In silico Design of Phosphonic Arginine and Hydroxamic Acid Inhibitors of Plasmodium falciparum M17 Leucyl Aminopeptidase with Favorable Pharmacokinetic Profile

We virtually design here new subnanomolar range antimalarials, inhibitors of plasmodium falciparum M17 Aminopeptidase (pfA-M17), by means of structure-based molecular design. Complexation QSAR models were elaborated for two training sets (6 methylphosphonic acids (APP) resp. 13 Hydroxamic Acid derivatives (AHO): QSARAPP. resp. QSARAHO) and a linear correlation was established between the computed Gibbs free energies of binding (GFE: ∆∆Gcom) and observed enzyme inhibition constants (Ki ) for each training set: QSARAPP: pKi =−0.1665×∆∆Gcom+7.9581, R=0.97 resp. QSARAHO: pKi =−0.4626×∆∆Gcom+8.1842, R=0.98. The predictive power of the QSAR models was validated with 3D-QSAR pharmacophore generation (PH4): PH4APP: pKi =0.99677×pKi– 0.00457, R=0.99 resp. PH4AHO: pKi =1.02016×pKi– 0.10478, R=0.99. Breakdown of computed pfA-M17:APPs resp. pfA-M17:AHOs interaction energy into each active site residue’s contribution provided additional helpful structural information to design new APP and AHO analogues in a consistent way. In a first step we designed a virtual library (VLAPP resp. VLAHO) from P1 and P’1 substitutions to explore both S1 and S’1 pockets. Further the VLs screened with the 3D-QSAR PH4s and the Ki pred of the best fit hits virtually evaluated with QSARAPP resp. QSARAHO models. This approach combining use of molecular modeling, PH4 and in silico VL screening helpfully provided valuable structural information for the synthesis of novel pfA-M17 inhibitors.


Introduction
Malaria, along with tuberculosis and HIV/AIDS are the major infectious diseases infecting hundred millions people each year at such a level that the United Nations raised their eradication as a Millennium Development Goal (MDG 6: "Combat HIV/AIDS, malaria and other diseases"). Recently UN MDG Report 2015 states: "The global malaria incidence rate has fallen by an estimated 37% and the mortality rate by 58%" [1]. This achievement is mainly attributed to the following facts: "More than 900 million insecticide-treated mosquito nets were delivered to malaria-endemic countries in sub-Saharan Africa between 2004 and 2014 …, //indoor residual spraying, //diagnostic testing and artemisinin-based combination therapies (ACT)" [1].
For one decade (2006 -2016), survival of millions people at risk of malaria infection relies on the strategy of "ACT", the strength of which remains unable to address the unstoppably growing spread of artemisinin-resistant Plasmodium falciparum (pf) [2,3] for long.
One alternative to ACT anti pf resistance strategy is the design of hybrid drugs where two antimalarial moieties are linked into a unique molecule, these moieties known for their confirmed antimalarial potency against two different parasite targets and operating according different mechanisms of action [4,5,6].
Another promising gate was opened through the identification and characterization of new parasite therapeutic targets such as metalloaminopeptidases (MAP), reported to play a crucial role in the parasite viability and survival [7,8,9]. Among them both M1 alanyl and M17 leucyl specific aminopeptidases (pfA-M1, pfA-M17) are validated ones, due to their mediation during the final stage of hemoglobin digestion were they split small peptide fragments into free amino acids and thereafter by their inhibition which is lethal to pf [10,11] they are worth targeting.
Phosphinate dipeptide analogue inhibitors of pfA-M17 reaching 14 µM (compound 1 in Figure 1) were identified first as lead antimalarials [7]. Thereafter phosphonic arginine inhibitors were designed (K i ≥ 11 nM) (compound 2 in Figure  1) probing the S1 pocket at the enzyme active site for a better insight into the interactions and structural requirements for nanomolar range potency [12]. X-rays crystal structure of pfA-M17 in complex with 2 confirmed these interactions particularly the coordination of 2 to the catalytic zinc ion through the aminophosphonate moiety [12]. More recently, the same authors extended their investigations by replacing the phosphonic acid Zinc Binding Group (ZBG) present in 2 with an aminohydroxamic acid analogue of aminophosphonic in order to probe the S1' pocket of pfA-M17 for additional affinity. This novel series allowed access to S1', improving the potency from 11 to 8 nM (compound 3, Figure 1) [13].
The structure activity relationship (SAR) from these studies and the availability of X-rays crystal structure pfA-M17 in complex with 2 and 3 respectively triggers an in silico attempt for the design of pfA-M17 inhibitors filling advantageously both S1 and S1' cavities.
In our present study, starting from high resolution 3D structure of pfA-M17 in complex with 2 and 3 respectively (PDB codes 4K3N and 4R76), we first elaborate two Molecular Mechanics Poisson-Boltzmann (MMPB) QSAR models of complexation of pfA-M17 first with aminophosphonic (APP) and then with aminohydroxamic (AHO) inhibitors respectively by computing the free energy of complexation (∆G com ) for the Training Set (TS) which were correlated with the biological activity to explain about 98% of the variation of K i values by that of ∆G com . Besides a three dimensional quantitative structure-activity relationship (3D-QSAR) pharmacophore model was used to prepare a four-feature pharmacophore (PH4) of the bound APPs and AHOs respectively, which then was matched against the molecular data of the designed analogues to confirm the compliance with each QSAR model. Moreover, the computed enzyme -ligand interaction energy map E int correlates well with K i ; thus allowing us to reach its breakdown to each active site residue contribution. From this last structural information we were able to select P1 and P1' suitable fragments as building blocks for a Virtual Library (VL) of pfA-M17 inhibitors. In order to prevent any toxicity issue and access good pharmacokinetic profile analogues, the VL was focused, prior to any screening, to those compounds the ADMET of which has 0 property descriptors that fall outside the range of values determined for 95% of known drugs out of 24 selected descriptors computed by the QikProp [14]. The predictability of the obtained QSAR models of inhibitor-enzyme binding cross-checked with a PH4 3D-QSAR pharmacophore model was used to screen the VL. The best Hit Fits from the PH4-based virtual screening of the VL have been in silico MMPB evaluated to yield a predicted inhibitory activity reaching the picomolar range for the most potent analogues. falciparum M17 Leucyl Aminopeptidase with Favorable Pharmacokinetic Profile

Training and Validation Sets
The training and validation sets of aminophosphonic and amino-hydroxamic analogues inhibitors of pfA-M17 used in this study were selected from literature [12,13]. The inhibitory potencies of these derivatives cover sufficiently broad range of activity to allow reliable QSAR models to be built (11 ≤ K i ≤ 1000001 nM and 8 ≤ K i ≤ 501000 nM respectively).

Model Building
Molecular models of free enzyme pfA-M17 (E), free inhibitors (I) and the enzyme-inhibitor complexes (E:I), were built from high-resolution crystal structure of the reference complex pfA-M17:APP1 and pfA-M17:AHO4 containing the APP1 and AHO4 respectively [12,13] (Protein Data Bank [25] entry codes 4K3N and 4R76 at a resolution of 2 Å and 2.5 Å respectively) using the graphical user interface available in Insight-II molecular modeling program [26]. The most recent detailed description of this method is in ref. [23,24].

Entropic Term
The most recent detailed description of the ligand vibrational entropic loss upon E:I complex formation and its use procedure in binding affinity calculation is available ref. [23].

Calculation of Binding Affinity
The calculation of binding affinity upon E:I complex formation according to the Molecular Mechanics and Poisson-Boltzmann scheme (MM-PB) is reported fully in refs. [23,24].

Pharmacophore Generation
Pharmacophore modeling procedure as reported in [24] has been used in this work.

Interaction Energy
To calculate the MM interaction energy (E int ) between falciparum M17 Leucyl Aminopeptidase with Favorable Pharmacokinetic Profile enzyme residues and the inhibitor, a protocol available in Discovery Studio 2.5 [27] that computes the non-bonded interactions (van der Waals and electrostatic terms) between defined sets of atoms, was used. The calculations were performed using CFF force field [27] with a dielectric constant of 4. The breakdown of E int into active site residue contributions (presented in% of total E int of pfA-M17 -APP and pfA-M17 -AHO respectively) reveals the significance of individual interactions and allows a comparative analysis, which leads to identification of affinity enhancing and unfavorable APP and AHO substitutions.

ADME-Related Properties
The ADME-related properties treatment in this work has been fully described in [24].

Virtual Library Generation
The analogue model building was performed with Molecular Operating Environment (MOE) program [28]. The library of analogues was enumerated by attaching the R-groups (fragments, building blocks) onto the APP and AHO scaffolds using the Quasar Combi Design module of MOE program [28]. Chemical reagents considered in this study were taken from the directories of chemicals available from the commercial suppliers of chemicals [29]. Each analogue was built as a neutral molecule in MOE program [28] and its molecular geometry was refined by molecular mechanics optimization using Discovery Studio 2.5 [27] smart minimizer with high convergence criteria (energy difference of 10 -4 kcal mol -1 , R. M. S. displacement of 10 -5 Å) and a dielectric constant of 4 using class II consistent force field CFF [27] as described in Molecular Mechanics section.

ADME and Pharmacophore-Based Library Focusing
The library focusing strategy based on ADME-related properties and the pharmacophore-based library focusing procedure is described in ref. [24].

In Silico Screening
The conformer with the highest mapping (four features) in each cluster was selected for virtual screening using the complexation model. The Gibbs Free Energy (GFE) computed upon E:I complex formation in a solvent was evaluated for each of them to reach the new analogues with favorable pharmacokinetic profile since they were ADMET selected during the first focusing step described in section § ADME-based library focusing. As indicated previously the GFE takes account of many components derived from ideal hydrogen bonds, perturbed ionic interactions, lipophilic inter-actions, contributions, due to the freezing of internal degrees of freedom and due to the loss of translational and rotational entropy of the ligand. The ∆∆G com was then used for prediction of pfA-M17 inhibitory potencies (K i pre ) of the focused virtual library of APP and AHO analogues by employing this parameter in a target-specific scoring function. The scoring function, specific for the pfA-M17 receptor of Plasmodium falciparum: pK i pre [pfA-M17] = a × ∆∆G com + b, was parameterized using the QSAR model described above. The workflow in Figure 3 summarizes the methodology of the Computer Aided Design of AHO analogues molecule.

Results and Discussion
A training set of 6 APPs ( Table 1) and another of 13 AHOs and validation set of 2 VHOs (Table 2) were selected from a series of compounds with experimentally determined activities from the same laboratory [12,13]. Their experimental activities (11 ≤ K i ≤ 1 000 001 nM and 8 ≤ K i ≤ 501 000 nM respectively) cover sufficiently large range to serve well for building of a reliable QSAR model of pfA-M17 inhibition.

QSAR Models
The relative Gibbs free energy of the enzyme (E), inhibitor (I), E:I complex formation, equation (6) [24], was computed for the complexes from in situ modification of the template inhibitors APP1 and AHO4 respectively in the binding site of pfA-M17 as described in the Methods section. Tables 3 and 4 respectively list the ∆∆G com of E:I binding and its components, (∆∆H MM , ∆∆G sol , ∆∆TS vib , ∆∆G com = ∆∆H MM +∆∆G sol -∆∆TS vib , see equation (6) [24]). At the bottom of each one of both tables the ratio of predicted and observed inhibition constants (pK i pre /pK i exp ) for the validation set of APPs and AHOs (not included into the training set) are listed. The statistical data of the regression are presented in Figure 4 (A, B and A', B' respectively) and listed in Table 5.
The free energy (∆∆G com ) and its enthalpic contribution (∆∆H MM ) in tables 3 resp. 4 are computed in an approximate way as indicated in methods section. The relevance of the one descriptor QSAR model derived from these energies is assessed through linear regression correlation with the experimental activity data (K i ). As we can see from table 5 the regression coefficient and Fischer F-test values are relatively high indicating a strong correlation between the binding model and the observed inhibitory potencies of the APPs resp. AHOs. The consistency of the model is evaluated through the ratio of predicted and observed inhibition constants ( exp / i pre i pK pK ) for the validation set of AHOs in table 4 (since APPs are in a very short number the whole set totally was reserved for the training set) is close to one confirming the high predictive power of the QSAR models. On the basis of these results, the regression equations and computed ∆∆G com quantities of novel APPs and AHOs analogues, sharing the same binding mode with the training set compounds, can be used to predict their pfA-M17 inhibitory concentrations pre i K . This Computational approach successfully narrows the filter and accelerates the process to new lead compounds compared with traditional synthesis approach. This was observed for peptide and peptidomimetic inhibitors of HIV-1 [16,17] and hepatitis C (HCV) [15] virus proteases design, for combinatorial design of bicyclic thymidine analogues [22] and complexation structure-based design of thymidine analogues [21,23] inhibitors of Mycobacterium Tuberculosis (MTb) thymidine monophosphate kinase, to provide insight into selectivity of peptidomimetic inhibitors with modified statine core for pf Plasmepsin II over human Cathepsin D [18]. Recently a complete process starting from the complexation QSAR model and its derived 3D-QSAR four features Pharmacophore (PH4) model have successfully served to screen a large Virtual Library of 1.6 million compounds reaching one hundred orally bioavailable pyrrolidine carboxamide inhibitors of MTb enoyl acyl carrier protein reductase (InhA) [24].

Free Energy
The binding modes of APPs and AHOs coming from the complexation models are illustrated in 3D depiction in Figure  5 and 6, the last one for the Connolly surface for APP1 and AHO1 respectively. The main interactions at the active sites reported from the reference X-rays structures ( Figure 2) are conserved, involving the key residues and S1 and S1' pockets. In order to get a better insight in these interactions, we build the enzyme -inhibitor interaction energy diagram.

Enzyme -Inhibitor Interaction Energy
Besides the complexation enthalpy, entropy and solvation energy, global interaction energy ∆E int can guide in the design of new potent analogues as exemplified by our design of new thymidine analogues [21] where analysis of breakdown of ∆E int into contribution of each active site residue to the interaction with the inhibitor (∆E int -ResX ) shed light on the appropriate substitutions able to improve stabilizing effect of specific residue (ResX) contribution to ∆E int and to affinity in fine. In this study, a 40% jump is observed in experimental activity from APP1 (11 nM, the most potent APPs) to AHO1 (8 nM, the best AHOs). Since strong correlation between global interaction energy ∆E int and K i exp has been noticed as well for APPs (Tables 6) and AHOs (Tables 7) from statistical data in table 8 and the plots in Figure 5, we can try to get insight into ∆E int breakdown (∆E int -ResX ) as provided in table 9 where ∆E int -ResX are grouped by type of interaction such as H-Bonds (∆E int -ResX:HB ), S1 (∆E int -ResX:S1 ) and S1' (∆E int -ResX:S1' ) pockets filling, for the most active inhibitors, APP1 and AHO1 respectively.
For ∆E int -ResX:HB , 17% of E int is HB-based in APP1 versus 25% in AHO1 for a 47% jump closely in the trend of K i exp . For ∆E int -ResX:S1 , 19% of E int with APP1 versus 15% with AHO1 for a -27% jump, reverse to the trend of K i exp . Analysis of table 9 confirms inability of APP1 to reach S1' pocket while AHO1 does. Indeed, for ∆E int -ResX:S1' , about 10% of E int is related to S1' residues' interaction with APP1 while 22% are involved with AHO1 reaching an exaggerated jump of 120%, nevertheless it's in the trend of K i exp . In the hypothesis that these three groups of interaction could explain the essentials of interaction between pfA-M17 and APP1 and AHO1 respectively, the 46% of E int for APP1 versus 62% for AHO1 corresponds to a 35% jump globally, close to 40% observed experimentally.
On the basis of this result, improvement of interaction with S1 pocket residues as for APPs and with S1' pocket as for AHOs will lead to the design of most active analogues.
In the optimization phase, the highest scoring PH4s according to their probability function cost calculated based on a simulated annealing algorithm, were retained. A total of 10 optimized hypotheses were kept all displaying four features for each pharmacophore.
The features, their internal coordinates and their mapping are displayed in Figure 8 (A, A' and B, B' respectively). The statistical data for the set of hypotheses (costs, RMSD, R) are listed in Tables 10 (for APP) and 11 (for AHO) for both PH4, while the statistical information on regression analysis correlation for the training set between estimated activities and experimental ones are listed in table 12.
The bound conformation of inhibitors at pfA-M17 active site reveals a great deal of information about affinity. The best way of exploiting them consists in generating the Pharmacophore (PH4) of activity. The process is described in the methods section. The reliability of the generated pharmacophore models was then assessed through the calculated cost parameters. As indicated in tables 10 and 11 for APPs resp. AHOs the overall costs ranged from 41.2 (Hypo1) to 48.2 (Hypo10) for APPs and from 492.1 (Hypo1) to 854.7 (Hypo10) for AHOs. The gap between the highest and lowest cost parameter was relatively small and matched well with the homogeneity of the generated hypotheses and the consistency of the training set. For the best APPs' PH4 model the fixed cost (26.3) was lower than the null cost (741.1) by ∆ = 714.8; in the same way the best AHOs' PH4 model fixed cost (23.4) was lower than the null cost (11552.9) by ∆ = 11529.5. This difference represents a chief indicator of the predictability of the PH4 model (∆ > 70 corresponds to a probability higher than 90% that the model represents a valid correlation [21]). In order to be statistically significant the hypotheses have to reach values as close as possible to the fixed cost and as far as possible from the null cost.  The differences ∆ ≥ 699.9 for the set of 10 APPs' PH4 hypotheses and ∆ ≥ 11060.8 for the set of 10 AHOs' PH4 hypotheses confirm the high quality of the pharmacophore models PH4 APP and PH4 AHO . Standard indicators such as the root-mean-square deviation (RMSD) between the hypotheses ranged from 1.009 to 1.709 for APPs and from 8.351 to 10.692 for AHOs while the squared correlation coefficient (R 2 ) occupied an interval from 0.998 to 0.994 for APPs and from 0.98 to 0.96 for AHOs. The PH4 hypothesis with the best RMSD and highest R 2 for APPs and for AHOs respectively was retained for further analysis; they have been selected from the detailed statistical data for the set of hypotheses (costs, RMSD, R 2 ) and listed in Tables 10 resp. 11. The geometries of the Hypo1 pharmacophore of pfA-M17 inhibition are displayed on Figure 8 for APPs (A) and AHOs (A'): features and coordinates between them for each APPs' and AHOs' PH4 respectively and with the best inhibitor APP1 and AHO1 mapped to them. Clearly both PH4 are different from each other and not superimposable: PH4 APP featuring S1 pocket filling while PH4 AHO targets S1 and S1' filling as exemplified by Figure 9. The consistency of the model is reinforced by the position of the PH4 APP with APP1 resp. PH4 AHO with AHO1 mapped to it at the enzyme active site as displayed in Figure  10. From this the mapping to the PH4s gives a reliable indication of the level of enzyme -ligand affinity at the pfA-M17 active site.    Figure 8 (B and B') displays their plot respectively for APPs and AHOs. For the AHOs validation set the ratio pK i pre /pK i exp has been computed for VHO1:1.03 and VHO2:1.04 both relatively close to one which attests the substantial predictive power of this regression for the best PH4 model. The selected PH4 models will serve in virtual screening of APPs and AHOs designed analogues libraries.

New pfA-M17 APP and AHO Inhibitors Analogues
The diversity library design of new APP and AHO analogues inhibitors of pfA-M17 was based on substitutions at one position for APP (in order to better fit S1 pocket at the enzyme active site) and two for AHO (in order to better fit S1 and S1' pockets). The 77 and 134 R-groups used are listed in tables 13 and 16 for APP and AHO respectively.  Table 8.    Table 9. Breakdown of pfA-M17 global interaction energy to each active site residue contribution for H-Bonds, S1 and S1' pockets in the case of the most active inhibitors, APP1 and AHO1 respectively.

Library Design
In order to identify more potent orally bioavailable pfA-M17 inhibitors we have built two virtual libraries; first for new APP with a variety of substitutions in one position, second for new AHO with substitutions in two positions. During the virtual library enumeration the R-groups listed in Table 13 were attached to position R 1 for APPs of the appropriate scaffold to form a combinatorial library of the size: 77 APP analogues. In the same way R 1 and R 2 positions of AHOs have been substituted with R-groups from Table 16, R 1 × R 2 = 84 × 50 = 4,200 AHO analogues. All analogues are matching the substitution pattern of the best inhibitor APP1 and AHO1 respectively.
These diversity libraries of APPs and AHOs analogues were generated from the substituted phosphonic arginine and hydroxamic acids listed in the databases of available chemicals [29]. Nowadays one of the criteria for the design of new antimalarials, in regards to the target population, is their oral bioavailability. In order to design a more focused library of a reduced size and increased content of drug-like and orally bioavailable molecules, we have introduced a set of filters and penalties, which can help to select smaller number of suitable APPs and AHOs that can be submitted to in silico screening. The initial virtual library was thus filtered in an ADME-based focusing step to remove compounds with expected poor oral bioavailability and low drug likeness. For AHOs only analogues with high predicted percentage of human oral absorption (HOA) in the gastrointestinal tract larger than 25% (25% is the highest value of poor oral bioavailability compounds, see legend (r) in table 20) [24] and compounds satisfying the Lipinski's rule of five [30] computed for the entire virtual library using QikProp software [14], were kept. Let's recall that this focusing was not applied for APPs analogues since they were used here to guide S1 pocket interaction improvement only, their ADME profile is already known to be not favorable as table 20 confirms for most of them.

In Silico Screening of Libraries of APPs and AHOs
The libraries of APPs and AHOs analogues were further screened for molecular structures matching to the 3D-QSAR PH4 pharmacophore model Hypo1 of pfA-M17 inhibition by APPs (PH4 APP ) and AHOs (PH4 AHO ) respectively.

Library of APPs
From the set of 77 analogues, only 16 best fitting analogues (PH4 APP hits) have been retained and submitted to screening with help of the complexation QSAR model. Computed Gibbs free energy upon complex formation with pfA-M17 and its components as well as predicted K i pre estimated from the correlation Equation (B), table 5, are given in table 14. For a short majority of new APP analogues, the estimated inhibitory potencies shown in Table 14 are better than that for the most active training set compound APP1 (K i exp = 11 nM [12]). The main goal of the design of APP analogues is to provide structural information, essentially appropriate substituent sup-posed to serve in discovery of new scaffold or to improve existing one such as AHOs' affinity.

Library of AHOs
From the set of 4200 analogues of AHOs 4128 mapped to at least 2 features, 3615 of which mapped to 4 features of the pharmacophore. Out of then, only 17 best fitting analogues (PH4 hits) have been retained and submitted to screening with help of the complexation QSAR model. Computed Gibbs free energy of complex formation with pf A-M17 and its component as well as predicted K i pre estimated from the correlation Equation (D), table 5, are given in tables 17 and 18. For a majority of new AHO analogues, the estimated inhibitory potencies shown in table 17 are better than that for the most active training set compound AHO1 (K i exp = 8 nM [13]).

Analysis of New Inhibitors
In order to identify which substituents on the APPs (Table  13) lead to new inhibitor candidates with the highest predicted potencies against pfA-M17, we have analyzed the frequency of occurrence of R 1 , groups in the APPs and R 1 , R 2 groups in the AHOs PH4 best fitting hit selected from the focused virtual library shown in tables 15 and 18.
For the 17 selected AHOs in R 1 position 14 R-groups are presents once: 6,7,10,12,24,26,44,47,50,54,74,77,81 and 82 except 10 and 44 present twice for fitting S1' pocket. R 2 position (for S1 pocket filling) is less spread since 10 R-groups are presents once: 93, 95, 112, 116, 118, 125 and 133 while 92 and 108 twice and 127 five times among them the best AHO new analogue A-81-127. A-10-112 analogue keeping the methyl extended 4-(1H-pyrazol-1-yl) phenyl group in S1 (as suggested from APPs best analogue) reached a predicted potency of 3.7 nM, in other words, twice more potent than AHO1. In order to improve better the potency, a larger R-group (#127) fitted better in S1 pocket while a CF 3 -phenyl occupies S1' pocket to reach the best-designed AHO analogue A-81-127 displaying a predicted inhibitory potency more than 40 times higher than for the AHO1. Figure 11 displays interactions of best designed APP (A) and AHO (B) analogues respectively with pfA-M17. As one can see, besides S1 and S1' pockets hydrophobic contacts, Hydrogen Bonds were conserved. In Figure 12, the AHO analogue (B) fits better in S1 and S1' pockets than the best active training set AHO1 (see Figure 6 and 7, B) justifying, in a large part, the predicted increase of inhibitory potency of A-81-127. Specifically in Figure 11

ADME Profiles of Designed APPs and AHOs
Two main requirements from WHO about new antimalarials is their low cost and oral bioavailability. About this last one; table 19 displays ADME profile of predicted best active APP analogues. As expected their%HOA (r) are lower than 32%. The range for 95% of drugs (< 25% -poor, > 80% high) clearly indicates almost poor oral bioavailability. In the same way, predicted apparent Caco-2 cell membrane permeability in Boehringer-Ingelheim scale, for which the range for 95% of drugs (< 25 poor, > 500 great) (n) in table 19 confirms their poor membrane permeability as indicated formerly [13]. Let's recall that APPs served to get insight into S1 pocket impact for affinity improvement. Oppositely, as indicated in Table 20, AHOs ADME profile for the best designed new analogues displays%HOA in a higher range for their majority. Only one of them, the less active in this short list, displays poor cell membrane permeability. For almost all of them drug likeness defined as the number of stars, namely the number of property descriptors (from 24 out of the full list of 49 descriptors of QikProp, ver. 3.7, release [14]) that fall outside of the range of values for 95% of known drugs, equals 0. For comparison purpose, the computed ADME profile of some current antimalarials is displayed at the bottom of table 19. As we can see, for almost all antimalarials of the Artemisinin Combined Therapy (ACT) initiative the number of stars is greater than 0 despite their high%HOA.

Conclusions
The studied enzyme pfA-M17 active site pockets S1 and S1' have been investigated separately as reported through various papers. The hydrophobic S1 suitably accommodates hPhe [31] while the larger amphipathic S1'pocket with hydrophilic neck and hydrophobic extended cavity was expected to improve the target inhibition [13]. Fitting R-group in S1' remained one of the main goals of this work. Accordingly our unrestrained effort to retrieve structural information from the crystal structure of pfA-M17 -APP1 and pfA-M17 -AHO1 complexes in order to elaborate reliable one descriptor QSAR models of pfA-M17 inhibition for APPs and AHOs inhibitors using training and validation sets with known inhibitory activities [12,13]. The unique descriptor, namely the computed Gibbs free energies (GFE) upon complex formation, correlated with observed inhibitory potencies. Since GFE is a combined descriptor involving enthalpic gas phase, solvation free energy and entropic contributions, a precise and deeper insight into S1 and S1' pockets filling have been retrieved from the model. A consistent way to crosscheck this structural information was the analysis of interactions between the enzyme active-site residues and the inhibitor. In this regard the breakdown of interaction energy to each active site residue contribution clearly ordered the pockets S1' >> S1 when comparing their affinity with APP1 and AHO1: a key information (Table 9) which directed our effort to design an initial diversity virtual combinatorial library of new analogues to be screened by the pharmacophore models derived from the GFE QSAR. The initial library filtered by a set of ADME-related descriptors to a focused one subsequently was screened by mapping of the analogues to the PH4 AHO pharmacophore to reach a selected library subset of potent and orally bioavailable AHOs. This subset of 17 best virtual hits was submitted to the computation of predicted pfA-M17 inhibitory potencies by the formerly elaborated complexation QSAR AHO model. Starting from a training set led by AHO1 (K i exp =8 nM), our best designed analogues reached predicted K i in the low nanomolar concentration range. The best designed AHOs analogues A.50.108 (K i pre = 0.4 nM), A-81-127 (K i pre = 0.2 nM), A-7-127 (K i pre = 0.4 nM) and A-44-92 (K i pre = 0.3 nM), see tables 17 and 18, are recommended for synthesis and subsequent activity evaluation in pfA-M17 inhibition assays and may lead to a discovery of novel potent orally bioavailable lethal and selective antimalarials.
Since pfA-M17 is a recent target the inhibition of which is at the beginning era, besides the list of 17 best virtual hits, the Structure Activity Relationship (SAR) information we provide about non potent analogues namely the non-suitable R-groups is of interest for medicinal chemists since they will know the analogues that are not worth synthesizing. More detailed such information is available on demand.
At the end pfA-M17 S1 and S1' pockets have to be targeted synergistically: while fitting S1' to improve affinity with pfA-M17 better than S1, the latter subsite, through hydrogen bonding the selective enzyme residue Tyr 493 keeps the inhibitor selective over human hA-M17 as exemplified by our best designed A-81-127.

Conflict of Interest
"The author (s) declare (s) that there is no conflict of interest regarding the publication of this manuscript."     Table 13. R-groups (fragments, building blocks, substituents) used in the design of the initial diversity library of APP.       ; g number of non-trivial (not CX3), non-hindered (not alkene, amide, small ring) rotatable bonds (range for 95% of drugs: 0 -15); h estimated number of hydrogen bonds that would be donated by the solute to water molecules in an aqueous solution. Values are averages taken over a number of configurations, so they can be non-integer (range for 95% of drugs: 0.0 -6.0); i estimated number of hydrogen bonds that would be accepted by the solute from water molecules in an aqueous solution. Values are averages taken over a number of configurations, so they can be non-integer (range for 95% of drugs: 2.0 -20.0); j logarithm of partitioning coefficient between n-octanol and water phases (range for 95% of drugs: -2 -6.5); k logarithm of predicted aqueous solubility, log S. S in mol dm -3 is the concentration of the solute in a saturated solution that is in equilibrium with the crystalline solid (range for 95% of drugs: -6.0 -0.5); l logarithm of predicted binding constant to human serum albumin (range for 95% of drugs: -1.5 -1.5); m logarithm of predicted brain/blood partition coefficient. Note: QikProp predictions are for orally delivered drugs so, for example, dopamine and serotonin are CNS negative because they are too polar to cross the blood-brain barrier (range for 95% of drugs: -3.0 -1. Ki pre was predicted from computed ∆∆Gcom using the regression equation shown in Table 5; q human oral absorption (1 -low, 2 -medium, 3 -high); r percentage of human oral absorption in gastrointestinal tract (<25% -poor, >80% high); (*) star indicating that the property descriptor value falls outside the range of values for 95% of known drugs.