STRUCTURE O FUNCTION O BIOINFORMATICS
The performance of ZDOCK and ZRANK
in rounds 6–11 of CAPRI
Kevin Wiehe,1Brian Pierce,1Wei Wei Tong,2Howook Hwang,1
Julian Mintseris,1and Zhiping Weng1,2*
1Bioinformatics Program, Boston University, Boston, Massachusetts 02215
2Department of Biomedical Engineering, Boston University, Boston, Massachusetts 02215
The CAPRI experiment gives the structural bioinformatics community an arena to
practically test protein–protein docking algorithms. It also acts as a catalyst of innova-
tion by presenting challenges that require new methods in protein–protein docking.
Our own experience in CAPRI has led us to such innovations1and has also served to
teach us how to tailor our docking approach to real-world docking problems. We par-
ticipated in the Rounds 1–5 of CAPRI,2,3with the goal of attaining at least one ‘‘ac-
ceptable’’ prediction for each target (predictions are classified by the CAPRI assessor as
high quality, medium quality, or acceptable quality).3We succeeded in doing so for 10
out of 16 targets with our fast Fourier transform (FFT)-based rigid-body protein dock-
ing algorithm, ZDOCK.4–6ZDOCK was trained and tested on two protein–protein
docking benchmarks7,8that have been made freely available to the docking community.
Recently we developed a new statistical potential called IFACE and implemented it in
the ZDOCK program.9In addition, we have developed a new reranking algorithm
called ZRANK to accurately score ZDOCK predictions.10In Rounds 6–11 of CAPRI,
we applied these two new improvements to our docking strategy, and here we describe
our performance. Our ability to rerank and refine predictions of various docking algo-
rithms was also tested in three rounds of the new scoring section of CAPRI and is
MATERIALS AND METHODS
Our general docking approach throughout all rounds of CAPRI has been simi-
lar.11,12However, each target in CAPRI entails a specific strategy based on the biologi-
cal data known prior to docking, as will be shown in the Results section. The general
approach involves first mining the literature for biological data about the interaction.
This information is applied in ZDOCK in two ways. During docking, ZDOCK can
downweight predictions with interfaces that lack residues of interest, a method we refer
to as ‘‘blocking.’’ After docking, predictions can also be ranked and filtered based on
distances between residues of interest. ZDOCK is typically run with 68 rotational sam-
pling and generates 54,000 predictions, a subset of which is then reranked using one of
our scoring algorithms. We can apply clustering to remove structural redundancy either
before or after the scoring step. A portion of the remaining clusters is visually inspected
to cull the list of predictions down to 10 submissions. We have recently developed a
new scoring function and a new refinement algorithm, the details of which follow.
Conflict of interest: The authors state that the docking programs used in this article are licensed to Accelrys Inc. through
Grant sponsor: NSF; Grant numbers: DBI-0133834, DBI-0116574.
*Correspondence to: Zhiping Weng, Bioinformatics Program, Boston University, Boston, MA 02215.
Received 31 May 2007; Revised 23 July 2007; Accepted 24 July 2007
Published online 5 September 2007 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/prot.21747
We present an evaluation of
our protein–protein docking
approach using the ZDOCK
and ZRANK algorithms, in
combination with structural
clustering and filtering, utiliz-
ing biological data in Rounds
6–11 of the CAPRI docking
experiment. We achieved at
least one prediction of ac-
ceptable accuracy for five of
six targets submitted. In addi-
tion, two targets resulted in
In the new scoring portion of
the CAPRI exercise, we were
able to attain at least one ac-
ceptable prediction for the
three targets submitted and
achieved three medium-accu-
racy predictions for Target 26.
Scoring was performed using
ZRANK, a new algorithm for
reranking initial-stage dock-
ing predictions using a weighted
energy function and no struc-
tural refinement. Here we
outline a practical and suc-
cessful docking strategy, given
limited prior biological knowl-
edge of the complex to be pre-
Proteins 2007; 69:719–725.
C 2007 Wiley-Liss, Inc.
Key words: protein
CAPRI; ZDOCK; ZRANK.
C 2007 WILEY-LISS, INC.
ZDOCK 2.4 and 3.0
We developed two new versions of ZDOCK that utilize
the IFACE statistical potential13to score protein docking
predictions, versus the atomic contact energies14that were
used in the previous version, ZDOCK 2.3. IFACEwas devel-
oped as a pair potential specifically geared toward detecting
transient protein–protein interfaces and was trained on a
nonredundant dataset of 150 transient complexes. It is
implemented in two different versions of ZDOCK.9In
ZDOCK 2.4, IFACE energies are substituted for atomiccon-
tact energy values, and the scoring is implemented in the
same method as in 2.3 by computing one additional FFT
with the average value of all atom pairs for each atom type.
In ZDOCK 3.0, atom pair energies are calculated explicitly
times slower than previous versions of ZDOCK because of
the computation time for the additional FFTs, but is more
accurate than ZDOCK 2.4. Testing on the Benchmark 2.0
showed considerable improvement in overall success rate
and number of hits, for both 2.4 and 3.0, over previous
ZDOCKversions. ZDOCK 2.4 preceded 3.0 in development
and we first applied it to targets 24 and 25 in CAPRI Round
ZRANK is a scoring method developed to quickly
rerank initial-stage docking predictions using an energy-
based potential.10In contrast to RDOCK, which was
previously developed by our lab, ZRANK does not
require structural minimization of the predictions; thus it
is able to process all 54,000 predictions from a ZDOCK
run using fine sampling. ZRANK was shown to lead to
significant improvements in success rates for two versions
of ZDOCK on Benchmark 2.0.10
Our overall performance in Rounds 6–11 is summar-
ized in Table I. We were able to attain at least one
acceptable classification for 5 out of the 6 targets. Two
targets, 25 and 27, resulted in predictions of ‘‘medium’’
accuracy and a third target, 26, missed this classification
by only 0.25-A˚interface RMSD. Additionally, we were able
to attain more than one acceptable prediction in Targets
25–27. Our improved success in the later rounds may be a
reflection of the upgrading of ourdocking and scoring algo-
rithms at that time. The details of the methods used for
each target and theirevaluation aredescribedas follows.
Target 20 was an unbound homology-model docking
case involving the prediction of the HemK/RF-1 complex.
HemK is a methyl transferase that methylates RF-1, a
prokaryotic ribosomal release factor, in a step that is crit-
ical in stop codon recognition.15From the literature, we
were able to determine a putative interface for this com-
plex. HemK recognizes the GGQ motif on a loop of RF-1
and methylates Gln 235.16S-Adenosyl-L-homocysteine of
HemK is the methyl donor17and the motif NPPY is
important for positioning the Gln substrate.18The
unbound structure for RF-1 was homology modeled with
the 3D-JIGSAW server,19using the X-ray structure of
RF-2 as a template, which had a sequence identity to RF-
1 of 40%. Our homology model lacked a loop conforma-
tion that exposed Gln 235. Therefore, we remodeled the
loop by first using PSI-BLAST20to identify a homolo-
gous loop in the NMR structure of a peptidyl-tRNA hy-
drolase domain (PDB code 1J26), and then selecting a
loop conformation from the 20 NMR models that maxi-
mally exposed Gln 235. We added this new loop confor-
mation into the RF-1 model and used Insight II21to
optimize side-chain positions. In addition, we removed
all but 81 residues (211–292) of RF-1 to isolate the fold
with the GGQ loop for docking. Blocking was used to
give preference to binding at the putative interface
hypothesized from the literature. Two independent runs
of ZDOCK 2.1 and ZDOCK 2.3 resulted in 108,000 pre-
dictions at 68 rotational sampling. A distance filter was
applied to this set of predictions. All predictions with a
distance greater than 10 A˚between the Cacarbon of the
methylation target Gln 235 of RF-1 and the Sdatom of
S-adenosyl-L-homocysteine bound to HemK were elimi-
nated. The full RF-1 model was added in place of the
RF-1 fragment used in docking, and any predicted
complexes with subsequent significant clashing were
discarded. Structural clustering was used to aid in the
manual inspection of structures, and the aforementioned
distance was emphasized when ranking the final predic-
tions. The top 10 predictions underwent a CHARMM22
energy minimization step to relieve any remaining inter-
Our top-ranked model made many of the appropriate
contacts for methylation, as seen in the crystal structure
of the complex.15However, because of the severe inac-
curacy in the loop conformation of our homology model
(13.13 A˚RMSD), we did not achieve any acceptable-
rated predictions. In addition, our docking of a fragment
of RF-1 (residues 211–292) instead of the entire RF-1
may have affected ZDOCK’s ability to identify near-
native predictions. The RF-1 fragment we used did not
include a region (residues 127–157) that is in contact
with HemK in the solved complex.
Target 21 (Sir1/Orc1)
For Target 21 we were asked to perform unbound–
unbound docking of the Orc1p subunit of the origin rec-
ognition complex of yeast with Sir1p, a silent informa-
K. Wiehe et al.
tion regulator protein. Orc1p interacts with Sir1p to
recruit Sir1p to a silencer.23Mutagenesis data from the
literature suggested that the interaction involved the SRD
region of Sir1p24and the H region of Orc1p.25Specifi-
cally, S491 and R493 in Sir1p24and K121 in Orc1p25
were shown to be crucial for binding. Regions distal to these
critical residues were blocked prior to docking, in order for
preferencetobe applied tothisputativeinterface.
For our best-rated prediction, we docked with ZDOCK
2.3 and filtered the results for predictions with close con-
tacts between the SRD and H regions, in general, and the
three aforementioned residues, in specific. The lowest
energy structure from ZDOCK 2.3 that met the distance
criteria was selected as our number one submission and
achieved an ‘‘acceptable’’ classification. This prediction is
shown in Figure 1(A).
Target 24 (Arf1/ArfBD homology model)
In Target 24, we predicted the complex formed
between the ARF1 binding domain (ArfBD) of ARGH-
GAP21, a Rho family GTPase-activating protein, and
Arf1, a small GTPase. This target required a homology
model to be built for ArfBD, using the PH domain of b-
spectrin (PDB: 1BTN), as a template with 30% sequence
identity. We used the five homology models that were
generously provided by Baker’s lab to the CAPRI com-
munity, with the ROBETTA server.26Because of a lack
of confidence in docking protein structures folded ab ini-
tio, we removed from all ArfBD models the 20-residue
C-terminal extension that is missing in the template
structure. Next, we used ZDOCK 2.4 to dock all five
ArfBD models to Arf1 in five independent docking runs.
The predictions were pooled and clustered to remove
structural redundancy. The binding sites on Ras super-
family G proteins such as Arf1 are well conserved at the
switch regions,27and we applied distance measures to
ensure contacts of ArfBD with the switches. Three loops
were also deemed important: the b3/b4 loop which is
involved in binding in structurally homologous PH do-
aHighest Fnat and lowest RMSD of any of the 10 submissions for each target.
bNumber of predictions with at least an acceptable CAPRI classification.
cVersion(s) of the ZDOCK program used for docking the target.
dLigand was homology modeled.
Docking predictions for Targets 21–27 of CAPRI Rounds 7–11. Orientations of
predicted (blue) and crystal structure (green) ligands after superposition of their
receptors (beige, only crystal structure receptor shown). (A) Target 21: Orc1p
(receptor) and Sir1p (ligand); (B) Target 24: Arf1 (receptor) and homology
model of ArfBD (ligand); (C) Target 25: Arf1 (receptor) and bound ArfBD
(ligand); (D) Target 26: TolB (receptor) and Pal (ligand); (E) Target 27: Hip2
(receptor) and Ubc9 (ligand). Figures were created using PyMOL.
ZDOCK/ZRANK Performance in CAPRI Rounds 6–11
main/Ras complexes,28,29the b1/b2 loop which is struc-
turally similar to b3/b4, and the b5/b6 loop which we
found to be highly conserved in multiple sequence align-
ments and was therefore potentially important to bind-
ing. Clusters were filtered and ranked according to the
distance between the switch regions of Arf1 and the
loops of interest in ArfBD.
Our best prediction achieved an acceptable classifica-
tion and came from a cluster in which the conserved
loop of ArfBD contacted switch 2 of Arf1 [Fig. 1(B)].
The 20-residue C-terminal region of ArfBD was improp-
erly folded in the homology model (16.40-A˚RMSD for
this region compared with 1.19-A˚RMSD in the remain-
ing regions) and was critical to binding as it contacted
switch 1 of Arf1 in the complex. Removing this region
prior to docking turned out to be a double-edged sword.
On the one hand, we did not dock the incorrect fold; on
the other hand, the contacts that it makes with Arf1 did
not contribute to ZDOCK scoring and therefore pre-
vented better predictions.
Target 25 (Arf1/bound ArfBD)
In Target 25, the structure of ArfBD was released from
the complex and we were asked to perform unbound–
bound docking of Arf1 and ArfBD. Because of missing
atoms in the bound structure of ArfBD it was possible to
eliminate some of the earlier identified regions of interest
from the binding site. Of the ArfBD loops of interest we
identified in Target 24, only the conserved b5/b6 loop
was totally resolved in the electron density map and
therefore we favored predictions with contacts to this
loop. The C-terminal extension that we removed prior to
docking in T24 was fully solved in the bound ligand
structure as a hinge loop and long helix.30We preferred
predictions involving this region in interactions, as the
temperature factors of this region in the bound structure
were very low, hinting at its participation in the binding
site. For our best predictions, missing side-chain atoms
were rebuilt using Insight II21and areas near missing
residues were blocked prior to docking. We then docked
with ZDOCK 2.4 and clustered to remove structural re-
dundancy. Clusters were selected for further inspection
based on distance for interactions involving the C-termi-
nal extension and the conserved loop of ArfBD and the
two switch regions of Arf1. Our best prediction achieved
a medium-accuracy classification with 1.51-A˚interface
RMSD and included over 81% of correct contacts [Fig.
1(C)]. Because of our high confidence in our binding
site model, we also had two predictions that achieved ac-
ceptable accuracy classifications (Table I).
Target 26 (TolB/Pal)
For Target 26, we predicted the complex of TolB and
Pal, given both unbound components. The TolB-Pal
complex is part of a supramolecular assembly of proteins
that is important for the structural integrity of the outer
membrane of Gram-negative bacteria.31TolB consists of
two domains, a b-propeller domain and a secondary do-
main. Extensive mutagenesis data was available, showing
that the face of the beta propeller domain of TolB distal
to the secondary domain was involved in binding, and
that, specifically, residues H246, A249, and T292 were
critical.32Additional literature implicated contacts at res-
idues 89–104 and 126–130 of Pal.33Given an abundance
of relevant biological data, we removed the secondary do-
main of TolB and applied a liberal blocking scheme to
place a heavy preference on predictions with contacts
between the b-propeller domain and the Pal residues
mentioned. We ran ZDOCK 3.0 and results were clus-
tered to remove structural redundancy. Clusters were
then ranked both by ZRANK and by the distance
between the three critical residues of TolB and residues
89–130 of Pal. Predictions were selected on the basis of
the best combination of these ranks. Two of our predic-
tions received acceptable CAPRI classifications, one
achieved an interface RMSD of 2.25 A˚and the other had
50% of correct contacts predicted [Fig. 1(D)].
Target 27 (Hip2/Ubc9)
For Target 27, we were asked to predict the complex of
Hip2 and Ubc9.34Ubc9 catalyzes a fusion between the
C-terminal residue of the signaling protein Sumo-1 and a
lysine of Hip2.35In addition to Hip2, Ubc9 also sumo-
lyates RanGap1. The structure of the Ubc9-RanGap1
complex has been solved, and we assumed sumoylation
of Hip2 would occur using the same catalytic groove on
Ubc9.36The X-ray structure of unbound Ubc9 has two
residues, Gln 126 and Asn 127, which partially block the
putative catalytic groove.37We chose to dock with the
Ubc9 structure from the Ubc9-RanGap1 complex in
which this groove is more accessible36and this structure
has an identical sequence as our target. In addition, we
noticed that Lys 14 of Hip2, which lays in the putative
catalytic groove and is fused to Sumo-1 via an isopeptide
bond between its Nf atom and the terminal carbon
atom, was not in an extended conformation as it is in
the Ubc9/RanGap1/Sumo-1 complex.38We rotated Lys
14 of Hip2 into a nonnative rotamer, in order for it to
be more accessible to Ubc9 during docking. Docking was
accomplished with ZDOCK 3.0 followed by reranking
with ZRANK. Cys 93 of Ubc9 is the catalytic residue
involved in the conjugation at the active site and must be
in close proximity to Lys 14 for the isopeptide bond to
form.39We clustered all predictions with less than 5 A˚dis-
tance between these two residues. Predictions were selected
after manual inspectionof the remaining clusters.
There are two possible interfaces proposed from the
crystal structure of the Ubc9-Hip2 complex and our pre-
dictions were evaluated against both. For predictions
K. Wiehe et al.
evaluated using the second interface (T27.2), we achieved
one ‘‘medium’’ classification and three acceptable classifi-
cations [Fig. 1(E)]. Our medium-classified prediction
had an interface RMSD of 1.86 A˚and our best acceptable
classified prediction correctly identified 49% of the true
We participated in the scoring rounds for Targets 25,
26, and 27. Our general strategy for these rounds was to
test ZRANK performance on the sets of structures to be
scored. In addition, for Targets 26 and 27 we utilized
RosettaDock40to produce refined structures for the top
10 models. The results from the CAPRI evaluation of our
submitted structures for these Targets are given in Table
II; we obtained acceptable predictions for Targets 25 and
27, and medium-level predictions for Target 26. The
Fnat, interface RMSD, and ligand RMSD of the top pre-
diction for each target (as evaluated by interface RMSD)
are also given in Table II. Details on the protocols and
results for each target are provided later.
For Target 25 (Arf1/Bound ArfBD), we were given 700
predicted complexes to score. We then used RosettaDock
to add hydrogens to all structures and scored them using
ZRANK. The top 10 models based on ZRANK score
were submitted; of these, we had one acceptable predic-
tion, which was ranked no. 7.
The next target (Target 26; TolB/Pal) included 1567
predictions to rescore. In this case, we added hydrogens
and scored all predictions as before, but also employed a
distance filter based on the predicted contacts discussed
in the docking portion of the Results section. In par-
ticular, we removed all predictions above a 15-A˚cutoff,
leaving 241 predictions. We found that several of the pre-
dictions that passed the filter were highly redundant
(based on observations from the residue distances and
ZRANK scores), and so we selected one representative
structure with the highest ZRANK score and removed
From the remaining predictions, we took the top 10
based on ZRANK score and refined each of these predic-
tions using RosettaDock (to produce 500 refined models
per prediction). The refined model with the best Rosetta-
Dock score from the set of 500 was used for submission.
Our submissions for this target included three medium
predictions and one acceptable prediction, for a total of
4 out of the 10 submitted models that were at least ac-
ceptable. The top submitted model had ligand RMSD of
3.04 A˚and interface RMSD of 1.11 A˚, which is an
improvement of 3.76 and 0.86 A˚in RMSD over those
models because of refinement (Table II; Marc Lensink,
personal communication). This indicates the success of
the RosettaDock refinement protocol in improving the
We employed a similar strategy for our scoring predic-
tions for Target 27.2 (Hip2/UBC9). For this target, we
utilized a residue distance measure based on the residues
described in the docking portion of the Results section,
and removed predictions above a 19-A˚cutoff, leaving
350 out of the original 1489 predictions. We selected the
top seven of the filtered predictions based on ZRANK
score, and also the top three from the total set of predic-
tions based on ZRANK score.
With these 10 predictions, we utilized a slightly differ-
ent protocol than for Target 26. We refined them using
RosettaDock as before (to produce 400 refined models
per prediction), and the refined models were selected by
ZRANK score (in contrast to Target 26 for which the
RosettaDock scores were used to select the models from
the refinement). This led to seven acceptable predictions
for the second evaluated interface (models ranked no. 1–
7), while the remaining three models submitted were
those from the unfiltered set. Five of the seven acceptable
models had improved structures after refinement, includ-
ing the best-submitted model. This model, which had a
ligand RMSD of 6.42 A˚and an interface RMSD of 2.39 A˚
was improved because of refinement by 3.28 and 0.46 A˚
ZR 1 Ros
ZR 1 Ros 1 ZR
aLigand and interface RMSD of top submitted prediction from native. For Targets 26 and 27.2 (for which refinement was performed), the amount of RMSD improve-
ment over the unrefined structure is given in parentheses.
bNumber of predictions with at least an acceptable CAPRI classification.
cScoring protocol employed; ZR, ZRANK; Ros, RosettaDock refinement.
ZDOCK/ZRANK Performance in CAPRI Rounds 6–11
for ligand and interface RMSDs, respectively (Table II;
Marc Lensink, personal communication). These RMSD
improvements are commensurate with those seen for Tar-
get 26, for which the RosettaDock score was used to
choose the refined model. This, as well as the overall
results for this target, indicate that using ZRANK in the
context of scoring sets of RosettaDock refined predictions
can lead to improved structures.
The goal of ZDOCK, as an initial-stage rigid-body dock-
ing algorithm, is to find at least one near-native structure
within a set of predictions. In CAPRI, this is extended to
finding at least one acceptable prediction within the 10
structures submitted for each target. Our performance in
Rounds 6–11 demonstrates that this goal was very nearly
reached, as we were able to find at least one acceptable in 5
out of 6 targets. Target 20 was the only target for which we
did not attain a good prediction and serves as an example
for the limitations of protein–protein docking when start-
ing with inaccurate homology models. The challenges of
homology docking can also be seen in the differing degrees
of accuracy in Targets 24 and 25. When the bound ArfBD
of Target 25 was substituted for its homology model in
Target 24, we were able to increase our correct contact per-
centage from 20 to 81% and our interface RMSD was cut
by more than half from 3.13 to 1.51 A˚. Overcoming the
challenges of homology docking will be an area of further
research and may require shAring techniques from the
similar problem of flexible docking.
Rounds 6–11 of CAPRI also served as another measure
to evaluate the evolution of our ZDOCK algorithm. As
seen in Table I, improvements in the ZDOCK algorithm
coincided with increases in the accuracy and a higher
number of good predictions we were able to achieve.
ZRANK was also added to our approach in the later
rounds and contributed to attaining successful predic-
tions for Targets 26 and 27.2.
Our scoring performance (Table II) indicates that
ZRANK can be successfully employed for rescoring many
predictions from a variety of sources. This is encouraging,
considering that ZRANKwas initially developed and tested
using predictions from one particular rigid-body docking
algorithm (ZDOCK), whereas the predictions in the scor-
ing round come from a variety of sources, and some of
these are not necessarily rigid-body docking. It was possi-
ble that refined or energy-minimized false-positive struc-
tures would appear more favorable than near-native rigid-
body predictions, particularly to an energy-based scoring
function such as ZRANK. As was seen in particular for
Target 27.2, the filtering step is quite useful for removing
false positives and complements the ZRANK scoring well.
Our results from the CAPRI scoring rounds for Targets
26 and 27.2 indicate that it is helpful to refine the pre-
dictions prior to submission. We were able to successfully
employ RosettaDock to generate the refined structural
models for these targets. In addition, for Target 27.2, we
rescored the refined models using ZRANK and found
this to achieve improved structures. This protocol is cur-
rently being explored in more detail to determine its
effectiveness and optimal usage. We are also considering
optimizing the ZRANK scoring function to evaluate and
compare refined docking models rather than just rigid-
body docking models.
Overall our performance in the Rounds 6–11 of CAPRI
demonstrates the progress of our approach in both pro-
tein–protein docking and scoring. We look forward to par-
ticipating in future rounds of CAPRI in order to test the
continued development of ZDOCK and ZRANK.
We thank the CAPRI organizers and evaluation team
for their tremendous efforts. In particular, we thank
Marc Lensink for giving us the scoring evaluation data
and answering many of our questions regarding the scor-
ing evaluation process. We also thank our system admin-
istrator, Mary Ellen Fitzpatrick, for her continued techni-
cal support. We are grateful for the resources of the sci-
entific computing facilities at Boston University.
1. Pierce B, Tong W, Weng Z. M-ZDOCK: a superior grid-based
approach for Cn symmetric multimer docking. Bioinformatics
2. Mendez R, Leplae R, De Maria L, Wodak SJ. Assessment of blind
predictions of protein–protein interactions: current status of dock-
ing methods. Proteins 2003;52:51–67.
3. Mendez R, Leplae R, Lensink MF, Wodak SJ. Assessment of CAPRI
predictions in Rounds 3–5 shows progress in docking procedures.
4. Chen R, Li L, Weng Z. ZDOCK: an initial-stage protein-docking
algorithm. Proteins 2003;52:80–87.
5. Chen R, Weng Z. A novel shape complementarity scoring function
for protein–protein docking. Proteins 2003;51:397–408.
6. Chen R, Weng Z. Docking unbound proteins using shape comple-
mentarity, desolvation, and electrostatics. Proteins 2002;47:281–294.
7. Chen R, Mintseris J, Janin J, Weng Z. A protein–protein docking
benchmark. Proteins 2003;52:88–91.
8. Mintseris J, Wiehe K, Pierce B, Anderson RJ, Chen R, Janin J, Weng
Z. Protein–protein docking benchmark 2.0: an update. Proteins
9. Mintseris J, Pierce B, Wiehe K, Anderson R, Chen R, Weng Z. Inte-
grating statistical pair potentials into protein complex prediction.
10. Pierce B, Weng Z. ZRANK: reranking protein docking predictions
with an optimized energy function. Proteins 2007;67:1078–1086.
11. Chen R, Tong W, Mintseris J, Li L, Weng Z. ZDOCK predictions
for the CAPRI challenge. Proteins 2003;52:68–73.
12. Wiehe K, Pierce B, Mintseris J, Tong WW, Anderson R, Chen R,
Weng Z. ZDOCK and RDOCK performance in CAPRI rounds 3, 4,
and 5. Proteins 2005;60:207–213.
13. Mintseris J, Weng Z. Optimizing protein representations with infor-
mation theory. Genome Inform 2004;15:160–169.
K. Wiehe et al.
14. Zhang C, Vasmatzis G, Cornette JL, DeLisi C. Determination of
atomic desolvation energies from the structures of crystallized pro-
teins. J Mol Biol 1997;267:707–726.
15. Graille M, Heurgue-Hamard V, Champ S, Mora L, Scrima N,
Ulryck N, van Tilbeurgh H, Buckingham RH. Molecular basis for
bacterial class I release factor methylation by PrmC. Mol Cell
16. Heurgue-Hamard V, Champ S, Engstrom A, Ehrenberg M, Bucking-
ham RH. The hemK gene in Escherichia coli encodes the N(5)-gluta-
mine methyltransferase that modifies peptide release factors. EMBO
17. Yang Z, Shipman L, Zhang M, Anton BP, Roberts RJ, Cheng X.
Structural characterization and comparative phylogenetic analysis of
Escherichia coli HemK, a protein (N5)-glutamine methyltransferase.
J Mol Biol 2004;340:695–706.
18. Schubert HL, Phillips JD, Hill CP. Structures along the catalytic
pathway of PrmC/HemK, an N5-glutamine AdoMet-dependent
methyltransferase. Biochemistry 2003;42:5592–5599.
19. Bates PA, Kelley LA, MacCallum RM, Sternberg MJ. Enhancement
of protein modeling by human intervention in applying the automatic
20. Altschul SF, Madden TL, Schaffer AA, Zhang J, Zhang Z, Miller W,
Lipman DJ. Gapped BLASTand PSI-BLAST: a new generation of pro-
tein database search programs. Nucleic Acids Res 1997;25:3389–3402.
21. Accelyrs. Insight II, version 2005. San Diego, CA: Accelyrs; 2005.
22. Brooks BR, Bruccoleri RE, Olafson BD, States DJ, Swaminathan S,
Karplus M. CHARMM: a program for macromolecular energy,
minimization, and dynamics calculations. J Comput Chem 1983;4:
23. Hou Z, Bernstein DA, Fox CA, Keck JL. Structural basis of the
Sir1-origin recognition complex interaction in transcriptional
silencing. Proc Natl Acad Sci USA 2005;102:8489–8494.
24. Gardner KA, Rine J, Fox CA. A region of the Sir1 protein dedicated
to recognition of a silencer and required for interaction with the
Orc1 protein in saccharomyces cerevisiae. Genetics 1999;151:31–44.
25. Zhang Z, Hayashi MK, Merkel O, Stillman B, Xu RM. Structure
and function of the BAH-containing domain of Orc1p in epigenetic
silencing. EMBO J 2002;21:4600–4611.
26. Kim DE, Chivian D, Baker D. Protein structure prediction and
analysis using the Robetta server. Nucleic Acids Res 2004;32:W526–
27. Vetter IR, Wittinghofer A. The guanine nucleotide-binding switch
in three dimensions. Science 2001;294:1299–1304.
28. Worthylake DK, Rossman KL, Sondek J. Crystal structure of the
DH/PH fragment of Dbs without bound GTPase. Structure 2004;
29. Rossman KL, Worthylake DK, Snyder JT, Siderovski DP, Campbell
SL, Sondek J. A crystallographic view of interactions between Dbs
and Cdc42: PH domain-assisted guanine nucleotide exchange.
EMBO J 2002;21:1315–1326.
30. Menetrey J, Perderiset M, Cicolari J, Dubois T, Elkhatib N, El Kha-
dali F, Franco M, Chavrier P, Houdusse A. Structural basis for
ARF1-mediated recruitment of ARHGAP21 to Golgi membranes.
EMBO J 2007;26:1953–1962.
31. Bonsor DA, Grishkovskaya I, Dodson EJ, Kleanthous C. Molecular
mimicry enables competitive recruitment by a natively disordered
protein. J Am Chem Soc 2007;129:4800–4807.
32. Ray MC, Germon P, Vianney A, Portalier R, Lazzaroni JC. Iden-
tification by genetic suppression of Escherichia coli TolB residues
important for TolB-Pal interaction. J Bacteriol 2000;182:821–
33. Clavel T, Germon P, Vianney A, Portalier R, Lazzaroni JC. TolB
protein of Escherichia coli K-12 interacts with the outer membrane
peptidoglycan-associated proteins Pal, Lpp and OmpA. Mol Micro-
34. Walker JR, Avvakumov GV, Xue S, Newman EM, Mackenzie F, Wei-
gelt J, Sundstrom M, Arrowsmith CH, Edwards AM, Bochkarev A,
Dhe-Paganon S. A novel and unexpected complex between the
SUMO-1-conjugating enzyme UBC9 and the ubiquitin-conjugating
enzyme E2-25 kDa, to be published.
35. Pichler A, Knipscheer P, Oberhofer E, van Dijk WJ, Korner R,
Olsen JV, Jentsch S, Melchior F, Sixma TK. SUMO modification of
the ubiquitin-conjugating enzyme E2-25K. Nat Struct Mol Biol
36. Bernier-Villamor V, Sampson DA, Matunis MJ, Lima CD. Structural
basis for E2-mediated SUMO conjugation revealed by a complex
between ubiquitin-conjugating enzyme Ubc9 and RanGAP1. Cell
37. Tong H, Hateboer G, Perrakis A, Bernards R, Sixma TK. Crystal
structure of murine/human Ubc9 provides insight into the variabili-
ty of the ubiquitin-conjugating system. J Biol Chem 1997;272:
38. Reverter D, Lima CD. Insights into E3 ligase activity revealed by a
SUMO-RanGAP1-Ubc9-Nup358 complex. Nature 2005;435:687–
39. Yunus AA, Lima CD. Lysine activation and functional analysis of
E2-mediated conjugation in the SUMO pathway. Nat Struct Mol
40. Gray JJ, Moughon S, Wang C, Schueler-Furman O, Kuhlman B,
Rohl CA, Baker D. Protein–protein docking with simultaneous
optimization of rigid-body displacement and side-chain conforma-
tions. J Mol Biol 2003;331:281–299.
ZDOCK/ZRANK Performance in CAPRI Rounds 6–11