Assessment of Nonnucleoside Inhibitors Binding to HIV-1 Reverse Transcriptase Using HYDE Scoring

In this study, 48 inhibitors were docked to 107 allosteric centers of human immunodeficiency virus 1 (HIV-1) reverse transcriptase from the Protein Data Bank (PDB). Based on the average binding scores, quantitative structure-activity relationship (QSAR) equations were constructed in order to elucidate directions of further development in the design of inhibitors. Such developments, informed by structural data, must have a focus on activity against mutated forms of the enzyme, which are the cause of the emergence of multidrug-resistant viral strains. Docking studies employed the HYDE scoring function. Two types of QSARs have been considered: One based on topological descriptors and the other on structural fragments of the inhibitors. Both methods gave similar results, indicating substructures favoring binding to mutated forms of the enzyme.


Introduction
Approximately 77 million people have become infected with human immunodeficiency virus-1 (HIV-1) since the first cases were reported in 1981 [1]. By the end of 2016, the World Health Organization estimated that approximately 36.7 million people were currently living with HIV-1 infection, including 1.8 million newly infected people in 2016 [2]. Over the past two decades, significant advances have been mounted to address the epidemic and significant progress has been made. However, because of toxicity, the rapid emergence of resistance, detrimental side effects caused by long-term drug treatment and issues associated with drug tolerability, there remains a need for new antiviral agents [3,4].
Due to well-characterized mechanisms of action and abundant structure information, reverse transcriptase (RT)-one of three essential enzymes encoded in the HIV-1 genome-represents a successful target for chemotherapeutic intervention [5,6]. RT inhibitors may be divided into nucleoside/nucleotide RT inhibitors (NRTIs/NtRTIs) and non-nucleoside RT inhibitors (NNRTIs) [7]. These inhibitors are widely used in highly active antiretroviral therapy (HAART) regimens, owing to their potent activity, high selectivity, and favorable pharmacokinetics [8]. NNRTIs in particular act by disrupting the normal functions of RT via binding to the NNRTI binding pocket (NNIBP), close to the polymerase active site [7]. So far, six NNRTIs have been licensed for use [9]. Nevirapine, delavirdine, and efavirenz are the first-generation NNRTIs with high potency [10]. However, central nervous system side effects, hepatotoxicity, poor resistance profile and low genetic barriers for viral resistance are the major treatment-limiting factors in their clinical application [9,11]. In particular, single mutants K103N, Y181C and double mutation K103N/Y181C are prevalent in clinical HIV-1 isolates [12,13]. Even

Results and Discussion
The procedure used in docking followed that previously reported for the FlexX scoring function [32]. In short, 48 ligands were docked to the allosteric cavity of 107 HIV-1 RT enzymes (available from the PDB [33]), and an average binding score for a given ligand was obtained separately for the wild type (wt) and mutated enzyme. Poses (Table 1) were inspected for correct orientation within the cavity. Average binding scores obtained for the wt and mutated enzymes were then compared. As illustrated in Figure 1, binding of either the wt enzyme or its mutated forms was generally random and did not correlate with the calculated binding energy. This observation indicates that activity against mutated forms of the enzyme is not governed by the strength of binding. A sole exception to the above observation is the result obtained for the DJZ ligand, which binds to the mutated forms of HIV-1 RT much stronger than to its wt structures.  Ligands that bind to the allosteric site impair enzyme action by a wedge mechanism; they hinder domain mobility, opening and closing access to the active site. Final allosteric site architecture is achieved upon binding of the ligand. This flexibility and the possible clash between protein and ligand was accounted for by using a large overlap volume (100 Å 3 ). Lack of systematic differences between binding to the wt versus the mutated enzyme indicates that activity against mutants is connected with the structural features of the ligand, rather than the energy of their binding. Interactions within the allosteric site are mostly associated with van der Waals forces and, to a lesser extent, hydrogen bonding [33]. Ligand H18 exhibited the strongest binding to all forms of the enzyme, however, as mentioned above, it is a DJZ ligand that shows the largest change in binding The signed difference in averaged HYDE score between binding to wild type enzyme vs. binding to mutants as a function of the binding score. Ligands that bind to the allosteric site impair enzyme action by a wedge mechanism; they hinder domain mobility, opening and closing access to the active site. Final allosteric site architecture is achieved upon binding of the ligand. This flexibility and the possible clash between protein and ligand was accounted for by using a large overlap volume (100 Å 3 ). Lack of systematic differences between binding to the wt versus the mutated enzyme indicates that activity against mutants is connected with the structural features of the ligand, rather than the energy of their binding. Interactions within the allosteric site are mostly associated with van der Waals forces and, to a lesser extent, hydrogen bonding [33]. Ligand H18 exhibited the strongest binding to all forms of the enzyme, however, as mentioned above, it is a DJZ ligand that shows the largest change in binding energy when moving from the wt to the mutated enzymes. Its success seems to come from hydrogen bonding to lysine 101 rather than frequently mutated lysine 103. Furthermore, its orientation within the pocket is improved by a stronger and stiffer hydrogen bond from piperidine to histidine, compared to a water molecule, and an additional hydrogen bond to the pyrimidine ring (compare panels of Figure 2). Pharmaceuticals 2019, 12, 64 4 of 14 energy when moving from the wt to the mutated enzymes. Its success seems to come from hydrogen bonding to lysine 101 rather than frequently mutated lysine 103. Furthermore, its orientation within the pocket is improved by a stronger and stiffer hydrogen bond from piperidine to histidine, compared to a water molecule, and an additional hydrogen bond to the pyrimidine ring (compare panels of Figure 2). The results obtained from docking were used in the QSAR studies. In order to understand the very general structural features affecting the binding affinity, two types of QSAR models were used. The first one was based on descriptors that had clear "chemical interpretation", giving suggestions for composition and physico-chemical properties in the new compound design. In this approach, we used the algorithms and the library of descriptors implemented in the SCIGRESS Suite software [34].
The best relationships were identified for both enzyme types. In both cases, they included the same four descriptors out of 383 evaluated descriptors: Both equations exhibited similar statistics characterized by an r 2 equal to 0.8028 and 0.7817 for the wt and mutant enzyme, respectively, as well as a F-ratio of 41.7281 and 36.7059, respectively. The standard deviation in the error predicted by leave-one-out cross-validation and associated r 2 were 3.8481 and 0.7419 for wt enzyme and 4.0614 and 0.7604 for mutants, respectively, suggesting that the equations were stable enough to be used for predictions. The quality of predictions is also illustrated in Figures 3 and 4, in which the energies of binding predicted by the QSAR are plotted against the values obtained from docking (blue circles). Yellow points represent results obtained by leave-oneout cross validation (CV).
As may be seen, both relationships are very similar; an increase of numerical values of the descriptors leads to a decrease in the affinity. Relative weights of the normalized coefficients of the QSAR equations collected in Table 2 imply similar influence on the binding affinity for all four parameters. Physico-chemical interpretation of the descriptors suggests that the inhibitors should be large, uniformly hydrophobic, and they should have a small number of carboxyl groups and nitrogen atoms. The results obtained from docking were used in the QSAR studies. In order to understand the very general structural features affecting the binding affinity, two types of QSAR models were used. The first one was based on descriptors that had clear "chemical interpretation", giving suggestions for composition and physico-chemical properties in the new compound design. In this approach, we used the algorithms and the library of descriptors implemented in the SCIGRESS Suite software [34].
The best relationships were identified for both enzyme types. In both cases, they included the same four descriptors out of 383 evaluated descriptors: Both equations exhibited similar statistics characterized by an r 2 equal to 0.8028 and 0.7817 for the wt and mutant enzyme, respectively, as well as a F-ratio of 41.7281 and 36.7059, respectively. The standard deviation in the error predicted by leave-one-out cross-validation and associated r 2 were 3.8481 and 0.7419 for wt enzyme and 4.0614 and 0.7604 for mutants, respectively, suggesting that the equations were stable enough to be used for predictions. The quality of predictions is also illustrated in Figures 3 and 4, in which the energies of binding predicted by the QSAR are plotted against the values obtained from docking (blue circles). Yellow points represent results obtained by leave-one-out cross validation (CV).
As may be seen, both relationships are very similar; an increase of numerical values of the descriptors leads to a decrease in the affinity. Relative weights of the normalized coefficients of the QSAR equations collected in Table 2 imply similar influence on the binding affinity for all four parameters. Physico-chemical interpretation of the descriptors suggests that the inhibitors should be large, uniformly hydrophobic, and they should have a small number of carboxyl groups and nitrogen atoms. The second attempt aimed at creating a QSAR was based on a fragment contribution approach using common substructures present in the training set. For this purpose, the ADMEWORKS ModelBuilder was used [35,36]. Due to the size of the training set, sets of 6 descriptors were chosen, as in our previous work [32]. The substructures contained in both sets are shown in Table 3.
The obtained r 2 value of less than 70% in the leave-one-out cross-validation did not encourage its use for direct prediction of the unknown compounds. Thus, it was rather the sign of the linear regression equations weight coefficients that were considered. These are a measure of the influence of a given substructure's contribution to the overall activity; negative values indicate improvement in binding, while positive values suggest that the corresponding substructures should be eliminated or their presence minimized. The obtained results are summarized in Table 4, where substructures with the positive contribution to binding are presented on a green background, while those which should be avoided are distinguished by a red background.  The second attempt aimed at creating a QSAR was based on a fragment contribution approach using common substructures present in the training set. For this purpose, the ADMEWORKS ModelBuilder was used [35,36]. Due to the size of the training set, sets of 6 descriptors were chosen, as in our previous work [32]. The substructures contained in both sets are shown in Table 3.
The obtained r 2 value of less than 70% in the leave-one-out cross-validation did not encourage its use for direct prediction of the unknown compounds. Thus, it was rather the sign of the linear regression equations weight coefficients that were considered. These are a measure of the influence of a given substructure's contribution to the overall activity; negative values indicate improvement in binding, while positive values suggest that the corresponding substructures should be eliminated or their presence minimized. The obtained results are summarized in Table 4, where substructures with the positive contribution to binding are presented on a green background, while those which should be avoided are distinguished by a red background.   As illustrated by Figures 5 and 6, the quality of predictions is again very good. The exceptional behavior of the DJZ inhibitor, noticed in docking results, is revealed by the significant difference between the prediction based on the full set of studied compounds and that of the leave-one-out result (point at -42.5 in Figure 6), suggesting its substantial difference in the mode of action or properties from the rest of the inhibitors. Also in agreement with the descriptor-based approach was As illustrated by Figures 5 and 6, the quality of predictions is again very good. The exceptional behavior of the DJZ inhibitor, noticed in docking results, is revealed by the significant difference between the prediction based on the full set of studied compounds and that of the leave-one-out result (point at -42.5 in Figure 6), suggesting its substantial difference in the mode of action or properties from the rest of the inhibitors. Also in agreement with the descriptor-based approach was the finding that the presence of both tertiary nitrogen atoms and aromatic rings should be minimized. A strong positive component of substructure 3 that correlates with the overall size of the molecule was again in line with the conclusion obtained from the 1/bond count descriptor in the relationship obtained by SCIGRESS. Specific activity against mutations appeared to come from two specific substructures, labeled 44 and 53 in Table 3. The presence of the carboxyl groups, on the other hand, appeared to lower the binding activity of the inhibitors.
Docking was performed using the HYDE scoring function [37] as implemented in the LeadIT software package [38]. The receptor was prepared using the graphical interface of the package. Both protein chains were selected, and the binding site was defined to include residues within 6.5 Å radius around the native ligand. A library of ligands was imported from the mol2 file. Protonation states corresponding to the aqueous solution were used. Soft docking (allowing for volume overlap up to 100 Å 3 ) was performed. For ligand base placement, the default hybrid Enthalpy and Entropy strategy was used. The clash factor was set to 0.6. Default values were used for other parameters. Scores of the top-ranked poses of each ligand were averaged for a given enzyme structure.
QSAR analysis was performed in two different modes. In both cases the training set included 46 structures, the end-point was the average docking score obtained from HYDE calculations after exclusion of results for AC7 and QO9. The size of the data set was too small to allow for splitting it into training and validation sets, so an internal leave-one-out cross-validation approach was chosen instead.
In the first approach, based on classical descriptors, several models were created using the Complete Topological QSAR method and regression equation, created by feature selection with the enhanced replacement method [ERM], as implemented in SCIGRESS Suite software [35]. In the second approach, QSAR was based on molecular fragments contribution. The common substructures were extracted from the training set. Altogether, the 96 most common fragments in all the dataset produced by ModelBuilder's extract substructures feature were used. The six most important features for each enzyme type were selected using the particle swarm optimization algorithm, with 1000 individuals run until convergence was observed (about 3000 iterations). The six most often used were selected as descriptors. The final model was created by applying multiple linear regression. . The receptor was prepared using the graphical interface of the package. Both protein chains were selected, and the binding site was defined to include residues within 6.5 Å radius around the native ligand. A library of ligands was imported from the mol2 file. Protonation states corresponding to the aqueous solution were used. Soft docking (allowing for volume overlap up to 100 Å 3 ) was performed. For ligand base placement, the default hybrid Enthalpy and Entropy strategy was used. The clash factor was set to 0.6. Default values were used for other parameters. Scores of the top-ranked poses of each ligand were averaged for a given enzyme structure.
QSAR analysis was performed in two different modes. In both cases the training set included 46 structures, the end-point was the average docking score obtained from HYDE calculations after exclusion of results for AC7 and QO9. The size of the data set was too small to allow for splitting it into training and validation sets, so an internal leave-one-out cross-validation approach was chosen instead.
In the first approach, based on classical descriptors, several models were created using the Complete Topological QSAR method and regression equation, created by feature selection with the enhanced replacement method [ERM], as implemented in SCIGRESS Suite software [35]. In the second approach, QSAR was based on molecular fragments contribution. The common substructures were extracted from the training set. Altogether, the 96 most common fragments in all the dataset produced by ModelBuilder's extract substructures feature were used. The six most important features for each enzyme type were selected using the particle swarm optimization algorithm, with 1000 individuals run until convergence was observed (about 3000 iterations). The six most often used were selected as descriptors. The final model was created by applying multiple linear regression. Docking was performed using the HYDE scoring function [37] as implemented in the LeadIT software package [38]. The receptor was prepared using the graphical interface of the package. Both protein chains were selected, and the binding site was defined to include residues within 6.5 Å radius around the native ligand. A library of ligands was imported from the mol2 file. Protonation states corresponding to the aqueous solution were used. Soft docking (allowing for volume overlap up to 100 Å 3 ) was performed. For ligand base placement, the default hybrid Enthalpy and Entropy strategy was used. The clash factor was set to 0.6. Default values were used for other parameters. Scores of the top-ranked poses of each ligand were averaged for a given enzyme structure.
QSAR analysis was performed in two different modes. In both cases the training set included 46 structures, the end-point was the average docking score obtained from HYDE calculations after exclusion of results for AC7 and QO9. The size of the data set was too small to allow for splitting it into training and validation sets, so an internal leave-one-out cross-validation approach was chosen instead.
In the first approach, based on classical descriptors, several models were created using the Complete Topological QSAR method and regression equation, created by feature selection with the enhanced replacement method [ERM], as implemented in SCIGRESS Suite software [35]. In the second approach, QSAR was based on molecular fragments contribution. The common substructures were extracted from the training set. Altogether, the 96 most common fragments in all the dataset produced by ModelBuilder's extract substructures feature were used. The six most important features for each enzyme type were selected using the particle swarm optimization algorithm, with 1000 individuals run until convergence was observed (about 3000 iterations). The six most often used were selected as descriptors. The final model was created by applying multiple linear regression. Docking was performed using the HYDE scoring function [37] as implemented in the LeadIT software package [38]. The receptor was prepared using the graphical interface of the package. Both protein chains were selected, and the binding site was defined to include residues within 6.5 Å radius around the native ligand. A library of ligands was imported from the mol2 file. Protonation states corresponding to the aqueous solution were used. Soft docking (allowing for volume overlap up to 100 Å 3 ) was performed. For ligand base placement, the default hybrid Enthalpy and Entropy strategy was used. The clash factor was set to 0.6. Default values were used for other parameters. Scores of the top-ranked poses of each ligand were averaged for a given enzyme structure.
QSAR analysis was performed in two different modes. In both cases the training set included 46 structures, the end-point was the average docking score obtained from HYDE calculations after exclusion of results for AC7 and QO9. The size of the data set was too small to allow for splitting it into training and validation sets, so an internal leave-one-out cross-validation approach was chosen instead.
In the first approach, based on classical descriptors, several models were created using the Complete Topological QSAR method and regression equation, created by feature selection with the enhanced replacement method [ERM], as implemented in SCIGRESS Suite software [35]. In the second approach, QSAR was based on molecular fragments contribution. The common substructures were extracted from the training set. Altogether, the 96 most common fragments in all the dataset produced by ModelBuilder's extract substructures feature were used. The six most important features for each enzyme type were selected using the particle swarm optimization algorithm, with 1000 individuals run until convergence was observed (about 3000 iterations). The six most often used were selected as descriptors. The final model was created by applying multiple linear regression. Docking was performed using the HYDE scoring function [37] as implemented in the LeadIT software package [38]. The receptor was prepared using the graphical interface of the package. Both protein chains were selected, and the binding site was defined to include residues within 6.5 Å radius around the native ligand. A library of ligands was imported from the mol2 file. Protonation states corresponding to the aqueous solution were used. Soft docking (allowing for volume overlap up to 100 Å 3 ) was performed. For ligand base placement, the default hybrid Enthalpy and Entropy strategy was used. The clash factor was set to 0.6. Default values were used for other parameters. Scores of the top-ranked poses of each ligand were averaged for a given enzyme structure.
QSAR analysis was performed in two different modes. In both cases the training set included 46 structures, the end-point was the average docking score obtained from HYDE calculations after exclusion of results for AC7 and QO9. The size of the data set was too small to allow for splitting it into training and validation sets, so an internal leave-one-out cross-validation approach was chosen instead.
In the first approach, based on classical descriptors, several models were created using the Complete Topological QSAR method and regression equation, created by feature selection with the enhanced replacement method [ERM], as implemented in SCIGRESS Suite software [35]. In the second approach, QSAR was based on molecular fragments contribution. The common substructures were extracted from the training set. Altogether, the 96 most common fragments in all the dataset produced by ModelBuilder's extract substructures feature were used. The six most important features for each enzyme type were selected using the particle swarm optimization algorithm, with 1000 individuals run until convergence was observed (about 3000 iterations). The six most often used were selected as descriptors. The final model was created by applying multiple linear regression. Docking was performed using the HYDE scoring function [37] as implemented in the LeadIT software package [38]. The receptor was prepared using the graphical interface of the package. Both protein chains were selected, and the binding site was defined to include residues within 6.5 Å radius around the native ligand. A library of ligands was imported from the mol2 file. Protonation states corresponding to the aqueous solution were used. Soft docking (allowing for volume overlap up to 100 Å 3 ) was performed. For ligand base placement, the default hybrid Enthalpy and Entropy strategy was used. The clash factor was set to 0.6. Default values were used for other parameters. Scores of the top-ranked poses of each ligand were averaged for a given enzyme structure.
QSAR analysis was performed in two different modes. In both cases the training set included 46 structures, the end-point was the average docking score obtained from HYDE calculations after exclusion of results for AC7 and QO9. The size of the data set was too small to allow for splitting it into training and validation sets, so an internal leave-one-out cross-validation approach was chosen instead.
In the first approach, based on classical descriptors, several models were created using the Complete Topological QSAR method and regression equation, created by feature selection with the enhanced replacement method [ERM], as implemented in SCIGRESS Suite software [35]. In the second approach, QSAR was based on molecular fragments contribution. The common substructures were extracted from the training set. Altogether, the 96 most common fragments in all the dataset produced by ModelBuilder's extract substructures feature were used. The six most important features for each enzyme type were selected using the particle swarm optimization algorithm, with 1000 individuals run until convergence was observed (about 3000 iterations). The six most often used were selected as descriptors. The final model was created by applying multiple linear regression. Docking was performed using the HYDE scoring function [37] as implemented in the LeadIT software package [38]. The receptor was prepared using the graphical interface of the package. Both protein chains were selected, and the binding site was defined to include residues within 6.5 Å radius around the native ligand. A library of ligands was imported from the mol2 file. Protonation states corresponding to the aqueous solution were used. Soft docking (allowing for volume overlap up to 100 Å 3 ) was performed. For ligand base placement, the default hybrid Enthalpy and Entropy strategy was used. The clash factor was set to 0.6. Default values were used for other parameters. Scores of the top-ranked poses of each ligand were averaged for a given enzyme structure.
QSAR analysis was performed in two different modes. In both cases the training set included 46 structures, the end-point was the average docking score obtained from HYDE calculations after exclusion of results for AC7 and QO9. The size of the data set was too small to allow for splitting it into training and validation sets, so an internal leave-one-out cross-validation approach was chosen instead.
In the first approach, based on classical descriptors, several models were created using the Complete Topological QSAR method and regression equation, created by feature selection with the enhanced replacement method [ERM], as implemented in SCIGRESS Suite software [35]. In the second approach, QSAR was based on molecular fragments contribution. The common substructures were extracted from the training set. Altogether, the 96 most common fragments in all the dataset produced by ModelBuilder's extract substructures feature were used. The six most important features for each enzyme type were selected using the particle swarm optimization algorithm, with 1000 individuals run until convergence was observed (about 3000 iterations). The six most often used were selected as descriptors. The final model was created by applying multiple linear regression.

Conclusions
The major aim of the studies presented herein was the identification of features improving activities of NNTRIs in general and in particular those specifically suited for the most common mutations of HIV-1 RT. The interpretation of the results of the fragment-based QSAR led to similar conclusions as those made for the topological QSAR: The presence of tertiary nitrogen atoms and nitriles should be avoided, while large, uniformly hydrophobic molecules are preferable. The presence of carboxyl groups appears to lower the binding affinity of the inhibitors to mutated forms of the enzyme, while the presence of moieties 44 and 53 (illustrated in Table 3) improves it.
Funding: This research was funded by grant number 2011/02/A/ST4/00246 from the Polish National Science Centre (NCN).

Conflicts of Interest:
The authors declare no conflict of interest.

Conclusions
The major aim of the studies presented herein was the identification of features improving activities of NNTRIs in general and in particular those specifically suited for the most common mutations of HIV-1 RT. The interpretation of the results of the fragment-based QSAR led to similar conclusions as those made for the topological QSAR: The presence of tertiary nitrogen atoms and nitriles should be avoided, while large, uniformly hydrophobic molecules are preferable. The presence of carboxyl groups appears to lower the binding affinity of the inhibitors to mutated forms of the enzyme, while the presence of moieties 44 and 53 (illustrated in Table 3) improves it.

Conclusions
The major aim of the studies presented herein was the identification of features improving activities of NNTRIs in general and in particular those specifically suited for the most common mutations of HIV-1 RT. The interpretation of the results of the fragment-based QSAR led to similar conclusions as those made for the topological QSAR: The presence of tertiary nitrogen atoms and nitriles should be avoided, while large, uniformly hydrophobic molecules are preferable. The presence of carboxyl groups appears to lower the binding affinity of the inhibitors to mutated forms of the enzyme, while the presence of moieties 44 and 53 (illustrated in Table 3

Conclusions
The major aim of the studies presented herein was the identification of features improving activities of NNTRIs in general and in particular those specifically suited for the most common mutations of HIV-1 RT. The interpretation of the results of the fragment-based QSAR led to similar conclusions as those made for the topological QSAR: The presence of tertiary nitrogen atoms and nitriles should be avoided, while large, uniformly hydrophobic molecules are preferable. The presence of carboxyl groups appears to lower the binding affinity of the inhibitors to mutated forms of the enzyme, while the presence of moieties 44 and 53 (illustrated in Table 3) improves it.

Conclusions
The major aim of the studies presented herein was the identification of features improving activities of NNTRIs in general and in particular those specifically suited for the most common mutations of HIV-1 RT. The interpretation of the results of the fragment-based QSAR led to similar conclusions as those made for the topological QSAR: The presence of tertiary nitrogen atoms and nitriles should be avoided, while large, uniformly hydrophobic molecules are preferable. The presence of carboxyl groups appears to lower the binding affinity of the inhibitors to mutated forms of the enzyme, while the presence of moieties 44 and 53 (illustrated in Table 3) improves it. GFA YKN NNC H20 AAP G73 5DV ZZE QO9

Conclusions
The major aim of the studies presented herein was the identification of features improving activities of NNTRIs in general and in particular those specifically suited for the most common mutations of HIV-1 RT. The interpretation of the results of the fragment-based QSAR led to similar conclusions as those made for the topological QSAR: The presence of tertiary nitrogen atoms and nitriles should be avoided, while large, uniformly hydrophobic molecules are preferable. The presence of carboxyl groups appears to lower the binding affinity of the inhibitors to mutated forms of the enzyme, while the presence of moieties 44 and 53 (illustrated in Table 3) improves it.

Conclusions
The major aim of the studies presented herein was the identification of features improving activities of NNTRIs in general and in particular those specifically suited for the most common mutations of HIV-1 RT. The interpretation of the results of the fragment-based QSAR led to similar conclusions as those made for the topological QSAR: The presence of tertiary nitrogen atoms and nitriles should be avoided, while large, uniformly hydrophobic molecules are preferable. The presence of carboxyl groups appears to lower the binding affinity of the inhibitors to mutated forms of the enzyme, while the presence of moieties 44 and 53 (illustrated in Table 3) improves it.