E. Perola et al. / Bioorg. Med. Chem. Lett. 24 (2014) 2177–2181
2179
dictated by two factors: (1) serum shift provides a direct measure
of the impact of protein binding on antibacterial activity; (2) at the
high levels of protein binding observed for our series, small varia-
tions in protein binding lead to relatively large variations in serum
shift, thus making measurements of the latter more discriminant.
As an example, if compound A is 99.5% protein bound and com-
pound B is 98% protein bound, the free fraction of compound B is
4-fold higher than the free fraction of compound A, and the corre-
sponding serum shift will be 4-fold lower. A 4-fold difference in
serum shift can be reliably measured, while the difference between
99.5% and 98% in protein binding is likely within experimental er-
ror. Serum shift data was available for 338 compounds sharing the
substructure in Figure 1 below.
Since the objective was to build a model for neutral compounds,
all the charged compounds (mostly basic amines) were removed,
and the initial attempts at model building were made on the
remaining set of 202 compounds. Additional experiments had
shown that binding to human serum albumin was the dominant
cause of the observed serum shift. In our initial approaches we ana-
lyzed the correlation between serum shift and global properties
such as molecular weight, ClogP and polar surface area. We then
attempted to combine ClogP, which gave a weak but tangible sig-
nal, with 2D or 3D similarity to warfarin and other known high
affinity ligands of albumin. We also built a binary model using sup-
port vector machine with different sets of 2D descriptors and we
docked our compounds into the benzodiazepine and warfarin
binding sites of albumin. Although some of these approaches pro-
vided hints on the factors affecting serum shift for this class of
compounds, none of them resulted in a convincing model that
could bear sufficient predictive power. At that time the two sub-
classes illustrated in Figure 2 were being further investigated.
A pairwise analysis across the two subclasses showed that com-
pounds carrying a pyrimidine ring at the 7 position of the benz-
imidazole core (class B) consistently showed a lower serum shift
relative to the corresponding 7-fluoropyridines (class A). This real-
ization prompted us to perform a separate analysis for each sub-
class. Serum shift data was available for 62 compounds in class A
and 47 in class B. Given the findings that emerged from the broader
set, we focused our analysis on the correlation between serum shift
and global properties such as ClogP and polar surface area (PSA).
The plots in Figure 3 illustrate the correlation between serum shift
and each of these two properties for class A compounds.
While there was no clear correlation between serum shift and
PSA, there was a clear dependency in this subclass between serum
shift and ClogP. A ClogP value of 3.1 provided a good separation
between high and low shifters: 79% of compounds with ClogP
<3.1 had serum shifts of 8 or less, while 97% of the compounds with
ClogP >3.1 had serum shifts of 16 or more. This trend suggested
that prospective compounds with ClogP >3.1 would almost cer-
tainly be high shifters, while compounds with ClogP <3.1 would
likely be low shifters. Figure 4 illustrates the correlation between
serum shift and the same two properties for class B compounds.
In this case both ClogP and PSA played a significant role, and
the plots clearly indicated that the vast majority of the compounds
with serum shift of 4 or less had ClogP <2.1 and PSA >120. A more
detailed analysis of the individual data showed that using the two
parameters in combination had the potential to minimize both the
false positives and the false negatives: 92% of the compounds with
ClogP <2.1 and PSA >120 had serum shifts of 4 or less, while 91% of
the compounds with ClogP >2.1 and/or PSA <120 had serum shifts
of 8 or more. This trend suggested that prospective compounds
meeting both criteria would almost certainly be low shifters, while
compounds violating at least one of the two would almost cer-
tainly be high shifters.
The findings of this analysis were applied to the design of pro-
spective analogs by prioritizing compounds that met the following
criteria:
C7-fluoropyridines (class A): ClogP <3.1 (expected serum shift
68)
C7-pyrimidines (class B): ClogP <2.1 and PSA >120 (expected
serum shift 64)
This approach, combined with the SAR-driven and structure-
based methods previously described for this series,9 led to the syn-
thesis of compounds with low serum shift and a greater potential
to be efficacious in an animal infection model. Overall, 61 addi-
tional compounds between the two subclasses were synthesized
according to the general procedure in Scheme 1. Additional details
on the synthesis of these two subseries can be found in our previ-
ous publications.9,18
R2
Low nanomolar potencies against gyrase B and topoisomerase
IV were generally maintained, and MIC values between 0.016 and
0.25 lg/ml against S. aureus were observed for 55 of the 61 com-
R1
N
NH
pounds. The serum shift values were consistent with the predic-
tions based on the above criteria in 46/61 cases (75.4%). In all the
remaining 15 cases the shifts were only 2-fold outside of the pre-
dicted range. This method proved especially effective in filtering
out compounds with high serum shift, as only two of the com-
pounds predicted to have high shift turned out to have low shift.
Overall, the compounds with low shift were predicted with 62%
accuracy while the compounds with high shift were predicted with
92% accuracy. Table 3 illustrates the structures of four optimized
compounds with their antimicrobial activities both in the absence
and in the presence of human serum.
HN
O
X = O, NH
X
Figure 1. General substructure for compounds in the benzimidazole series with
available serum shift data.
R
R
F
N
7
7
As both models included an upper limit for ClogP of low serum
shifters, their application led to the design of compounds with re-
duced lipophilicity. It has been observed that reducing the lipophil-
icity of compounds within a series often leads to a reduction in
intrinsic clearance, which in turn can result in increased levels of
exposure in vivo.1,19 Such an effect could have been a contributing
factor to the desired improvements in efficacy for our series. How-
ever, a clear correlation between ClogP and intrinsic clearance for
the synthesized compounds could not be established, and no corre-
lation was observed between the same parameter and in vivo
exposure upon IV or oral dosing (data not shown).
N
N
N
N
NH
O
NH
O
HN
HN
HN
HN
Class A
Class B
Figure 2. Substructures of the benzimidazole urea subclasses investigated in this
study.