Connectivity and Binding-Site Recognition: Applications

Relevant to Drug Design

CHRISTOPHER J. R. ILLINGWORTH,1 PAUL D. SCOTT,2 KEVIN E. B. PARKES,3 CHRISTOPHER R. SNELL,3

MATTHEW P. CAMPBELL,1 CHRISTOPHER A. REYNOLDS1
1
Department of Biological Sciences, University of Essex, Wivenhoe Park, Colchester
CO4 3SQ, United Kingdom
2
Department of Computational Science and Electronic Engineering, University of Essex,
Wivenhoe Park, Colchester CO4 3SQ, United Kingdom
3
Medivir UK Ltd., Chesterford Research Park, Little Chesterford, Essex CB10 1XL,
United Kingdom

Received 24 June 2009; Revised 18 December 2009; Accepted 16 March 2010

DOI 10.1002/jcc.21561
Published online in Wiley InterScience (www.interscience.wiley.com).

Abstract: Here, we describe a family of methods based on residue–residue connectivity for characterizing binding
sites and apply variants of the method to various types of protein–ligand complexes including proteases, allosteric-
binding sites, correctly and incorrectly docked poses, and inhibitors of protein–protein interactions. Residues within
ligand-binding sites have about 25% more contact neighbors than surface residues in general; high-connectivity resi-
dues are found in contact with the ligand in 84% of all complexes studied. In addition, a k-means algorithm was
developed that may be useful for identifying potential binding sites with no obvious geometric or connectivity fea-
tures. The analysis was primarily carried out on 61 protein–ligand structures from the MEROPS protease database,
250 protein–ligand structures from the PDBSelect (25%), and 30 protein–protein complexes. Analysis of four pro-
teases with crystal structures for multiple bound ligands has shown that residues with high connectivity tend to have
less variable side-chain conformation. The relevance to drug design is discussed in terms of identifying allosteric-
binding sites, distinguishing between alternative docked poses and designing protein interface inhibitors. Taken
together, this data indicate that residue–residue connectivity is highly relevant to medicinal chemistry.
q 2010 Wiley Periodicals, Inc. J Comput Chem 00: 000–000, 2010

Key words: connectivity; ligand-binding sites; k-means; docking; allosteric-binding sites; protein–protein interface
inhibitors; molecular chaperones; local connectivity

Introduction methods, which more successfully predict protein–protein inter-

The centrality of binding sites to medicinal chemistry has led to
a multitude of different approaches for their identification in a
protein structure and their subsequent exploitation, including
methods based on structure,1–5 sequence conservation,6–10 and
interaction energies.11–14 Here, we describe a method based on
identifying residue–residue contacts and apply it to various types
of protein–ligand–related problems including proteases, residue
conformational flexibility, correctly and incorrectly docked
poses, allosteric-binding sites, and inhibitors of protein–protein
interactions. With reference to the protein–protein interface
inhibitors, we note that it is generally easier to detect ligand-
binding sites than protein binding sites,15 and indeed, although
the methods described here do detect the binding site for a num-
ber of recently published protein–protein inhibitors, they are
poor at predicting protein–protein interfaces. The methods
described here are completely complementary to sequence-based
faces,6–10,16,17 but may have some relationship to other geome-
try-based methods, because residues at the bottom of a binding
site cleft will be likely to form a greater number of residue–resi-
due contacts than residues around the lip of the cleft or residues
distant to the cleft. Similarities may also exist with energetic
methods based on the use of probes11,13,14 because when the
probe is deep within a cleft it will have more close contacts,
giving a higher score. However, the approach described here is
essentially independent of these methods and so although it has
Additional Supporting Information may be found in the online version of
this article.
Correspondence to: C. A. Reynolds; e-mail: c.a.reynolds@essex.ac.uk
Contract/grant sponsors: BBSRC, Medivir (UK), Royal Society for Fund-
ing (Theo Murphy Blue Skies Award)

q 2010 Wiley Periodicals, Inc.

2 Illingworth et al. • Vol. 00, No. 00 • Journal of Computational Chemistry
merit in its own right, it may also be used in combination with
other methods in consensus approaches.
The key role played by highly connected residues in the tran-
sition state for in vitro folding,18–21 as determined, for example,
by F-value analysis of a series of soluble proteins, raises the
possibility that such highly connected residues may also play an
important role in stable ground-state protein structures. For
example, Vendruscolo et al.20 used F-value analysis to show
that a cluster of well-connected residues in acylphosphatase
made a key contribution to the folding transition state. Within
ground-state structures, such highly connected residues are
expected to possess greater than average stability as their greater
number of neighbors are likely to restrict both their natural ther-
mal movement and also their movement in response to external
stimuli such as ligand binding, substrate transformation, or oli-
gomerization. Here, we therefore test the hypothesis that ligand
binding to residues with a high local connectivity is expected to
be favorable; one rationale for this is that there will be compara-
tively less entropy loss on binding compared with binding to
residues that are not so highly constrained. This should be
reflected by ligand-binding sites having a higher than average
connectivity to surrounding residues. Here, our approach, which
examines the local properties of a residue, is distinct from other
graph-theoretical approaches in which the key property being
measured is a function of all the residues in the protein.5,19,21
Some additional evidence for a link between binding-site resi-
dues and connectivity comes from Brownian dynamics studies,
where the catalytic residues of a number of proteins were found
to have high force constants,22,23 implying severe constraints on
their movement. Here, this connectivity effect will be analyzed
using a number of different metrics of connectivity for a series
of protein–ligand complexes. We show that ligand-binding sites,
but not protein interfaces, have a greater number of residue–resi-
due contacts than surface residues in general; the relevance of
these findings for molecular recognition, drug design in general,
and virtual screening in particular is discussed.
Methods
Connectivity and Selection of Test Systems
To test the basic hypothesis of a link between residue–residue
connectivity and ligand-binding sites, the number of residue–res-
idue contacts of a residue was counted according to two differ-
ent methods. In the first (inclusive) method, the number of resi-
due–residue contacts was defined as the number of neighboring
residues at least three residues in the sequence from the initial
residue that are in contact within 5.5 Å; this is the method used
below unless otherwise stated. In the second (grouped) method,
the number of contacts with distinct parts of the protein chain
was counted. According to this second method, residues that
contact within 5.5 Å were identified as in the first method. Next,
pairs of contacting residues within six residues of each other in
the chain were grouped and counted as a single contact. This
finds the number of disjoint parts of the protein chain that con-
tact any one residue, distinguishing between a residue that has N
contacts all lying close together and a residue that has N con-
Journal of Computational Chemistry
tacts distributed throughout the sequence. In general, the number
of residue–residue contacts of a residue is referred to as the con-
nectivity of that residue. (Here, it is important to note that the
local connectivity defined below is more complex than the sim-
ple measure given here in that it gives a higher connectivity to
residues in high-connectivity regions).
To study the effect of connectivity on ligand binding, a set
of 331 proteins were considered including 61 proteases taken
from MEROPS24 and 270 proteins taken from the PDBSelect
25% list,25,26 each of length at least 100 residues, and with
some ligand-binding residues. For NMR structures, the first
structure in the ensemble was chosen. The ligand-binding resi-
dues for the PDBSelect set were identified using the LPC data-
base,27 accessed via the RCSB protein databank,28 whereas for
the other proteins they were defined as residues within 5.5 Å of
the primary ligand. Only surface-accessible residues were con-
sidered, and these were identified as having at least 5% surface-
accessible area as determined using the program NAccess.29
The calculations were performed to test the hypothesis that
residues in ligand-binding sites have greater than average con-
nectivity. For each of the two measures of connectivity
described above, the maximum and average connectivity over
the ligand-binding site were compared to the maximum and av-
erage connectivity over the whole of the protein surface for both
the MEROPS and the PDBSelect sets of proteins. For the MER-
OPS set, the calculations were repeated, with catalytic residues,
identified from the Catalytic Site Atlas,30 replacing ligand-bind-
ing site residues.
To represent the results in a graphical way, the number of
residue–residue contacts for the residues of a protein was nor-
malized over the range [0,1] to produce images of the protein in
line with B-factor plots, with blue representing low connectivity
and red high connectivity.
Binding-Site Prediction: Allosteric-Binding Sites,
Rhodopsin, and Protein–Protein Interfaces
To examine the application of connectivity to identifying alloste-
ric-binding sites, calculations were performed for 11 examples
from 10 different systems identified as having allosteric-binding
sites.31–41
For each of these systems, the average connectivity was cal-
culated for residues on the protein surface and for residues in
the primary and allosteric-binding sites.
Extending the basic approach to residue–residue contacts
described above, two extensions of the method for semiauto-
matic binding-site prediction were derived using residue–residue
contacts to predict the location of binding sites; these are the
local connectivity method and the k-means approach.
Local Connectivity Method
In the first method, referred to as the local connectivity method,
a local measure identifying points of high connectivity was
defined. For each of the heavy atoms in a protein, the set of resi-
dues with at least one heavy atom within 5.5 Å of that atom
was found (this set of residues will include the residue of the
original atom), and the average number of residue–residue con-
DOI 10.1002/jcc
 Connectivity and Binding-Site Recognition
tacts of these residues was calculated, this measure being
referred to as the local connectivity of the atom. To identify
high-connectivity residues, the atom within the entire protein
with the highest local connectivity was then identified. A residue
was defined as having high connectivity if it contained at least
one atom with a local connectivity at least 90% of the maximal
local connectivity for the atoms of the protein. A particular
advantage of this measure is that it explicitly ensures that a
high-connectivity residue in a high-connectivity region scores
higher than a high-connectivity residue in a low-connectivity
region. To identify ligand-binding sites, only high-connectivity
residues on the surface of the protein were considered.
The local connectivity method was applied to the proteins
from the MEROPS set. For each protein, surface residues with
high connectivity were identified and examined for whether or
not they were in the ligand-binding site. This method is designed
to identify key residues in the binding site rather than the entire
binding site. In some cases, none of the surface residues was
identified as being high connectivity. Where this was the case,
the percentage cutoff was lowered in increments of 5% until a
nonempty set of surface residues was found. For comparison,
the same proteins were examined combining a similar percent-
age cutoff measure with the ‘‘closeness centrality’’ method
applied by Amitai et al.,5 which takes into account the network
properties of the protein as a whole. In an implementation simi-
lar to that of the local connectivity method (details given in Sec-
tion 1 of Supporting Information), binding-site residues were
predicted, and a comparison was made of the relative perform-
ance of the methods.
The local connectivity method was applied to a group of sys-
tems representing areas of current pharmacological interest.
These included rhodopsin, the prototypical G-protein–coupled
receptor (GPCR)42 because of its prominence as a representative
of possibly the most common drug target.43–46 Six structures
taken from five systems in which small molecules have been
demonstrated to inhibit the formation of protein–protein interac-
tion47 were also studied [PDB codes: 1PY2 (IL-2—protein part-
ner is IL-2R), 1R6N (N-terminal transactivation domain of
human papillomaviruses E2 helicase—protein partner is the E1
helicase), 1RV1 and 1T4E (MDM2—protein partner is p53),
1Y2F (ZipA—protein partner is FtsZ), and 2YXJ (Bcl-xL—pro-
tein partner is BAK). To account for potential conformational
changes induced in ligand binding, a parallel set of structures, in
which the structure of the protein had been determined in the
absence of the ligand, was subjected to the same analysis (Pdb
codes: 1M47, 1R6K, and 1Z1M (residues 25–109), 1Z1M (resi-
dues 16–111), 1F7W, and 1MAZ).
k-Means Method
In a second method for identifying binding sites, referred to as
the k-means method, an adaptation of the k-means algorithm48
was implemented. This identifies residues in the binding site
based on a global measure of connectivity across the whole of
the protein surface. The method divides the protein surface into
clusters, with cluster centers, or seed points, determined by a
weighted averaging of connectivity scores of residues in each
cluster. The k-means method was applied to the proteins in the
Journal of Computational Chemistry
3
MEROPS set to find seed points, which correspond to points of
high connectivity on the protein surface, and was compared to
the local connectivity measure described above. Full details of
the method are given in Section 2 of Supporting Information.
Connectivity and Side-Chain Conformation
To test the hypothesis that a high number of residue–residue
contacts leads to a loss of residue flexibility, four example sets
of X-ray structures with different ligands bound to an identical
protein were taken from the work of Fairlie et al.49 The exam-
ples taken were endothiapepsin with 20 different bound inhibi-
tors, porcine pancreatic elastase with 15 different bound inhibi-
tors, thermolysin with 12 different bound inhibitors, and papain
with eight different bound inhibitors. Dihedral angles in the side
chains of each residue were used to divide residues into confor-
mational clusters, based on an extensive analysis of amino acid
conformation.50 In each set of structures, the variation in confor-
mation was then described in terms of entropy, given by eq. (1).
X
S¼
pi log pi ; (1)
clusters
where pi is the observed probability of the side chain of a resi-
due being in cluster i. Full details are given in Section 3 of Sup-
porting Information.
We note that the entropy measured here differs from the
sequence entropy, another property which has been linked with
residues of functional importance.51 Here, a value of S equal to
zero indicates that the conformation of the side chain of a resi-
due is roughly constant across the set of structures, whereas
increasing variance in the conformation of a residue leads to
increasing values of S.
Docking
Redocking of the X-ray crystallographic ligand was carried out
for nine systems using autodock 4.52 A large docking box was
used that covered approximately a quarter of the protein. The
Spearman rank correlation coefficient was used to determine
whether the root mean square displacement (RMSD) of the
docked ligand relative to the crystallographic coordinates corre-
lated with connectivity of the residues contacting the ligand.
Protein–Protein Interactions
Thirty protein complexes (listed in Section 4 of Supporting In-
formation) were retrieved from the PQS Protein Quaternary
Structure server53 to enhance the probability that the dimeric
interface is genuine rather than an artifact of nonspecific crystal
packing. The connectivity of the protein was determined over
the surface residues as above. In addition, the protein was exam-
ined by the evolutionary trace (ET) method using in-house
code,8,9 and the results indicated that 13 of 60 had only one
interface, whereas the ET results of the remaining proteins indi-
cated the presence of additional interfaces or binding sites
besides the crystallographic dimer interface. These 13 structures
were studied to determine whether connectivity could be used to
DOI 10.1002/jcc
4 Illingworth et al. • Vol. 00, No. 00 • Journal of Computational Chemistry

identify the protein-interaction interface. The ET method sorts

residues into bins according to residue conservation, and this
data were converted into a measure analogous to connectivity,
the ET score, defined according to the formula

ETscore ðiÞ ¼ N
binðiÞ þ 1; (2)

where N is the total number of bins from the ET method (here,

20 bins were used), and i is a residue in the protein that is sorted
by the ET method into bin i. The ET score is large if the residue
i is highly conserved, and small if the residue i is poorly con-
served.

Connectivity and Residue Type

An investigation was carried out into the dependence of connec-

tivity on residue type. It has been observed that catalytic resi-
dues in the active sites of enzymes are more likely to be charged
or polar than residues elsewhere in the protein.54 Bulkier amino
acids might reasonably be expected to have a higher connectiv-
ity score, and if these amino acids were also more likely to be
involved in interactions with ligands that would result in a bias
toward higher connectivity in ligand-binding sites. For both of
the MEROPS and PDBSelect sets of proteins, the number of
each of the 20 amino acids occurring in the binding site, on the
protein surface, and in the protein as a whole was determined.
Mean connectivity scores were calculated for each of the 20
amino acids, considering residues on the surface of proteins
across each of the two sets in turn. These mean connectivity
scores were used to calculate ‘‘expected’’ connectivity scores for
the ligand-binding sites of each protein, calculated by counting
the number of occurrences of each amino acid in the binding
site, summing the mean connectivity scores for each amino acid,
and dividing by the number of residues in the binding site.
These ‘‘expected’’ scores were compared to observed connectiv-
ity values.
Software
The software is available from the authors at ftp://ftp.essex.ac.
uk/pub/oyster/Connectivity/Connectivity.tar.
Figure 1. The distribution of connectivity (normalized by the total
number of neighbors in the protein set) for 270 proteins from
PDBSelect, for all surface residues (dashed) and for residues that
are in contact with a ligand (solid).
is on an average about 25% higher in the binding site than
across the surface as a whole. This is a somewhat surprising
result as the volume occupied by the ligand will ensure that the
connectivity of the binding site is inevitably reduced as contacts
to the ligand are not included. Here, we note that, because of its
grouping of contacts, the second method of measuring connec-
tivity gives significantly lower connectivity values. However, for
both sets, the difference between the average surface and aver-
age binding-site connectivities is greater when this second
(grouped) method is used, indicating a positive contribution of
distant contacts to the high connectivity of ligand-binding sites.
The individualized results for proteins in the MEROPS dataset
are shown in Table 2 and indicate that for 59 of these 61 pro-
Table 1. Mean Connectivity for the MEROPS and PDBSelect Datasets
Containing 61 and 270 Proteins, Respectively.
Connectivity

Results Mean highest Mean average

Connectivity and Ligand Binding Dataset/Methods Surface Binding site Surface Binding site Da

A clear difference was observed between the connectivity of res- MEROPS1 15.4 13.9 6.6 8.2 3.5
idues in the binding site and other residues on the protein sur- MEROPS2 6.4 5.7 2.6 3.4 4.0
face. Figure 1 shows the relative distribution of residue connec- PDBSelect1 14.8 13.0 6.5 7.6 2.7
tivity for residues from the 270 proteins in the PDBSelect data- PDBSelect2 5.9 5.2 2.5 3.1 2.9
set, for all surface residues, and for residues in the binding site.
Residues in the binding site generally have higher connectivity, The numbers 1 and 2 indicate the method of determining connectivity. 1
indicates the inclusive method, 2 indicates the grouped method. Columns
with comparatively fewer having a connectivity between 0 and
2 and 3 show the mean value of the maximal residue connectivity. Col-
7, and comparatively more having a connectivity between 7 and umns 4 and 5 show the mean value of the average residue connectivity of
13. For each protein, the average connectivity for residues in the residues on the surface as a whole and in the binding site of proteins in
surface and for surface residues in the binding site was calcu- each set. Column 6 gives the difference between columns 4 and 5.
lated. Table 1 presents the statistics for these values under the a
Difference between surface and binding-site average in standard devia-
different connectivity measures and shows that the connectivity tions of the surface figure.

Journal of Computational Chemistry DOI 10.1002/jcc

 Connectivity and Binding-Site Recognition 5

Table 2. Connectivity on the Protein Surface and at the Ligand-Binding Table 2. (Continued)
Site for 61 Proteins from the MEROPS Database.
Maximum connectivity Mean connectivity
Maximum connectivity Mean connectivity
Protein Whole surface Binding site Whole surface Binding site
Protein Whole surface Binding site Whole surface Binding site
1uk4 17 14 6.52 6.92
1a16 15 15 6.67 9.00 1wht 18 18 6.96 9.58
1afq 16 16 6.47 7.90 2rmp 17 14 6.62 8.27
1ajq 17 15 6.22 11.10 3apr 14 13 6.64 8.09
1ao0 19 19 6.41 7.27 4aig 14 13 6.40 6.94
1arc 16 13 6.64 8.71 5sga 16 14 6.93 7.81
1ayu 15 12 6.77 7.58
1b6a 14 14 6.87 9.55 Column 3 is denoted in bold when the maximum connectivity is at the
1bh6 14 13 6.88 7.52 binding site, and column 5 is denoted in bold when the mean connectiv-
1bil 13 12 6.00 6.29 ity of the biding site is greater than mean connectivity of the surface as
1bmq 15 13 6.18 7.56 a whole (including the binding site).
1bqy 15 13 6.45 7.85
1bru 17 12 6.48 9.69
1bxo 15 14 6.66 8.45 teins, the mean connectivity of the binding site is higher than
1c24 14 14 6.87 10.20 the mean over the whole of the protein surface (the exceptions
1cgh 18 16 6.58 9.41 being 1cmx and 1d4l). A similar result to that for binding-site
1cmx 11 9 4.94 3.25 residues was obtained for catalytic residues. Of 58 proteins in
1cv8 16 12 6.48 8.06 the MEROPS set for which catalytic residue information was
1cvr 15 12 6.83 8.64 available, the average connectivity of the catalytic residues was
1cvz 13 12 6.58 7.61 higher than that for the protein as a whole (including surface
1czi 18 17 7.90 8.33
and buried residues) in 54 cases (Supporting Information Table
1d4l 12 9 4.87 4.80
5). A number of distinctly different proteins in Tables 1, 2, and
1dmt 16 15 7.46 9.78
1dy9 13 12 6.42 7.33 5 share common ligands or ligand types. For example, analysis
1eag 15 13 6.74 7.77 of the PDBSelect set revealed 18 proteins that bound ATP.
1f2o 15 14 6.82 9.17 Analysis of the proteins against the CATH database showed that
1fh0 16 16 7.04 8.19 each contained either a two- or three-layer a-beta sandwich do-
1fjs 16 16 7.01 8.46 main. The observation of higher connectivity applied in each of
1ft7 16 16 7.16 9.75 the 18 cases, regardless of the fold. Six of these proteins had the
1fxy 15 14 6.50 8.32 Walker A box motif (www.expasy.ch/prosite/PS00017); for each
1ga6 16 13 6.81 8.20 of these proteins, the average connectivity for residues in the
1gec 14 13 6.82 8.41
motif was higher than the average for the surface.
1gmy 19 14 6.95 8.20
Representative plots of connectivity for the systems with pdb
1gvu 14 14 6.57 8.45
1hne 15 14 6.50 8.88 codes 1c24 (methionine aminopeptidase), 1s4v (cysteine endo-
1hpg 16 13 6.70 7.44 peptidase), 1cvr (gingipain R), and 1cmx ((ubiquitin C-terminal
1i76 14 13 6.29 8.06 hydrolase) are shown in Figure 2 (lhs)—these four examples
1kug 17 17 6.61 7.50 were chosen to be as diverse as possible in terms of the quality
1lf2 16 13 6.48 8.04 of the results obtained. For the first two of these systems, the
1ls5 14 13 6.74 7.93 higher degree of connectivity of the ligand-binding site can
1lyb 15 12 5.56 6.47 clearly be seen. For the system 1cvr (Fig. 2c), higher connectiv-
1m4h 15 14 6.68 8.32 ity is revealed by the statistics, though it is not so obviously visi-
1me4 15 14 6.59 7.81
ble. The system 1cmx (Fig. 2d) is one of the two examples from
1mu0 16 16 7.06 12.00
the 61 MEROPS proteins for which the connectivity in the bind-
1n1m 16 15 7.09 11.00
1nqc 15 14 6.88 8.33 ing site was lower than for the surface as a whole. Examination
1nw9 14 14 5.86 6.92 of this system shows two ligands binding in a low-connectivity
1onx 15 15 6.90 9.92 region. However, because one of these is bound covalently, there
1pfx 16 14 6.69 8.94 are clearly additional principles affecting the binding.
1pwu 17 15 6.84 7.69
1qjj 14 14 6.31 9.10 Connectivity and Residue Type
1qrp 16 14 6.62 8.14
1qs8 16 13 6.63 7.93 Summing across all residues in all proteins in the MEROPS set,
1s2k 17 14 6.52 8.61 the average connectivity for a residue on the surface of a pro-
1s4v 15 12 6.55 8.36 tein, averaged across all residues, was found to be 6.67. Eleven
1smr 17 15 6.60 7.33
amino acids had average connectivity scores higher than this
(continued) (Gln, Val, Arg, Ile, His, Leu, Cys, Met, Phe, Tyr, and Trp,

Journal of Computational Chemistry DOI 10.1002/jcc

6 Illingworth et al. • Vol. 00, No. 00 • Journal of Computational Chemistry
Figure 2. Connectivity plotted over the surface of (a) 1c24, (b)
1s4v, (c) 1cvr, and (d) 1cmx; red indicates high connectivity, and
blue indicates low connectivity. The ligand is displayed as a black
stick diagram. Clusters of residues (varying colors) and high-connec-
tivity points (white) generated by the k-means algorithm for the pro-
teins (e) 1c24, (f) 1s4v, (g) 1cvr, and (h) 1cmx. The ligand is shown
in a contrasting color (blue or yellow, stick diagram).
scoring on an average between 6.9 and 11.3), and analysis of
the ligand-binding sites showed a prevalence of these residues.
In total, these 11 higher connectivity residues accounted for
49% of the residues in the ligand-binding sites, but for only
38% of residues in the surface as a whole. Although appearing
to support the hypothesis that higher connectivity in binding
sites is the result of differing residue composition, this factor did
not account for the higher connectivity values observed in the
binding sites. Using the average connectivity scores for each of
the amino acids, an expected value was generated for the con-
nectivity of the binding sites of each of the proteins in the
MEROPS set, based on their residue composition. Over this set,
the mean connectivity score of residues in the binding site was
Journal of Computational Chemistry
24% higher than the expected value, with a higher connectivity
score occurring in 56 of 61 cases. Similar results were obtained
for proteins in the PDBSelect set, as described in Section 5 of
Supporting Information.
Ligand Docking and Binding-Site Prediction
Binding-Site Prediction
When the local connectivity method for binding-site prediction
was applied to the MEROPS set, 190 high-connectivity residues
were identified, an average of just over three per protein. Of
these, 94 or 49% were in the ligand-binding site, with at least
one point in the binding site in 51 of 61 cases. By comparison,
an implementation of the ‘‘closeness centrality’’ method5 applied
to the same set of structures also identified a residue in the bind-
ing site in 51 of 61 cases. The structures for which the respec-
tive methods failed were different for the two algorithms, with
only two binding sites not identified by either method, suggest-
ing that the two approaches are complementary; the local con-
nectivity method did how identify more catalytic residues (Sup-
porting Information Table S1).
The k-means algorithm gave an alternative method of binding-
site prediction. Representative plots showing output from the
k-means algorithm for the MEROPS systems with pdb codes
1c24, 1s4v, 1cvr, and 1cmx are shown in Figure 2 (rhs). For 1c24
and 1cvr, Figures 2e and 2g, a high-connectivity seed point is
present in the ligand-binding site. The result for 1cvr is particu-
larly significant, as the ligand-binding site is not obvious from a
visual inspection of the connectivity of the protein surface, but is
identified by the k-means algorithm. In cases where the average
connectivity in the binding site is lower than the average for the
protein surface, such as in 1cmx, Figure 2h, the k-means algorithm
is extremely unlikely to identify the correct ligand-binding site.
In general, the ligands did not necessarily bind exclusively to
a single k-means cluster but rather tended to bind to residues in
two or more clusters. This is a reasonable result as the determi-
nation of the number of clusters in k-means analyses is some-
what arbitrary and the MEROPS peptide-like ligands tend to be
large. Thus, the binding of a ligand to more than one cluster
should not be taken to imply multiple experimental binding
modes. For 37 of the 61 proteins, a seed point was found to
occur in the set of ligand-contacting residues, which is quite
high considering that the average size of these clusters was 71
residues (range 17–249). This compared to an expected value of
no more than 18 of 61 if seed points were chosen at random
(calculation details given in Section 6 of Supporting Informa-
tion), confirming the association between high-connectivity
points and ligand binding. Thus, although the local connectivity
method of binding-site identification is superior, the k-mains
approach may nevertheless be helpful in niche applications: for
example, in proteins where there is no obvious region of high
connectivity (Figs. 2c and 2g).
Ligand Docking
The correlations between the RMSD score for the docked pose
and the connectivity for the residues contacting the ligand show
DOI 10.1002/jcc
 Connectivity and Binding-Site Recognition 7

two distinct types of behavior. Where a large docking box is

used so that the ligand may bind to multiple distinct regions of
the protein, there is a high correlation, showing that connectivity
may be used as a filter to help determine the correct binding
site. However, where the ligand docks to a single binding site,
there is a low correlation, showing that connectivity may not be
used to determine the correct pose. Thus, Supporting Informa-
tion Table S6 shows the Spearman rank correlation coefficient
calculated for the correlation between the RMSD score and con-
nectivity for the residues contacting the ligand for each of the
nine dockings. In Supporting Information Table S6, where the
different ligand poses are in the same binding site (1ett and
1pph), and in Supporting Information Table S7, the correlation
coefficient values are of low significance, but of the remaining
seven dockings in Supporting Information Table S6, six have a
strong correlation between high connectivity and low RMSD,
the exception being 1lgr. Figures 3a and 3b show the strong cor-
relation for the two ligands with the highest correlation coeffi-
cients (0.78 and 0.83 for 1dwb and 4ts1, respectively). The
results for AMP binding to glutamine synthetase, pdb code 1lgr,
are shown in Figure 3c, and here a region of high connectivity
can be seen 13–17 Å from the location of the ligand in the crystal
structure, and further examination of the structure suggests that
this represents an alternative binding site. Glutamine synthetase
binds both ATP and glutamate simultaneously in a single binding
channel. In the docking, the AMP ligand was sometimes incor-
rectly docked in the glutamate site, about 14 Å from its location
in the crystal structure. A later study of the same enzyme is
recorded in the pdb file 1fpy,55 in which phosphinothrycin occu-
pies the glutamate binding site. Thus, in all cases studied, includ-
ing 1lgr, high connectivity is associated with a genuine binding
site, albeit a binding site for a different ligand in the case of 1lgr.
This observation raises the possibility that connectivity could also
be used to identify allosteric-binding sites.

Allosteric-Binding Sites

Examination of the allosteric-binding sites showed higher than
average connectivity in the majority of cases. Details of the
results for these systems are shown in Table 3. In each of the 11
cases, the primary ligand-binding site had a higher connectivity
than did the surface in general. Results for the allosteric ligands
identified a binding-site connectivity greater than that for the
protein surface in 9 of 11 cases. This suggests that allosteric-
binding sites are more difficult to identify using connectivity
than the primary binding site, although connectivity may be of
use in many cases. In fructose-1,6-bisphosphatase (1q9d, 2fhy),
human mitochondrial NAD(P)1-dependent malic enzyme
(1gz3), and ribonucleotide reductase R1 protein (4r1r), the
enzyme is an oligomer and the higher connectivity of the allo-
steric-binding site includes contributions from both monomers.35
Thus, in two of cases where the allosteric-binding site did not
have high connectivity, inclusion of a second protein chain
reversed this result. For each of the allosteric-binding sites, simi-
lar results were obtained regardless of whether the results were
performed on the bound cases or the unbound cases.
The results on rhodopsin (detailed below) are also relevant in
this respect because allosteric ligands are of interest for systems
Figure 3. The relationship between connectivity and RMSD for var-
ious ligands redocked to their native crystal structure using auto-
dock: (a) 4ts1, (b) 1dwb, and (c) 1lgr. A large docking box was
used in these docking experiments. Images of docked poses and
filtered docked poses for 4TS1 and 1LGR are given in Supporting
Information Figure 1.
such as the muscarinic subfamily,56,57 where the allosteric-bind-
ing site is associated with the extracellular region. For class C
GPCRs where the main ligand binds to the N-terminus, the cav-
ity within the helical bundle, which has high connectivity (see
below), becomes the allosteric-binding site.58 In addition, for
rhodopsin there is a high connectivity region in the cytoplasmic
domain that has recently been shown to bind a small molecule
(see below).
GPCRs
Connectivity analysis of rhodopsin showed a markedly increased
connectivity in the ligand-binding site, with an average score of
11.5, compared with an average for the surface of 7.4. Three of

Journal of Computational Chemistry DOI 10.1002/jcc

8 Illingworth et al. • Vol. 00, No. 00 • Journal of Computational Chemistry

Table 3. Connectivity (conn.) Data for the Surface, Primary Binding Site, and Allosteric-Binding Site.

Primary ligand Allosteric ligand

Connectivity Connectivity

No. pdb Ligand Protein Binding site Ligand Protein Binding site

1 1q9da Fructose 7.0 9.3 GC252-354 7.0 5.3

2fhyb Fructose 7.6 9.3 GC252-354 7.6 7.7
2 1iwhc Heme 7.1 8.6 Benzafibrate 7.1 6.6
3 1gz3 Tartronate 7.4 10.6 Fumarate 7.4 10.6
4 4rlra GDP 7.1 7.6 dTTP 7.1 6.6
4rlrb GDP 7.1 7.6 dTTP 7.1 7.2
5 2wu1 NAG-6-phosphate 6.9 8.2 NAG-6-phosphate 6.9 6.4
6 3hrf ATP 6.4 7.5 P48 6.4 7.9
7 3ijo Glutamate 6.6 11.2 Althiazide 6.6 8.8
8 3ddw/3ceh C27H22ClN3O4 7.3 8.5 AVE5688 7.4 8.3
9(i) 3h1p/1shj DEVD_CHO 6.6 8.1 DICA 6.5 7.1
9(ii) 3h1p/1shl DEVD_CHO 6.6 8.1 FICA 6.6 7.8
10 2wmq/3jvr C9H11N3O2S 6.5 6.9 C16H13Cl2N3O2 6.2 6.6

There are 11 allosteric ligands for the 10 systems listed.

a
Monomer.
b
Dimer.
c
Generally, the k-means seed point was not part of the binding site except for the main binding site for 1iwh and the
allosteric-binding site for 2fhy.

the six highest connectivity residues are in extracellular loop 2,
underlying the importance of this loop for the stability of rho-
dopsin-like GPCRs. Using the local connectivity method, high-
connectivity residues were identified at the 85% and 90% cutoff
levels. At the 90% level all residues found were adjacent to the
main ligand-binding site within the helical bundle. At the 85%
level, although most residues were close to the ligand, four addi-
tional residues were identified, these being Tyr10 (N-terminus),
Leu72, Leu77 (interface between helix 1 and helix 2), and
Thr93 (interface between helix 2 and helix 3). Leu72 is on the
intracellular face of the receptor and may participate in binding
G-protein, but its mutation to cysteine has negligible effects on
G-protein activation59; the equivalent residue is in contact with
an octyl glucoside molecule in the squid rhodopsin structure.60
However, Leu 72 has been implicated in small molecule bind-
ing, namely chlorin to Meta-II state rhodopsin61 (Klein-Seethara-
man, personal communication). It is possible that the high-con-
nectivity sites on the lipid-facing regions of the GPCR are asso-
ciated with protein–protein interactions9,16,62 (see below).
Small-Molecule Inhibition of Protein–Protein Interfaces
Results of the application of the local connectivity method are
shown in Table 4. In the structures with bound ligands, between

Table 4. High-Connectivity Point Data for Protein–Protein Interface Systems, from Structures Derived with
and Without Ligand Bound.

Ligand bound Ligand unbound

PDB Cutoff (%) #Points #Points correct PDB Cutoff (%) #Points #Points correct

1PY2 85 1 0 1M47 85 6 1
1R6N 90 3 0 1R6K 90 7 3
1RV1 90 1 1 1Z1Ma 90 4 2
1T4E 90 1 1 1Z1Mb 90 7 3
1Y2F 90 2 1 1F7W 90 4 2
2YXJ 85 3 1 1MAZ 80 2 1

The cutoff describes the maximal percentage (in increments of 5%) such that at least one residue on the protein sur-
face contained an atom with a least this percentage of the maximal local connectivity for the protein as a whole.
#Points describe the number of such residues, and #points correct describes the number of these residues in the
ligand-binding site.
a
Residues 25–109.
b
Residues 16–111.

Journal of Computational Chemistry DOI 10.1002/jcc

 Connectivity and Binding-Site Recognition
Figure 4. High-connectivity residues identified in the structures of
human MDM2 (a) with bound benzadiazepinedione and (b) without
a bound ligand, and of interleukin-2 (c) with bound small molecule
SP4206 and (d) without a bound ligand. The structure of the protein
is shown in cartoon format. Residues containing high-connectivity
points are shown in spacefill mode and colored green (residues in
the binding site) and red (residues not in the binding site). The
ligand is shown in stick format and colored in yellow.
1 and 3, high-connectivity points were identified in each protein,
with a high-connectivity point in the binding site in four of six
cases. A total of 36% of high-connectivity points were in the
binding sites. Repeating the method on the unbound structures
gave improved results. At least one high-connectivity point was
found to occur in the binding site in each case, with 40% of
high-connectivity points occurring in the binding sites. Figure 4
shows the locations of high-connectivity residues for two of the
structures that were examined. For the system 1T4E, a single
high-connectivity residue was identified, located in the ligand-
binding site. Repeating the connectivity analysis on the unligated
structure gave seven high-connectivity residues, of which three
were in the binding site. Of the remaining high-connectivity res-
idues, three were adjacent to the high-connectivity residues in
the binding site, whereas the remaining residue was on the other
side of the protein. The IL-2 system (1PY2) is an example of a
structure where the binding site is not located by the method—a
single high-connectivity residue is identified, away from the
binding site. Such high-connectivity residues could indicate
alternative functions, such as other binding sites or regions
important for protein folding.63 Thus, in the unligated IL-2
structure, six high-connectivity residues are identified, of which
Journal of Computational Chemistry
9
residue 96 is in the binding site of the 1PY2 ligand; of the
remaining five high-connectivity residues, residue 45 is a bind-
ing site for other ligands (PDB codes 1M48, 1M4B, 1QVN, and
1PW6), residue 92 is a lock residue for a closed loop,63 residue
20 is adjacent to a lock residue (lock residues are inevitably in
high-connectivity regions19,20,63), and residues 120 and 122 are
at the interface with the receptor (PDB code 2B5I).
Protein–Protein Interactions
The average connectivity in the protein–protein interface was
5.9 and is thus slightly lower than the average of 6.4 for the sur-
face as a whole. Of the 60 protein monomers studied, only 14
had a higher than average connectivity in the interface region,
and only 16 had a higher than expected connectivity based on
residue composition. This demonstrates a significant difference
between ligand-binding sites and protein–protein interaction
sites, namely that although high connectivity is associated with
ligand-binding sites, it is not necessarily associated with pro-
tein–protein interfaces.
In contrast, the average evolutionary trace score (ETscore)
value for the interface of a monomer was 11.2 compared with
10.3 for the surface as a whole. Forty-six of the 60 monomers had
a higher average ETscore in the interface region than for the surface
as a whole. When only the 13 monomers with a single high-
ETscore patch were considered, the overall results were not signifi-
cantly different (Supporting Information Table S2). Three of the
13 had a higher average connectivity in the interface region than
for the surface as a whole, whereas 10 of the 13 had a higher
average ETscore in the interface than for the surface as a whole.
Connectivity and Side-Chain Conformation
The relationship between connectivity and the extent of confor-
mational entropy resulting from ligand binding is reported in
Table 5. In each of the sets, the residues with low connectivity
(i.e., in the bottom third of residues ordered by connectivity
score) are likely to have nonzero conformational entropy over a
range of proteases (with probability 70% in the case of the set
1eed). Conversely, residues with high connectivity (i.e., in the
top third of residues) are likely to have zero conformational en-
tropy (with probability 67% in the case of the set 1eed).
Discussion
There are several approaches for determining binding sites in
proteins. These include data mining of sequence information
within multiple sequence alignments, as exemplified by the ET
method, sequence entropy and correlated mutations,6–10 search-
ing for suitable pockets based on geometric criteria,1–3 and
methods based on energy criteria,12,13 amongst which GRID is
the classic method.11 Such studies have shown that it is gener-
ally easier to detect ligand-binding sites than protein binding
sites.15 The method has advantages over sequence-based meth-
ods in that it is not reliant on either having sufficient homolo-
gous sequences or on the assumption that all homologous
sequences share the same fold at the point of interest. Neverthe-
DOI 10.1002/jcc
10 Illingworth et al. • Vol. 00, No. 00 • Journal of Computational Chemistry
Table 5. Normalized Proportions of Residues with Zero and Nonzero
Conformational Entropy (i.e., Residues That Do or Do Not Retain the
Same Conformation Across a Variety of Structures) Grouped by
Connectivity Score for Four Sets of Identical Protein Structures.
Connectivity
Lead PDB/# ligands Entropy Low Medium
1eed Zero 0.30 0.53
/20 Nonzero 0.70 0.47
1bma Zero 0.20 0.58
/15 Nonzero 0.80 0.42
1thl Zero 0.37 0.61
/12 Nonzero 0.63 0.39
1pad Zero 0.25 0.50
/8 Nonzero 0.75 0.50
The results show, for example, that if a residue in the set 1eed has low
connectivity, there is a 70% chance that it will have nonzero entropy;
here, the other 20 ligands bind to proteins identical to 1eed albeit with
different ligands and PDB codes.
less, the method described here is essentially independent of
these other methods and so may also be used in combination.
Thus, of the 61 proteins from the MEROPS database that
were studied, 59 were found to have a higher connectivity for
the ligand-binding site than for the protein surface as a whole.
The exceptions prove that a high connectivity is not an essential
property of a ligand-binding site, but the statistics as a whole,
carried out across a wide spectrum of different proteins, con-
firms that high local connectivity is a desirable property. Else-
where, network analysis has indicated that global connectivity is
associated with binding sites,5 but no explanation as to why was
given. Here, we propose that the most likely interpretation of
this is the role of a high connectivity in reducing the entropy
change of the enzyme on binding. Because of the interactions
involved in ligand binding, residues next to a bound ligand are
likely to have relatively low conformational entropy. The en-
tropy change in binding is thus more favorable if the residues in
their unbound state also have low (conformational) entropy. We
suggest that interactions with adjacent residues (i.e., those in
close physical proximity) provide this ordering in the binding
site. This conjecture is supported by the data on conformational
entropy in Table 5 because residues with high-connectivity
scores are more likely to present a consistent side-chain confor-
mation across a range of structures.
The results obtained bear some similarity to work independ-
ently carried out on Brownian dynamics-derived force con-
stants22,23 that revealed an association between residues with
high force constants and catalytic residues. A correlation was
noted between the force constant and the number of ‘‘neighbors
of neighbors’’ (c.f. contacts of contacts) of a residue.23 We have
shown a very strong correlation between the number of
‘‘neighbors of neighbors’’ of a residue and the square of the
number of contacts (results not shown).
The results also bear some similarity to work on protein fold-
ing, which indicates that high connectivity plays an important
role in the transition structure for folding18–20 for small two-state
Journal of Computational Chemistry
proteins folding in vitro. In this respect, therefore, it would seem
that ligand binding has some features in common with protein
folding. This work may therefore offer some insight into the
mechanism of molecular chaperones64 because it indicates that
ligands binding to high-connectivity regions may help to stabi-
lize the fold if these high-connectivity regions are indeed key to
High folding.63
0.67 Protein–Protein Interactions
0.33
0.72 The analysis of connectivity over a protein surface is potentially
0.28 problematical as it is not necessarily known a priori how many
0.55 interfaces reside on the protein surface. The application of ET
0.45 analysis to identify proteins that are likely to have a single inter-
0.72 face is therefore a useful control. Nevertheless, the connectivity
0.28 results are similar for both the set of 60 proteins with potentially
two or more interfaces and the smaller control set of 13 proteins
with only one predicted interface, namely that connectivity is
not particularly useful for identifying protein–protein interaction
sites and that in this regard the ET method is superior (see Sec-
tion 4 of Supporting Information). However, because sequence-
based methods such as entropy51 and ET6 can identify both
ligand-binding sites and protein-interaction sites, but cannot nec-
essarily distinguish between the two, the ET method could be
used in combination with connectivity to assess the nature of the
binding site. More significantly, this result indicates a fundamen-
tal difference between ligand-binding sites and protein-interac-
tion sites that is related to the fold of the protein as the former
but not the latter is likely to have high connectivity. This funda-
mental difference is probably one contributing factor toward the
observation that small-molecule ligands are more likely than
proteins to bind with high affinity. The design of ligands to in-
hibit large protein–protein interfaces is notoriously difficult, but
the use of connectivity to identify the small proportion of pro-
tein–protein interfaces that have the correct connectivity features
over a smaller area that are conducive for ligand binding may
assist in this process.
Application to Drug Design
Insights into the link between connectivity and ligand binding
may prove useful in the identification of binding sites on a pro-
tein surface. Coloring residues according to their connectivity
score provide a visual means by which potential binding sites
may be found. Alternatively, the k-means method gives precise
points on the protein surface at which binding may take place.
As can be seen in Figure 2, the algorithm can be effective in
identifying binding sites even when they are not obvious from
the protein geometry or from a superficial examination of the
residue connectivity (Fig. 2g). Thus, the system 1cvr is an exam-
ple of a system where the ligand-binding site is identified by the
k-means method, in which the ligand is not noticeably in a cleft
in the protein surface. Both methods do, however, give a general
guide to docking location, rather than a precise indication.
Attempts to rescore ligands from a docking run according to
the connectivity score of the residues they contact were success-
ful where the docking used a large docking box and results of
the docking run occupied different potential binding sites. Here,
DOI 10.1002/jcc
 Connectivity and Binding-Site Recognition
connectivity was useful in assessing whether the ligand had
docked to a valid binding site, even if the binding site is not the
correct one for that ligand. However, the method was not useful
in distinguishing between different ligand poses in the same
binding site, presumably because the variations in connectivity
were much less.
It remains, however, that for the general identification of
ligand-binding sites, the connectivity score can prove a powerful
and effective tool. This may be particularly useful in the search
for allosteric-binding sites. Here, we have analyzed 11 systems
for which allosteric-binding sites have recently been identified
by X-ray crystallography.31–34,58 In all of these cases, although
residues close to the primary ligand (or functional group) consis-
tently had higher connectivity than the surface in general, allo-
steric-binding sites were more difficult to identify, although pos-
itive results were obtained in most cases.
The link between connectivity and conformational entropy of
residues suggests that connectivity could be used to enhance vir-
tual screening experiments involving a flexible receptor. Prior
knowledge of which residues in a receptor are likely to retain the
same or a restricted set of conformations could result in the
design of more efficient docking experiments and result in consid-
erable savings in computational time. As has been shown here,
residues of higher connectivity are significantly more likely to be
rigid over a range of receptor structures and therefore make likely
candidates to be fixed in a flexible docking simulation.
Targeting drugs to inhibit protein–protein interactions offers
huge therapeutic potential but is nevertheless extremely difficult,
despite recent encouraging successes.47,65–67 Given that protein
interfaces generally possess fewer druggable topological fea-
tures,47 it may be necessary to probe multiple protein conforma-
tions, e.g., as generated by X-ray crystallography (here) or by
molecular dynamics,68 to identify possible binding sites, includ-
ing hidden binding sites, that could be exploited to inhibit pro-
tein–protein interactions. Although the results obtained were
mixed in quality, in the six systems studied here, it was suffi-
cient to study the protein monomers to identify possible binding
sites because in each system at least one of the high-connectivity
points identified was located in the binding site.
Conclusions
We have developed a family of methods based on residue–resi-
due connectivity for characterizing binding sites and have
applied them to various problems in computational medicinal
chemistry. The basic method allows high-connectivity binding
sites to be visualized, as in Figures 2a–2d. The k-means algo-
rithm (Figs. 2e–2h) automatically detects centers of clusters of
high connectivity and may be particularly useful when there is
no obvious region of high connectivity. The algorithm for deter-
mining residues with high local connectivity may also be used
in automatic mode, as illustrated in Figure 4.
Taken together, the results presented here have confirmed a
role for high connectivity in ligand-binding sites.
It is possible that ligands preferentially bind to regions of
high connectivity to minimize the loss of entropy on binding
that would occur if the motion of flexible side chains were to be
Journal of Computational Chemistry
11
restricted. In this respect, ligand binding seems to utilize princi-
ples that are also common to protein folding, though these prin-
ciples do not apply in the same way to the prediction of pro-
tein–protein interactions. We have shown that connectivity can
be used to indicate the expected degree of side-chain movement
on ligand binding; it can also be used to search for allosteric-
binding sites and to delineate poses that have been docked to a
valid binding site from those that have not. The low connectivity
observed at most protein–protein interfaces may underlie the dif-
ficulty in designing protein interface inhibitors, but, nevertheless,
connectivity can be used to indicate possible binding sites for
protein–protein interface inhibitors.
Acknowledgment
The authors acknowledge Garrett Morris for a copy of autodock
and for helpful discussions.
References
1. Laskowski, R. A.; Luscombe, N. M.; Swindells, M. B.; Thornton, J.
M. Protein Sci 1996, 5, 2438.
2. Hendlich, M.; Rippmann, F.; Barnickel, G. J Mol Graph Model
1997, 15, 359.
3. Venkatachalam, C. M.; Jiang, X.; Oldfield, T.; Waldman, M. J Mol
Graph Model 2003, 21, 289.
4. Wangikar, P. P.; Tendulkar, A. V.; Ramya, S.; Mali, D. N.;
Sarawagi, S. J Mol Biol 2003, 326, 955.
5. Amitai, G.; Shemesh, A.; Sitbon, E.; Shklar, M.; Netanely, D.;
Venger, I.; Pietrokovski, S. J Mol Biol 2004, 344, 1135.
6. Lichtarge, O.; Bourne, H. R.; Cohen, F. E. Proc Natl Acad Sci USA
1996, 93, 7507.
7. Sowa, M. E.; He, W.; Slep, K. C.; Kercher, M. A.; Lichtarge, O.;
Wensel, T. G. Nat Struct Biol 2001, 8, 234.
8. Dean, M. K.; Higgs, C.; Smith, R. E.; Bywater, R. P.; Snell, C. R.;
Scott, P. D.; Upton, G. J.; Howe, T. J.; Reynolds, C. A. J Med
Chem 2001, 44, 4595.
9. Thummer, R. P.; Campbell, M. P.; Dean, M. K.; Frusher, M. J.;
Scott, P. D.; Reynolds, C. A. J Mol Neurosci 2005, 26, 113.
10. Filizola, M.; Olmea, O.; Weinstein, H. Protein Eng 2002, 15, 881.
11. Goodford, P. J. J Med Chem 1985, 28, 849.
12. Laurie, A. T.; Jackson, R. M. Bioinformatics 2005, 21, 1908.
13. Harris, M. R.; Kihlen, M.; Bywater, R. P. J Mol Recognit 1993, 6, 111.
14. Glick, M.; Robinson, D. D.; Grant, G. H.; Richards, W. G. J Am
Chem Soc 2002, 124, 2337.
15. Burgoyne, N. J.; Jackson, R. M. Bioinformatics 2006, 12, 1335.
16. Vohra, S.; Chintapalli, S. V.; Illingworth, C. J.; Reeves, P. J.; Mulli-
neaux, P. M.; Clark, H. S.; Dean, M. K.; Upton, G. J.; Reynolds, C.
A. Biochem Soc Trans 2007, 35, 749.
17. Gouldson, P. R.; Dean, M. K.; Snell, C. R.; Bywater, R. P.; Gkou-
tos, G.; Reynolds, C. A. Protein Eng 2001, 14, 759.
18. Paci, E.; Lindorff-Larsen, K.; Dobson, C. M.; Karplus, M.; Vendrus-
colo, M. J Mol Biol 2005, 352, 495.
19. Vendruscolo, M.; Dokholyan, N. V.; Paci, E.; Karplus, M. Phys Rev
E 2002, 65, 061910.
20. Vendruscolo, M.; Paci, E.; Dobson, C. M.; Karplus, M. Nature
2001, 409, 641.
21. Dokholyan, N. V.; Li, L.; Ding, F.; Shakhnovich, E. I. Proc Natl
Acad Sci USA 2002, 99, 8637.
22. Sacquin-Mora, S.; Lavery, R. Biophys J 2006, 90, 2706.
DOI 10.1002/jcc
12 Illingworth et al. • Vol. 00, No. 00 • Journal of Computational Chemistry
23. Sacquin-Mora, S.; Laforet, E.; Lavery, R. Proteins 2007, 67, 350.
24. Rawlings, N. D.; Morton, F. R.; Barrett, A. J. Nucleic Acids Res
2006, 34, D270.
25. Hobohm, U.; Sander, C. Protein Sci 1994, 3, 522.
26. Hobohm, U.; Scharf, M.; Schneider, R.; Sander, C. Protein Sci 1992,
1, 409.
27. Sobolev, V.; Sorokine, A.; Prilusky, J.; Abola, E. E.; Edelman, M.
Bioinformatics 1999, 15, 327.
28. Berman, H. M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T. N.;
Weissig, H.; Shindyalov, I. N.; Bourne, P. E. Nucleic Acids Res
2000, 28, 235.
29. Hubbard, S. J.; Thornton, J. M. NACCESS Computer Program; Uni-
versity College London: London, 1993.
30. Porter, C. T.; Bartlett, G. J.; Thornton, J. M. Nucleic Acids Res
2004, 32, D129.
31. Choe, J. Y.; Nelson, S. W.; Arienti, K. L.; Axe, F. U.; Collins, T.
L.; Jones, T. K.; Kimmich, R. D.; Newman, M. J.; Norvell, K.;
Ripka, W. C.; Romano, S. J.; Short, K. M.; Slee, D. H.; Fromm, H.
J.; Honzatko, R. B. J Biol Chem 2003, 278, 51176.
32. Shibayama, N.; Miura, S.; Tame, J. R.; Yonetani, T.; Park, S. Y.
J Biol Chem 2002, 277, 38791.
33. Yang, Z.; Lanks, C. W.; Tong, L. Structure 2002, 10, 951.
34. Eriksson, M.; Uhlin, U.; Ramaswamy, S.; Ekberg, M.; Regnstrom,
K.; Sjoberg, B. M.; Eklund, H. Structure 1997, 5, 1077.
35. von Geldern, T. W.; Lai, C.; Gum, R. J.; Daly, M.; Sun, C.; Fry, E.
H.; bad-Zapatero, C. Bioorg Med Chem Lett 2006, 16, 1811.
36. Rudino-Pinera, E.; Rojas-Trejo, S. P.; Calcagno, M. L.; Horjales, E.
(in press); PDB code 2WU1.
37. Hindie, V.; Stroba, A.; Zhang, H.; Lopez-Garcia, L. A.; Idrissova,
L.; Zeuzem, S.; Hirschberg, D.; Schaeffer, F.; Jorgensen, T. J. D.;
Engel, M.; Alzari, P. M.; Biondi, R. M. Nat Chem Biol 2009, 5,
758.
38. Ptak, C. P.; Ahmed, A. H.; Oswald, R. E. Biochemistry 2009, 48,
8594.
39. Anderka, O.; Loenze, P.; Klabunde, T.; Dreyer, M. K.; Defossa, E.;
Wendt, K. U.; Schmoll, D. Biochemistry 2008, 47, 4683.
40. Hardy, J. A.; Lam, J.; Nguyen, J. T.; O’Brien, T.; Wells, J. A. Proc
Natl Acad Sci USA 2004, 101, 12461.
41. Vanderpool, D.; Johnson, T. O.; Ping, C.; Bergqvist, S.; Alton, G.;
Phonephaly, S.; Rui, E.; Luo, C.; Deng, Y. L.; Grant, S.; Quenzer,
T.; Margosiak, S.; Register, J.; Brown, E.; Ermolieff, J. Biochemis-
try 2009, 48, 9823.
42. Li, J.; Edwards, P.; Burghammer, M.; Villa, C.; Schertler, G. F. X.
J Mol Biol 2004, 343, 1409.
43. Bondensgaard, K.; Ankersen, M.; Thogersen, H.; Hansen, B. S.;
Wulff, B. S.; Bywater, R. P. J Med Chem 2004, 47, 888.
44. Brauner-Osborne, H.; Wellendorph, P.; Jensen, A. A. Curr Drug
Targets 2007, 8, 169.
Journal of Computational Chemistry
45. Schlyer, S.; Horuk, R. Drug Discovery Today 2006, 11, 481.
46. Higgs, C.; Reynolds, C. A. In Theoretical Biochemistry—Processes
and Properties of Biological Systems, Vol.9 ( Theoretical and Com-
putational Chemistry Series); Eriksson, L., Ed.; Elsevier: Amster-
dam, 2001; pp. 341–376.
47. Wells, J. A.; McClendon, C. L. Nature 2007, 450, 1001.
48. MacQueen, J. Some Methods for Classification and Analysis of Mul-
tivariate Observations; University of California Press: California,
1967; pp. 281–297.
49. Fairlie, D. P.; Tyndall, J. D.; Reid, R. C.; Wong, A. K.; Abbenante,
G.; Scanlon, M. J.; March, D. R.; Bergman, D. A.; Chai, C. L.;
Burkett, B. A. J Med Chem 2000, 43, 1271.
50. Lovell, S. C.; Word, J. M.; Richardson, J. S.; Richardson, D. C.
Proteins 2000, 40, 389.
51. Elcock, A. H.; McCammon, J. A. Proc Natl Acad Sci USA 2001,
98, 2990.
52. Morris, G. M.; Goodsell, D. S.; Huey, R.; Olson, A. J. J Comput-
Aided Mol Des 1996, 10, 293.
53. Henrick, K.; Thornton, J. M. Trends Biochem Sci 1998, 23, 358.
54. Bartlett, G. J.; Porter, C. T.; Borkakoti, N.; Thornton, J. M. J Mol
Biol 2002, 324, 105.
55. Gill, H. S.; Eisenberg, D. Biochemistry 2001, 40, 1903.
56. Huang, X. P.; Ellis, J. Mol Pharmacol 2007, 71, 759.
57. May, L. T.; Leach, K.; Sexton, P. M.; Christopoulos, A. Annu Rev
Pharmacol Toxicol 2007, 47, 1.
58. Malherbe, P.; Kratochwil, N.; Knoflach, F.; Zenner, M. T.; Kew, J. N.;
Kratzeisen, C.; Maerki, H. P.; Adam, G.; Mutel, V. J Biol Chem
2003, 278, 8340.
59. Klein-Seetharaman, J.; Hwa, J.; Cai, K.; Altenbach, C.; Hubbell, W.
L.; Khorana, H. G. Biochemistry 1999, 38, 7938.
60. Murakami, M.; Kouyama, T. Nature 2008, 453, 363.
61. Balem, F.; Yanamala, N.; Klein-Seetharaman, J. Photochem Photo-
biol 2009, 85, 471.
62. Simpson, L. M.; Taddese, B.; Wall, I. D.; Reynolds, C. A. Curr
Opin Pharmacol 2010, 10, 30.
63. Chintapalli, S. V.; Yew, B. K.; Upton, G. J. G.; Illingworth, C. J.
R.; Reeves, P. J.; Parkes, K. E.; Snell, C. R.; Reynolds, C. A.
J Comput Chem 2010, DOI: 10.1002/jcc/21562.
64. Bernier, V.; Lagace, M.; Lonergan, M.; Arthus, M. F.; Bichet, D.
G.; Bouvier, M. Mol Endocrinol 2004, 18, 2074.
65. Arkin, M. R.; Randal, M.; DeLano, W. L.; Hyde, J.; Luong, T. N.;
Oslob, J. D.; Raphael, D. R.; Taylor, L.; Wang, J.; McDowell, R. S.;
Wells, J. A.; Braisted, A. C. Proc Natl Acad Sci USA 2003, 100,
1603.
66. Yin, H.; Hamilton, A. D. Angew Chem Int Ed Engl 2005, 44, 4130.
67. Fry, D. C. Biopolymers 2006, 84, 535.
68. Schames, J. R.; Henchman, R. H.; Siegel, J. S.; Sotriffer, C. A.; Ni,
H. H.; McCammon, J. A. J Med Chem 2004, 47, 1879.
DOI 10.1002/jcc