Proteochemometrics analysis of substrate interactions with dengue virus NS3 proteases

Article abstract of DOI:10.1016/j.bmc.2008.08.081

The prime side specificity of dengue protease substrates was investigated by use of proteochemometrics, a technology for drug target interaction analysis. A set of 48 internally quenched peptides were designed using statistical molecular design (SMD) and assayed with proteases of four subtypes of dengue virus (DEN-1-4) for Michaelis (K_m) and cleavage rate constants (k_cat). The data were subjected to proteochemometrics modeling, concomitantly modeling all peptides on all the four dengue proteases, which yielded highly predictive models for both activities. Detailed analysis of the models then showed that considerably differing physico-chemical properties of amino acids contribute independently to the K_m and k_cat activities. For k_cat, only P1′ and P2′ prime side residues were important, while for K_m all four prime side residues, P1′-P4′, were important. The models could be used to identify amino acids for each P′ substrate position that are favorable for, respectively, high substrate affinity and cleavage rate.

Full text of DOI:10.1016/j.bmc.2008.08.081

Bioorganic & Medicinal Chemistry 16 (2008) 9369–9377
Contents lists available at ScienceDirect
Bioorganic & Medicinal Chemistry
journal homepage: www.elsevier.com/locate/bmc
Proteochemometrics analysis of substrate interactions with dengue virus
NS3 proteases
Peteris Prusis a, , Maris Lapins a, , Sviatlana Yahorava a, Ramona Petrovska a, Pornwaratt Niyomrattanakit b,
Gerd Katzenmeier b, Jarl E. S. Wikberg a,
*
a Department of Pharmaceutical Biosciences, Uppsala University, Box 591 BMC, SE751 24 Uppsala, Sweden
b Laboratory of Molecular Virology, Institute of Molecular Biology and Genetics, Mahidol University, Salaya Campus, Phutthamonthon 4 Road, Nakornpathom 73170, Thailand
a r t i c l e i n f o
a b s t r a c t
Article history:
The prime side specificity of dengue protease substrates was investigated by use of proteochemometrics,
a technology for drug target interaction analysis. A set of 48 internally quenched peptides were designed
using statistical molecular design (SMD) and assayed with proteases of four subtypes of dengue virus
(DEN-1–4) for Michaelis (Km) and cleavage rate constants (kcat). The data were subjected to proteochemo-
metrics modeling, concomitantly modeling all peptides on all the four dengue proteases, which yielded
highly predictive models for both activities. Detailed analysis of the models then showed that consider-
ably differing physico-chemical properties of amino acids contribute independently to the Km and kcat
activities. For kcat, only P10 and P20 prime side residues were important, while for Km all four prime side
residues, P10–P40, were important. The models could be used to identify amino acids for each P0 substrate
position that are favorable for, respectively, high substrate affinity and cleavage rate.
Received 15 March 2008
Revised 7 August 2008
Accepted 20 August 2008
Available online 13 September 2008
Keywords:
Dengue proteases
Proteochemometrics
Substrate library
Peptide library
Ó 2008 Elsevier Ltd. All rights reserved.
Library design
Statistical molecular design
Molecular recognition modeling
1. Introduction
it. Dengue hemorrhagic fever and shock syndrome are severe forms
with hemorrhaging that may cause dramatic loss of blood pres-
Dengue fever has been known for more than two hundred
years. The disease is caused by the dengue virus, a member of Fla-
viviridae family, which is transmitted to humans by mosquitoes of
the species Stegomiya aegypti (formerly Aedes). There are four clo-
sely related but antigenically distinct dengue virus serotypes
(DEN-1–4) for which immunity to one serotype does not protect
against infection by another.1,2 Infections occur primarily in the
tropics, where the virus threatens a large portion of the population.
Its global distribution is comparable to that of malaria, and an esti-
mated 2.5 billion people live in areas at risk for epidemic transmis-
sion.3–5
Dengue causes a spectrum of clinical symptoms ranging from
mild, uncomplicated dengue fever to the severe forms of dengue
hemorrhagic fever and dengue shock syndrome.4–7 Dengue fever
can cause aches, pains, headaches, and high fever, and is some-
times called ‘‘breakbone fever” because of the pain associated with
sure. Nearly 5% of the ꢀ1 million dengue hemorrhagic fever cases
occurring each year are fatal.4
Accordingly, there is considerable interest in developing thera-
peutics against dengue. Vaccine candidates for all four serotypes
derived from live attenuated or chimeric viruses are in clinical
trial.8 However, there is presently neither useful vaccine nor anti-
viral drug available.
The genomes of the dengue viruses consist of an 11-kb single
positive-stranded RNA that encodes 3 structural (C, prM, and E)
and seven non-structural proteins (NS1, NS2A, NS2B, NS3, NS4A,
NS4B, and NS5).9 The correct processing of these proteins is essen-
tial for virus replication and requires host proteases such as signa-
lase and furin10 and a two-component viral protease, NS2B/NS3.9,11
The dengue virus NS3 protease is an attractive target for devel-
opment of therapeutic inhibitors of dengue. The enzyme is vital for
the post-translational proteolytic processing of the dengue poly-
protein precursor and is essential for viral replication and matura-
tion of infectious virons.11–13 The NS3 protease catalyzes the
cleavage of the viral polyprotein precursor in the non-structural re-
gion, in cis at the NS2A/NS2B and NS2B/NS3 junctions and in trans
at the NS3/NS4A and NS4B/NS5 sites11,14–16 (as well as at addi-
tional sites within the viral capsid protein, NS2A, NS4A, and within
NS3 itself).17–20 A trypsin-like protease domain with a classical
Abbreviations: SMD, statistical molecular design; Km, Michaelis constant; kcat
,
conversion rate constant; DEN-1–4, dengue virus serotypes 1–4; Abz, o-aminoben-
zoic acid; nY, 3-nitrotyrosine; PCA, principal component analysis; PLS, partial least-
squares projections to latent structures.
* Corresponding author. Tel.: +46 18 471 4238; fax: +46 18 55 9718.
E-mail address: Jarl.Wikberg@farmbio.uu.se (J.E.S. Wikberg).
These authors contributed equally to this work.
0968-0896/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.bmc.2008.08.081

9370
P. Prusis et al. / Bioorg. Med. Chem. 16 (2008) 9369–9377
serine protease catalytic triad (His51, Asp75, and Ser135) was orig-
inally identified in the N-terminal region of the 69 kDa NS3 pro-
tein15 and the minimum sequence which supports protease
activity was mapped to 167 residues of NS3.21 The protease activ-
ity is enhanced by the NS2B protein (at least 40 amino acids of it
are required), which acts as cofactor for the NS3 protease.14,22,23
The NS3 protease-cleavage sites in the viral polyprotein have
pairs of dibasic amino acids (i.e., RR, RK, KR, or, at times, QR) at
the P2 and P1 positions, and small non-branched amino acid at
the P10 position.9,12,24–27 Our earlier study suggested that basic
amino acids at the P3 and P4 positions improved substrate cleav-
age27, and another study by Li et al.26 recently suggested that ali-
phatic amino acids may be favorable at the P4 position. Our
previous work showed that removal of the methionine at the P40
position in a peptide which spanned both P4–P1 and P10–P40 of
the capsid protein cleavage site in DEN-2, namely Abz–
RRRR;SAGM–nY–NH2 peptide (i.e., resulting in Abz–RRRR;SAG–
nY–NH2), led to improved kcat activity.27 Li and colleagues investi-
gated the specificity of the P0 side by comparing fluorescence-de-
tected initial velocities for pools of peptide substrates, holding
one amino acid fixed while varying the other three positions,
which resulted in mixtures of 19 amino acids at each position.26
The results of this study, which included all four viral subtypes
of the protease, showed that serine was the most preferred amino
acid in the P10 and P30 positions of the substrate, whereas the P20
and P40 positions tolerated a broad diversity of amino acids. Two
earlier studies were also dedicated to the design of dengue prote-
ase inhibitors, essentially based on peptidomimetics mirroring the
dibasic P1–P2 site of dengue protease substrates.28,29
positions, similarly to the capsid protein cleavage site in the
DEN-2 virus. Based on the predictions of this QSAR model, we then
identified sets of amino acids for the P10–P40 positions that should
ensure cleavability of the substrates. The following sets of amino
acids were identified: A, N, D, G, H, and S for the P10 position; A,
N, D, G, H, L, P, S, T, and W for the P20 position; C, G, P, S, T, and
V for the P30 position; and N, D, C, H, L, F, or none for the P40 posi-
tion (thus giving a total of 6 ꢁ 10 ꢁ 6 ꢁ 7 = 2520 possible amino
acid combinations for P10–P40).
An informative peptide library was created from the above sets
by applying the principles of statistical molecular design,37,38 using
D-optimal design as implemented in MODDE 6.0 software (Umet-
rics AB, Sweden). This design technique maximizes the diversity in
the set of objects comprised by the design matrix X. The optimality
criterion in generating D-optimal design is the determinant of the
X0X matrix, which is an overall measure of the information in X.39
Applying D-optimal design gave a set of 48 peptides (peptides Nr
1–48 in Table 1) that covered the maximal volume of the whole
multivariate space of possible amino acid combinations. The qual-
ity of design was characterized by the following parameters:
log(detX0X) = 24.7 and G-efficiency = 66.1.40 Hereafter, this library
of 48 peptides is referred to as the work set.
The validity of the proteochemometric models was assessed by
an external test set for which we used eight previously prepared
peptides having the same general structure as those above.27 The
prime side sequences of these peptides had been derived from
the native cleavage sites in the dengue polyprotein, indicated as
peptides Nr 49–56 shown in Table 1.
We have developed a novel approach to study ligand–protein
interactions, proteochemometrics.30,31 Proteochemometrics ana-
lyzes experimentally determined interaction data for series of li-
gands with respect to series of proteins, correlating interaction
data to the physico-chemical and/or structural descriptions of both
the ligands and proteins. Proteochemometric models are thereby
created, which reveal the properties of both interaction partners
that determine their activities and specificities. Proteochemomet-
rics has been successfully applied to various classes of G-protein
coupled receptors,32–34 antibodies35 and aspartate proteases.36
The aim of the present study was to extend our knowledge on
the requirements for dengue virus NS3 protease substrates. To this
end, we investigated the P0 side specificity of substrates for the NS3
protease from all four serotypes using a large set of internally
quenched peptides, separately analyzing the rate of substrate
cleavage (kcat) and substrate affinity (Km) by proteochemometric
modeling.
2.2. Kinetic characterization of substrate library
For the four subtypes of the dengue protease, we determined
values of Km and kcat using the work set of 48 substrates (see Sec-
tion 4 for peptide synthesis, expression of proteases, and determi-
nation of kinetic constants). The data obtained represented the
averages from at least three independent measurements for each
protease–substrate pair and are summarized in Table 1. Only one
peptide failed to be cleaved by all four proteases. Moreover, the
DEN-1 protease failed to cleave one additional substrate, while
the DEN-3 protease failed to cleave seven substrates. Some sub-
strates gave marked substrate inhibition, which precluded calcula-
tions of Km and kcat values. Interestingly, there was no apparent
correlation between the Michaelis constant and the cleavage rate.
Thus, for the DEN-1, DEN-3, and DEN-4 proteases there were no
correlations whatsoever between the logarithmically transformed
Km and kcat values (r2 < 0.05), while for the DEN-2 protease there
was a marginally negative correlation (r2 = 0.14). These results thus
suggest that mechanisms for substrate binding and cleavage are
independent for the dengue proteases.
2. Results and discussion
2.1. Design of substrate library
2.3. Results of proteochemometric modeling
To investigate the roles of the P10–P40 positions of dengue pro-
tease substrates, we designed a library of internally quenched pep-
tides having the general structures Abz–RRRR;XXXX–nY–NH2
(Abz, o-aminobenzoic acid; nY, 3-nitrotyrosine) and Abz–
RRRR;XXX–nY–NH2. In preliminary studies, we deduced a set of
amino acids for each of the prime sequence positions that had
great potential to result in active substrates for the DEN-2 prote-
ase. These amino acids were identified from tests of a series of
32 octapeptides with amino acids selected from the P4–P1 and
P10–P40 positions of natural DEN-2 polyprotein cleavage sites.
Cleavage kinetics of these peptides by the DEN-2 protease were
investigated and the obtained initial reaction velocity values were
subjected to quantitative structure–activity relationship (QSAR)
modeling (data not shown). According to this analysis, the peptides
that were most readily cleaved had four arginines in the P4–P1
As detailed in Section 4, each of the four P10–P40 residues of the
substrates was characterized by 6 quantitative descriptors repre-
senting hydrophilicity (zz1), size (zz2), polarity (zz3), charge
(C7.4), rigidity (t1-Rig), and flexibility (t2-Flex), while the prote-
ases were described by four binary descriptors. The substrate
descriptors were confined in the ‘S block’ and the protease descrip-
tors in the ‘P block’ and were mean-centered and block-scaled to
make them comparable. Separate proteochemometric models were
constructed for rate of substrate cleavage (kcat) and substrate bind-
ing (Km) (see Section 4 for details). We first created linear models
using only S and P blocks. These descriptors were later comple-
mented by various square- and cross-terms (S2, S ꢁ P, and S ꢁ S
blocks) to identify non-linearities in substrate–protease interac-
tions. The predictive ability of each model was evaluated by 7-fold

P. Prusis et al. / Bioorg. Med. Chem. 16 (2008) 9369–9377
9371
Table 1
Kinetic constants obtained with DEN-1–4 proteases using novel peptide substrates modified within the P10–P40 sequence with the general structure Abz–RRRR;XXXX–nY–NH2
Nr
P10–P40 sequence
DEN-1
)
DEN-2
)
DEN-3
)
DEN-4
)
kcat (minꢂ1
Km
(lM)
kcat (minꢂ1
Km
(lM)
kcat (minꢂ1
Km
(l
M)
kcat (minꢂ1
Km (lM)
1
2
3
4
5
6
7
8
9
APCN
HHGN
DDGN
SNSN
NWTN
GTVN
AGPN
SHCD
APGD
HWSD
GGTD
AAVD
NDVD
NLPD
HNCC
NACC
SWGC
AHSC
DSTC
SGVC
GDPC
DGCH
ANGH
NPSH
GATH
HLVH
HSPH
STPH
GWCF
GLGF
DTSF
2.38
0.24
0.27
3.93
0.87
1.16
1.54
4.11
NC
0.15
4.29
2.11
0.36
0.59
SIb
11.43
9.10
0.24
0.08
0.07
1.68
0.30
0.33
0.41
0.74
0.04
0.07
0.43
0.75
0.10
0.22
0.14
0.28
0.39
0.26
0.07
0.98
0.18
0.09
0.09
0.08
0.94
0.14
0.07
0.94
NC
0.90
0.06
0.08
0.21
0.29
2.00
0.22
0.10
0.22
0.48
1.78
0.07
0.07
SI
5.41
4.34
79.59
6.25
6.02
6.50
0.67
0.11
NCa
2.22
0.39
0.58
0.62
1.71
NC
3.87
10.09
2.86
1.15
0.54
11.26
2.27
4.41
4.47
4.70
0.62
0.65
3.51
6.81
1.52
2.02
0.94
2.23
2.39
1.83
0.84
4.54
1.26
0.47
3.57
1.35
4.69
1.08
0.25
4.68
NC
4.94
7.06
170.98
15.41
13.73
10.50
12.56
26.33
89.51
14.37
10.41
11.76
19.15
9.77
16.34
9.37
12.57
10.83
11.31
4.18
11.40
19.55
14.86
12.93
18.26
43.17
16.71
8.81
3.21
2.87
5.63
12.81
4.26
13.24
7.02
7.02
5.18
4.66
14.10
5.77
17.23
89.05
9.76
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
17.19
49.72
31.42
106.17
36.84
NC
1.48
0.92
NC
0.32
0.17
0.61
SI
18.64
21.17
12.21
36.89
85.93
33.48
3.28
3.03
4.79
4.21
23.08
7.06
11.52
9.05
10.81
21.47
7.79
16.29
7.25
14.75
53.49
10.04
13.01
0.83
SI
SI
0.46
SI
8.81
SI
15.23
0.20
1.75
0.42
0.17
0.61
0.11
1.42
NC
0.05
0.87
NC
1.41
0.04
SI
0.37
0.46
2.87
SI
1.30
0.08
1.07
2.45
NC
0.17
2.97
0.10
0.14
0.57
0.86
0.44
3.34
0.89
0.63
0.77
4.11
0.72
1.80
NA
11.12
19.72
41.51
20.84
31.40
26.92
18.86
1.21
0.77
1.86
0.37
4.95
0.37
0.38
1.88
NC
3.19
0.21
SI
0.73
1.84
8.17
SI
4.13
0.53
2.21
6.33
0.34
0.29
SI
0.18
0.54
1.37
1.69
1.20
9.98
1.94
3.60
2.77
16.91
4.91
9.18
13.60
31.95
22.89
22.33
13.41
17.43
8.83
16.66
18.06
19.56
42.25
12.26
36.69
4.33
15.02
12.29
26.75
10.84
5.33
16.60
11.28
8.11
0.36
1.34
1.21
NAc
21.17
32.42
24.69
16.76
HPTF
ADTF
NSVF
SAPF
36.54
24.75
20.03
48.56
28.86
22.21
11.48
1.45
4.58
0.82
2.69
14.39
0.99
1.07
5.96
1.03
0.77
2.36
5.09
1.32
16.93
4.01
3.32
5.13
27.67
5.11
13.15
13.34
13.28
7.16
9.57
24.81
6.92
14.88
75.69
87.96
8.48
11.25
21.78
9.38
8.51
8.15
22.65
7.18
10.84
10.68
15.20
8.22
ATCL
7.31
7.17
NGGL
HDSL
GSSL
21.04
112.33
15.38
17.39
54.56
33.75
25.42
25.06
25.92
29.50
18.43
10.44
4.96
28.85
17.31
SLTL
DHTL
DNPL
SSC–d
HTG–
DLS–
NHT–
AWV–
GPP–
SAGM
SWPL
AGVL
GTGN
SAG–
AAGM
SAAM
SAGA
75.26
7.43
11.30
53.82
24.74
11.49
16.92
25.5
12.58
23.85
22.11
28.39
17.90
16.34
23.27
0.08
0.14
0.17
0.27
0.17
1.86
0.63
0.34
0.29
2.54
0.55
1.56
1.62
6.55
19.65
6.07
2.24
5.15
3.44
5.54
2.19
3.05
3.61
2.46
3.33
3.06
15.39
69.33
23.82
18.19
18.70
24.86
11.97
10.59
20.87
20.90
12.13
13.85
14.33
10.49
a
b
c
NC, minimal or no cleavage.
SI, substrate inhibition.
NA, not measured.
–, denotes that there is no amino acid in the P40 position.
d
cross-validation, here referred to as q2, and a modified cross-vali-
dation with seven randomly formed groups of substrates, here re-
ferred to as q2substr, as well as by external predictions using the eight
test set substrates, resulting in the q2ext estimate (see Section 4 for
details). Modeling results are summarized in Table 2 for kcat and in
Table 3 for Km. Linear models could approximate kcat and Km with
good accuracy, explaining 83% of the variation in kcat and 69% in
Km. Models also showed high predictive ability as assessed by
the q2 measure. However, qs2ubstr and, especially, q2ext were substan-
tially lower than the q2, indicating that the linear models were
Table 2
Results of proteochemometric modeling of log kcat data using different combinations
of substrate and protease descriptors
Descriptor blocks
r2
q2
q2substr
q2ext
S, P
0.83
0.91
0.94
0.93
0.76
0.88
0.84
0.86
0.62
0.80
0.80
0.74
0.44
0.81
0.78
0.76
S, P, S2
S, P, S2, S ꢁ P
S, P, S2, S ꢁ S

9372
P. Prusis et al. / Bioorg. Med. Chem. 16 (2008) 9369–9377
Table 3
the four serotypes of dengue. For all three models, the difference
r2 ꢂ q2 was in the range 0.03–0.12 while the difference r2 ꢂ q2substr
varied over the narrow range of 0.10–0.14, which indicated that
the model would be highly reliable for interpretations.
Results of proteochemometric modeling of log Km data using different combinations
of substrate and protease descriptors
Descriptor blocks
r2
q2
q2substr
q2ext
Interpretations of models were based on the analysis of partial
least-squares projections to latent structures (PLS) regression
equations. For a model using S, P, and S2 descriptor blocks, the
resulting regression equation can be expressed as follows:
S, P
0.69
0.75
0.80
0.78
0.59
0.68
0.68
0.66
0.54
0.65
0.66
0.63
0.40
0.54
0.57
0.45
S, P, S2
S, P, S2, S ꢁ P
S, P, S2, S ꢁ S
4
24
X
X
ꢀ
ꢀ
poor predictors of activities from chemical properties of com-
pletely new substrates.
y ¼ y þ
ðcoeffP ꢁ ðxP ꢂ xPÞ þ
ðcoeffS ꢁ ðxS ꢂ xSÞ þ coeffS2
P¼1
S¼1
Squared-terms of substrate descriptors (S2 block) were then
added, which gave substantial improvements of the models. For
the Km model all three assessments of the predictive ability in-
creased by 0.09–0.14 (Table 3), while the kcat model improved even
more (q2 = 0.88, q2substr = 0.80, and qe2xt = 0.81) (Table 2). Thus, if the
linear model suffered in the ability to predict activities for new
substrates, the predictive ability became excellent according to
all assessments used herein by the inclusion of square-terms, S2.
One may therefore assume that strong non-linearities are present
in the substrate property–activity relationships. For example, an
optimal degree of hydrophilicity, size, or other property of sub-
strate residues might exist that yields the most efficient molecular
interactions with the proteases (vide infra for further interpreta-
tions of the models).
Addition of substrate–protease cross-terms (S ꢁ P block) im-
proved the Km model slightly, but had a negative influence on
the predictive ability of the kcat model. Moreover, including
cross-terms for substrate descriptors (S ꢁ S block) also resulted
in some decrease in the predictive ability for both Km and kcat. This
lack of relevance of substrate cross-terms suggested that substrate
prime side residues interact mostly in a non-cooperative manner,
and, in turn, implied that the most preferable amino acids would
be found separately for each prime side position of the substrates
of the dengue proteases. Moreover, the lack of improvement of
the kcat model with the addition of the substrate–protease cross-
terms showed that substrate cleavages by all the four dengue pro-
teases are governed by closely similar processes, although the
slight improvement of the Km model indicated that the different
dengue protease forms demonstrated some binding preferences
among the different substrates.
ꢁ ðx2S ꢂ x2S ÞÞ
The sign and value of a PLS coefficient for a descriptor directly
showed the influence of the represented property on the modeled
activity, y. Interpretation of coefficients for square-terms is some-
what more intricate. Understanding of them is grounded on the
fact that they become negative (after centering) when the values
of original descriptors are close to their mean. On the other hand,
square-terms get highly positive values if the values of the original
descriptors are far from the mean. By comparing the coefficients
for the square-term together with the original descriptor one
may judge the degree of a non-linearity between the represented
molecular property and the modeled activity.
Regression coefficients of the PLS models are represented
graphically in Figure 2, where Figure 2A shows coefficients from
the kcat model and Figure 2B from the Km model (Since a higher
affinity for the substrate by the protease is indicated by a lower
Km constant, a positive coefficient for a descriptor in the Km model
denotes that the described molecular property correlates nega-
tively with substrate binding, and vice versa).
As is revealed from the coefficients for the protease subtype
descriptors of the kcat model (Fig. 2A), the DEN-4 protease shows,
in general, higher substrate cleavage rates, while the DEN-2 prote-
ase has the slowest rates. The negative coefficient for DEN-2 in the
Km model (Fig. 2B) indicates, on the other hand, that this protease
subtype possesses overall higher affinity for the substrates. DEN-1
shows, on average, higher cleavage rates, while its substrate bind-
ing affinities are relatively low. Taken together, the data suggest
that there is no apparent correlation between the coefficients for
the protease descriptors for Km and kcat, a result that corroborates
the lack of correlation between Km and kcat values mentioned
above.
Comparison of the overall patterns for substrate descriptors in
Figure 2A and B reveals major differences in the contribution of
substrate properties to the protease kcat and Km activities. The Km
model has assigned substantially large PLS coefficients to several
descriptors for each of the four prime site residue (Fig. 2B). This
means that all prime side residues influenced the binding of sub-
strates by the dengue proteases. By contrast, for the kcat model,
some descriptors for the P10 and P20 positions gained considerably
large PLS coefficients, while at the P30 and P40 positions, the coeffi-
cients for all descriptors were minor (Fig. 2A). Accordingly, only the
P10 and P20 positions substantially influence the rate of substrate
cleavage.
The performances of kcat and Km models using S, P, and S2 blocks
are illustrated graphically in Figure 1. For only eight interaction
pairs (i.e., 4% of the data) the predictions by the kcat model were
more than 0.5 logarithmic units wrong, as assessed by cross-vali-
dation and external validation (Fig. 1A). For the Km model only four
such mispredictions (2%) occurred (Fig. 1B).
2.4. Interpretation of models
With several highly predictive models for substrate binding and
rate of substrate cleavage available, we next assessed the reliability
of each model for interpretation by examining the differences be-
tween the goodness of fit and the predictive ability. Reliable inter-
pretations can be attained if r2 does not exceed the q2 by more than
0.2–0.3 units; a larger difference is a sign of chance correlations to
irrelevant descriptors, or of the presence of outliers in the data
set.37 As seen from Tables 2 and 3, the smallest margins between
r2 and q2 are shown by the models comprising substrate descrip-
tors, protease descriptors, and square-terms of substrate descrip-
tors (i.e., S, P, and S2 descriptor blocks); these models were
accordingly used for interpretation of the roles of substrate and
protease properties for substrate cleavage and binding. We also
used the Km model that in addition included the S ꢁ P descriptor
block to investigate substrate selectivity between proteases of
Further analysis of the PLS coefficients at the P10 position of the
kcat model revealed large negative values for most of the substrate
square-terms. This result implies that the optimal amino acid(s) for
the P10 position reside within the modeled amino acids’ ‘physico-
chemical property space’, and that large deviation from the center
of this space negatively influence substrate turnover. For the origi-
nal descriptors (i.e., non-squared descriptors), large negative coef-
ficients are assigned to size and rigidity. Thus, these two properties
mainly cause a reduction in the rate of substrate cleavage. By con-
trast, the coefficient for the polarity descriptor (zz3) is close to zero,

P. Prusis et al. / Bioorg. Med. Chem. 16 (2008) 9369–9377
9373
Figure 1. Correlation of predicted versus observed log(kcat) (A) and log(Km) (B) values derived from the proteochemometric models using substrate and protease descriptors,
and square-terms of substrate descriptors. Predictive ability is assessed by cross-validation leaving out 1/7 of substrates at a time (gray diamonds) and by external validation
(black circles). Shown is also model fit (small unfilled squares). Indicated by the oblique gray lines is the plot area where prediction errors do not exceed 0.5 logarithmic units.
indicating that broad variations are acceptable at P10 for high kcat
values. The hydrophilicity and flexibility descriptors (i.e., zz1 and
t2-Flex) yielded some negative correlation with kcat, but the effects
are nonlinear as is indicated by the large coefficients for their
square-terms. Accordingly, very hydrophobic amino acids or amino
acids that lack any flexible features are not advantageous at P10.
Detailed analysis of the various molecular properties at the P10 po-
sition identified Ser as the best amino acid for high kcat. Serine en-
sured, on average, 3-fold and 5-fold higher cleavage rates than Ala
and Gly, respectively, the two amino acids closest to Ser for achiev-
ing high kcat (These conclusions are in fair agreement with earlier
findings27).
and flexibility (i.e., zz1 and t2-Flex) descriptors, whereas the
most negative influence arose from rigidity features (i.e., t1-
Rig) of the residue. By comparing the values of all descriptors
and square-terms for all ten amino acids present in the P20
among our substrates, we found that flexible or small amino
acids (e.g., Ala, Gly, Leu) should be beneficial at this position.
Moreover, as indicated by the positive coefficient for the charge
descriptor, the acidic Asp is inferior to Asn at this position.
Among the identified less favorable residues at P20 were Thr
and Trp. This result was unexpected since these amino acids
are present at the P20 position in some native cleavage sites of
the dengue viruses. However, our test substrates containing
Thr and Trp at the P20 position indeed showed relatively low kcat
activities, as predicted by the model, and thus supported these
interpretations.
The pattern of coefficients for descriptors at the P20 position
(kcat model) was completely different. Here, the largest positive
coefficients were given to square-terms of the hydrophilicity

9374
P. Prusis et al. / Bioorg. Med. Chem. 16 (2008) 9369–9377
A: kcat model
1
4
3
2
P1’
P2’
P3’
P4’
Protease
subtype
Substrate properties
B: Km model
1
3
4
2
P1’
P2’
P3’
P4’
Protease
subtype
Substrate properties
Figure 2. Regression coefficients of proteochemometric models of log(kcat) (A) and log(Km) (B) data. In each panel, shown are regression coefficients for the four protease
subtype descriptors (on the left) and descriptors of six properties for the four substrate residues, P10–P40. Solid black bars represent regression coefficients for original
descriptors (Note that negative values of hydrophilicity descriptors are seen for hydrophobic amino acids whereas positive values are seen for hydrophilic amino acids;
negative values of size descriptor characterize small, compact amino acids, while positive values are seen among large, bulky amino acids; for polarity descriptor, a negative
value indicates that the amino acid is electrophilic, whereas a positive value indicates that it is electronegative. The charge descriptor reflects the charge of the amino acid.
Positive values of the rigidity and flexibility descriptors indicate the presence of, respectively, a high fraction of rigid and flexible fragments in an amino acid, while negative
values indicate the lack of such fragments.) and gray bars represent regression coefficients for their square-terms.
The coefficients for descriptors at the P30 and P40 positions were
small, thus showing that the model did not identify any apprecia-
ble correlations between properties of these positions and the kcat
activity.
taining aspartic acid at this position. Several other P20 descriptors
(zz1, zz3, and t2-Flex) and their cross-terms obtained positive coef-
ficients. These combined results showed that hydrophilicity, elec-
tronegativity, and flexibility tended to correlate with Km in a
nonlinear manner. By comparing the predicted Km values when
all ten amino acids were tried at P20, we can conclude that this po-
sition allows a broad range of amino acids, with the most favorable
being Trp, closely followed by His, Thr, Ala, and Gly.
However, the patterns were very different for the coefficients of
the Km model. Here all four P0 positions gained considerably large
PLS coefficients for one or other of the descriptors (Fig. 2B). The
highest absolute values of coefficients were obtained by the charge
descriptors of the P20 and P40 sites. The negative values of these
coefficients identified the acidic Asp and Glu residues as unfavor-
able for high protease binding affinity. On the other hand, the
square-terms of these two descriptors attained positive coeffi-
cients, which showed that correlation between charge and Km is
not linear over the whole range of values (e.g., His is not much pre-
ferred over a neutral amino acid at P40). A large positive coefficient
is also assigned to the square-term of the hydrophilicity descriptor
at P40, indicating that extremely hydrophobic or hydrophilic amino
acids have here a negative impact on substrate binding. Although
the coefficients for polarity and flexibility are rather small at P40,
these data suggest that Ala, Gly, and Cys (and perhaps Pro) could
improve the Km activity.
The patterns of coefficients for the P10 and P20 positions are
quite similar (Fig. 2B). However, when interpreting the require-
ments at these positions one should take into account that we
had a wider selection of amino acids for the P20 position than for
the P10 position; for the latter, the chemical space covered was
more compact (Preliminary studies suggested that hydrophobic
and electrophilic amino acids were not appropriate at P10 for
obtaining cleavable substrates). Despite these limitations, our
interpretation for the P10 position revealed that Ala, Gly, and Ser,
which were beneficial for high kcat activity, also provided good
binding affinity. On the other hand, the coefficients for charge
and polarity descriptors revealed that the acidic and electronega-
tive amino acid Asp was the least appropriate at the P10 position
for tight enzyme binding.
To complete the interpretations, we investigated the PLS coeffi-
cients for substrate–protease cross-terms in the Km model built on
the S, P, S2, and S ꢁ P descriptor blocks, which was also the model
that showed the highest q2substr and qe2xt values. Three of the four
highest PLS coefficient values were here given to the cross-terms
formed between the DEN-2 descriptor and the flexibility descrip-
tors at P30 and P20, and the rigidity descriptor at P10. Hence, we
may conclude that the DEN-2 protease deviates most from other
subtypes in its substrate binding preferences. For the P30 position,
For the P30 position, the most important were descriptors of
polarity and hydrophilicity, and also their square-terms. Careful
analysis of these coefficients and the descriptor values for all ami-
no acids in the work set indicated some preference for Cys over the
other amino acids including Gly, Ser, Thr, Val, and Pro, the five ami-
no acids that are present at P30 in native substrate cleavage sites of
dengue. Thus, the model finds that P30 in native substrates is not
optimal for protease binding affinity.
At the P20 position, the charge descriptor gave an important
contribution to Km, explaining the low activity of substrates con-

P. Prusis et al. / Bioorg. Med. Chem. 16 (2008) 9369–9377
9375
particularly large selectivity differences would arise, according to
these results, for Cys-containing substrates, and such a substrate
would be less favorable for the DEN-2 protease. A high PLS coeffi-
cient was also given to the cross-term of the DEN-3 descriptor with
the charge descriptor at P40, indicating that Asp has a less negative
influence on the binding of the substrate to DEN-3 than for the
other protease subtypes.
Because the majority of cross-terms obtained very low PLS coef-
ficients, they played minor roles in the modeling. The negligible in-
crease of r2 and q2 values of PLS model upon including of cross-
terms supports this view. Thus, we may conclude that DEN-1–4
proteases share similar substrate recognition mechanisms. This
implies that a protease inhibitor may be designed which is simul-
taneously capable of targeting all four subtypes of the dengue
virus.
by comparing initial velocity values for a single concentration
of peptide mixtures, and Ser was found to be the optimal amino
acid for P10 and P30. Our results confirm that Ser at P10 is highly
beneficial for both kinetic constants, but this was not the case at
P30. Our substrates that included Ser at P30 did not show supe-
rior kcat or Km activity. However, one should keep in mind that
each measurement in the study by Li et al. was obtained using
a mixture of 19 ꢁ 19 ꢁ 19 amino acid combinations at three of
the four P0 side positions. Artifacts could then arise because for
some peptides, a bond hydrolysis at a position other than the
predicted scissile bond could occur (i.e., at the bond after the
two arginines at P2–P1).
To our knowledge, our study is the first to report a quantitative
analysis for the contributions of physico-chemical properties of
residues of peptide substrates with respect to the kinetics of den-
gue proteases. Our models might be used to modify the substrates
by natural or synthetic amino acids to optimize separately kcat and
Km constants for the DEN-1-4 proteolytic enzymes. Design and
evaluation of high affinity uncleavable peptides which might serve
as templates for peptidomimetic inhibitors of dengue should be a
task of further studies.
3. Conclusions
Rather than applying a large combinatorial library approach, we
elected here to analyze data of individual substrates, evaluating 6–
10 of the most promising amino acids for each of the P10–P40 resi-
dues. To make the task affordable, we applied statistical molecular
design. By using D-optimal design we created a balanced and rep-
resentative substrate library of a reasonable size. For the modeling,
we encoded the varying P10–P40 positions of substrates by descrip-
tors of amino acids, representing essentially all physico-chemical
properties that could be important for substrate–protease interac-
tions. The proteochemometric models included concomitantly all
substrates and all four dengue proteases. Models where thoroughly
validated for their predictive ability (q2substr for the kcat model being
0.80 and for the Km model, 0.65) and reliability of interpretations
(the difference r2 ꢂ qs2ubstr being as low as 0.11 and 0.10 for, respec-
tively, the kcat and Km models). We may thus conclude that for all
statistical aspects of model validation, these models are highly
reliable.
It is common to characterize enzyme kinetics using kcat/Km
ratios only. However, if one aims to find a peptide with both a
low Km and low kcat, which might serve as a lead peptide in
the design of a protease inhibitor, analysis of the kcat/Km ratio
is not sufficient. From the raw data on the kinetic activities of
the peptide substrate series synthesized herein, it was evident
that there was no relationship between substrate affinity and
cleavage rate. The proteochemometric modeling revealed that
the values of kcat and Km of the dengue proteases are affected
by different properties of the substrate amino acids. Thus, the
proteases demonstrated a high rate of cleavage for peptides hav-
ing small amino acids (Ser, Gly, Ala) at the P10 position and
small or flexible amino acids at the P20 position, while the P30
and P40 positions had little effect on substrate turnover. By con-
trast, all four P0 positions of the substrate significantly influenced
Km of the proteases. Thus, for high affinity (low Km), a P10 amino
acid should be small and moderately hydrophilic, while acidic
residue is highly unfavorable. For the P20 position, the most
favorable was Trp; however, this position allowed a broader
diversity of amino acids. For the P30 position, Cys was more ben-
eficial than the amino acids present in native cleavage sites of
the dengue polyproteins. For the P40 position, the best affinity
would be obtained with Ala, Gly, or Cys.
4. Experimental section
4.1. Synthesis of peptides
Internally quenched fluorescent substrates were based on the
Abz (o-aminobenzoic acid)/3-nitrotyrosine (Tyr(3-NO2) or nY)
pair and were prepared by solid-phase synthesis using an auto-
mated multiple peptide synthesizer (MultiPep; Intavis Bioanalyt-
ical Instruments AG, Koeln, Germany) using the automated
standard protocol optimized for Fmoc chemistry. Reagents were
purchased from Fluka (Sigma–Aldrich, St. Louis, MO, U.S.A.), Ap-
plied Biosystems (Foster City, CA, U.S.A.), Bachem (Bubendorf,
Switzerland) and Novabiochem (Calbiochem-Novabiochem AG,
Laufelfingen, Switzerland). The following amino acid derivatives
were used: Fmoc-Ala-OH (where Fmoc is fluoren-9-ylmethoxy-
carbonyl), Fmoc-Arg(Pbf)-OH (where Pbf is 2,2,4,6,7-pentame-
thyldihydrobenzofuran-5-sulphonyl), Fmoc-Asn(Trt)-OH (where
Trt is trityl), Fmoc-Asp(Ot-Bu)-OH (where t-Bu is t-butyl), Fmoc-
Cys(Trt)-OH, Fmoc-His(Trt)-OH, Fmoc-Gln(Trt)-OH, Fmoc-Gly-
OH, Fmoc-Leu-OH, Fmoc-Lys(Boc)-OH (where Boc is t-butoxycar-
bonyl), Fmoc-Met-OH, Fmoc-Phe-OH, Fmoc-Pro-OH, Fmoc-Ser(t-
Bu)-OH, Fmoc-Thr(t-Bu)-OH, Fmoc-Trp(Boc)-OH, Fmoc-Val-OH,
Fmoc-Tyr(3-NO2)-OH, and Boc-Abz-OH. PyBOP (benzotriazole-1-
yl-oxy-tris-pyrrolidino-phosphonium hexafluorophosphate) was
used as the activating reagent and Rink Amide MBHA resin
(capacity 0.64 mmol/g) as the polymeric support [Rink Amide
MBHA resin is 4-(20,40-dimethoxyphenyl-Fmoc-aminomethyl)-
phenoxyacetamido-norleucyl-4-methylbenz-hydrylamine resin].
Peptides were characterized by HPLC and their structures were
confirmed by MS. Analytical HPLC was performed on a Waters
(Milford, MA, U.S.A.) system (Millenium32 workstation, 2690 Sep-
aration Module, 996 Photodiode Array Detector) equipped with a
Vydac RP C18 90 Å reversed-phase column (2.1 mm ꢁ 250 mm).
The purity of raw peptides according to HPLC was above 80%,
and they were used for kinetic experiments after freeze-drying.
Molecular mass measurements were performed on a PerkinElmer
(PerkinElmer Life and Analytical Sciences, Boston, MA, U.S.A.)
instrument PE SCIEX API 150EX with TurboIonSpray ion source.
Freeze-drying was carried out at 0.01 bar (1 bar = 100 kPa) on a
Lyovac GT2 freeze-dryer (Steris Finn-Aqua, Tuusula, Finland)
equipped with a Trivac D4B (Leybold Vacuum GmbH, Cologne,
Germany) vacuum pump and a liquid nitrogen trap. All other
chemicals were reagent grade from Sigma.
These interpretations support largely the results of our previ-
ous study, which provided measured kcat and Km data for combi-
nations of prime and non-prime side sequences of natural
cleavage sites of DEN-2.27 However, our results are not directly
comparable to the recent study by Li et al. who investigated
pools of substrates of DEN-1–4 proteases.26 In the latter study,
the P0 site specificities of dengue substrates were determined

9376
P. Prusis et al. / Bioorg. Med. Chem. 16 (2008) 9369–9377
4.2. Expression and purification of dengue proteases
tions of both the peptide substrates and the proteases. Proteases
were described by four indicator variables, each representing the
protease from one serotype, DEN-1–4. For example, if the activity
measurement was done using the DEN-1 protease, the values of
the four descriptors were 1, 0, 0, and 0; if the activity measurement
was done using the DEN-2 protease, the values of the four descrip-
tors were 0, 1, 0, and 0, and so on.
pTrcHis plasmids containing a recombinant NS2B(H)-NS3pro
sequence from one of the dengue virus proteases 1–4 were trans-
formed into Escherichia coli C41(DE3) as described previously.27
Transformants were grown in Luria broth (LB) medium supple-
mented with ampicillin (100
about 0.5, isopropyl-1-thio-b-
l
D
g/mL) at 37 °C. When OD600 reached
-galactopyranoside was added to a
The substrates were characterized by descriptors for each var-
ied amino acid within the library, representing the following six
molecular properties: hydrophilicity (represented by zz1 descrip-
tor), size (zz2), polarity (zz3), charge (C7.4), rigidity (t1-Rig), and
flexibility (t2-Flex). The zz-scales, zz1-zz3 were proposed earlier
by Sandberg et al.41 and had been obtained by principal compo-
nent analysis (PCA)42 of 26 measured and computed physico-
chemical properties of 87 natural and artificial amino acids.
Zz1-zz3 are the three first principal components; accordingly,
they represent the largest variations within all analyzed phys-
ico-chemical properties and are orthogonal (i.e., uncorrelated).
In the zz1-scale, hydrophobic amino acids are represented by
negative values and hydrophilic amino acids by positive values.
In zz2, negative values represent small, compact amino acids
while positive values represent large, bulky amino acids. In zz3,
a negative value indicates that the amino acid is electrophilic
whereas a positive value indicates that it is electronegative.
The C7.4 descriptor was proposed by Gottfries43 and represents
the proportion of the amino acid side chains that are ionized
at pH = 7.4 (i.e., for anionic Asp and Glu, C7.4 = ꢂ1, for cationic
His, +0.5 and for Arg and Lys, +1, and for all other amino acids
C7.4 = 0). Descriptors t1-Rig and t2-Flex were also calculated
by Gottfries43 by applying PCA to a set of constitutional descrip-
tors of amino acids, which included number of rings, rigid bonds
and rotatable bonds, rigid fragments, flexible and partially flexi-
ble chains, length of the longest flexible chain, and so forth; t1-
Rig and t2-Flex are first two principal components of these and
are accordingly uncorrelated to each other. In this way, each of
the P10–P40 positions was described by six descriptors of amino
acid properties, yielding totally 24 substrate descriptors (prote-
ase descriptors will be referred to as P block, and substrate
descriptors as S block). During modeling, we found that the rela-
tionships between activities and substrate descriptors were only
partially linear. To account for this discrepancy, we calculated
squares of mean-centered substrate descriptors, and used these
24 square-terms as an additional descriptor block (here termed
S2 block). Moreover, we also investigated possible cooperation
of substrate and protease properties. For this purpose sub-
strate–protease and substrate–substrate cross-terms were com-
puted by multiplication of mean-centered descriptors. Thus,
two additional blocks were obtained: S ꢁ P comprising
24 ꢁ 4 = 96 descriptors and S ꢁ S comprising 24 ꢁ 23/2 = 276
descriptors.
final concentration of 0.1 mM, and the culture was grown at
37 °C for 8 h. Cells were harvested by centrifugation (5000 g,
10 min, 4 °C), resuspended in 20 mL of lysis buffer A (100 mM
Tris–HCl, pH 7.5, 300 mM NaCl), and lysed with a sonicator
(MPS) 6 V, 5 ꢁ 30 s. The lysate was sedimented by centrifugation
(10,000g, 30 min, 4 °C), and the pellet with inclusion bodies was
washed two times with lysis buffer containing 1% Triton X-100.
The above pellet was then suspended in 15 mL of denaturing
buffer B (100 mM Tris–HCl, pH 8.0, 300 mM NaCl, 8 M urea),
homogenized with a Ultra-Turrax T25 tissue disperser (Labassco)
and centrifuged (10,000g, 30 min, 4 °C), whereafter the superna-
tant was loaded on a Hitrap chelating column (Pharmacia) equili-
brated with denaturing buffer. The column was washed with 10
column volumes of denaturing buffer containing 20 mM imidazole
and eluted at a flow rate of 0.5 mL/min with denaturing buffer con-
taining 200 mM imidazole. Fractions of 1 mL were collected, and
aliquots were analyzed for the presence of NS2B(H)-NS3pro by so-
dium dodecyl sulfate (SDS)–polyacrylamide gel electrophoresis
(PAGE) on 15% polyacrylamide gels.
Peak fractions were pooled and loaded on a Superdex 200 HR
10/30 gel filtration column (Pharmacia). The column was eluted
with denaturing buffer at a flow rate of 0.3 mL/min and the
fractions containing NS2B(H)-NS3pro, as analyzed by SDS–PAGE,
were pooled and diluted with the same buffer to 0.5 mg/mL.
Refolding of the protein was initiated by stepwise dialysis of
1 mL samples with a dialysis tubing (cutoff, 8 kDa) at 4 °C against
three changes of 100 mM Tris–HCl, pH 8.0–300 mM NaCl
(200 mL), and one change against 200 mL of 100 mM Tris–HCl,
pH 9.0–50 mM NaCl (buffer C). The dialysate was centrifuged
(10,000g, 10 min, 4 °C) and the protein concentration was deter-
mined with a Bradford protein assay kit (Bio-Rad). Preparations
of the NS2B(H)-NS3pro protein were stored at ꢂ20 °C in 100 mM
Tris–HCl, pH 9.0–50 mM NaCl-50% glycerol.
4.3. Kinetic characterization of peptides on dengue proteases
Fluorogenic assays were carried out with the 56 substrates pre-
pared above by using a POLARstar OPTIMA 96 well plate reader
(BMG Labtech GmbH). Substrates cleavage over time was followed
by monitoring the emission at 420 nm upon excitation at 320 nm.
Final reaction volumes were 100
lL and contained 50 mM Tris–
HCl, pH 9.0 and 20% glycerol and the incubation temperature
was 37 °C. Enzyme concentrations were varied over the range of
150–200 nM, depending on the substrate used. The substrate con-
4.5. Data pre-processing
centrations were varied between 1 and 100 lM (12 dilution points
for each substrate). For each substrate concentration the initial
rates of enzyme cleavage were computed using ‘‘first order rate
equation with offset” equation as implemented in program GraFit,
version 5 (Sigma). From the obtained initial rates, the Km and kcat
constants were computed according to Michaelis–Menten enzyme
kinetics equation using the GraFit version 5 program. Measure-
ments were repeated at least three times for each protease–sub-
strate pair.
All descriptors were mean-centered and scaled to unit vari-
ance prior to further use. Block-scaling was also applied to pre-
vent the situation that the large amount of cross-terms would
dominate the correlations and the few protease descriptors then
would become almost indiscernible. In block scaling, all variables
of each block are multiplied by some constant value, the block
scaling factor. Herein, for each of the five descriptor blocks
(i.e., P, S, S2, S ꢁ P, and S ꢁ S) the scaling factor was assigned
pffiffiffi
as 1= k, where k is the number of descriptors in the given block.
In this way, we ensured that all blocks were equally weighted in
the modeling. The two activities, Km and kcat, were logarithmi-
cally transformed and mean-centered prior to applying further
calculations.
4.4. Numerical description of proteases and substrates
For the sake of the proteochemometric modeling, the measured
Km and kcat activity data were correlated to quantitative descrip-

P. Prusis et al. / Bioorg. Med. Chem. 16 (2008) 9369–9377
9377
4.6. Correlation by partial least-squares projections to latent
structures
References and notes
1. Halstead, S. B. Science 1988, 239, 476.
2. Jacobs, M. G.; Young, P. R. Curr. Opin. Infect. Dis. 1998, 11, 319.
3. Monath, T. P. Proc. Natl. Acad. Sci. U.S.A. 1994, 91, 2395.
4. Gubler, D. J.; Clark, G. G. Emerg. Infect. Dis. 1995, 1, 55.
5. Guzman, M. G.; Kouri, G. Lancet Infect. Dis. 2002, 2, 33.
6. Gubler, D. J. Clin. Microbiol. Rev. 1998, 11, 480.
7. Gubler, D. J. Trends Microbiol. 2002, 10, 100.
8. Edelman, R. Clin. Infect. Dis. 2007, 45(Suppl. 1), S56.
9. Chambers, T. J.; Hahn, C. S.; Galler, R.; Rice, C. M. Annu. Rev. Microbiol. 1990, 44,
649.
10. Stadler, K.; Allison, S. L.; Schalich, J.; Heinz, F. X. J. Virol. 1997, 71, 8475.
11. Falgout, B.; Pethel, M.; Zhang, Y. M.; Lai, C. J. J. Virol. 1991, 65, 2467.
12. Zhang, L.; Mohan, P. M.; Padmanabhan, R. J. Virol. 1992, 66, 7549.
13. Ramachandran, M.; Sasaguri, Y.; Nakano, R.; Padmanabhan, R. Methods
Enzymol. 1996, 275, 168.
14. Preugschat, F.; Yao, C. W.; Strauss, J. H. J. Virol. 1990, 64, 4364.
15. Chambers, T. J.; Weir, R. C.; Grakoui, A.; McCourt, D. W.; Bazan, J. F.; Fletterick,
R. J.; Rice, C. M. Proc. Natl. Acad. Sci. U.S.A. 1990, 87, 8898.
Descriptors were correlated to the logarithms of the Km and kcat
activities of the 48 work set peptides by partial least-squares pro-
jections to latent structures (PLS)44 using the Unscrambler 9.7 soft-
ware (CAMO Software AS, Norway). PLS is a widely used method
for finding a quantitative relationship between a set of descriptors
(X data) and one or several responses (Y data). This is achieved by
simultaneously projecting the X and Y matrices onto lower dimen-
sionality variable space (PLS components) with an additional con-
straint to maximize the covariance between projections of X and Y.
For each response, PLS derives a regression equation, where regres-
sion coefficients show the direction and magnitude of the influence
of descriptors on the response (for detailed descriptions see Refs.
45 and 46).
16. Preugschat, F.; Strauss, J. H. Virology 1991, 185, 689.
17. Arias, C. F.; Preugschat, F.; Strauss, J. H. Virology 1993, 193, 888.
18. Lin, C.; Amberg, S. M.; Chambers, T. J.; Rice, C. M. J. Virol. 1993, 67, 2327.
19. Lobigs, M. Proc. Natl. Acad. Sci. U.S.A. 1993, 90, 6218.
4.7. Validation of models
20. Teo, K. F.; Wright, P. J. J. Gen. Virol. 1997, 78, 337.
21. Li, H.; Clum, S.; You, S.; Ebner, K. E.; Padmanabhan, R. J. Virol. 1999, 73, 3016.
22. Clum, S.; Ebner, K. E.; Padmanabhan, R. J. Biol. Chem. 1997, 272, 30715.
23. Falgout, B.; Miller, R. H.; Lai, C. J. J. Virol. 1993, 67, 2034.
24. Chambers, T. J.; Grakoui, A.; Rice, C. M. J. Virol. 1991, 65, 6042.
25. Wengler, G.; Czaya, G.; Farber, P. M.; Hegemann, J. H. J. Gen. Virol. 1991, 72, 851.
26. Li, J.; Lim, S. P.; Beer, D.; Patel, V.; Wen, D.; Tumanut, C.; Tully, D. C.; Williams, J.
A.; Jiricek, J.; Priestle, J. P.; Harris, J. L.; Vasudevan, S. G. J. Biol. Chem. 2005, 280,
28766.
27. Niyomrattanakit, P.; Yahorava, S.; Mutule, I.; Mutulis, F.; Petrovska, R.; Prusis,
P.; Katzenmeier, G.; Wikberg, J. Biochem. J. 2006, 397, 203.
28. Yin, Z.; Patel, S. J.; Wang, W. L.; Wang, G.; Chan, W. L.; Rao, K. R.; Alam, J.;
Jeyaraj, D. A.; Ngew, X.; Patel, V.; Beer, D.; Lim, S. P.; Vasudevan, S. G.; Keller, T.
H. Bioorg. Med. Chem. Lett. 2006, 16, 36.
We used thorough validations to find the optimal complexity
(i.e., number of components) of the PLS models and to assess their
reliability for interpretations and predictions. The goodness of fit of
a PLS model was assessed by r2, the fraction of the explained var-
iance of the response. The predictive ability was characterized by
q2, the fraction of the predicted variance of the response assessed
by cross-validation.45
In the current study, cross-validation was performed in two
ways. First, the dataset was randomly divided into seven groups.
In this way, we assessed the predictive ability for new substrate–
protease combinations. Cross-validation groups were thereafter
rearranged so that all four activity measurements for each of 48
substrates were assigned to the same cross-validation group. This
was done to avoid overly optimistic assessments of the predictive
ability in case the four protease subtypes shared similar substrate
interaction profiles (Results from conventional cross-validation are
here referred to as q2, and results from modified variant of cross-
validation as q2substr). Finally, we also estimated the external predic-
tive ability of models by performing predictions for the eight test
set substrates (The assay data for the test set substrates were ob-
tained independently using the same assay method as for the work
set substrate and are given in Table 1). The thus obtained q2ext esti-
mate differs from q2substr essentially by the fact that P10–P40 of the
test set peptides resemble native cleavage sites of DEN-1–4 poly-
proteins and, as a result, these peptides generally show higher
activities than the average for the work set. Thus, q2ext measure is
preferable for the assessment of the ability of models to generalize
towards high activities.
29. Chanprapaph, S.; Saparpakorn, P.; Sangma, C.; Niyomrattanakit, P.;
Hannongbua, S.; Angsuthanasombat, C.; Katzenmeier, G. Biochem. Biophys.
Res. Commun. 2005, 330, 1237.
30. Lapinsh, M.; Prusis, P.; Gutcaits, A.; Lundstedt, T.; Wikberg, J. E. S. Biochim.
Biophys. Acta 2001, 1525, 180.
31. Wikberg, J. E. S.; Lapinsh, M.; Prusis, P. Proteochemometrics:
A Tool for
Modelling the Molecular Interaction Space. In Chemogenomics in Drug
Discovery—A Medicinal Chemistry Perspective; Kubinyi, H., Müller, G., Eds.;
Wiley-VCH: Weinheim, 2004; pp 289–309.
32. Lapinsh, M.; Prusis, P.; Uhlén, S.; Wikberg, J. E. S. Bioinformatics 2005, 21, 4289.
33. Lapinsh, M.; Veiksina, S.; Uhlén, S.; Petrovska, R.; Mutule, I.; Mutulis, F.;
Yahorava, S.; Prusis, P.; Wikberg, J. E. S. Mol. Pharmacol. 2005, 67, 50.
34. Prusis, P.; Uhlen, S.; Petrovska, R.; Lapinsh, M.; Wikberg, J. E. S. BMC
Bioinformatics 2006, 7, 167.
35. Mandrika, I.; Prusis, P.; Yahorava, S.; Shikhagie, M.; Wikberg, J. E. S. Protein Eng.
Des. Sel. 2007, 20, 301.
36. Kontijevskis, A.; Prusis, P.; Petrovska, R.; Yahorava, S.; Mutulis, F.; Mutule, I.;
Komorowski, J.; Wikberg, J. E. S. PLoS Comput. Biol. 2007, 3, e48.
37. Eriksson, L.; Johansson, E. Chemom. Intell. Lab. 1996, 34, 1.
38. Lundstedt, T. E.; Seifert, E.; Abramo, L.; Thelin, B.; Nystrom, A.; Pettersen, J.;
Bergman, R. Chemom. Intell. Lab. 1998, 42, 3.
39. De Aguiar, P. F.; Bourguignon, B.; Khots, M. S.; Massart, D. L.; Phan-Than-Luu, R.
Chemom. Intell. Lab. 1995, 2, 199.
40. Eriksson, L.; Johansson, E.; Kettaneh-Wold, N.; Wikström, N.; Wold, S. Design of
Experiments, Principles and Applications; Umetrics AB: Umeå, 2000.
41. Sandberg, M.; Eriksson, L.; Jonsson, J.; Sjöström, M.; Wold, S. J. Med. Chem. 1998,
41, 2481.
Acknowledgments
42. Wold, S.; Esbensen, K.; Geladi, P. Chemom. Intell. Lab. 1987, 2, 37.
43. Gottfries, J. Chemom. Intell. Lab. 2006, 83, 148.
44. Wold, S.; Sjöström, M.; Eriksson, L. Chemom. Intell. Lab. 2001, 58, 109.
45. Wold, S. PLS for Multivariate Linear Modeling. In Chemometric Methods in
Molecular Design; van de Waterbeend, H., Ed.; VCH: Weinheim, Germany,
1995; pp 195–218.
We thank Ewelina Fogelström for technical assistance. Support
was obtained by SIDA-Swedish Research Links (348-2004-5993)
and the Swedish VR (04X-05957) and by basic science grant BRG
49800008 (to G.K.) and a Royal Golden Jubilee scholarship from
the Thailand Research Fund (TRF).
46. Geladi, P.; Kowalski, B. R. Anal. Chim. Acta 1986, 185, 1.