J.-w. Wu et al.
Bioorganic&MedicinalChemistryLettersxxx(xxxx)xxx–xxx
temperature produced corresponding N,N-bis-sulfonylated inter-
mediates 8g-8s, which were in turn monodesulfonylated to furnish the
desired N-pyridinylmethanesulfonamides (9g-9s) by treatment with
10% aqueous NaOH at 0 °C to room temperature. Finally, regioselective
S-alkylation of 5 with 9g-9s in the presence of KI as catalyst and K2CO3
as base in DMF at 130 °C produced the final desired products 1f-1s.
The results of in vitro inhibitory assay of 19 synthesized compounds
(1a-1s) as well as lesinurad as positive control against human URAT1
ploration in this study started with the carboxylic acid bioisosteres of
compound 1. Thus, we replaced the carboxylic acid moiety with those
bioisosteres explored (1a-1g), only the N-(pyridine-3-yl)sulfonamide
nucleus is more potent than the parent carboxylic acid as indicated by
the 2.9-fold improvement in IC50 value of 1g vs 1 (IC50 = 0.032 μM for
1g vs 0.094 μM for 1).
correlation coefficient (R2), lowest total cost (97.988), highest cost
difference (72.207), and a low root mean squared deviation (RMSD)
values. The correlation coefficient value (R2) and RMSD values for
Hypo 1 were 0.9477 and 1.6004, respectively.
The predictive ability of Hypo 1 on the training set compounds was
shown in Table 4. In accordance with the Hypo1 activity values, 17 out
of 19 compounds in the training set were predicted within their ex-
perimental activity scale except 1b and 1f. The error value is the ratio
between the estimated and experimental activities. An absolute value of
error below 10 indicates that the estimated activity was below one
order of magnitude. None of the 19 training-set compounds had an
absolute value of error above 4. Fit value can provide information in
understanding the chemical meaning of the pharmacophore hypothesis
by overlaping the molecule chemical features in the pharmacophore.
The fit value of the most active compound (1g) of the training set was
10.45 while the least active compound (1d) was 7.27. All these datas
suggested that Hypo 1 was a reliable model with high predictive ability.
The pharmacophore mapping of the most active (1g) compound, the
parent compound (1) and the least active compound (1d) were shown
in Figure 3A, B and C, respectively. Obviously, compound 1g mapped
well on three hypothetical features, while compound 1d did not map on
to two of the hypothetical features, particularly HBA and HBD, sig-
nifying the importance of these features. Moreover, the parent com-
pound 1 mapped on two hypothetical features (i.e. HBD and HY fea-
tures). As shown in Figure 3A and B, both 1 and 1g mapped the feature
bioisostere. By comparison with compound 1, 1g had proved to be one
of the best carboxylic acid bioisosteres. Therefore, Hypo1 is a reliable
model that accurately estimates the experimental activity of the
training-set compounds.
In the second round of SAR exploration, we in turn focused on the
fine tuning of the substituents around N-(pyridine-3-yl)sulfonamide
nucleus discovered in the first round. Thus, compounds 1h-1s were
designed. Unfortunately, none of these compounds displayed more
potent URAT1 inhibitory activity compared with 1g, indicating that the
N-(pyridine-3-yl)sulfonamide can not tolerate any substitution.
In summary, two round of SAR exploration led to the discovery of a
highly active novel inhibitor 1g, which was 225- and 3-fold more po-
tent than the parent compounds lesinurad and 1 (IC50 = 0.032 μM for
1g vs 7.20 μM for lesinurad and 0.094 μM for 1), respectively. However,
the IC50 values for these 1,2,4-triazole-5-substituted carboxylic acid
bioisosteres (1a-1s) ranged from 0.032 μM to 365 μM, which span four
orders of magnitude. With the aim to obtain additional information of
the relationship between structure and activity and to discover more
active URAT1 inhibitors, 3D-QSAR studies were carried out.
Finally, the best pharmacophore model, Hypo1, was further vali-
dated by cost analysis, Fischer randomization and leave-one-out
methods.
Based on the biological activity values (IC50) of the investigated
compounds (1a-1s), 3D-QSAR pharmacophore models were established
within Accelrys Discovery Studio 2.5 software using HypoGen32–34
module. As the training set compounds, 1a-1s were in accord with the
selection rule in HypoGen: at least 16 diverse inhibitors (a total of 19
compounds including 1a-1s) to ensure statistical significance and the
biological activities span at least four orders of magnitude
(0.032 μM ∼ 365 μM). All compounds in training set were classified
into four activity scales: highly active (IC50 ≤ 200 nM, +++), active
The quality of a pharmacophore model is evaluated primarily by
using two theoretical cost calculations that are represented in bit units.
One is the “null cost” representing the highest cost of a pharmacophore
model with no features; this value estimates every activity as the
averaged activity data from the training-set compounds. The second
cost is the “fixed cost,” which represents the simplest model that fits all
the data perfectly. The total cost should always be far from the null cost
and near the fixed cost when developing a meaningful model. The null
cost of the ten established pharmacophore models was 170.195 bits,
and the fixed cost was 62.375 bits. The analysis of ten generated
pharmacophore models indicated that the total cost value for Hypo 1 is
the closest to the fixed cost value than other models. The cost difference
between the null cost and total cost value of Hypo 1 is 72.207 bits
(Table 3). So Hypo 1 exhibited strong predictive capacity (If the dif-
ference is > 60 bits, the model has brilliant ability to fit all the data;
when the difference is 40–60 bits, there is a 75–90% chance that it
represents the data well; if the difference is under 40, it does not fit all
the maximum threshold of 17.40. Cost analysis confirms that the hy-
potheses had rational correlations.
(200 nM < IC50 ≤ 2 μM,
++),
moderately
active
(2 μM <
IC50 ≤ 20 μM, +), inactive (IC50 > 20 μM, -). This classification is
helpful to generate and evaluate the pharmacophore model with abroad
range of activities quickly.35 Structures of the training set (1a-1s) were
imported into Discovery Studio and energy minimized to the closest local
minimum using the generalized CHARMM force field by Minimized Li-
gands algorithm. Meanwhile, conformational ensembles of the local
minimized structure with a maximum limit of 255 conformers per
molecule were generated within Conformation Generation module by
best methods and using an energy threshold of 20 kcal/mol.36 A list of
pharmacophore feature types for the training set, such as hydrogen
bond acceptors (HBA), hydrogen bond doners (HBD), hydrophobic (HY)
and ring aromatic (RA), were represented in Edit and Cluster Features
module. The minimum numbers of each feature type were set to 1 and
the maximum numbers of them were set to 3. With the help of 3D QSAR
Pharmacophore Generation tool, the pharmacophore models were es-
tablished by three major steps: the constructive phase (generation of the
common pharmacophores among the active training molecules), the
subtractive phase (elimination of the pharmacophores that are common
to most of the inactive molecules) and the optimization phase (attempt
to improve the score by applies small perturbation to the pharmaco-
phores created in constructive and subtractive phases).37
Secondly, to verify whether the pharmacophore model had a strong
correlation between the structure of the training set compounds (1a-1s)
and the biological activities, a Fischer randomization test was carried
out.41 The parameters of the randomized pharmacophore model were
the same as those used to generate the original one. 19 different ran-
domizations were generated to achieve a 95% confidence level that the
best pharmacophore Hypo 1 was not generated by chance.42 The total
costs of Hypo 1 and 19 randomized pharmacophore models were shown
in Figure 4. Apparently, none of these 19 new hypotheses had lower
cost values than the true hypothesis. This test shows that the Hypo 1
was not generated by chance, and that there is a probability of at least a
95% that it shows a valid correlation between chemical structure and
URAT1 inhibitory activity.43
As shown in Table 3, HypoGen provided the top 10 scoring hy-
potheses, Hypo1 was the best one of them because it had the highest
4
Wu, Jing-wei
