High-affinity binders selected from designed ankyrin repeat protein libraries.

Abstract

We report here the evolution of ankyrin repeat (AR) proteins in vitro for specific, high-affinity target binding. Using a consensus design strategy, we generated combinatorial libraries of AR proteins of varying repeat numbers with diversified binding surfaces. Libraries of two and three repeats, flanked by 'capping repeats,' were used in ribosome-display selections against maltose binding protein (MBP) and two eukaryotic kinases. We rapidly enriched target-specific binders with affinities in the low nanomolar range and determined the crystal structure of one of the selected AR proteins in complex with MBP at 2.3 A resolution. The interaction relies on the randomized positions of the designed AR protein and is comparable to natural, heterodimeric protein-protein interactions. Thus, our AR protein libraries are valuable sources for binding molecules and, because of the very favorable biophysical properties of the designed AR proteins, an attractive alternative to antibody libraries.

Full text

Page 1

ARTICLES
Repeat proteins are ubiquitous binding molecules fundamental to
many biological processes1–3. Their modular architecture is presum-
ably the key to their evolutionary success4.Repeat proteins are charac-
terized by consecutive homologous structural units (repeats), which
stack to form an elongated protein domain with a continuous
hydrophobic core3. In principle, this architecture allows their binding
specificities to evolve not only by point mutations but also by inser-
tion,deletion or shuffling of repeats5.This evolutionary strategy might
enable repeat proteins to acquire new functions by adjusting their sur-
face without jeopardizing their overall topology. AR proteins are one
prominent repeat protein family illustrating the binding versatility of
repeat proteins.They occur throughout all phyla and mediate protein-
protein interactions in the nucleus or cytoplasm,or while anchored to
the membrane or when secreted into the extracellular space6.AR pro-
teins are built from stacked, 33 amino acid repeats, each forming a
β-turn that is followed by two antiparallel α-helices and a loop reach-
ing the β-turn of the next repeat7. In most known complexes, the
β-turn and the first α-helix mediate the interactions with the target,
and different numbers of adjacent repeats are involved in binding7.
The reported target binding affinities of natural AR proteins are in the
low nanomolar range8,9.
In biotechnology and biomedical research,antibodies and fragments
thereof are the most widely used specific, high-affinity binding mole-
cules.Antibodies can be generated against essentially any target either
by immunization or by using natural or rationally designed antibody
librariesin vitro10,11.Yet,many antibodies have relatively low expression
yields, a tendency to aggregate and a dependence on disulfide bonds
for stability. An ideal alternative protein scaffold would have none of
these drawbacks while still exhibiting the same affinity and specificity
as antibodies.Previous attempts to generate alternative binding mole-
cules relied on either loop or surface randomization of protein scaf-
folds,which are typically small and always fixed in dimension12–14.
We present here the results of a different strategy15to design and
select alternative protein scaffolds, relying on the modularity of AR
proteins (Fig. 1). We generated combinatorial libraries of consensus-
designed AR proteins of varying sizes (that is,varying repeat numbers)
with randomized potential interaction surfaces. Unselected library
members are very well expressed, soluble, thermodynamically stable
and show the typical AR domain fold16,17. Here, we show the success-
ful selection of binding molecules from these libraries, proving that
our design strategy works. We selected specific binders with high
affinities for the Escherichia coli maltose binding protein (MBP) and
two eukaryotic mitogen-activated protein kinases (MAPKs).The crys-
tal structure of one of the selected binders in complex with MBP was
determined, revealing atomic level insights into the mode of target
interaction of this class of designed binding molecules.
RESULTS
Designed AR protein libraries
We designed a consensus AR module consisting of six diversified
potential interaction residues (which can be any amino acid except
cysteine,glycine and proline) and 27 framework residues (26 are fixed
and one is allowed to be asparagine, histidine or tyrosine). This mod-
ule was designed from sequence alignments and structural analyses
(Fig.1)16.The randomized potential interaction residues are located in
the β-turn and the first α-helix of the AR module.We cloned varying
numbers of this repeat module between capping repeats, which are
special terminal repeats of AR domains shielding the hydrophobic
1Biochemisches Institut, Universität Zürich, Winterthurerstrasse 190, CH-8057 Zürich, Switzerland. 2These authors contributed equally to this work. Correspondence
should be addressed to A.P. (plueckthun@bioc.unizh.ch) or M.G.G. (gruetter@bioc.unizh.ch).
Published online 18 April 2004; doi:10.1038/nbt962
High-affinity binders selected from designed ankyrin
repeat protein libraries
H Kaspar Binz1,2,Patrick Amstutz1,2,Andreas Kohl1,2,Michael T Stumpp1,Christophe Briand1,Patrik Forrer1,
Markus G Grütter1& Andreas Plückthun1
We report here the evolution of ankyrin repeat (AR) proteins in vitro for specific, high-affinity target binding. Using a consensus
design strategy, we generated combinatorial libraries of AR proteins of varying repeat numbers with diversified binding surfaces.
Libraries of two and three repeats, flanked by ‘capping repeats,’ were used in ribosome-display selections against maltose
binding protein (MBP) and two eukaryotic kinases. We rapidly enriched target-specific binders with affinities in the low
nanomolar range and determined the crystal structure of one of the selected AR proteins in complex with MBP at 2.3 Å
resolution. The interaction relies on the randomized positions of the designed AR protein and is comparable to natural,
heterodimeric protein-protein interactions. Thus, our AR protein libraries are valuable sources for binding molecules and,
because of the very favorable biophysical properties of the designed AR proteins, an attractive alternative to antibody libraries.
NATURE BIOTECHNOLOGY VOLUME 22 NUMBER 5 MAY 2004
575
© 2004 Nature Publishing Group http://www.nature.com/naturebiotechnology

Page 2

ARTICLES
core (Fig.1).We thereby increased the size of the potential interaction
surface and potentiated its diversity. This strategy provided combina-
torial libraries of designed AR proteins with distinct repeat numbers.
The libraries were named N2C and N3C, indicating proteins had an
N-terminal capping repeat,two and three,respectively,designed (and
randomized) AR modules and a C-terminal capping repeat (Fig. 1).
We used the N2C and N3C libraries for the selections, because AR
proteins of this length are very abundant in nature6.Unselected mem-
bers of these libraries were expressed in soluble form at about 200 mg/l
inE.coli shake flask cultures.These proteins were monomeric,showed
circular dichroism (CD) spectra indistinguishable from natural AR
proteins16and the AR fold was confirmed by a crystal structure of an
unselected library member17. Similar to designed proteins in other
consensus repeat protein studies18,19, our designed AR proteins
showed high thermodynamic stability during unfolding induced
by heat16and denaturants17. Hence, the consensus-designed AR pro-
teins are stable scaffolds with large and modular potential interaction
surfaces.The theoretical diversities of these libraries are 5.2·1015(N2C)
and 3.8·1023(N3C)16. The DNA libraries used in the selections con-
tained at least 1010individual members each as estimated from the
amount of ligated library DNA. The library diversities were further
increased in subsequent PCR cycles.
Ribosome-display selection against MBP
We chose E. coli MBP as the first target protein for evaluating our
libraries,because it can be obtained in large amounts in pure form and
because its structure is known (Protein Data Bank (PDB) entry
1LLS)20. We did the ribosome-display selections21with biotinylated
MBP bound to neutravidin in microtiter plates. An enrichment of
binders was observed after the second selection round both for the
N2C and the N3C libraries.We performed a total of four to five selec-
tion rounds before analyzing single,selected library members.
We screened individual selected AR proteins for MBP binding by an
enzyme-linked immunosorbent assay (ELISA) using crude E. coli
extracts. Of 60 N3C AR proteins screened, 18 gave a specific signal
(signal/background ≥ 10), compared to 4 of 56 N2C molecules.
Sequencing of the 18 MBP binding N3C AR proteins revealed that
they could be divided into at least three sequence groups (see
Supplementary Fig. 1 online). However, identical clones were never
found and considerable diversity was left,indicating that an even more
stringent selection pressure could be applied. For the N2C clones the
sequence analysis was less conclusive because of the limited data set.
Nevertheless, some repeats of N2C molecules showed striking
sequence similarities to repeats of the selected N3C proteins. In both
576
VOLUME 22 NUMBER 5 MAY 2004 NATURE BIOTECHNOLOGY
a
b
c
Figure 1 Construction of designed AR protein libraries. (a) Sequences of
the N-terminal capping AR, the designed AR module and the C-terminal
capping AR. The secondary structure elements are indicated above the
sequences. The designed AR module consists of 26 defined framework
residues, six randomized potential interaction residues (red x, any of the
20 natural amino acids except cysteine, glycine or proline) and one
randomized framework residue (z, any of the amino acids asparagine,
histidine or tyrosine). The designed AR module was derived via sequence
and structure consensus analyses16. (b) Schematic representation of the
library generation of designed AR proteins. Note that this assembly is
represented on the protein level, whereas the real library assembly is on
the DNA level. By assembling an N-terminal capping AR (green), varying
numbers of the designed AR module (blue) and a C-terminal capping AR
(cyan), combinatorial libraries of designed AR proteins of different repeat
numbers were generated (side chains of the randomized potential interaction
residues are shown in stick-mode in red). (c) Ribbon representation of the
selected MBP binding AR protein off7 (colors as in b). This binder is derived
from a library consisting of a N-terminal capping AR, three designed AR
modules and a C-terminal capping AR. This figure was made with MolMol49.
a
b
Figure 2 Expression, purification and SPR analysis of selected AR proteins.
(a) Expression and purification of the selected MBP binders mbp3_16 (1),
off7 (2) and mbp3_5 (3). At OD600 = 0.6, the noninduced (– Ind.) cultures
were induced with 0.5 mM IPTG and grown for 4 h at 37 °C (+ Ind.). After
cell lysis, the AR proteins are in the soluble fraction (Sol.). Insol., insoluble
fraction. The proteins were then purified in a single IMAC purification step
(Purif.). (b) BIAcore analysis of off7. Different concentrations of off7 (0, 2,
5, 10, 20, 50 and 100 nM) were applied to a flow cell with immobilized
MBP for 2 min, followed by washing with buffer flow. The global fit is
indicated in the figure by red dashed lines (see Table 1 for the extracted
kinetic data).
© 2004 Nature Publishing Group http://www.nature.com/naturebiotechnology

Page 3

ARTICLES
N2C and N3C molecules, aromatic residues appeared frequently and
thus seem to be important in the selected sequences (see
Supplementary Fig.1 online).
Framework mutations were randomly scattered and occurred at low
frequencies but were present in all molecules (on average, 3.8 amino
acid mutations per N3C and 2.4 mutations per N2C molecule; see
Supplementary Fig. 1 online). Each library member went through at
least 450 PCR cycles (library generation and selection), which might
explain this finding. The mutations are mostly of a conservative
nature, and thus the fold of the AR domains is most probably not
affected by the alteration of the framework (see description of the
crystal structure below).
Selected AR proteins show high affinity and specificity
The selected AR proteins were expressed at high levels in soluble form
in the cytoplasm ofE.coli (up to 200mg/l) and purified to homogene-
ity by a single immobilized metal ion affinity chromatography (IMAC)
purification step (Fig.2a).We screened 21 clones (4 N2C,17 N3C) by
surface plasmon resonance (SPR). Using the purified AR proteins at
1 µM,we first compared the on- and off-rates of MBP binding.Three
AR proteins with slow off-rates were analyzed at multiple concentra-
tions and evaluated with a global kinetic fit (N2C:mbp3_16;N3C:off7
and mbp3_5) (Fig.2band Table 1).off7,a selected N3C molecule,had
the highest affinity for MBP (KD= 4.4 nM). The N2C molecule
mbp3_16 had a dissociation constant of KD=17nM.Hence,both N2C
and N3C molecules can be selected to bind MBP with high affinity.
Clones that went through five selection rounds had higher affinities
for MBP than clones that were selected through four rounds. In SPR
experiments, the N3C library member off7 was specific and did not
cross-react with phage lambda protein D22,streptavidin or the amino-
glycoside-3′-phosphotransferase APH(3′)-IIIa, a bacterial kanamycin
resistance protein23.
To further investigate specificity,ELISA experiments were done with
purified MBP-binding AR proteins (Fig. 3). mbp3_16, off7 and
mbp3_5 are specific for MBP and do not interact with phage lambda
protein D22, APH23or neutravidin (Fig. 3a). In a competition ELISA
experiment, the binding of off7 to immobilized MBP could be inhib-
ited by preincubation with free MBP (Fig. 3b). The affinity estimated
from these experiments was consistent with the SPR measurements
(50% inhibition at 10 nM). The unselected N3C AR library member
E3_516,17did not interact with MBP, indicating that the designed AR
domain scaffolds per se do not bind MBP (Fig.3b).
At a concentration of 15 µM, off7 is monomeric (as indicated by
size exclusion chromatography) and shows a CD spectrum identical to
that of E3_5 (ref. 16), which has an AR domain fold17(data not
shown).
Selection of specific high-affinity MAPK binders
To further evaluate the potential of our AR protein libraries, we
chose the eukaryotic protein kinases JNK2 and p38 as our next target
proteins (see ref. 24 and references therein). We did a total of four
ribosome-display selection rounds with the N2C library before com-
paring single, selected library members. Screening 15 clones each, we
obtained ten ELISA-positive JNK2 binders and ten ELISA-positive
p38 binders. The sequence of one representative member for each of
these target-specific groups is given in Supplementary Figure 1online.
These MAPK binders share many features of the selected MBP
binders.They have affinities in the low nM range (Table 1);they can be
expressed at high levels in soluble form in the cytoplasm of E. coli and
purified to homogeneity by a single IMAC purification step (data not
shown); their randomized positions are enriched in aromatic amino
acids (see Supplementary Fig. 1 online). To investigate their target
specificity, we did ELISA experiments with purified, selected AR
NATURE BIOTECHNOLOGY VOLUME 22 NUMBER 5 MAY 2004
577
Table 1 Kinetic binding data of selected clones determined by
surface plasmon resonance
Target Clone name (length)kon [M–1s–1]koff [s–1]KD [M]
MBPoff7 (N3C)
mbp3_5 (N3C)
mbp3_16 (N2C)
JNK2_2_3 (N2C)
p38_2_3 (N2C)
4.2·105
2.0·105
6.0·105
9.7·105
9.5·105
1.9·10–3
4.4·10–3
1.0·10–2
2.0·10–3
3.5·10–3
4.4·10–9
22·10–9
17·10–9
2.1·10–9
3.7·10–9
JNK2
p38
a
b
c
Figure 3 ELISAs with selected AR proteins. (a) Specificity of MBP binders.
The interaction of the proteins mbp3_16, off7 and mbp3_5 (each 50 nM;
control with no AR protein) with immobilized MBP, pD, APH and neutra-
vidin is shown. (b) Competition ELISA illustrating the interaction between
the selected AR protein off7 and MBP. off7 (5 nM) was incubated with
varying concentrations of free MBP before binding on immobilized MBP.
The binding to MBP of off7 can be specifically inhibited by increasing
concentrations of free MBP in solution. An unselected AR protein of the N3C
library (*) showed no interaction with MBP (100 nM of E3_5)16,17, giving a
signal identical to that of the control (no AR protein on immobilized MBP).
(c) Specificity comparison of an MBP, a JNK2 and a p38 binder. The
interaction of 100 nM each of the proteins off7 (binds MBP), JNK2_2_3
(binds JNK2), p38_2_3 (binds p38) and E3_5 (unselected N3C library
member) with MBP, JNK2, p38, APH, pD and BSA is shown. Note that in
all representations the background binding of the detection antibodies has
not been subtracted.
© 2004 Nature Publishing Group http://www.nature.com/naturebiotechnology

Page 4

ARTICLES
proteins (Fig. 3c). All binders were highly specific for the target pro-
teins that were used to select them.Most importantly,the AR proteins
allow perfect discrimination between the homologous MAPKs (51%
identity and 59% similarity between JNK2 and p38 on the amino acid
level in our format).
Structure determination of off7 in complex with MBP
To validate our AR randomization scheme and to analyze the selected
interaction at the atomic level, we determined the crystal structure of
one binder (off7) in complex with MBP (see Methods). The phasing
problem was solved by molecular replacement using E3_5,a designed
N3C AR protein17as a search model without using the phases of the
larger MBP. Hence,when used in cocrystallization studies,AR proteins
may serve as valuable tools to obtain first phases and finally the struc-
ture of its binding partner. The results of the data collection and ref-
inement are shown in Supplementary Table 1online. As can be seen in
the crystal structure (Figs. 4 and 5), the AR protein binds the open
form of MBP20creating an elongated complex. For the AR protein,
clear electron density starts at Ser12,but no or only very weak electron
density was observed for the N-terminal His6tag.The electron density
is clear until the second-to-last amino acid
(Leu168). In the complex, the AR protein has
the typical AR domain fold and is highly sim-
ilar to the known structure of the designed
AR protein E3_5 (root-mean-square devia-
tion of the Cαatoms (r.m.s.d.Cα) < 1 Å)17.
The main differences between the two struc-
tures are found in the β-turn region of repeat
module 1 (second repeat),where the Cαchain
shows a maximal r.m.s.d.Cαof 2.1 Å. The
H-bond network,which most probably stabi-
lizes the β-turn region17, is conserved and
undisturbed.off7 has three framework muta-
tions (K16R, N74D and H125Y) not present
in the library design,but they are conservative
and in accordance with the AR domain fold,
and only Y125 is involved in MBP binding
(see below).
MBP was found in the open conformation
with no ligand bound20. A superposition on
PDB entry 1LLS20shows very few differences
(r.m.s.d.Cα< 0.9 Å). The N-terminal His6
tag and the C terminus of MBP are not
defined in the electron density. Clear density
extends from Gly19 to Gly387.
Analysis of the interaction of off7 with
MBP
The interaction of the AR protein off7 with
MBP was analyzed as described in the
Methods section. The only direct interaction
between off7 and MBP in the crystal lattice is
the selected heterodimer interface (Fig. 4).
Crystal packing contacts are mainly between
adjacent off7 molecules and between adjacent
MBP molecules. The heterodimer interface is
formed by the concave randomized surface of
the AR protein off7 (611 Å2buried surface)
and a slightly larger convex surface on the
MBP (656 Å2buried surface), resulting in a
total buried surface area of 1,267 Å2(Fig. 5).
The off7/MBP complex is further characterized by six H-bonds and a
planarity index of 2.125. The details of the interaction are listed in
Table 2 and in Supplementary Table 2 online.
The interaction of off7 involves residues from all three randomized
repeat modules, although the randomized repeat modules 2 and 3
(constituting repeats 3 and 4 in the protein because of the capping
repeats) contribute more to the binding than the randomized repeat
module 1 does. In total, 9 out of 18 randomized potential interaction
residues are involved in the binding to MBP.The interface is character-
ized by a large number of aromatic residues,which account for 73% of
the buried surface area on off7 (Fig. 4 and Supplementary Table 2
online). Among these residues, the tyrosines cover 28% of the buried
surface area and are involved in four H-bonds (Tyr56, Tyr81, Tyr89
and Tyr125). In the interface (Fig. 4), tyrosines have a dual role being
both H-bond formers and hydrophobic contact mediators.Besides the
four tyrosines, Trp90 and Asp110 also form H-bonds. Three frame-
work residues (Leu86, Asp110 and Tyr125) form part of the interac-
tion surface.Two (Asp110 and Tyr125) form H-bonds to MBP (Fig.4)
and the third (Leu86) is engaged in hydrophobic interactions.
Interestingly, Tyr125 is a framework mutation (H125Y). With the
578
VOLUME 22 NUMBER 5 MAY 2004 NATURE BIOTECHNOLOGY
ad
b
c
Figure 4 Crystal structure of the designed AR protein off7 in complex with MBP. (a,b) Two perpendicu-
lar views of the complex are shown. MBP is on the left (blue), off7 on the right (ochre). The interaction
residues are highlighted in stick-mode in red (off7) and blue (MBP), respectively. (c) A close stereo
view on the H-bond pattern in similar view as in b. Note that in this representation only residues
involved in H-bonds (green dashed lines) are shown. For orientation, some residues involved in
H-bonding are labeled. Figure 4a–c were made with MolMol49. (d) Ligplot48representation of the
interaction between MBP (chain B, blue) and off7 (chain A, red). H-bonds (in green) including the
H-bond distances as well as residues and atoms involved in hydrophobic contacts (indicated by red
or blue rays) are shown.
© 2004 Nature Publishing Group http://www.nature.com/naturebiotechnology

Page 5

ARTICLES
exception of the high percentage of aromatic amino acids,the residue
composition of the off7 interface is comparable to other AR protein
complexes. However, the small and functionally biased set of only
seven structurally analyzed natural AR protein complexes limits the
comparison to their average amino acid composition in the binding
interface.
The interaction surface of MBP is located on helices H7, H9 and
H17 (Fig.4 and Supplementary Table 2 online).Four lysines (Lys151,
Lys154,Lys214 and Lys216) are involved in four H-bonds and together
form approximately 60% of the buried surface area on the MBP sur-
face (Fig. 4). Every lysine of MBP is in contact with a tyrosine of off7
resulting in three H-bonds and five hydrophobic contacts (see
Supplementary Table 2 online). In general, the surface of the MBP is
rather negatively charged except for the spot where off7 binds, where
there is a positively charged surface patch formed by the four lysines.
A comparison to natural protein-protein interactions
We compared the off7/MBP complex to all available complex crystal
structures of AR proteins, to natural heterodimer complexes, to anti-
body-antigen complexes and to an affibody complex (see Methods
section and Table 2). The off7/MBP interaction is comparable to
natural protein-protein interactions25.The 611Å2buried surface area,
one H-bond per 100 Å2buried surface area and the planarity value of
2.1 of off7, are all within the normal parameters of heterodimer
complexes (Table 2)25. The buried surface area is at the lower limit of
antibody-antigen complexes26, whereas the affinity is comparable to
high affinity monovalent antibody-antigen binding. Otherwise, all
values seem to fit the standard parameters for antibody-antigen com-
plexes quite well, but are also similar to those of heterodimer com-
plexes. Natural AR protein interactions can be very diverse in size,
composition and in the structural elements involved (Fig. 5 and
Table 2). Our randomization scheme of the β-turn and the first
α-helix of the designed AR modules was based on crystal structures of
natural AR complexes, where it was clear that the AR scaffold is
directly used for binding16. Because 50% of the randomized target
interaction residues that were selected do indeed interact with MBP,
the interaction mode is very similar to that of natural AR proteins,
such as the GABPβ1 (Fig. 5), which partly inspired our library
design16. In comparison to the natural AR proteins, off7 has a slightly
smaller buried surface area but a higher H-bond density (Table 2).
Recently the structure of an affibody, another designed binding
molecule,based on the staphylococcal three-helix bundle protein A,in
complex with its target protein was published27.The complex shows a
slightly larger buried surface area than the off7/MBP complex but with
a comparable number of H-bonds (Table 2). The affibody has a
thousand-fold lower affinity for its target28than the AR protein off7
does for MBP (KD= 6 µM vs.KD= 4.4 nM,respectively).This is pro-
bably because this affibody is in a molten globule state and assumes a
defined structure only upon binding29,leading to a loss of entropy that
reduces the overall free energy of binding and thus the observed affinity.
DISCUSSION
We designed AR protein libraries of varying repeat numbers using a
consensus design strategy15–17. Here we show the successful selection
of binding molecules from these libraries. The properties of the
designed AR proteins perfectly match the criteria for alternative scaf-
folds. They are expressed at a high level in soluble form, are
monomeric and do not contain any cysteines16.Unlike previously pre-
sented scaffolds (for reviews see refs.12,13),which typically use either
randomized loops or a randomized surface on a given protein scaffold
for binding, AR proteins use both β-turns and a randomized surface.
Most importantly, they are not restricted in dimension because of
their modular architecture (Fig. 1). Thus, the interaction surface can
be adapted by adding more repeat modules. The favorable properties
of the molecules in the starting libraries seem to positively influence
both the speed of selection and its outcome. In only four to five
ribosome-display selection rounds, we were able to enrich the pool in
specific,high-affinity protein binders from the N2C and N3C libraries,
which appears to be faster than ribosome-display selections from anti-
body single-chain Fv libraries30.
The selected AR proteins retain the advantageous properties of the
designed AR proteins, being expressed in high amounts in soluble
form and free of cysteines (Fig.2).The selected clones specifically rec-
ognize the target protein against which they were selected and do not
cross-react with other proteins as shown by ELISA (Fig.3).The affini-
ties of the selected clones are in the low nanomolar range (Table 1 and
Fig.2),the association rates are in the typical range for protein-protein
interactions (that is,105–106M–1s–1)31and the dissociation rates are in
the range of 10–2to 2·10–3s–1(Table 1). We anticipate that these off-
rates can probably be improved further by a diversification step fol-
lowed by an off-rate ribosome-display selection round32. Such a
diversification step could involve not only classical error-prone PCR,
but also, more interestingly, strategies purely amenable to repeat pro-
teins such as repeat shuffling or repeat elongation.
The selected AR protein sequences contain a high number of aro-
matic residues (see Supplementary Fig. 1 online) and in the case of
off7,seven aromatic amino acids are involved in MBP binding (Fig.4),
including four prominent tyrosines (Supplementary Table 2 online).
Interestingly, high tyrosine content has also been noted in antibody
binding sites26.The dual interaction role of tyrosine as H-bond former
and hydrophobic contact mediator (see Results)33is probably the
NATURE BIOTECHNOLOGY VOLUME 22 NUMBER 5 MAY 2004
579
a
b
c
Figure 5 Open sandwich illustrations of the interaction surfaces of AR
proteins and their targets. (a–c) GRASP50shape complementarity represent-
ations of the interactions between off7 and MBP (shape complementarity,
0.739) (a), GABPβ1 and GABPα (PDB entry 1AWC; shape complementarity,
0.665) (b), and p18INK4cand CDK6 (PDB entry 1G3N; shape complement-
arity, 0.688) (c), respectively. The complex is shown on the left with the AR
proteins in a backbone worm representation (α-helices in blue, β-turns in
green) and the target protein in a surface representation. The open sandwich
surface representations are shown in the middle (AR proteins) and on the
right (targets). The contact areas are stained according to the shape
complementarities from orange (low) to red (high).
© 2004 Nature Publishing Group http://www.nature.com/naturebiotechnology

Page 6

ARTICLES
reason for this accumulation.In addition,aromatic residues are gener-
ally enriched in protein-protein interaction interfaces25.
Apart from the high content of aromatic residues,the crystal struc-
ture of the off7/MBP complex reveals an interaction interface that is
comparable to that found in natural heterodimer and antibody-
antigen complexes (Table 2)25.It also shows that the AR protein binds
its target with the randomized amino acids (Fig.4and Supplementary
Table 2 online), hence validating our randomization scheme. In the
crystal structure of the complex, off7 shows the typical AR domain
fold with a high similarity to the unselected N3C library member E3_5
(ref. 17). In solution, the uncomplexed, monomeric off7 shows a CD
spectrum virtually identical to that of E3_5 (ref. 16). Thus, the back-
bone of off7 does not seem to be rearranged in a substantial way upon
binding to MBP but rather seems to interact in a key-to-lock mecha-
nism. Such a rigid-body interaction may be advantageous both for
affinity (low entropic costs upon binding) and specificity (conforma-
tional restriction).p18,a natural AR protein interacting with CDK4/6
(Fig. 5), also has a low r.m.s.d.Cαof < 0.9 Å between the complexed
(PDB entry 1G3N) and the uncomplexed (PDB entry 1IHB) AR pro-
tein. However, the interaction with CDK4/6 leads to structural alter-
ations in the target34. In contrast to those of the rigid AR domain
scaffolds,the binding site of antibodies seem to be able to adopt differ-
ent conformations. The loops of the complementarity-determining
region of typical protein binding antibodies can undergo substantial
changes upon binding35and the loop flexibility might even be used to
accomplish multispecificity36.
Here, we have successfully validated our AR proteins as designed
binding molecules using MBP and two MAPKs as model targets, and
thus have introduced a binding molecule with very favorable proper-
ties. The combination of the high expression level of designed AR
proteins (200 mg/l soluble protein in shake flasks),their high thermo-
dynamic stability (9.5 to 21 kcal/mol17), the absence of cysteines, the
fast enrichment of binders (four selection rounds),their low nanomo-
lar affinities along with high specificities and their modular architec-
ture compares favorably with reports on antibodies and other
alternative scaffolds (such as protein A28,lipocalins14,fibronectin37or
green fluorescent protein38). Our results open the door for a number
of applications. Apart from being useful in capturing molecules on
protein chips or in affinity purification—typical applications for
designed binding molecules—designed AR proteins are especially
suited for intracellular applications.Their high stability,the absence of
disulfide bonds and the selectable high affinities are optimal prerequi-
sites for intracellular inhibitors39,where antibodies are less than ideal.
Generally,AR proteins are also conceivable in therapeutic applications,
which are currently a domain of recombinant antibodies. As we have
shown here, it is possible to both cocrystallize a target protein with a
designed binding AR protein and determine its crystal structure with
the help of the AR protein.Hence,designed binding AR proteins could
be used in cocrystallization and structure determination of proteins dif-
ficult to crystallize,similar to what has been shown with antibodies40.
METHODS
Molecular biology. Unless stated otherwise,all experiments were done accord-
ing to protocols found in reference 41. Enzymes and buffers were from New
England Biolabs (NEB) or Fermentas. All PCR reactions were done using the
proofreading Vent-polymerase (NEB).
Vectors used in antigen production. The different vectors that were prepared
for the present study are described in detail in the Supplementary Methods
online. pQEMBP (GenBank accession no.AY327141) was used for the expres-
sion of His-tagged,nonbiotinylated MBP.pAT224 (AY327139) was used for the
expression of His-tagged, biotinylated MBP. pAT222 (AY327137) was used for
the production of His-tagged, biotinylated pD. pAT222_JNK2 and
pAT222_p38 were used for the production of His-tagged, biotinylated JNK2
and p38, respectively. All pAT222 and pAT224 constructs carry an Avi tag for
biotinylation at the N terminus and a His6tag at the C terminus. pBirAcm
(Avidity) was used for in vivo biotinylation.
Antigen production and purification. The biotinylated proteins pD, MBP,
JNK2 and p38 (plasmids pAT222, pAT224, pAT222_JNK2 and pAT222_p38)
were produced using in vivobiotinylation with plasmid pBirAcm in E.coliXL-1
Blue (Stratagene) according to the protocols of Avidity and QIAgen. Efficient
biotinylation was confirmed by ELISA and blotting with a streptavidin-alkaline
phosphatase conjugate (Roche) and mass spectrometry.Nonbiotinylated MBP
for the ELISA analysis and crystallization was produced in the same way as the
AR proteins16using pQEMBP in E.coliXL-1Blue.The protein purification was
carried out as described16.
Ribosome-display vector (pRDV; AY327136). The cloning of pRDV is
described in detail in the Supplementary Methods online. pRDV contains all
flanking DNA regions necessary for ribosome display: the T7-promoter, the
ribosomal binding site and an in-frametolA gene spacer.Hence,by simple liga-
tion of the DNA encoding the combinatorial library into pRDV and by a PCR
using this ligation mix as template, all features necessary for ribosome display
are added to the library.The use of pRDV has the advantage that it always pro-
vides error-free library flanking regions and that it saves a number of working
steps compared to the standard PCR approach for library generation21.
Generation of combinatorial libraries. The AR protein library generation has
been described16.We changed that protocol in this study in that all ARs,that is,
580
VOLUME 22 NUMBER 5 MAY 2004 NATURE BIOTECHNOLOGY
Table 2 Comparison between AR protein complexes and other protein-protein interactions
PDB entryResolution
∆ASAain Å2
No. H-bondsNo. H-bonds/100 Å2∆ASAa
No. of salt bridgesPlanarityNo. of bridging H2O
1awc
1bi7
1blx
1g3n
1ikn
1oy3
1ycsb
1svx (off7/MBP)
1lp1 (Affibody)
Antibody-antigen complexesc
Heterodimeric protein-protein
complexes c
2.2
3.4
1.9
2.9
2.3
2.1
2.2
2.3
2.3
–
–
853.62
1205.5
845.4
843.3
763.96
1593.9
713.7
611.2
848.9
777 ± 135
983 ± 582
5
7
0.58
0.29
1.30
1.42
0.26
0.44
1.00
1.00
0.71
1.1 ± 0.5
1.1 ± 0.5
0
4
1
1
n.d.
n.d.
1
0
0
–
–
2.30
2.50
2.50
2.20
2.73
4.01
2.60
2.10
2.00
2.2 ± 0.4
2.8 ± 0.9
3
0
11
12
12
0
02
7
7
6
6
–
–
10
1
0
2
–
–
aSurface area per molecule occluded upon complex formation. b1ycs uses a different binding surface than the other AR proteins in this table. cAccording to ref. 25.
© 2004 Nature Publishing Group http://www.nature.com/naturebiotechnology

Page 7

ARTICLES
the N-terminal capping AR, the designed repeat module and the C-terminal
capping AR were used as PCR products for the assembly of the libraries.In the
present study,both the N- and the C-terminal capping repeat were amplified by
PCR from cloned and verified sequences to reduce sequence errors.In this way,
AR protein libraries consisting of an N-terminal capping AR, two or three
designed AR modules and a C-terminal capping AR (N2C and N3C libraries)
were assembled. To convert the libraries to the ribosome-display format, they
were amplified by PCR using oligonucleotides EWT4 (5′-TTCCTCATGAGAG
GATCGCATCACCATCACCATCACGGATCCGACCTGGG-3′) and WTC4 (5′-
TTTGGGAAGCTTTTGCAGGATTTCAGC-3′) and ligated into pRDV using
the restriction enzymes BspHI (or NcoI for pRDV) and HindIII. The ligation
product was purified using QIAquick (QIAgen) columns.The purified ligation
served as template for a PCR using oligonucleotides RDVf1 (5′-CCTTTTGCT
CACATGACCCG-3′) and tolAk (5′-CCGCACACCAGTAAGGTGTGCGGTT
TCAGTTGCCGCTTTCTTTCT-3′). Thereby, combinatorial N2C and N3C
DNA libraries were generated in the ribosome-display format.
Ribosome display.The PCR-amplified libraries were transcribed and selections
were done as described21.For the selection,the biotinylated antigen was immo-
bilized as follows: neutravidin (66 nM,100 µl/well; Pierce) in TBS150 (50 mM
Tris HCl,pH 7.4,150 mM NaCl) was immobilized on a Maxisorp plate (Nunc)
by overnight incubation at 4 °C.The wells were then blocked with 300 µl 0.5%
BSA (Fluka) in TBS150 for 1 h at 23 °C.Biotinylated antigen (100 µl,1 µM) in
TBS150 with 0.5% BSA was allowed to bind for 1 h at 4 °C. Before the ribo-
some-display round, the wells were extensively washed with washing buffer
WBT (50mM Tris acetic acid,pH7.5,150mM NaCl,50mM Mg(CH3COO–)2,
0.05% Tween 20). A ribosome-display round consisted of two 30-min pre-
panning steps on neutravidin and a 1h binding step on the target protein.After
washing, RNA purification and reverse transcription (with oligonucleotide
tolAk), a first PCR was done using oligonucleotides T7B (5′-ATACGAAAT
TAATACGACTCACTATAGGGAGACCACAACGG-3′) and tolAk. This RT-
PCR product was purified on an agarose gel and reamplified in a second PCR
using the same oligonucleotides. The second PCR product served as template
for the next round of ribosome display. The number of RT-PCR cycles was
reduced from 40 to 30 to 25 in the first three rounds to monitor the enrichment
of binders.Binders were analyzed after four or five rounds.
Analysis of selected binders. From the selected DNA pools,the AR open read-
ing frame was amplified by PCR and cloned into pQE30 (QIAgen) via
BamHI/HindIII (oligonucleotides:
CTGGG-3′ and WTC4). The DNA sequences were determined using standard
techniques. The sequences of the MBP binding proteins off7 (AY326424),
MBP3_5 (AY326425) and MBP3_16 (AY326426) have been deposited in
GenBank. The amino acid sequences of all sequenced clones are listed in
Supplementary Figure 1 online. For the ELISA screening, the crude extract of
0.6 ml protein expression cultures was used (expression according to QIAgen).
The cell pellets were lysed with 50 µl B-Per (Pierce) and the lysates were mixed
with 250µl TBS500 (50mM Tris HCl,pH8.0,500mM NaCl) each.For quanti-
tative ELISA, BIAcore, CD, analytical gel-filtration and crystallization, single,
selected library members were produced on a liter scale and purified as
described16. CD spectroscopy and analytical gel-filtration were done as
described16using 15 µM protein in TBS150 (pH 7.4).
EWT3:5′-TTCCGCGGATCCGAC-
ELISA.Biotinylated antigens were immobilized on neutravidin-coated plates as
described above. For the screening of the pools, 100 µl of the above crude
extracts were applied to wells with or without immobilized antigen for 1 h at
4 °C.After extensive washing with TBS150,binding was detected with an anti-
RGS-His antibody (QIAgen; detects only the RGS-His6-tag of the AR protein,
not the His6-tag of the antigen),an anti-mouse-IgG-alkaline phosphatase con-
jugate (Pierce) and p-nitrophenylphosphate (Fluka).Quantitative ELISAs were
done in the same manner, except purified protein was used (see Fig. 3). For
competition ELISA, the purified AR protein off7 was incubated with varying
amounts of free MBP before (4 °C, 100 min) and during the binding reaction
(see Fig.3).
Surface plasmon resonance (SPR). SPR was measured using a BIAcore 3000
instrument (BIAcore). The running buffer was 20 mM HEPES, pH 7.4,
150 mM NaCl and 0.005% Tween 20. A streptavidin SA chip (BIAcore) was
used with 480 RU biotinylated MBP immobilized (440 RU JNK2 and 450 RU
p38, respectively). The interactions were measured at a flow of 60 µl/min with
5 min buffer flow,2 min injection of MBP-binding AR protein in varying con-
centrations (10 pM to 200 nM) and an off-rate measurement of 40 min with
buffer flow. The signal of an uncoated reference cell was subtracted from the
measurements.Inhibition BIAcore measurements gave results similar to that of
the kinetic analyses (data not shown). The p38 and JNK2 binders were meas-
ured similarly,but with an injection time of 3 min.The kinetic data of the inter-
action were evaluated with a global fit using BIAevaluation 3.0 (BIAcore),
Scrubber (BioLogic software) and Clamp42.
Complex purification and crystallization. MBP and the selected MBP-binding
AR proteins off7, mbp3_5 and mbp3_16 were produced as described above.
The cell pellets of 1-liter bacterial culture of each MBP and off7 (or mbp3_5 or
mbp3_16) were pooled and then lysed using an Emulsiflex C5 (Avestin) fol-
lowed by additional sonication.The proteins were purified using an IMAC col-
umn as described16, followed by a preparative Superdex-75 (Amersham
Pharmacia) size exclusion chromatography step in 10mM Tris HCl,pH7.6 and
100 mM NaCl. For every protein mixture, the peak fraction with the smallest
molecular weight containing both MBP and the AR protein in equimolar
amounts, as determined by SDS-PAGE, was collected and used for crystalliza-
tion.Light scattering of these fractions was measured as described17.It showed
a monodisperse particle distribution for all three complexes.For the off7/MBP
complex (61.4 kDa calculated mass for the 1:1 complex) an average radius of
3.6nm,equivalent to a hydrated particle of 95kDa,was estimated,which corre-
sponds to a nonhydrated particle of 65 kDa. The off7/MBP complex crystal-
lized readily and was further analyzed. The protein complex was concentrated
to 26 mg/ml for crystallization. Initial crystallization screening was done in
96-well, sitting drop, square well crystallization plates (Greiner Bio-One). The
reservoirs were filled with 100 µl reservoir solution using an 8-channel pipette
from a 2 ml 96-well, deep-well block into the crystallization plate. Using an
8-channel pipette,2 µl of reservoir solution were pipetted in the crystallization
well and mixed with 2 µl of protein solution. The initial crystals were refined
using standard techniques. The crystals used for data collection grew in about
2–3 weeks in 30% PEG 6000,0.1 M Tris HCl pH 8–9,100 mM NaCl in a hang-
ing drop experiment with 500µl reservoir,2µl protein solution mixed with 2µl
water and 2µl reservoir solution.For data collection the crystals were soaked in
the mother liquid with 10% ethylene glycol for about 30 s to 1 min and flash
frozen in a cryostream at 100 K.
Data collection, reduction, structure solution and refinement. Data were col-
lected at the European Synchrotron Radiation Facility beamline ID14-1. The
data were processed using MOSFLM, SCALA and TRUNCATE43. The crystal
belonged to space group P21, with a Matthews coefficient of VM = 2.1 Å3/Da,
corresponding to an estimated water content of 39%.
The crystal structure was determined by molecular replacement using the
program AmoRe44, with the structure of the unselected N3C library member
E3_5 (PDB entry 1MJ017) as a search model. A conventional AmoRe protocol
(rotation, translation, rigid body refinement) was applied yielding a solution.
This information was used to obtain a first electron density.At this point only
the AR protein was clearly visible in the electron density.A solvent flipping pro-
tocol was then applied to modify the map45.MBP in its open form (PDB entry
1LLS)20was positioned in the resulting electron density using program O46.
Because about a third of MBP was visible, manual building would have been
feasible as well. The rest of the model building was carried out using the pro-
gram O46, the structure refinement was done in CNS45resulting in a final
model with an R-factor of 19.5% and an Rfree-factor of 24.9% (Supplementary
Table 1 online).
Analysis ofthe complexes.The structural analysis of the complexes was done as
suggested25. H-bonds and hydrophobic interactions were calculated with
HBPLUS47, LIGPLOT48and DIMPLOT48using the default settings. Other
parameters were calculated using the protein-protein interaction server
http://www.biochem.ucl.ac.uk/bsm/PP/server/index.html25or CNS45.
The atomic coordinates of the described complex were deposited in the PDB
(PDB-ID:1SVX).
Note: Supplementary information is available on the Nature Biotechnology website.
NATURE BIOTECHNOLOGY VOLUME 22 NUMBER 5 MAY 2004
581
© 2004 Nature Publishing Group http://www.nature.com/naturebiotechnology

Page 8

ARTICLES
ACKNOWLEDGMENTS
We thank the members of the Plückthun and Grütter laboratories for valuable
discussions and David L.Zechel for the critical reading of the manuscript.H.K.B.
was supported by a pre-doctoral fellowship of the Roche Research Foundation.
M.T.S.was in receipt of a pre-doctoral scholarship from the Fonds der Chemischen
Industrie and the Bundesministerium für Bildung und Forschung.This work was
supported by the Swiss National Center of Competence in Research in structural
biology and the Swiss Krebsliga.
COMPETING INTERESTS STATEMENT
The authors declare competing financial interests (see the Nature Biotechnology
website for details).
Received 24 December 2003; accepted 14 January 2004
Published online at http://www.nature.com/naturebiotechnology/
1. Groves, M.R. & Barford, D. Topological characteristics of helical repeat proteins. Curr.
Opin. Struct. Biol. 9, 383–389 (1999).
2. Andrade, M.A., Perez-Iratxeta, C. & Ponting, C.P. Protein repeats: structures, func-
tions, and evolution. J. Struct. Biol. 134, 117–131 (2001).
3. Kobe, B. & Kajava, A.V. When protein folding is simplified to protein coiling: the con-
tinuum of solenoid protein structures. Trends Biochem. Sci. 25, 509–515 (2000).
4. Marcotte, E.M., Pellegrini, M., Yeates, T.O. & Eisenberg, D. A census of protein
repeats. J. Mol. Biol. 293, 151–160 (1999).
5. Ellis, J., Dodds, P. & Pryor, T. Structure, function and evolution of plant disease
resistance genes. Curr. Opin. Plant Biol. 3, 278–284 (2000).
6. Bork, P. Hundreds of ankyrin-like repeats in functionally diverse proteins: mobile
modules that cross phyla horizontally? Proteins: Struct. Funct. Genet. 17, 363–374
(1993).
7. Sedgwick, S.G. & Smerdon, S.J. The ankyrin repeat: a diversity of interactions on a
common structural framework. Trends Biochem. Sci. 24, 311–316 (1999).
8. Suzuki, F. et al. Functional interactions of transcription factor human GA-binding pro-
tein subunits. J. Biol. Chem. 273, 29302–29308 (1998).
9. Malek, S., Huxford, T. & Ghosh, G. IκBα functions through direct contacts with the
nuclear localization signals and the DNA binding sequences of NF-κB. J. Biol. Chem.
273, 25427–25435 (1998).
10. Winter, G. & Milstein, C. Man-made antibodies. Nature 349, 293–299 (1991).
11. Plückthun, A. et al. in Antibody Engineering (eds. McCafferty, J., Hoogenboom, H.J.
& Chiswell, D.J.) 203–252 (IRL Press, Oxford, 1996).
12. Nygren, P.-Å. & Uhlén, M. Scaffolds for engineering novel binding sites in proteins.
Curr. Opin. Struct. Biol. 7, 463–469 (1997).
13. Skerra, A. Engineered protein scaffolds for molecular recognition. J. Mol. Recognit.
13, 167–187 (2000).
14. Skerra, A. Imitating the humoral immune response. Curr. Opin. Chem. Biol. 7,
683–693 (2003).
15. Forrer, P., Stumpp, M.T., Binz, H.K. & Plückthun, A. A novel strategy to design bind-
ing molecules harnessing the modular nature of repeat proteins. FEBS Lett. 539, 2–6
(2003).
16. Binz, H.K., Stumpp, M.T., Forrer, P., Amstutz, P. & Plückthun, A. Designing repeat
proteins: well-expressed, soluble and stable proteins from combinatorial libraries of
consensus ankyrin repeat proteins. J. Mol. Biol. 332, 489–503 (2003).
17. Kohl, A. et al. Designed to be stable: crystal structure of a consensus ankyrin repeat
protein. Proc. Natl. Acad. Sci. USA 100, 1700–1705 (2003).
18. Mosavi, L.K., Minor, D.L. Jr. & Peng, Z.-y. Consensus-derived structural determinants
of the ankyrin repeat motif. Proc. Natl. Acad. Sci. USA 99, 16029–16034 (2002).
19. Main, E.R., Xiong, Y., Cocco, M.J., D’Andrea, L. & Regan, L. Design of stable alpha-
helical arrays from an idealized TPR motif. Structure (Camb) 11, 497–508 (2003).
20. Rubin, S.M., Lee, S.Y., Ruiz, E.J., Pines, A. & Wemmer, D.E. Detection and charac-
terization of xenon-binding sites in proteins by 129Xe NMR spectroscopy. J. Mol. Biol.
322, 425–440 (2002).
21. Hanes, J. & Plückthun, A. In vitro selection and evolution of functional proteins by
using ribosome display. Proc. Natl. Acad. Sci. USA 94, 4937–4942 (1997).
22. Yang, F. et al. Novel fold and capsid-binding properties of the lambda-phage display
platform protein gpD. Nat. Struct. Biol. 7, 230–237 (2000).
23. Hon, W.C. et al. Structure of an enzyme required for aminoglycoside antibiotic resist-
ance reveals homology to eukaryotic protein kinases. Cell 89, 887–895 (1997).
24. Forrer, P., Tamaskovic, R. & Jaussi, R. Enzyme-linked immunosorbent assay for the
measurement of JNK, ERK and p38 kinase activities. Biol. Chem. 379, 1101–1111
(1998).
25. Jones, S. & Thornton, J.M. Principles of protein-protein interactions. Proc. Natl.
Acad. Sci. USA 93, 13–20 (1996).
26. Lo Conte, L., Chothia, C. & Janin, J. The atomic structure of protein-protein recogni-
tion sites. J. Mol. Biol. 285, 2177–2198 (1999).
27. Högbom, M., Eklund, M., Nygren, P.-Å. & Nordlund, P. Structural basis for recogni-
tion by an in vitro evolved affibody. Proc. Natl. Acad. Sci. USA 100, 3191–3196
(2003).
28. Eklund, M., Axelsson, L., Uhlén, M. & Nygren, P.-Å. Anti-idiotypic protein domains
selected from protein A-based affibody libraries. Proteins: Struct. Funct. Genet. 48,
454–462 (2002).
29. Wahlberg, E. et al. An affibody in complex with a target protein: structure and cou-
pled folding. Proc. Natl. Acad. Sci. USA 100, 3185–3190 (2003).
30. Hanes, J., Schaffitzel, C., Knappik, A. & Plückthun, A. Picomolar affinity antibodies
from a fully synthetic naive library selected and evolved by ribosome display. Nat.
Biotechnol. 18, 1287–1292 (2000).
31. Wodak, S.J. & Janin, J. Structural basis of macromolecular recognition. Adv. Protein
Chem. 61, 9–73 (2002).
32. Jermutus, L., Honegger, A., Schwesinger, F., Hanes, J. & Plückthun, A. Tailoring in
vitro evolution for protein affinity or stability. Proc. Natl. Acad. Sci. USA 98, 75–80
(2001).
33. Mian, I.S., Bradwell, A.R. & Olson, A.J. Structure, function and properties of antibody
binding sites. J. Mol. Biol. 217, 133–151 (1991).
34. Jeffrey, P.D., Tong, L. & Pavletich, N.P. Structural basis of inhibition of CDK-cyclin
complexes by INK4 inhibitors. Genes Dev. 14, 3115–3125 (2000).
35. Sundberg, E.J. & Mariuzza, R.A. Molecular recognition in antibody-antigen com-
plexes. Adv. Protein Chem. 61, 119–160 (2002).
36. James, L.C., Roversi, P. & Tawfik, D.S. Antibody multispecificity mediated by confor-
mational diversity. Science 299, 1362–1367 (2003).
37. Xu, L. et al. Directed evolution of high-affinity antibody mimics using mRNA display.
Chem. Biol. 9, 933–942 (2002).
38. Zeytun, A., Jeromin, A., Scalettar, B.A., Waldo, G.S. & Bradbury, A.R. Fluorobodies
combine GFP fluorescence with the binding characteristics of antibodies. Nat.
Biotechnol. 21, 1473–1479 (2003).
39. Cattaneo, A. & Biocca, S. The selection of intracellular antibodies. Trends Biotechnol.
17, 115–121 (1999).
40. Ostermeier, C., Iwata, S., Ludwig, B. & Michel, H. Fv fragment-mediated crystalliza-
tion of the membrane protein bacterial cytochrome c oxidase. Nat. Struct. Biol. 2,
842–846 (1995).
41. Sambrook, J., Fritsch, E.F. & Maniatis, T. Molecular cloning: A laboratory manual,
edn. 2 (Cold Spring Harbor Press, New York, 1989).
42. Myszka, D.G. & Morton, T.A. CLAMP: a biosensor kinetic data analysis program.
Trends Biochem. Sci. 23, 149–150 (1998).
43. Collaborative Computational Project, Number 4. The CCP4 Suite: Programs for
Protein Crystallography. Acta Crystallogr. D 50, 760–763 (1994).
44. Navaza, J. Acta Crystallogr. A 50, 157–163 (1994).
45. Brünger, A.T. et al. Crystallography & NMR system: A new software suite for macro-
molecular structure determination. Acta Crystallogr. D 54, 905–921 (1998).
46. Jones, T.A., Zou, J.Y., Cowan, S.W. & Kjeldgaard, M. Improved methods for building
protein models in electron density maps and the location of errors in these models.
Acta Crystallogr. A 47, 110–119 (1991).
47. McDonald, I.K. & Thornton, J.M. Satisfying hydrogen bonding potential in proteins.
J. Mol. Biol. 238, 777–793 (1994).
48. Wallace, A.C., Laskowski, R.A. & Thornton, J.M. LIGPLOT: a program to generate
schematic diagrams of protein-ligand interactions. Protein Eng. 8, 127–134 (1995).
49. Koradi, R., Billeter, M. & Wüthrich, K. MOLMOL: a program for display and analysis
of macromolecular structures. J. Mol. Graph. 14, 51–55 (1996).
50. Nicholls, A., Sharp, K.A. & Honig, B. Protein folding and association: insights from
the interfacial and thermodynamic properties of hydrocarbons. Proteins: Struct.
Funct. Genet. 11, 281–296 (1991).
582
VOLUME 22 NUMBER 5 MAY 2004 NATURE BIOTECHNOLOGY
© 2004 Nature Publishing Group http://www.nature.com/naturebiotechnology