side chains with ≤3 or ≥10 carbons were also active against CQ-resistant P. falciparum (IC₅₀s, ≤25 nM). Conversely, none of the 12 AQs active against CQ-susceptible P. falciparum that had 4- to 8-carbon side chains were active against CQ-resistant P. falciparum (P, <0.001 by Fisher's exact test). However, side chains of ≤3 or ≥10 carbons were not sufficient for activity against CQ-resistant P. falciparum. Examples of AQs with side chains of ≤3 or ≥10 carbons that were not active against CQ-resistant P. falciparum included AQs with terminal amines such as p-Me-piperazine (AQ-128), di-o-Me-piperidine (AQ-88), p-Me-piperidine (AQ-104), N-isopropylamine (AQ-137, AQ-138), N-monoethylamine (AQ-72), N-mono-propylamine (AQ-139), N-isobutylamine (AQ-153), and NH₂ (AQ-50).

Optimization of QSAR models.

Descriptors ranked by mean PLS contributions were added stepwise until q²_prediction reached a maximum. The q², number of components, number of default PLS descriptors, number of optimized PLS descriptors, SDEP, and predictive q² data are shown in Fig. 1 to 3 for CQ-susceptible and -resistant parasites. The final contributions to each QSAR model are shown in Fig. 4 for both CQ-susceptible and CQ-resistant P. falciparum strains.

Fig. 4.

Effects of descriptor optimization on PLS contributions from QSAR analyses of activities against P. falciparum. HB, hydrogen bond.

Optimization of the descriptors had a profound effect (2- to 3-fold) on the q² values for the CQ-susceptible parasites and a much smaller effect for the CQ-resistant parasites (Fig. 1). Of all the AQs tested, only compound 18 was profoundly inactive against the CQ-susceptible parasite, giving a narrower range of IC₅₀ data than that for the CQ-resistant parasite (Table 2). For both data sets, the 2-D fingerprint and 3-D descriptors yielded similar predictive q² values and standard errors of prediction (Fig. 1 to 3). However, application of the same optimizations to the default descriptors without normalization produced little improvement in the QSAR statistics (data not shown). A potential antimalarial drug must diffuse to, and accumulate in, the acidic food vacuole of P. falciparum to be active—crossing several cell membrane-aqueous interfaces. Accumulation in the acidic vacuole is driven by a gradient of ≈2 pH units. Once the putative drug is inside the parasite food vacuole, protonation prevents it from diffusing out, thus allowing pharmacologically active concentrations of drug to accumulate. Since this transport requires both aqueous and lipid solubility, migration, and protonation, we augmented the QSAR descriptors with calculated pK_a, log P, and log D values. However, the addition of these descriptors did not enhance the predictive ability of these models.

Final QSAR model.

The HQSAR, UFP, CoMFA, and CoMSIA methods and descriptors each produced QSAR models with useful predictabilities (Fig. 1 to 3). Inspection of the contributions to the final QSAR models showed that descriptor refinement by iterative optimization increased the importance of QSAR contributions from the minor descriptors in the SYBYL default PLS analyses (Fig. 4) while simultaneously increasing the predictability of the models (Fig. 3). HQSAR and UFP have a computational speed advantage over 3-D QSAR because they do not require either molecular models or alignments. However, the 3-D QSAR models provided more-intuitive visual data for interpreting the QSAR. The CoMSIA method, with its hydrophobicity and hydrogen bond donor and acceptor fields, provided a visually richer model than the CoMFA method, which employs only steric and electrostatic fields, although both produced highly predictive models for CQ-susceptible and -resistant parasites. To highlight the aspects of the QSAR model that favor activity against CQ-resistant versus CQ-susceptible parasites, we generated a differential CoMSIA model by subtracting the PLS coefficients of the susceptible model from those of the resistant model. Figure 5 shows the iterative descriptor-optimized differential CoMSIA QSAR coefficient (coefficient × SD) contours together for 2 AQ analogues active against CQ-resistant parasites (the long-side-chain analogue AQ-40 [black] and the short-side-chain analogue AQ-13 [red]) versus an inactive AQ (CQ [blue]). Differential factors that favor or disfavor activity against CQ-resistant parasites are shown in green and yellow for steric interactions, red and blue for electrostatic interactions, grey and brown for hydrophobic interactions, cyan and light blue for hydrogen bond acceptor interactions, and purple and pink for hydrogen bond donor interactions, respectively. Activity against CQ-resistant parasites is associated with either long (≥10-carbon) or short (≤3-carbon) side chains. A hydrophobic moiety in position 7 and a hydrophilic moiety on the side chain near the quinoline rings are also associated with greater activity against CQ-resistant parasites. In contrast, the model tolerates wide variation in the size and composition of the terminal tertiary amine on the side chain among AQs active against CQ-resistant parasites. The CoMSIA contours were dominated by hydrophobic contributions (brown and gray contours in Fig. 5B). These results suggest that hydrophilicity with short side chains should produce greater activity against CQ-resistant parasites and, thus, that future synthesis of aminoquinolines should explore the inclusion of hydrophilic moieties in this region.

Fig. 5.

Differential optimized CoMSIA QSAR contours for two AQ analogues active against CQ-resistant parasites and one analogue inactive against CQ-resistant parasites. The active long-side-chain and short-side-chain analogues (AQ-40 and AQ-13) are represented in black and red, respectively; the inactive intermediate-length side-chain analogue (CQ) is shown in blue. (A) Steric and electrostatic factors. Favored and disfavored steric interactions appear green and yellow, respectively. Favored and disfavored electrostatic interactions appear red and blue. (B) Hydrophobicity, hydrogen bond (HB) acceptors, and HB donors. For hydrophobicity, favored and disfavored interactions appear white (gray) and brown. For hydrogen bond acceptors, favored and disfavored interactions appear cyan and light blue. For hydrogen bond donors, favored and disfavored interactions appear purple and pink. Favored contributions are shown at a contour percentile of 80%; disfavored interactions are shown at a contour percentile of 20%.

Conclusions.

The results reported here demonstrate that it is possible to develop substantial numbers of AQs active against both CQ-susceptible and CQ-resistant P. falciparum (59 of 108 AQs [55%]). We have identified four features associated with activity against both CQ-resistant and CQ-susceptible parasites (an AQ ring without alkyl substitutions, a halogen at position 7, and two protonatable nitrogens at position 1 and the distal end of the side chain). In contrast, only one feature (a side chain length of ≤3 or ≥10 carbons) was identified as a determinant of activity against CQ-resistant P. falciparum by QSAR and the examination of 2-D structures in relation to antiparasite activity (Fig. 6).

Fig. 6.

Summary of SAR features for activity against CQ-resistant P. falciparum.

The advantage of the iterative descriptor-optimized QSAR strategy reported here is that it tests the abilities of QSAR models to predict the antiparasite activities of unknown AQs. In order to prevent descriptors with greater magnitudes from dominating other descriptors, the algorithm normalizes each descriptor to a mean of 0 with an SD of 1. Iterative retention of the strongest descriptors found by ranking PLS contributions from separate, unvalidated PLS analyses during cross-validation (to identify the optimum number of principal components) permitted refinement of the QSAR models by identifying the optimal number of descriptors necessary to predict the activities of the unknown, test set compounds (Table 3). For both biological data sets, optimization increased the importance of the minor descriptors and increased the values of q² and the predictive q² (Fig. 1 to 3; Table 3).