modeling antibody-antigen complexes by information-driven ... · structure resource modeling...
TRANSCRIPT
Resource
Modeling Antibody-Antige
n Complexes byInformation-Driven DockingGraphical Abstract
Hyper variable
loops
BindingEpitopeHyper
variable loops
Res
idue
sel
ectio
nIn
form
atio
n-dr
iven
dock
ing
Highlights
d Accurate prediction of antibody-antigen structure is still a
challenge
d Hypervariable loops of antibodies can be used to bias the
modeling process
d Antigen epitope information, even loosely defined, is valuable
for docking
d HADDOCK shows better performance and generates higher-
accuracy models
Ambrosetti et al., 2020, Structure 28, 119–129January 7, 2020 ª 2019 Elsevier Ltd.https://doi.org/10.1016/j.str.2019.10.011
Authors
Francesco Ambrosetti,
Brian Jimenez-Garcıa,
Jorge Roel-Touris,
Alexandre M.J.J. Bonvin
In Brief
Ambrosetti et al. demonstrate that, for the
modeling of antibody-antigen
complexes, using information about
hypervariable loops and, when available,
a loose definition of the epitope, improves
the docking results. In this context,
HADDOCK, which directly uses this
information to guide docking, performs
better than the other three software
applications tested.
Structure
Resource
Modeling Antibody-Antigen Complexesby Information-Driven DockingFrancesco Ambrosetti,1,2 Brian Jimenez-Garcıa,2 Jorge Roel-Touris,2 and Alexandre M.J.J. Bonvin2,3,*1Department of Physics, Sapienza University, Piazzale Aldo Moro 5, 00184 Rome, Italy2Faculty of Science – Chemistry, Computational Structural Biology Group, Bijvoet Center for Biomolecular Research, Utrecht University,
Padualaan 8, 3584 CH Utrecht, The Netherlands3Lead Contact
*Correspondence: [email protected]
https://doi.org/10.1016/j.str.2019.10.011
SUMMARY
Antibodies are Y-shaped proteins essential for im-mune response. Their capability to recognize anti-gens with high specificity makes them excellenttherapeutic targets. Understanding the structuralbasis of antibody-antigen interactions is thereforecrucial for improving our ability to design efficientbiological drugs. Computational approaches suchas molecular docking are providing a valuable andfast alternative to experimental structural character-ization for these complexes. We investigate herehow information about complementarity-determiningregions and binding epitopes can be used to drive themodeling process, and present a comparative studyof four different docking software suites (ClusPro,LightDock, ZDOCK, and HADDOCK) providing spe-cific options for antibody-antigen modeling. Theirperformance on a dataset of 16 complexes is re-ported. HADDOCK, which includes information todrive the docking, is shown to perform best in termsof both success rate and quality of the generatedmodels in both the presence and absence of informa-tion about the epitope on the antigen.
INTRODUCTION
Antibodies are essential components of the immune response.
Their capability to recognize antigens with high specificity along
with their modular anatomy, which facilitates their design and en-
gineering, makes them excellent therapeutic targets (Kaplon and
Reichert, 2019). For the design of efficient biological drugs
(Morea et al., 2000) based on antibodies it is crucial to properly
understand the structural basis of antibody-antigen interactions.
Antibodies are Y-shaped proteins usually composed of two
pairs of identical polypeptide chains named light and heavy
chains. On the basis of their structural and sequence variability,
it is possible to identify variable and constant domains, more
specifically one variable and one constant domain for the light
chain and one variable and three or more constant domains for
the heavy one. The variable domain is composed of a very well
conserved framework containing six hypervariable loops (HV
loops), three from the light chain and three from the heavy chain.
These are part of the so-called complementarity-determining re-
gions (CDRs) (Wu, 2004). These regions, and in particular the HV
loops, are crucial for antigen recognition and specificity (Novotny
et al., 1983). The position of HV loops is known a priori and can
be inferred given only the antibody sequence (Al-Lazikani et al.,
1997). Despite the majority of the antibody-binding residues be-
ing included in the CDRs (Kunik et al., 2012; MacCallum et al.,
1996), it has been shown that residues that fall outside these
CDRs can also play a crucial role in the antigen recognition pro-
cess (Narciso et al., 2011). Various experimental methods have
been proposed to investigate the role of each residue in the an-
tigen recognition process such as, for example, hydrogen/
deuterium (H/D) exchange (Lim et al., 2017), mutagenesis anal-
ysis (Fontayne et al., 2007), or the classical structural methods
such as nuclear magnetic resonance (NMR) and X-ray crystal-
lography. These provide various levels of information about the
key antibody and antigen residues involved in the interaction,
but most of them require high cost and effort. Computational ap-
proaches able to predict antibody-antigen structures would offer
a valuable and fast alternative. For the antibody, the residues
involved in the binding, the so-called paratope residues, can
be predicted quite accurately through various computational ap-
proaches (Krawczyk et al., 2013; Kunik et al., 2012; Liberis et al.,
2018; Olimpieri et al., 2013). The identification or prediction of the
set of antigen residues that are recognized by the antibody is,
however, the most challenging part. Although several methods
have been reported (Ansari and Raghava, 2010; Jespersen
et al., 2017; Krawczyk et al., 2014; Kringelum et al., 2012;
Liang et al., 2010; Qi et al., 2014; Rubinstein et al., 2009; Sela-
Culang et al., 2015), epitope prediction remains an open issue
(Ponomarenko and Bourne, 2007). In this context, docking ap-
proaches could present a valuable alternative to the available
epitope prediction methods, provided near-native solutions
can be generated and recognized.
Different docking algorithms have been developed over the
years to predict the three-dimensional (3D) structure of biological
complexes starting from the free, unbound structures of the
components (Moreira et al., 2010). These approaches rely on
the generation of thousands of possible complex conformations
ormodels, which are successively ranked according to a specific
scoring function to identify or predict themodels that are close to
the real conformation (near-native solutions). Most protein-pro-
tein docking algorithms do not consider possible conformational
changes occurring upon binding (rigid-body docking). This is the
case for software such as ClusPro (Kozakov et al., 2017) and
ZDOCK (Chen and Weng, 2002) that are based on the fast
Structure 28, 119–129, January 7, 2020 ª 2019 Elsevier Ltd. 119
Figure 1. Summary of the Three Docking
Scenarios Used in This Work
The first case represents the situation in which no
previous information about the epitope is known, so
the docking is performed exploring the whole sur-
face of the antigen while for the antibody the HV
loops are provided. In the second scenario, the
antibody HV loops and a loose epitope definition
corresponding to the antigen residues within 9 A
from the antibody are used to drive the docking.
Finally, in the third scenario the real interfaces of
both the antigen and the antibody (defined at 4.5 A
distance) are used.
Fourier transform search algorithm (Katchalski-Katzir et al.,
1992). In most cases, however, protein flexibility is a crucial fac-
tor to be considered (Kotev et al., 2016). Approaches that allow
for flexibility of side chains and/or backbone have also been
developed, such as ATTRACT (De Vries et al., 2015), LightDock
(Jimenez-Garcıa et al., 2018), SwarmDock (Torchala et al., 2013),
SnugDock (Sircar and Gray, 2010), and HADDOCK (De Vries
et al., 2010). The former three do so by using normal modes,
the latter two by allowing some flexibility along side chains and
the backbone during a refinement stage.
Docking methods can be classified into two categories ac-
cording to the sampling strategy applied during the simulation
(excluding the template-based docking approaches when a ho-
mologous interface can be identified, a less relevant approach
for antibody-antigen complexes). The first class includes algo-
120 Structure 28, 119–129, January 7, 2020
rithms that do not take into account any
previous information about the putative
binding interfaces and perform an exhaus-
tive search of the interaction space, the
so-called ab initio docking methods
(Ritchie, 2008). Nonetheless, in many
cases experimental information can be
used during the scoring step to select
near-native models. The second class
consists of docking approaches where
sampling is driven by experimental data,
coming, for example, from mutagenesis,
mass spectrometry (MS) (H/D exchange
and/or crosslinking experiments), NMR
analysis, or predicted information about
the binding interface. These fall under the
so-called information-driven or integrative
modeling approaches (Rodrigues and
Bonvin, 2014). The performances of dock-
ing methods is continuously assessed by
the Critical Assessment of Predicted Inter-
actions (CAPRI) experiment (Janin et al.,
2003; Mendez et al., 2003), stimulating re-
searchers’ efforts toward the development
of more accurate docking and scoring
algorithms.
All docking approaches rely on the avail-
ability of 3D structures or models of the
components. Since antibodies have a
very conserved framework and the confor-
mation of five out of six loops can be quite reliably predicted
(Chothia et al., 1989), modeling methods specific for antibodies
are able to generate reasonably accurate structures (Leem et al.,
2016; Lepore et al., 2017; Weitzner et al., 2017; Yamashita et al.,
2014). The main problems in this field are related to predicting
the conformation of the H3 loop, which remains challenging
due to its high structural and length variability (Shirai et al.,
1996; Weitzner et al., 2015). Methods specifically tailored to pre-
dict its conformation have been developed (Choi and Deane,
2010; Messih et al., 2014) to improve the accuracy of the anti-
body modeling systems.
Despite great progress in predicting protein-protein com-
plexes, docking of antibody-antigen complexes is still chal-
lenging (Pedotti et al., 2011; Ponomarenko and Bourne, 2007;
Vajda, 2005) due to the inherent properties of their interfaces
Table 1. Classification of Docking Models in the Classes ***, **,
and * according to Fnat and Either Ligand RMSD or i-RMSD
Measures
Class Fnat Ligand RMSD (A) i-RMSD (A)
High (***) R0.5 %1.0 or %1.0
Medium (**) R0.3 %5.0 or %2.0
Acceptable (*) R0.1 %10.0 or %4.0
(Lo Conte et al., 1999; Sela-Culang et al., 2013). In this work,
we present an assessment of the performance of ClusPro,
HADDOCK, LightDock, and ZDOCK in predicting antibody-anti-
gen structures. All of these software packages allow the use in
various ways of a priori knowledge, e.g., the hypervariable loops,
into the modeling process to drive or limit the sampling and/or
score the docking models. For the antigen, we use different
levels of information to define the epitope. Using a set of 16 anti-
body-antigen complexes corresponding to the new entries from
the docking benchmark 5 (BM5) (Vreven et al., 2015) for which
unbound structures are available, we compare the performance
of the various docking software applications following several
strategies to include information about the binding regions.
RESULTS
Antibody-antigen docking was performed following three
different scenarios in order to mimic different levels of informa-
tion that can be obtained about the antibody and antigen
residues involved in the binding. The first scenario (HV-Surf) in-
cludes information about the antibody (HV loops) but no infor-
mation about the epitope; in the second scenario (HV-Epi 9), a
vague definition of the epitope is provided based on all residues
within 9 A from the antibody in the reference structure; finally,
the third scenario (Real interface) represents the ideal case
whereby both interfaces are well characterized (see Figure 1).
Further information about the scenarios can be found in STAR
Methods. This information was used differently in the various
docking software depending on their ability to deal with it. In
short (for details see STAR Methods), HADDOCK follows a
data-driven sampling strategy whereby the information is
encoded into ambiguous restraints to drive the docking;
LightDock uses the information both to limit the sampling to
specific regions and in scoring, while ClusPro and ZDOCK
include this information in their scoring functions in order to
select the correct models.
Docking Performance: Single StructureWe analyzed the performance of the four docking methods in
predicting antibody-antigen complexes in terms of success
rate calculated as the percentage of cases in which at least
one acceptable, medium, or high-quality model (see Table 1) is
found in the top N ranked solutions. The success rate for the
top 1, 5, 10, 20, 50, and 100 is shown in Figure 2 for each docking
method and scenario as described in STAR Methods. The first
panel refers to the HV-Surf scenario and the second to the HV-
Epi 9, and the third shows the success rate obtained using the
real interface information in the docking. The latter represents
the gold standard achievable by each docking approach, i.e.,
the best accuracy that can be reached for this dataset given a
perfect interface definition (but no specific contacts) and starting
from the unbound structures of the components.
In the absence of any kind of information about the epitope (HV
loops—surface; top row in Figure 2) the overall performance is
rather low for all methods. HADDOCK reaches a success rate
of 25% in the top 1, which is higher than ClusPro (6.2%), ZDOCK
(6.2%), and LightDock (0%). Note that considering the limited
size of the benchmark, a difference of 6.2% only corresponds
to one more successfully predicted complex. The differences
are smaller for the top 10 (the typical number ofmodels evaluated
in CAPRI), with HADDOCK and ZDOCK leading with 31.2%, fol-
lowed by ClusPro with 18.7%. Considering the top 100, in this
scenario LightDock outperforms the other methods with a suc-
cess rate of 68.7%. This is linked to the fact that LightDock is
based on a very effective sampling strategy, but the scoring func-
tion used is not accurate for this type of complex.
By providing a low-accuracy definition of the epitope region
(HV-Epi 9; middle row in Figure 2), the success rate increases
significantly. For example, HADDOCK and ClusPro are able to
predict correct models for 75% and 68.7%, respectively, of the
cases already in the top 5 (43.8% in the top 1 for both), followed
by ZDOCK (56.3%) and LightDock (37.4%).
The bottom row in Figure 2 (Real interface) shows the results
when both interfaces are perfectly characterized such that the
exact residues involved in the binding are used in the docking.
In this case, HADDOCK ranks acceptablemodels in the top 1 po-
sition for all 16 complexes of the dataset (100% success rate),
while ZDOCK, ClusPro, and LightDock reach success rates of
81.2%, 75%, and 31.2%, respectively.
ClusPro offers an automated masking of non-CDR regions for
antibodies. We had, however, manually defined the HV loops for
consistency between all software. We therefore evaluated
whether this affected the docking performance by repeating
the docking using the automated masking of non-CDR regions
for all docking scenarios. The automatic identification of the
CDR regions in the first scenario led to a higher success rate
(e.g., top 10 increased from 18.7% to 31.3%) (see Figure S1).
On the other hand, in the second scenario the automated mask-
ing of non-CDR regions, while overall leading to an improvement
of the quality of the generated models, resulted in a decrease of
the success rate (except for the top 10 of scenario 2 that
increased from 67.7% to 75%). For the third scenario, as ex-
pected, driving the docking using the automatic masking of the
non-CDR regions rather than the real interface led to a significant
decrease in both the success rate (e.g., top 10 drops from 100%
to 87.5%) and quality of the generated models.
Overall, Figure 2 shows that HADDOCK is performing best in
every scenario. Even in the cases where ClusPro, LightDock,
and ZDOCK are able to reach comparable results (e.g., top 50
HV-Epi 9 scenario), the quality of the generatedmodels is usually
not as good as those produced by HADDOCK. This can be
attributed to the different strategies of using information between
the various software, with HADDOCK directly using restraints
during the sampling/refinement stages, and not only for filtering
and/or scoring as is the case in the other software.
Figure 3 shows the performance for each complex. Com-
plexes are grouped into two classes, rigid and medium, accord-
ing to the classification provided in docking benchmark BM5,
which is based on the interface root-mean-square deviation
Structure 28, 119–129, January 7, 2020 121
06.2
06.26.2
012.56.2
0
2512.5
0
2512.5
0
2512.5
0
18.8
25
0
37.5
31.2
0
37.5
31.2
0
37.5
50
0
37.5
56.2
0
37.5
56.2
0
56.2
18.8
0
68.8
25
0
75
25
0
75
25
0
75
25
0
75
25
012.512.5
0
31.2
0
31.2
6.2
31.2
6.2
37.5
6.2
37.512.5
12.512.5
18.8
12.5
37.5
25
12.5
43.8
18.8
12.5
43.8
25
18.8
43.8
25
18.8
43.8
37.5
18.8
37.5
43.8
18.8
43.8
37.5
18.8
50
31.2
18.8
50
31.2
25
50
25
31.2
43.8
25
0 06.2
06.2
06.2
25
06.2
31.2
012.5
56.2
012.5
06.2
31.2
0
18.8
37.5
0
18.8
56.2
0
18.8
75
0
18.8
75
0
31.2
012.5
37.5
0
25
43.8
0
43.8
37.5
0
56.2
31.2
0
62.5
31.2
06.2
06.2
25
06.2
25
06.2
25
0
18.8
18.8
0
18.8
25
012.56.2
0
18.8
37.5
0
31.2
31.2
0
37.5
37.5
0
50
37.5
0
62.5
31.2
0
31.2
50
6.2
56.2
31.2
6.2
62.5
25
6.2
7512.5
6.2
81.212.5
12.5
7512.5
CLUSPRO HADDOCK LIGHTDOCK ZDOCK
HV − Surf
HV − Epi 9
Real interface
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
Succ
ess
rate
Quality: Acceptable Medium High
Figure 2. ClusPro, HADDOCK, LightDock, and ZDOCKSuccess Rate for the Three Scenarios Described in ThisWork as a Function of the Top
1, 5, 10, 20, 50, and 100 Ranked Models
The top row (HV-Surf) shows the success rate using the antibody HV loops and the entire antigen surface as restraints. The second row represents the success
rate achieved by driving the docking with the antibody HV loops and a loose epitope definition using a 9 A cutoff. The third row shows the docking results using the
true interfaces (defined at 4.5 A). The color coding indicates the quality of themodels according to CAPRI criteria (see STARMethods). See also Tables S1–S4 and
Figures S1–S4.
(i-RMSD) and the fraction of non-native residue contacts (Vreven
et al., 2015).
Overall, all methods achieve good results for the most rigid
structures (PDB: 3EOA, 3MXW, 4G6M, and 3RVW)when a vague
definition of the epitope is provided. HADDOCK is able to pro-
vide for 3MXW and 4G6M sub-angstrom high-quality models
as top-ranking models. Although HADDOCK and LightDock
allow flexibility of the molecules, the difference in terms of accu-
racy between them and ClusPro or ZDOCK for rigid and medium
complexes is not striking. This is probably due to the rather
limited conformational change that occurs upon binding of the
antibody to the antigen (average i-RMSD of 0.97 ± 0.50 A for
all entries in this dataset, with minimum and maximum values
of 0.39 A and 1.86 A, respectively) (Vreven et al., 2015). For
each complex and scenario, the i-RMSD and fraction of native
contacts (Fnat) values for the first acceptable and best models
for all of the software used in this study are reported in Tables
S1–S4.
HADDOCK Performance: Cluster BasedMany approaches perform a clustering after docking in order to
group together similar models and simplify the analysis. This has
been demonstrated to significantly improve the accuracy of the
scoring. The most widely used parameter to measure similarities
among different structures is the positional RMSD. The fraction
of common contacts (FCC) has been introduced as a fast and
122 Structure 28, 119–129, January 7, 2020
valuable alternative to classical RMSD-based methods (Ro-
drigues et al., 2012). FCC clustering is used by default in
HADDOCK to cluster the docking models using a default cutoff
of 0.6. This has been optimized on classical protein-protein sys-
tems. Taking into account the result of the cluster analysis, it is
possible to express the success rate as the percentage of cases
in which there is at least one acceptable, medium, or high-quality
model in the top four cluster members of the top 1, 2, 3, 4, and 5
clusters. In this work clusters were ranked by the average
HADDOCK score of their top 4 models (the default scoring
scheme of the HADDOCK server [De Vries et al., 2010]). The
cluster-based success rate of HADDOCK is shown in Figure 4
for the three different scenarios.
Comparing Figures 4 and 2, and in particular the success rate
for top 1 and top 5, one can clearly see how cluster-based
scoring increases the success rate of HADDOCK when informa-
tion about the epitope is provided, but reduces it when no infor-
mation on the antigen is available and the entire antigen surface
is used to drive the docking. This is due to the fact that the
sampling around the entire surface of the antigen leads to the
generation ofmany possible different conformations. This results
in many local minima of the energy landscape, which the
HADDOCK scoring function is not able to distinguish properly.
Also, a slightly lower number of models do fall into clusters in
this case as illustrated by the clustering coverage calculated
as the fraction of clustered models with average values of
HV − Surf HV − Epi 9 Real interface
Medium
Rigid
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
3L5W3EO13HI6
3V6Z3G6D
3EOA3MXW4G6M3RVW4G6J
3HMX4GXU4DN44FQI
2W9E2VXT
CLUSPROHV − Surf HV − Epi 9 Real interface
Medium
Rigid
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
3L5W3EO13HI6
3V6Z3G6D
3EOA3MXW4G6M3RVW4G6J
3HMX4GXU4DN44FQI
2W9E2VXT
HADDOCK
HV − Surf HV − Epi 9 Real interfaceM
ediumR
igid
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
3L5W3EO13HI6
3V6Z3G6D
3EOA3MXW4G6M3RVW4G6J
3HMX4GXU4DN44FQI
2W9E2VXT
LIGHTDOCKHV − Surf HV − Epi 9 Real interface
Medium
Rigid
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
T1 T5 T10 T20 T50 T100
3L5W3EO13HI6
3V6Z3G6D
3EOA3MXW4G6M3RVW4G6J
3HMX4GXU4DN44FQI
2W9E2VXT
ZDOCK
Quality: Low Acceptable Medium High
Figure 3. Performance of ClusPro, HADDOCK, and ZDOCK as a Function of the Number of Ranked Models per Complex and per Scenario
The dataset is split into rigid and medium complexes according to the docking benchmark 5 (BM5) definition. For each class, the complexes are sorted from
bottom to top in increasing interface Ca-RMSD between the reference and the superimposed unbound structures. The color coding indicates the quality of the
models according to CAPRI criteria (see STAR Methods).
0.83 ± 0.10, 0.93 ± 0.04, and 0.99 ± 0.003, respectively, for the
HV-Surf, HV-Epi 9, and Real interface scenarios. However,
even with a rather loose definition of the epitope (HV-Epi 9 sce-
nario), clustering leads to an improvement in scoring perfor-
mance from 43.8% to 56.3% for top 1. These results indicate
that different scoring strategies should ideally be followed de-
pending on the availability (or not) of epitope information.
Sampling PerformanceDocking involves two different steps, the sampling for the gener-
ation of thousands of models and the scoring to select the best
(near-native) models according to a specific scoring function.
Most software includes the information about the binding inter-
face at the scoring stage, but HADDOCK is the only system
that uses this information to drive the sampling (the information
is encoded into an additional energy term that generates forces
to drive the minimization and molecular dynamics steps). The ef-
fect of these different strategies can be noticed by calculating
the number of acceptable, medium, or high-quality models
generated out of the total number of produced models. This
number is summarized for each software suite in Figure 5 (see
also Tables S5–S8). One can clearly see how the driving strategy
implemented in HADDOCK leads to the generation of a much
higher number of goodmodels when information about the inter-
face is provided (HV-Epi 9 and Real interface scenarios). There
is, however, the danger that no single acceptable model might
be generated in the case of bad information. The other software,
ClusPro, LightDock, and ZDOCK, use the interface information
only at the scoring level (except for LightDock, which filters start-
ing swarms to sample around the provided binding site). These
have the advantage that they perform an exhaustive search of
the interaction space, but this comes at the cost of a small
Structure 28, 119–129, January 7, 2020 123
012.56.2
0
18.8
0
18.8
0
25
0
25
12.5
25
18.8
12.5
37.5
25
12.5
37.5
37.5
18.8
37.5
37.5
18.8
37.5
37.5
18.8
43.8
37.5
18.8
43.8
37.5
18.8
43.8
37.5
18.8
43.8
37.5
18.8
43.8
37.5
HADDOCK
HV − Surf
HV − Epi 9
Real interface
T1 T2 T3 T4 T5
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
Succ
ess
rate
Quality: Acceptable Medium High
Figure 4. HADDOCKCluster-Based Success Rate for the Three Docking Scenarios as a Function of the Top 1, 2, 3, 4, and 5 Ranked Clusters
The color coding indicates the quality of the models according to CAPRI criteria (see STAR Methods).
number of near-native models generated. In this case, the
scoring becomes crucial in identifying the acceptable models.
H3 Loop Modeling PerformanceAs already mentioned, the H3 loop of antibodies is the most
important loop involved in antigen recognition. Its accurate
modeling is still a challenge due to its high structural and
sequence variability. Of the four docking software applications
used in this work, two allow for conformational changes during
the docking, namely HADDOCK and LightDock. We analyzed
their capability of inducing the appropriate conformational
changes of the loop upon binding with the antigen. For this,
we superimposed the antibody framework residues of the
bound and unbound structure and calculated the RMSD of H3
(H3unbound). We then repeated the same procedure for each
docking model compared with the native complex (H3model).
For both HADDOCK and LightDock, models produced from
the different scenarios were merged and split into correct
(i-RMSD % 4A) and wrong models (i-RMSD > 4A). Figure 6
shows the distribution of H3model versus H3unbound for correct
and wrong models. Values below the diagonal correspond to
an improvement of the conformation of the H3 loop. Overall,
for HADDOCK (Figure 6A) the flexible refinement tends to in-
crease the RMSD of the H3 loop for complexes that show a
low H3 conformational change upon binding but, in contrast,
for complexes undergoing larger conformational changes of
H3unbound, the refinement leads to improvement in the H3 confor-
mation, especially in the scenarios where information about the
124 Structure 28, 119–129, January 7, 2020
epitope is provided (HV-Epi 9 and Real interface), with a
maximum observed improvement of 1.25 A. In the case of
LightDock (Figure 6B), the final selected H3 loop conformation
from normal modes remains very close to the unbound form,
with no remarkable changes in terms of RMSD.
To further investigate the impact of the HADDOCK flexible
refinement stage on the H3 loop conformation, we analyzed
the Fnat that H3 makes at the rigid-body docking stage (H3it0)
and after flexible refinement (H3water). Figure 7 plots H3water
versus H3it0 for the three different scenarios discussed in this
work, taking into account the quality of the models. In this
case, all points above the diagonal correspond to an improve-
ment in Fnat after flexible refinement. Figure 7 clearly shows
that for most cases the flexible refinement improves the number
of native contacts made by H3, with a maximum improvement
observed of 0.72. This is more evident for the last two scenarios
(HV-Epi 9 and Real interface), indicating that an accurate selec-
tion of the native interface is crucial in improving the H3 confor-
mation during the simulation. A deeper analysis of the impact of
HADDOCK flexible refinement on the quality of the models is re-
ported in terms of distribution of i-RMSD and Fnat differences in
Figures S5 and S6.
DISCUSSION
This work aimed at comparing four well-established protein-pro-
tein docking software suites (ClusPro, HADDOCK, LightDock,
and ZDOCK) in their accuracy for predicting antibody-antigen
Rigid Medium
HV − Surf
HV − Epi 9
Real interface
3EOA
3MXW
4G6M
3RVW
4G6J
3HMX
4GXU
4DN4
4FQI
2W9E
2VXT
3L5W
3EO1
3HI6
3V6Z
3G6D
25
50
25
50
25
50
Perc
enta
ge
CLUSPRORigid Medium
HV − Surf
HV − Epi 9
Real interface
3EOA
3MXW
4G6M
3RVW
4G6J
3HMX
4GXU
4DN4
4FQI
2W9E
2VXT
3L5W
3EO1
3HI6
3V6Z
3G6D
25
50
75
100
25
50
75
100
25
50
75
100
Perc
enta
ge
HADDOCK
Rigid MediumH
V − SurfH
V − Epi 9R
eal interface
3EOA
3MXW
4G6M
3RVW
4G6J
3HMX
4GXU
4DN4
4FQI
2W9E
2VXT
3L5W
3EO1
3HI6
3V6Z
3G6D
10
20
10
20
10
20
Perc
enta
ge
LIGHTDOCKRigid Medium
HV − Surf
HV − Epi 9
Real interface
3EOA
3MXW
4G6M
3RVW
4G6J
3HMX
4GXU
4DN4
4FQI
2W9E
2VXT
3L5W
3EO1
3HI6
3V6Z
3G6D
10
20
10
20
10
20
Perc
enta
geZDOCK
Quality: Acceptable Medium High
Figure 5. Percentages of Acceptable, Medium and High-Quality Models Generated by Each Software per Complex and per Scenario
Complexes are split into rigid and medium categories according to the Docking Benchmark5 definition. Note that the y axis scales are different for each docking
method for better readability. The color coding indicates the quality of the models according to CAPRI criteria (see STAR Methods). See also Tables S5–S8.
complexes assuming different amounts of available experi-
mental information. Alternative docking methods do exist, such
as SwarmDock and SnugDock, but these were not considered
in this work for various reasons. The SwarmDock server does
not allow the user to automate the runs for high-throughput
docking, requiring manual selection of each residue in the HV
loops and epitope. Furthermore, the option in SwarmDock to
specify specific interface residues is still experimental and not
yet properly tested (Dr. Paul Bates, personal communication).
For these reasons, we did not include it in this work. With regard
to SnugDock, it requires the two molecules to be preoriented
before optimizing them, which prevents making a fair compari-
son with the docking scenarios considered in this work.
While for the antibody a proxy of the binding interface can be
extracted from the sequence and in particular the HV loops (Al-
Lazikani et al., 1997; Novotny et al., 1983; Sela-Culang et al.,
2013), predicting the epitope on the antigen is amore challenging
problem. Despite many efforts to develop reliable methods to
predict the antibody-specific epitopes, most approaches pub-
lished to date are still very limited in their accuracy. Driving a
docking process or filtering docking poses using predicted epi-
topes can thus be detrimental to the accuracy of the complex
structure prediction. There are, however, experimental methods
such asH/D exchange detected byMS that can providemore ac-
curate information. In this work, we focus on defining the best
way of including available epitope information to guide the
modeling process. To this end, we investigated three different
scenarios that mimic different levels of knowledge about the
epitope: In the first scenario, no information is available and the
sampling/scoring must involve the entire surface of the antigen;
in the second case, a loose definition of the epitope is assumed
(a larger surface than the real interface); finally, the third scenario
Structure 28, 119–129, January 7, 2020 125
A
B
Figure 6. H3 Loop RMSD (A) from the Bound Conformation for the Docked Models (H3model) versus the Starting Unbound Conformation
(H3unbound)
HADDOCK models (A) and LightDock models (B). Correct and wrong models were defined according to their i-RMSD from the reference structure using a 4 A
cutoff.
represents the ideal case whereby both antibody and antigen in-
terfaces are well characterized. HADDOCK, which has been
developed to make use of available information, is performing
best in terms of both success rate and quality of the generated
models in all of the described scenarios. Analysis of the
HADDOCKscoring performance per scenario indicates that clus-
tering is beneficial provided some reasonably accurate informa-
tion is available for the epitope. If this is not the case, a single
structure-based scoring approach might perform better, but still
with rather low success rate (31.2% for both top 5 and top 10).
In this work, the unbound forms of antibodies were used to
perform the docking, but often in reality only the antibody
sequence is known; therefore, we assessed HADDOCK accu-
racy using antibody models generated with the PIGSpro web-
server (https://cassandra.med.uniroma1.it/pigspro/) (Lepore
126 Structure 28, 119–129, January 7, 2020
et al., 2017). There are many other tools with which to model an-
tibodies, but it is outside the scope of this work to compare their
impact on the docking results. Results and method are reported
in Figure S2. The use of antibody models results in a significant
decrease of the success rate in all scenarios. For example, the
top 10 success rate using antibody models drops from 31.2%,
75%, and 100% to 18.7%, 68.7%, and 75%, respectively, for
the first, second, and third scenarios. This will of course depend
on the antibody modeling strategy, but it is outside the scope of
this work to compare different modeling software/servers.
The sampling used in this work for HADDOCK differs from the
default settings. While for the first scenario this is the recommen-
ded setting for cases with limited or no information on binding
sites (this was also the sampling used in the BM5 benchmarking
[Vreven et al., 2015]), the sampling in the two other scenarios
HV − Surf HV − Epi 9 Real interface
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
H3it0
H3 w
ater
Correct models Wrong models
Figure 7. Comparison of the Fraction of Native Contacts (Fnat) Made by the H3 Loop after the Rigid-Body Docking Stage (H3it0) (horizontal
axis) and after Flexible/Water Refinement (H3water) (vertical axix) of the HADDOCK Runs
Values above the diagonal correspond to an improvement in Fnat. See also Figures S5 and S6. Correct and wrongmodels were defined according to their i-RMSD
from the reference structure using a 4A cutoff.
(5,000/400/400) differs from the default settings (1,000/200/200).
For comparison, we reran the docking for these two scenarios
using default sampling settings. The results, presented in Fig-
ure S3, show that even with default settings, given a reasonable
definition of the epitope (HV-Epi 9 scenario), good results are ob-
tained with default sampling, with 87.5% top 10 success rate
versus 75% for the increased sampling. For the third scenario,
the reduced sampling does not seem to have an impact on the
success rate but rather slightly affects the quality of the gener-
ated models.
Another difference to note for HADDOCK between scenarios
2 and 3 is the definition of the interface information on the
epitope side: while for the true interface (scenario 3) the inter-
face residues were treated as ‘‘active’’ in the definition of inter-
action restraints, they were defined as passive in scenario 2.
To verify whether this might have affected the performance,
we repeated scenario 3 defining the epitope residues as pas-
sive. The results, presented in Figure S4, show that the passive
definition of the antigen interface residues leads to a decrease
of the success rate (e.g., from 100% to 87.6% for the top 10),
while the quality of the generated models does not change
significantly.
We further analyzed the capability of HADDOCK and
LightDock, the only two software suites that allow flexibility of
the components among the tested ones, in improving the H3
loop conformation. In fact, it has been largely demonstrated
that the H3 loop is crucial for antigen recognition, but its
modeling represents one of the biggest pitfalls in the antibody
modeling field. Only when the H3 loop underwent a large confor-
mational change upon binding did HADDOCK succeed in
improving its conformation (measured in terms of RMSDs) to-
ward the bound form, while the LightDock models did not
show any significant conformational changes. While the induced
conformational changes are rather limited in terms of RMSDs,
the flexible refinement stages of HADDOCK do lead to a signifi-
cant increase in the number of native contacts made by H3, and
this effect is more evident when information about the epitope is
provided to the system. This is relevant, since it will allow a better
identification/prediction of key interactions.
One of the main benefits of this work is to offer researchers a
clear overview about the state of the art of antibody-antigen
structure prediction (for the software considered) and the various
strategies that can be followed depending on the available infor-
mation. Provided that a vague definition of the epitope can be
obtained, reasonably accurate models can be obtained, with
HADDOCK performing best among the four software applica-
tions compared. Finally, our analysis also indicates that there
remain multiple opportunities for improvements, especially in
modeling conformational changes, with the H3 loop as a partic-
ular challenge, but also in scoring, considering that all software
achieved fairly good performance in the top 100, although
this significantly dropped in most cases when only the top 10
or lower were considered.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d LEAD CONTACT AND MATERIALS AVAILABILITY
d METHOD DETAILS
B Dataset
B Docking Scenarios
B Docking Settings
B HADDOCK Clustering Parameters
B Evaluation Criteria
d DATA AND CODE AVAILABILITY
Structure 28, 119–129, January 7, 2020 127
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.str.
2019.10.011.
ACKNOWLEDGMENTS
We wish to thank Pier Paolo Olimpieri (Department of Physics, Sapienza Uni-
versity) and Panagiotis Koukos (Bijvoet Center for Biomolecular Research,
Utrecht University) for useful comments and discussions. This work has
been carried out with the financial support of the European Union Horizon
2020 BioExcel (project #675728 and #823830) and EOSC-hub (project
#777536) projects and the Dutch Foundation for Scientific Research (NWO)
(TOP-PUNT grant 718.015.001).
The FP7 WeNMR (project #261572), H2020 West-Life (project #675858),
and the EOSC-hub (project #777536) European e-Infrastructure projects are
acknowledged for the use of the HADDOCK web portal, which makes use of
the EGI infrastructure with the dedicated support of CESNET-MetaCloud,
INFN-PADOVA, NCG-INGRID-PT, TW-NCHC, SURFsara, and NIKHEF, and
the additional support of the national GRID Initiatives of Belgium, France, Italy,
Germany, the Netherlands, Poland, Portugal, Spain, UK, Taiwan, and the US
Open Science Grid.
AUTHOR CONTRIBUTIONS
F.A., B.J.G., and J.R.T. performed the computational analysis. A.M.J.J.B. and
F.A. conceived and designed and interpreted the data. A.M.J.J.B. and F.A.
wrote the paper with contributions from the other authors.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: March 22, 2019
Revised: July 3, 2019
Accepted: October 18, 2019
Published: November 11, 2019
REFERENCES
Al-Lazikani, B., Les, A.M., and Chothia, C. (1997). Standard conformations for
the canonical structures of immunoglobulins. J. Mol. Biol. 273, 927–948.
Ansari, H.R., and Raghava, G.P. (2010). Identification of conformational B-cell
Epitopes in an antigen from its primary sequence. Immunome Res. 6, 6.
Brenke, R., Hall, D.R., Chuang, G.Y., Comeau, S.R., Bohnuud, T., Beglov, D.,
Schueler-Furman, O., Vajda, S., and Kozakov, D. (2012). Application of asym-
metric statistical potentials to antibody-protein docking. Bioinformatics 28,
2608–2614.
Chen, R., and Weng, Z. (2002). Docking unbound proteins using shape
complementarity, desolvation, and electrostatics. Proteins 47, 281–294.
Choi, Y., and Deane, C.M. (2010). FREAD revisited: accurate loop structure
prediction using a database search algorithm. Proteins 78, 1431–1440.
Chothia, C., Lesk, A.M., Tramontano, A., Levitt, M., Smith-Gill, S.J., Air, G.,
Sheriff, S., Padlan, E.A., Davies, D., Tulip, W.R., et al. (1989). Conformations
of immunoglobulin hypervariable regions. Nature 342, 877–883.
Lo Conte, L., Chothia, C., and Janin, J. (1999). The atomic structure of protein-
protein recognition sites. J. Mol. Biol. 285, 2177–2198.
Dominguez, C., Boelens, R., and Bonvin, A.M.J.J. (2003). HADDOCK: a pro-
tein-protein docking approach based on biochemical or biophysical informa-
tion. J. Am. Chem. Soc. 125, 1731–1737.
Fontayne, A., De Maeyer, B., De Maeyer, M., Yamashita, M., Matsushita, T.,
and Deckmyn, H. (2007). Paratope and epitope mapping of the antithrombotic
antibody 6B4 in complex with platelet glycoprotein Iba. J. Biol. Chem. 282,
23517–23524.
Hubbard, S.J., and Thornton, J.M. (1993). NACCESS, Computer Program
(University College, London). http://wolf.bms.umist.ac.uk/naccess/.
128 Structure 28, 119–129, January 7, 2020
Janin, J., Henrick, K., Moult, J., Eyck, L.T., Sternberg, M.J.E., Vajda, S.,
Vakser, I., and Wodak, S.J. (2003). CAPRI: a critical assessment of
PRedicted interactions. Proteins 52, 2–9.
Jespersen, M.C., Peters, B., Nielsen, M., and Marcatili, P. (2017). BepiPred-
2.0: improving sequence-based B-cell epitope prediction using conforma-
tional epitopes. Nucleic Acids Res. 45, W24–W29.
Jimenez-Garcıa, B., Roel-Touris, J., Romero-Durana, M., Vidal, M., Jimenez-
Gonzalez, D., and Fernandez-Recio, J. (2018). LightDock: a new multi-scale
approach to protein-protein docking. Bioinformatics 34, 49–55.
Jimenez-Garcıa, B., Vidal, M., and Roel-Touris, J. (2019). Brianjimenez/
LightDock: Release 0.5.6 (Zenodo). https://doi.org/10.5281/zenodo.2537146.
Kaplon, H., and Reichert, J.M. (2019). Antibodies to watch in 2019. MAbs 11,
219–238.
Katchalski-Katzir, E., Shariv, I., Eisenstein, M., Friesem, A.A., Aflalo, C., and
Vakser, I.A. (1992). Molecular surface recognition: determination of geometric
fit between proteins and their ligands by correlation techniques. Proc. Natl.
Acad. Sci. U S A 89, 2195–2199.
Kotev, M., Soliva, R., and Orozco, M. (2016). Challenges of docking in large,
flexible and promiscuous binding sites. Bioorg. Med. Chem. 24, 4961–4969.
Kozakov, D., Hall, D.R., Xia, B., Porter, K.A., Padhorny, D., Yueh, C., Beglov,
D., and Vajda, S. (2017). The ClusPro web server for protein-protein docking.
Nat. Protoc. 12, 255–278.
Krawczyk, K., Baker, T., Shi, J., and Deane, C.M. (2013). Antibody i-Patch pre-
diction of the antibody binding site improves rigid local antibody-antigen dock-
ing. Protein Eng. Des. Sel. 26, 621–629.
Krawczyk, K., Liu, X., Baker, T., Shi, J., and Deane, C.M. (2014). Improving
B-cell epitope prediction and its application to global antibody-antigen dock-
ing. Bioinformatics 30, 2288–2294.
Kringelum, J.V., Lundegaard, C., Lund, O., and Nielsen, M. (2012). Reliable B
cell epitope predictions: impacts of method development and improved
benchmarking. PLoS Comput. Biol. 8, e1002829.
Kunik, V., Ashkenazi, S., and Ofran, Y. (2012). Paratome: an online tool for sys-
tematic identification of antigen-binding regions in antibodies based on
sequence or structure. Nucleic Acids Res. 40, W521–W524.
Leem, J., Dunbar, J., Georges, G., Shi, J., and Deane, C.M. (2016).
ABodyBuilder: automated antibody structure prediction with data-driven ac-
curacy estimation. MAbs 8, 1259–1268.
Lepore, R., Olimpieri, P.P., Messih, M.A., and Tramontano, A. (2017). PIGSPro:
prediction of immunoGlobulin structures v2. Nucleic Acids Res. 45, W17–W23.
Liang, S., Zheng, D., Standley, D.M., Yao, B., Zacharias, M., and Zhang, C.
(2010). EPSVR and EPMeta: prediction of antigenic epitopes using support
vector regression and multiple server results. BMC Bioinformatics 11, 381.
Liberis, E., Velickovic, P., Sormanni, P., Vendruscolo, M., and Lio, P. (2018).
Parapred: antibody paratope prediction using convolutional and recurrent
neural networks. Bioinformatics 34, 2944–2950.
Lim, X.X., Chandramohan, A., Lim, X.Y.E., Crowe, J.E., Lok, S.M., and Anand,
G.S. (2017). Epitope and paratope mapping reveals temperature-dependent
alterations in the dengue-antibody interface. Structure 25, 1391–1402.e3.
MacCallum, R.M., Martin, A.C.R., and Thornton, J.M. (1996). Antibody-antigen
interactions: contact analysis and binding site topography. J. Mol. Biol. 262,
732–745.
McLachlan, A.D. (1982). Rapid comparison of protein structures. Acta
Crystallogr. Sect. A 38, 871–873.
Mendez, R., Leplae, R., De Maria, L., and Wodak, S.J. (2003). Assessment of
blind predictions of protein-protein interactions: current status of docking
methods. Proteins 52, 51–67.
Messih, M.A., Lepore, R., Marcatili, P., and Tramontano, A. (2014). Improving
the accuracy of the structure prediction of the third hypervariable loop of the
heavy chains of antibodies. Bioinformatics 30, 2733–2740.
Meyer, P.A., Socias, S., Key, J., Ransey, E., Tjon, E.C., Buschiazzo, A., Lei, M.,
Botka, C., Withrow, J., Neau, D., et al. (2016). Data publication with the struc-
tural biology data grid supports live analysis. Nat. Commun. 7, 10882.
Morea, V., Lesk, A.M., and Tramontano, A. (2000). Antibodymodeling: implica-
tions for engineering and design. Methods 20, 267–279.
Moreira, I.S., Fernandes, P.A., and Ramos, M.J. (2010). Protein-protein dock-
ing dealing with the unknown. J. Comput. Chem. 31, 317–342.
Morin, A., Eisenbraun, B., Key, J., Sanschagrin, P.C., Timony, M.A., Ottaviano,
M., and Sliz, P. (2013). Collaboration gets the most out of software. Elife 2,
e01456.
Narciso, J.E.T., Uy, I.D.C., Cabang, A.B., Chavez, J.F.C., Pablo, J.L.B.,
Padilla-Concepcion, G.P., and Padlan, E.A. (2011). Analysis of the antibody
structure based on high-resolution crystallographic studies. Nat. Biotechnol.
28, 435–447.
Novotny, J., Bruccoleri, R., Newell, J., Murphy, D., Haber, E., and Karplus, M.
(1983). Molecular anatomy of the antibody binding site. J. Biol. Chem. 258,
14433–14437.
Olimpieri, P.P., Chailyan, A., Tramontano, A., and Marcatili, P. (2013).
Prediction of site-specific interactions in antibody-antigen complexes: the
proABC method and server. Bioinformatics 29, 2285–2291.
Pedotti, M., Simonelli, L., Livoti, E., and Varani, L. (2011). Computational dock-
ing of antibody-antigen complexes, opportunities and pitfalls illustrated by
influenza hemagglutinin. Int. J. Mol. Sci. 12, 226–251.
Pierce, B.G., Hourai, Y., and Weng, Z. (2011). Accelerating protein docking in
ZDOCK using an advanced 3D convolution library. PLoS One 6, e24657.
Ponomarenko, J.V., and Bourne, P.E. (2007). Antibody-protein interactions:
benchmark datasets and prediction tools evaluation. BMC Struct. Biol. 7, 64.
Qi, T., Qiu, T., Zhang, Q., Tang, K., Fan, Y., Qiu, J.,Wu, D., Zhang,W., Chen, Y.,
Gao, J., et al. (2014). SEPPA 2.0-More refined server to predict spatial epitope
considering species of immune host and subcellular localization of protein an-
tigen. Nucleic Acids Res. 42, W59–W63.
Ritchie, D. (2008). Recent progress and future directions in protein-protein
docking. Curr. Protein Pept. Sci. 9, 1–15.
Rodrigues, J.P.G.L.M., and Bonvin, A.M.J.J. (2014). Integrative computational
modeling of protein interactions. FEBS J. 281, 1988–2003.
Rodrigues, J.P.G.L.M., Trellet, M., Schmitz, C., Kastritis, P., Karaca, E.,
Melquiond, A.S.J., and Bonvin, A.M.J.J. (2012). Clustering biomolecular com-
plexes by residue contacts similarity. Proteins 80, 1810–1817.
Rubinstein, N.D., Mayrose, I., Martz, E., and Pupko, T. (2009). Epitopia: a web-
server for predicting B-cell epitopes. BMC Bioinformatics 10, 287.
Sela-Culang, I., Kunik, V., and Ofran, Y. (2013). The structural basis of anti-
body-antigen recognition. Front. Immunol. 4, 302.
Sela-Culang, I., Ashkenazi, S., Peters, B., and Ofran, Y. (2015). PEASE: pre-
dicting B-cell epitopes utilizing antibody sequence. Bioinformatics 31,
1313–1315.
Shirai, H., Kidera, A., and Nakamura, H. (1996). Structural classification of
CDR-H3 in antibodies. FEBS Lett. 399, 1–8.
Sircar, A., and Gray, J.J. (2010). SnugDock: paratope structural optimization
during antibody-antigen docking compensates for errors in antibody homol-
ogy models. PLoS Comput. Biol. 6, e1000644.
Torchala, M., Moal, I.H., Chaleil, R.A.G., Fernandez-Recio, J., and Bates,
P.A. (2013). SwarmDock: a server for flexible protein-protein docking.
Bioinformatics 29, 807–809.
Vajda, S. (2005). Classification of protein complexes based on docking diffi-
culty. Proteins 60, 176–180.
Vreven, T., Moal, I.H., Vangone, A., Pierce, B.G., Kastritis, P.L., Torchala, M.,
Chaleil, R., Jimenez-Garcıa, B., Bates, P.A., Fernandez-Recio, J., et al. (2015).
Updates to the integrated protein-protein interaction benchmarks: docking
benchmark version 5 and affinity benchmark version 2. J. Mol. Biol. 427,
3031–3041.
De Vries, S.J., Van Dijk, M., and Bonvin, A.M.J.J. (2010). The HADDOCK web
server for data-driven biomolecular docking. Nat. Protoc. 5, 883–897.
De Vries, S.J., Schindler, C.E.M., Chauvot DeBeauchene, I., and Zacharias, M.
(2015). A web interface for easy flexible protein-protein docking with
ATTRACT. Biophys. J. 108, 462–465.
Weitzner, B.D., Dunbrack, R.L., and Gray, J.J. (2015). The origin of CDR H3
structural diversity. Structure 23, 302–311.
Weitzner, B.D., Jeliazkov, J.R., Lyskov, S., Marze, N., Kuroda, D., Frick, R.,
Adolf-Bryfogle, J., Biswas, N., Dunbrack, R.L., and Gray, J.J. (2017).
Modeling and docking of antibody structures with Rosetta. Nat. Protoc. 12,
401–416.
Wu, T.T. (2004). An analysis of the sequences of the variable regions of Bence
Jones proteins and myeloma light chains and their implications for antibody
complementarity. J. Exp. Med. 132, 211–250.
Yamashita, K., Ikeda, K., Amada, K., Liang, S., Tsuchiya, Y., Nakamura, H.,
Shirai, H., and Standley, D.M. (2014). Kotai Antibody Builder: automated
high-resolution structural modeling of antibodies. Bioinformatics 30,
3279–3280.
Zhou, H., and Zhou, Y. (2009). Distance-scaled, finite ideal-gas reference state
improves structure-derived potentials of mean force for structure selection
and stability prediction. Protein Sci. 11, 2714–2726.
Van Zundert, G.C.P., Rodrigues, J.P.G.L.M., Trellet, M., Schmitz, C., Kastritis,
P.L., Karaca, E., Melquiond, A.S.J., Van Dijk, M., De Vries, S.J., and Bonvin,
A.M.J.J. (2016). The HADDOCK2.2 web server: user-friendly integrative
modeling of biomolecular complexes. J. Mol. Biol. 428, 720–725.
Structure 28, 119–129, January 7, 2020 129
STAR+METHODS
KEY RESOURCES TABLE
REAGENT or RESOURCE SOURCE IDENTIFIER
Software and Algorithms
ClusPro (Kozakov et al., 2017) https://cluspro.org/
HADDOCK (De Vries et al., 2010) https://haddock.science.uu.nl/services/HADDOCK2.2
LightDock (Jimenez-Garcıa et al., 2019) https://github.com/brianjimenez/lightdock
ZDOCK (Pierce et al., 2011) http://zdock.umassmed.edu
PIGSpro (Lepore et al., 2017) https://cassandra.med.uniroma1.it/pigspro/
Deposited Data
Raw and analyzed data This paper https://data.sbgrid.org/dataset/686/
Others
Docking Benchmark v5 (Vreven et al., 2015) http://zlab.umassmed.edu/benchmark/
LEAD CONTACT AND MATERIALS AVAILABILITY
All data gathered for this work are publicly available and can be find at: https://data.sbgrid.org/dataset/686/.
Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Alexandre M.J.J.
Bonvin ([email protected]).
METHOD DETAILS
DatasetThe dataset used in this work includes 16 complexes, all with available unbound structures, which represent the new antibody-an-
tigen entries of the protein-protein benchmark version 5.0 (Vreven et al., 2015). These were selected because none were used for
training/scoring optimization of any of the docking software considered in this work. Antibody structures were renumbered using
an in-house R script. Only the variable domain was used for the docking. Antibodies and antigens were each randomly translated
and rotated in order to avoid any bias related to the starting orientation (this is required since the structures in the docking benchmark
are pre-oriented onto their reference bound complex).
Docking ScenariosAll methods allow the user to provide information about the binding interface. Nevertheless, only HADDOCK applies a purely data-
driven sampling strategy in order to create models in agreement with the provided information. LightDock uses this information both
to limit the sampling to specific regions and in scoring, while ClusPro and ZDOCK include this information into their scoring functions
in order to select the correct models. In this case the methods are able to assign a better score to the models that better satisfy the
given restraints.
For each complex three different docking runs were performed in order to represent different scenarios corresponding to different
amounts of available experimental information (Figure 1):
(1) HV - Surf: No information about the epitope residues are provided to the docking algorithm. Only knowledge of the antibody
HV loops, defined according to the Chothia numbering scheme (Al-Lazikani et al., 1997), is used in the docking. For HADDOCK
this was complemented by all solvent-exposed antigen residues defined by selecting those with a relative accessible surface
area (RSA) R 40% (calculated with NACCESS (Hubbard and Thornton, 1993)).
(2) HV – Epi 9: This scenario represents the case in which only a vague definition of the epitope region is provided. To mimic this
situation the epitopewas defined by selecting all antigen residueswithin 9A from the antibody. The docking runwas performed
by providing to the docking algorithms the HV loops residues and the 9A epitope.
(3) Real interface: In this ideal scenario both antibody and the antigen interfaces are well characterized. All interface residues
selected using a distance cutoff of 4.5A were given to the docking software.
Docking SettingsFour docking methods were compared in this work: ClusPro (Kozakov et al., 2017), HADDOCK (De Vries et al., 2010), LightDock (Ji-
menez-Garcıa et al., 2019) and, ZDOCK (Pierce et al., 2011).
e1 Structure 28, 119–129.e1–e2, January 7, 2020
TheClusProwebserver (https://cluspro.org) was used in the AntibodyMode (Brenke et al., 2012) using default settings. Information
was provided in the form of attractive residues.
ZDOCK predictions were obtained using version 3.0.2. The sampling was set to 2000 models. ZDOCK allows the user to assign a
highly unfavourable contact energy to the residues which are known not to be involved in the binding. Accordingly, all residues not
included in the defined interfaces were blocked.
For LightDock we used release 0.5.6 (Jimenez-Garcıa et al., 2019) of the software which provides a mechanism for including res-
idue restraints. At the receptor level, the surface swarms used in the simulation are filtered according to the Euclidean distance of the
restraints on the provided receptor residues: Only the ten closest swarms for each receptor residue restraint are kept. An additional
energy term is added to the scoring function (DFIRE (Zhou and Zhou, 2009) in this work) that accounts for the percentage of satisfied
restraints. The predictions are filtered with a minimum 40% cutoff of satisfied restraints, at both receptor and ligand levels. For the
remaining parameters default settings were: ANM enabled (10 first non-trivial modes for both receptor and ligand), 400 swarms
before filtering by restraints, 200 glowworms per swarm and 100 simulation steps.
Finally, HADDOCK version 2.2 was used with default settings except that the rigid-body (it0) sampling was increased to 5000
models for the HV – Epi9 and the Real interface run and to 10000 for the HV – Surf scenario. The flexible (it1) and water refinement
sampling were set to 400 models for all scenarios (Dominguez et al., 2003). This corresponds to an increased sampling compared to
the default settings. In general the least information is available to drive the docking in HADDOCK the larger the sampling should be.
The docking was performed using the web server version of HADDOCK (Van Zundert et al., 2016) (https://haddock.science.uu.nl). In
the case of the Real interface scenario, the random removal of restraints (by default 50% of restraints are randomly discarded for
each docking trial) was turned off. The information about the binding interface was encoded in the form of active and passive res-
idues: The antibody HV loops and paratope residues were provided as active, while, for the antigen, the surface and the epitope res-
idues, selected using a 9A cutoff were defined as passive for the first two scenarios. For the third, ideal scenario, the true interface
epitope residues selected at 4.5A distance cutoff were classified as active. The distinction between active and passive means that an
active residue not at the interface (defined as the union of active and passive residues of the partner molecule) will result in an energy
penalty while this is not the case for passive residues.
In all the methods, the antibody was treated as the receptor partner and the antigen as ligand.
HADDOCK Clustering ParametersBy default, HADDOCK performs a cluster analysis. In this work clustering was based on the Fraction of Common Contacts (FCC)
(Rodrigues et al., 2012) using 0.6 as cutoff and 4 as minimum cluster size. The clusters were sorted according to the average
HADDOCK score of the best 4 model of each cluster, from the lowest HADDOCK score to the highest.
Evaluation CriteriaDocking models were classified as high (***), medium (**) of low (*) quality according to the CAPRI criteria (Janin et al., 2003; Mendez
et al., 2003) (see Table 1) based on their similarities with the native structure by calculating the interface root mean square deviation
(i-RMSD), the ligand root mean square deviation (L-RMSD) and the fraction of native contacts (Fnat). Briefly, the i-RMSD is calculated
on the backbone atoms of all interface residues of the native complex defined using a 10A cutoff and the L-RMSD is calculated by
superimposing on the backbone atoms of the antibody and calculating the RMSD of the antigen backbone atoms. Finally, Fnat is
calculated as number of native contacts in a docking model divided by the total number of contacts in the reference structure.
Fnat has been calculated using is house scripts while fitting and RMSD calculations were performed using the McLachlan algorithm
(McLachlan, 1982) as implemented in the program ProFit (http://www.bioinf.org.uk/software/profit/) from the SBGrid distribution
(Morin et al., 2013).
DATA AND CODE AVAILABILITY
All models generated using the four software and the different scenarios, together with their quality statistics and scores have been
deposited into the SBGrid data repository (Meyer et al., 2016) and can be find at: https://data.sbgrid.org/dataset/686/. Links to the
various software used in this work are provided in the Key Resources Table.
Structure 28, 119–129.e1–e2, January 7, 2020 e2
Structure, Volume 28
Supplemental Information
Modeling Antibody-Antigen Complexes
by Information-Driven Docking
Francesco Ambrosetti, Brian Jiménez-García, Jorge Roel-Touris, and Alexandre M.J.J.Bonvin
Supplementary material
Modelling antibody-antigen complexes by information-driven docking
F. Ambrosettia,b, B. Jiménez-Garcíab, J. Roel-Tourisb, A.M.J.J. Bonvinb*
aDepartment of Physics, Sapienza University, Piazzale Aldo Moro 5, 00184, Rome, Italy,
bFaculty of Science - Chemistry, Computational Structural Biology Group, Bijvoet Center for
Biomolecular Research, Utrecht University, Utrecht, The Netherlands
Lead Contact: [email protected]
Table S1, related to Figure 2: Rank, i-RMSD and Fnat values of the first acceptable and the best model of HADDOCK
HV - Surf HV - Epi 9 Real interface
First acceptable model Best model First acceptable model Best model First acceptable model Best model
Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat
2VXT 14 1.77 0.53 14 1.77 0.53 2 1.86 0.57 26 1.54 0.61 1 1.82 0.66 312 1.41 0.71
2W9E 69 3.92 0.29 111 2.38 0.25 21 4.08 0.23 264 2.21 0.3 1 2.61 0.55 41 2.33 0.57
3EO1 387 3.62 0.24 387 3.62 0.24 85 3.26 0.3 115 2.48 0.46 1 2.4 0.74 233 1.81 0.74
3EOA 234 3.7 0.33 234 3.7 0.33 2 3.36 0.56 37 0.66 0.86 1 0.9 0.84 356 0.68 0.81
3G6D 4 2.32 0.65 155 2.23 0.6 3 2.17 0.71 6 2.03 0.73 1 2.35 0.81 2 2.13 0.83
3HI6 38 4.67 0.17 359 2.32 0.34 18 4.85 0.15 213 2.51 0.42 1 2.54 0.54 65 2.14 0.52
3HMX - - - 323 8.04 0.01 1 1.32 0.81 26 1.27 0.65 1 1.16 0.84 291 0.92 0.93
3L5W - - - 173 4.37 0.03 1 2.45 0.34 40 2.45 0.24 1 2.52 0.31 210 1.94 0.86
3MXW 1 1.05 0.83 69 0.94 0.75 1 0.94 0.78 55 0.78 0.72 1 0.93 0.86 291 0.78 0.83
3RVW 316 2.04 0.51 316 2.04 0.51 1 3.89 0.45 254 1.37 0.59 1 2.55 0.57 76 2.07 0.69
3V6Z 90 3.62 0.31 277 2.73 0.27 51 3.61 0.31 219 2.57 0.3 1 2.46 0.47 299 2.25 0.48
4DN4 1 4.44 0.19 82 1.27 0.67 2 1.74 0.74 168 1.19 0.72 1 1.61 0.77 228 1.18 0.86
4FQI - - - 363 13.31 0 1 3.43 0.32 21 2.53 0.3 1 2.82 0.6 58 2.01 0.54
4G6J 1 1.55 0.74 9 1.17 0.71 1 1.3 0.74 11 1.09 0.73 1 1.27 0.71 91 0.97 0.84
4G6M 1 2.36 0.53 4 1.68 0.62 1 0.94 0.91 56 0.83 0.84 1 0.95 0.91 25 0.76 0.95
4GXU - - - 345 54.42 0 4 1.72 0.77 8 1.39 0.74 1 1.29 0.83 174 0.99 0.85
Table S2, related to Figure 2: Rank, i-RMSD and Fnat values of the first acceptable and the best model of ClusPro
HV - Surf HV - Epi 9 Real interface
First acceptable model Best model First acceptable model Best model First acceptable model Best model
Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat
2VXT 12 1.99 0.49 12 1.99 0.49 1 2.7 0.43 1 2.7 0.43 1 1.93 0.33 1 1.93 0.33
2W9E 11 3.09 0.5 24 2.65 0.54 4 2.36 0.62 6 2.35 0.59 3 1.12 0.73 3 1.12 0.73
3EO1 - - - 16 7.8 0.06 19 3.87 0.22 19 3.87 0.22 1 2.45 0.48 1 2.45 0.48
3EOA - - - 26 10.09 0 18 2.36 0.58 18 2.36 0.58 1 1.95 0.7 1 1.95 0.7
3G6D 2 2.42 0.63 2 2.42 0.63 1 2.26 0.58 1 2.26 0.58 1 2.26 0.58 1 2.26 0.58
3HI6 - - - 10 6.58 0.05 - - - 19 5.64 0.09 7 4.15 0.17 7 4.15 0.17
3HMX 9 1.13 0.81 9 1.13 0.81 2 2.25 0.54 2 2.25 0.54 1 1.24 0.78 1 1.24 0.78
3L5W - - - 10 5.99 0 1 1.82 0.72 1 1.82 0.72 1 1.47 0.79 1 1.47 0.79
3MXW 1 1.57 0.74 1 1.57 0.74 1 1.51 0.66 1 1.51 0.66 1 1.68 0.78 1 1.68 0.78
3RVW - - - 30 7.91 0.22 1 2.39 0.57 1 2.39 0.57 1 1.57 0.67 1 1.57 0.67
3V6Z - - - 12 13.06 0 23 3.88 0.3 23 3.88 0.3 1 2.65 0.41 1 2.65 0.41
4DN4 - - - 15 4.52 0.16 3 1.98 0.7 3 1.98 0.7 1 1.97 0.65 1 1.97 0.65
4FQI - - - 11 10.07 0 1 2.98 0.26 30 2.79 0.24 3 3.05 0.26 24 2.67 0.32
4G6J 12 1.22 0.69 12 1.22 0.69 2 4.18 0.29 3 1.37 0.74 3 3.7 0.35 6 2.2 0.65
4G6M - - - 11 9.52 0.02 1 2.66 0.57 4 2.11 0.74 1 2.66 0.57 3 1.93 0.72
4GXU - - - 10 10.84 0.02 17 3.63 0.42 27 2.81 0.59 1 2.46 0.65 5 2.45 0.7
Table S3, related to Figure 2: Rank, i-RMSD and Fnat values of the first acceptable and the best model of LightDock
HV - Surf HV - Epi 9 Real interface
First acceptable model Best model First acceptable model Best model First acceptable model Best model
Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat
2VXT - - - 100 6.08 0.03 6 4.07 0.25 6 4.07 0.25 6 2.74 0.38 6 2.74 0.38
2W9E 88 2.28 0.36 88 2.28 0.36 28 2.54 0.25 28 2.54 0.25 20 2.11 0.41 20 2.11 0.41
3EO1 288 4.13 0.69 288 4.13 0.69 18 5.47 0.2 54 3.32 0.09 4 5.52 0.35 148 3.35 0.26
3EOA 17 5.03 0.16 17 5.03 0.16 9 4.84 0.37 9 4.84 0.37 7 2.56 0.65 75 2.31 0.44
3G6D 18 6.09 0.35 142 3.71 0.21 12 3.36 0.37 12 3.36 0.37 1 2.69 0.5 11 2.56 0.62
3HI6 - - - 55 5.78 0.03 2 5.26 0.12 10 3.06 0.42 8 5.37 0.22 53 5.04 0.2
3HMX 86 7.92 0.14 339 5.04 0.35 8 9.68 0.23 14 4.01 0.22 1 7.15 0.14 61 5.42 0.19
3L5W 79 3.03 0.52 79 3.03 0.52 20 2.98 0.55 20 2.98 0.55 26 3.94 0.38 129 3.41 0.38
3MXW 4 2.93 0.45 4 2.93 0.45 1 2.97 0.42 1 2.97 0.42 1 2.38 0.57 2 1.76 0.71
3RVW 64 5.74 0.16 198 3.66 0.69 2 4.5 0.43 17 4.38 0.22 1 3.85 0.41 85 2.26 0.49
3V6Z 104 3.56 0.31 178 3.05 0.42 23 3.55 0.36 23 3.55 0.36 66 2.71 0.42 102 2.61 0.33
4DN4 62 4.47 0.23 65 4.46 0.28 22 4.91 0.12 44 3.36 0.44 2 3.84 0.21 14 1.84 0.7
4FQI 26 4.17 0.2 26 4.17 0.2 3 4.15 0.22 3 4.15 0.22 1 4.65 0.2 43 4.39 0.04
4G6J 14 2.7 0.47 14 2.7 0.47 2 2.24 0.56 2 2.24 0.56 18 6.37 0.27 102 3.97 0.13
4G6M 11 5.11 0.21 6 4.31 0.09 1 4.28 0.1 9 2.35 0.43 2 1.62 0.57 7 1.17 0.78
4GXU 145 11.36 0.17 145 11.36 0.17 - - - 12 12.57 0 - - - 19 7.58 0.08
Table S4, related to Figure 2: Rank, i-RMSD and Fnat values of the first acceptable and the best model of ZDOCK
HV - Surf HV - Epi 9 Real interface
First acceptable model Best model First acceptable model Best model First acceptable model Best model
Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat Rank i-RMSD[Å]
Fnat
2VXT 5 4.01 0.2 360 1.71 0.42 11 3.93 0.18 340 1.7 0.51 1 1.69 0.62 1 1.69 0.62
2W9E 78 3.92 0.12 201 2.57 0.55 81 2.75 0.5 81 2.75 0.5 1 2.86 0.5 186 1.44 0.39
3EO1 120 3.31 0.37 314 2.99 0.2 4 3.9 0.11 102 2.21 0.35 1 3.42 0.37 2 2.43 0.33
3EOA 171 4.16 0.12 199 3.17 0.19 4 4.16 0.12 42 1.26 0.77 1 1.26 0.77 1 1.26 0.77
3G6D 4 3.88 0.38 183 2.01 0.65 3 2.86 0.6 9 2.01 0.65 1 2.86 0.6 56 2.01 0.65
3HI6 2 3.46 0.35 247 2.67 0.29 14 3.46 0.35 254 2.26 0.51 1 3.11 0.22 71 2.22 0.32
3HMX 335 1.22 0.68 335 1.22 0.68 1 1.27 0.76 1 1.27 0.76 1 1.08 0.91 1 1.08 0.91
3L5W 4 3.67 0.28 394 1.37 0.76 21 2.96 0.17 316 1.36 0.62 35 1.12 0.69 35 1.12 0.69
3MXW 1 1.1 0.77 1 1.1 0.77 1 1.1 0.77 1 1.1 0.77 1 1.1 0.77 1 1.1 0.77
3RVW - - - 394 8.04 0.14 2 2.94 0.55 20 1.36 0.76 1 2.2 0.61 86 0.81 0.71
3V6Z - - - 362 9.81 0 - - - 231 4.4 0.14 1 3.7 0.33 34 2.83 0.34
4DN4 35 2.83 0.49 312 1.91 0.65 7 2.17 0.58 56 1.05 0.7 1 2.17 0.58 11 1.05 0.7
4FQI - - - 69 5.63 0.24 1 3.38 0.34 315 2.27 0.52 1 3.38 0.34 16 1.6 0.74
4G6J 332 2.29 0.39 332 2.29 0.39 40 3.68 0.32 290 2.12 0.58 3 1.39 0.65 6 1.2 0.73
4G6M 141 3.33 0.24 254 1.86 0.45 3 1.95 0.48 87 1.72 0.81 1 1.86 0.45 3 0.81 0.91
4GXU - - - 339 12.15 0.03 5 3.73 0.48 303 2.39 0.48 2 2.06 0.59 3 1.38 0.47
Table S5, related to Figure 5: Percentage of acceptable, medium and high quality with respect
to the total number of models generated by HADDOCK for the three scenarios discussed in
this work: HV – Surf, HV – Epi 9 and the Real interface.
HV - Surf HV - Epi 9 Real interface
Pdb * ** *** * ** *** * ** ***
2VXT 0 0.25 0 12.5 68.75 0 0 100 0
2W9E 1.75 0 0 6.25 0 0 92 0 0
3EO1 0.25 0 0 7 0 0 62.5 37.5 0
3EOA 0.25 0 0 13.25 6 4.75 0 31 68.75
3G6D 0.25 0.75 0 2 5.5 0 96.5 3.5 0
3HI6 12.25 1.25 0 29.25 3.75 0 0.75 95.75 0
3HMX 0 0 0 6 1.25 0 0 99.25 0.75
3L5W 0 0 0 9.25 0 0 52 1.75 0
3MXW 3 13.75 1 8.75 73 18 0 6.75 93.25
3RVW 0.25 0 0 53 19.5 0 99.75 0.25 0
3V6Z 0.75 0 0 2 0 0 100 0 0
4DN4 10 11.5 0 21.75 42.5 0 0.25 99.75 0
4FQI 0 0 0 11.75 0 0 100 0 0
4G6J 1.5 6.25 0 8.5 16.75 0 0 98.25 1.75
4G6M 0.5 0.5 0 6.25 14 1 0 54 46
4GXU 0 0 0 2.25 3.25 0 0.5 99.25 0.25
Table S6, related to Figure 5: Percentage of acceptable, medium and high quality with respect
to the total number of models generated by ClusPro for the three scenarios discussed in this
work: HV – Surf, HV – Epi 9 and the Real interface.
HV - Surf HV - Epi 9 Real interface
Pdb * ** *** * ** *** * ** ***
2VXT 0 3.7 0 10 0 0 0 7.14 0
2W9E 7.69 0 0 30.77 0 0 16.67 8.33 0
3EO1 0 0 0 4 0 0 3.85 3.85 0
3EOA 0 0 0 3.85 0 0 10.53 5.26 0
3G6D 3.33 0 0 4.17 4.17 0 5 5 0
3HI6 0 0 0 0 0 0 4.76 0 0
3HMX 0 3.33 0 8.7 0 0 0 6.25 0
3L5W 0 0 0 14.29 4.76 0 21.43 7.14 0
3MXW 8 4 0 12.5 6.25 0 15.38 7.69 0
3RVW 0 0 0 9.09 0 0 5.56 5.56 0
3V6Z 0 0 0 3.33 0 0 4.55 0 0
4DN4 0 0 0 8.33 4.17 0 18.75 6.25 0
4FQI 0 0 0 10 0 0 15.38 0 0
4G6J 0 3.33 0 11.54 3.85 0 10 5 0
4G6M 0 0 0 9.52 4.76 0 16.67 5.56 0
4GXU 0 0 0 6.9 0 0 18.75 0 0
Table S7, related to Figure 5: Percentage of acceptable, medium and high quality with respect
to the total number of models generated by LightDock for the three scenarios discussed in this
work: HV – Surf, HV – Epi 9 and the Real interface.
HV - Surf HV - Epi 9 Real interface
Pdb * ** *** * ** *** * ** ***
2VXT 0 0 0 5.26 0 0 1.18 1.18 0
2W9E 0.98 0.49 0 2.3 0 0 1.99 0.5 0
3EO1 0.34 0.34 0 5.41 0 0 5.69 0.95 0
3EOA 0.64 0 0 5 0 0 3.23 1.29 0
3G6D 2.15 0 0 4.48 0 0 4.94 0.76 0
3HI6 0 0 0 7.5 2.5 0 6.82 0 0
3HMX 1.42 0 0 13.64 0 0 3.28 0 0
3L5W 1.18 0 0 2.82 0 0 4.88 0 0
3MXW 0.55 0 0 1.52 0 0 3.61 0.52 0
3RVW 1.73 0 0 13.16 0 0 10.29 1.71 0
3V6Z 1.65 0 0 3.28 0 0 2.55 0.43 0
4DN4 3.9 0 0 4 0 0 7.17 0.45 0
4FQI 2.37 0 0 14.29 0 0 23.53 0 0
4G6J 3.52 0.44 0 9.3 1.16 0 3.24 0 0
4G6M 2.04 0 0 7.14 1.43 0 4.78 2.39 0
4GXU 0.61 0 0 0 0 0 0 0 0
Table S8, related to Figure 5: Percentage of acceptable, medium and high quality with respect
to the total number of models generated by ZDOCK for the three scenarios discussed in this
work: HV – Surf, HV – Epi 9 and the Real interface.
HV - Surf HV - Epi 9 Real interface
Pdb * ** *** * ** *** * ** ***
2VXT 1.15 0.15 0 1.25 0.15 0 2 0.2 0
2W9E 1.45 0 0 2.35 0 0 5.75 0.5 0
3EO1 0.4 0 0 0.7 0 0 1.6 0 0
3EOA 0.65 0 0 2.1 0.35 0 2.7 0.45 0
3G6D 1.3 0.1 0 1.6 0.15 0 1.8 0.1 0
3HI6 0.8 0 0 0.7 0.05 0 1.35 0.1 0
3HMX 0.3 0.1 0 0.7 0.2 0 1.4 0.35 0
3L5W 3.85 0.05 0 6.15 0.25 0 4.3 0.4 0
3MXW 1.85 0.35 0 2.3 0.35 0 2.75 0.35 0
3RVW 0.05 0.05 0 3 0.25 0 2.7 0.3 0.05
3V6Z 0 0 0 0.45 0 0 1.45 0 0
4DN4 5.3 0.55 0 7.55 0.85 0 7.6 0.9 0
4FQI 0 0 0 2.35 0 0 2.75 0.05 0
4G6J 0.05 0.05 0 1 0.05 0 1.35 0.15 0
4G6M 0.95 0.05 0 2.05 0.3 0 2.1 0.15 0.05
4GXU 0 0 0 0.45 0 0 0.8 0.2 0
Figure S1, related to Figure 2: ClusPro success rate for three scenarios described in this work
as a function of the top 1, 5, 10, 20, 50 and 100 ranked models. Docking was performed using
antibody mode and automatic mask of non-CDR regions. For comparison, the results using the
HV loops selected according to the Chothia numbering scheme are shown in grayscale (those
are the ones shown in Figure 1).
06.2
012.56.2
0
18.812.5
0
25
18.8
0
31.2
18.8
0
31.2
18.8
06.2
06.26.2
012.56.2
0
2512.5
0
2512.5
0
2512.5
0
256.2
0
37.5
25
0
50
25
0
50
31.2
0
50
31.2
0
50
31.2
0
18.8
25
0
37.5
31.2
0
37.5
31.2
0
37.5
50
0
37.5
56.2
0
37.5
56.2
0
25
25
0
43.8
37.5
0
50
37.5
0
50
37.5
0
50
37.5
0
50
37.5
0
56.2
18.8
0
68.8
25
0
75
25
0
75
25
0
75
25
0
75
25
CLUSPRO
HV − SurfHV − Epi 9
Real interface
T1 T5 T10 T20 T50 T100
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
Succ
ess
rate
Quality: Acceptable−CDRMedium−CDRHigh−CDRAcceptable−HVMedium−HVHigh−HV
Figure S2, related to Figure 2: HADDOCK success rate for three scenarios described in this
work as a function of the top 1, 5, 10, 20, 50 and 100 ranked models. Docking were performed
using and ensemble of 10 antibody models, generated by refining, using the HADDOCK water
stage, the structure obtained with the PIGSpro webserver, and the antigen free forms. For
comparison, the results using the unbound experimental antibody structures are shown in
grayscale (those are the ones shown in Figure 1). The PIGSpro models were generated with
default template selection parameters and H3LoopPred option for the H3 modelling. All of the
structures with 100% of sequence identity with the targeted antibody were blacklisted to avoid
their use as template for the modelling.
06.2
06.26.2
06.212.5
06.2
18.8
012.5
18.8
012.5
37.5
012.512.5
0
31.2
0
31.2
6.2
31.2
6.2
37.5
6.2
37.512.5
012.5
0
18.8
37.5
0
31.2
37.5
0
37.5
37.5
0
50
31.2
0
50
37.5
12.512.5
18.8
12.5
37.5
25
12.5
43.8
18.8
12.5
43.8
25
18.8
43.8
25
18.8
43.8
37.5
0
31.2
31.2
0
37.5
25
6.2
37.5
31.2
6.2
56.2
18.8
6.2
56.2
18.8
6.2
62.512.5
18.8
37.5
43.8
18.8
43.8
37.5
18.8
50
31.2
18.8
50
31.2
25
50
25
31.2
43.8
25
HADDOCK
HV − SurfHV − Epi 9
Real − interface
T1 T5 T10 T20 T50 T100
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
Succ
ess
rate
Quality: Acceptable−ModMedium−ModHigh−ModAcceptable−UnMedium−UnHigh−Un
Figure S3, related to Figure 2: HADDOCK success rate as function of top 1, 5, 10, 20, 50
and 100 ranked models. Docking is presented for the second and third scenarios described in
this work using the default sampling settings (1000/200/200). For comparison, the results using
the increased sampling are shown in grayscale (those are the ones shown in Figure 1).
0
37.5
18.8
6.2
43.8
37.5
18.8
37.5
31.2
18.8
43.8
31.2
18.8
43.8
37.5
18.8
50
31.2
12.5
12.5
18.8
12.5
37.5
25
12.5
43.8
18.8
12.5
43.8
25
18.8
43.8
25
18.8
43.8
37.5
12.5
43.8
37.5
12.5
56.2
31.2
18.8
56.2
25
25
50
25
31.2
50
18.8
31.2
50
18.8
18.8
37.5
43.8
18.8
43.8
37.5
18.8
50
31.2
18.8
50
31.2
25
50
25
31.2
43.8
25
HADDOCK
HV − Epi 9Real − interface
T1 T5 T10 T20 T50 T100
0
25
50
75
100
0
25
50
75
100
Succ
ess
rate
Quality: Acceptable−RedMedium−RedHigh−RedAcceptableMediumHigh
Figure S4, related to Figure 2: HADDOCK success rate as function of the top 1, 5, 10, 20,
50 and 100 ranked models. Docking was performed by providing to the system the antibody
paratope residues defined with a distance cutoff of 4.5Å from the antigen as active and the
antigen residues within 4.5Å from the antibody as passive. For comparison, the results obtained
by defining the antigen residues as active in HADDOCK are shown in grayscale (those are the
ones shown in Figure 1).
12.5
50
18.8
12.5
50
25
18.8
43.8
25
18.8
50
18.8
18.8
50
18.8
25
50
18.8
18.8
37.5
43.8
18.8
43.8
37.5
18.8
50
31.2
18.8
50
31.2
25
50
25
31.2
43.8
25
HADDOCK
Real − interface
T1 T5 T10 T20 T50 T100
0
25
50
75
100
Succ
ess
rate
Quality: Acceptable−PassMedium−PassHigh−PassAcceptableMediumHigh
Figure S5, related to Figure 7: Distributions of i-RMSD differences between the three
HADDOCK stages for the three docking scenarios described in this work. The distributions
are calculated from all 16 complexes considering 400 models in each case. The left panels
show the difference between it1 and it0 (it1 – it0) while the right panes show the difference
between Water and it1 (Water – it1). Negative values represent an improvement of the models
in terms of i-RMSD values.
it1−it0
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
0.0
0.5
1.0
1.5
i−RMSD [Å]
Den
sity
Water−it1H
V − Surf
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
0.0
2.5
5.0
7.5
10.0
i−RMSD [Å]
Den
sity
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
0.0
0.5
1.0
1.5
i−RMSD [Å]
Den
sity
HV − Epi 9
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
0
2
4
6
8
i−RMSD [Å]
Den
sity
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
0.0
0.5
1.0
1.5
2.0
i−RMSD [Å]
Den
sity
Real interface
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
0
2
4
6
8
i−RMSD [Å]
Den
sity
Figure S6, related to Figure 7: Distributions of Fnat difference between the three HADDOCK
stages for the three docking scenarios described in this work. The distributions are calculated
from all 16 complexes considering 400 models in each case. The left panels show the difference
between it1 and it0 (it1 – it0) while the right panes show the difference between Water and it1
(Water – it1). Positive values indicate an improvement in the fraction of native contacts made
by the two molecules.
it1−it0
−0.2 0 0.2 0.4 0.6 0.80
20
40
60
Fnat
Den
sity
Water−it1H
V − Surf
−0.2 0 0.2 0.4 0.6 0.80
30
60
90
Fnat
Den
sity
−0.2 0 0.2 0.4 0.6 0.80
5
10
15
Fnat
Den
sity
HV − Epi 9
−0.2 0 0.2 0.4 0.6 0.80
10
20
30
Fnat
Den
sity
−0.2 0 0.2 0.4 0.6 0.80
1
2
3
4
5
Fnat
Den
sity
Real interface
−0.2 0 0.2 0.4 0.6 0.80
5
10
Fnat
Den
sity