Turning a protein kinase on or off from a singleallosteric site via disulfide trappingJack D. Sadowsky, Mark A. Burlingame, Dennis W. Wolan, Christopher L. McClendon,Matthew P. Jacobson, and James A. Wells1

Departments of Pharmaceutical Chemistry and Cellular and Molecular Pharmacology, University of California, 1700 Fourth Street,San Francisco, CA 94158

Contributed by James A. Wells, February 10, 2011 (sent for review September 28, 2010)

There is significant interest in identifying and characterizingallosteric sites in enzymes such as protein kinases both forunderstanding allosteric mechanisms as well as for drug discovery.Here, we apply a site-directed technology, disulfide trapping, tointerrogate structurally and functionally how an allosteric siteon the Ser/Thr kinase, 3-phosphoinositide-dependent kinase 1(PDK1)—the PDK1-interacting-fragment (PIF) pocket—is engagedby an activating peptide motif on downstream substrate kinases(PIFtides) and by small molecule fragments. Bymonitoring pairwisedisulfide conjugation between PIFtide and PDK1 cysteine mutants,we defined the PIFtide binding orientation in the PIF pocket ofPDK1 and assessed subtle relationships between PIFtide position-ing and kinase activation. We also discovered a variety of smallmolecule fragment disulfides (<300 Da) that could either activateor inhibit PDK1 by conjugation to the PIF pocket, thus displayinggreater functional diversity than is displayed by PIFtides conju-gated to the same sites. Biochemical data and three crystal struc-tures provided insight into the mechanism of action of the bestfragment activators and inhibitors. These studies show that disul-fide trapping is useful for characterizing allosteric sites on kinasesand that a single allosteric site on a protein kinase can beexploited for both activation and inhibition by small molecules.

fragment-based drug discovery ∣ kinase allostery

Allosteric sites on protein kinases are difficult to find, charac-terize, and target with small molecules. Whereas the devel-

opment of active site inhibitors has benefited greatly from theexistence of a natural small molecule substrate, ATP, there arevirtually no natural small molecule starting points from whichallosteric kinase modulators can be built. Thus, most allostericmodulators have been discovered serendipitously, generallythrough high-throughput screening (HTS), which is inherentlyinefficient. Confirming the binding sites and mechanisms ofaction of allosteric modulators also tends to be more laboriousrelative to ATP-mimetic inhibitors. Prior to initiating HTScampaigns, it would be beneficial to have a simple method thatpermits directed interrogation of kinase allosteric sites withsmall compounds and facile assessment of the effects of bindingon protein structure and function.

We envisioned that a site-directed approach developed pre-viously by our laboratory, disulfide trapping (or tethering), wouldallow more straightforward characterization and interrogation ofallosteric sites on protein kinases with small molecules. Disulfidetrapping involves screening disulfide-containing compounds fortheir ability to form amixed disulfide with a natural or engineeredcysteine residue near a site of interest on a protein target (1, 2).Because the approach is site-directed, disulfide trapping has pro-ven valuable as a tool to validate known, suspected, and orphanallosteric sites on protein surfaces as small molecule targets(3–6), but has not yet been applied to kinase allosteric sites.

Here, we use disulfide trapping to target an allosteric site onthe surface of 3-phosphoinositide-dependent kinase 1 (PDK1)—the so-called PDK1-interacting fragment (PIF) pocket. PDK1 is amember of the AGC family of Ser/Thr kinases, which includes

protein kinase A (PKA), B (PKB/Akt), and C (PKC) isozymes.In PKA, PKB, and PKC the site analogous to the PIF pocketbinds to a specific hydrophobic motif (HM) at the C terminusof these kinases (7). In PDK1, which lacks its own HM segment,the PIF pocket serves a dual purpose: It recruits downstream sub-strate kinases that possess an HM segment (e.g., PKC, S6K, RSK,and others) and, upon HM binding, allosterically stimulatesPDK1 activity (8–11). Short peptides (12–25 residues) derivedfrom the HM of several kinases (PIFtides) can activate PDK1by about four- to sevenfold, although their mode of interactionwith PDK1 has not yet been established (8, 9). Recently, smallmolecules that mimic PIFtides and similarly activate PDK1 bybinding the PIF pocket have been discovered and characterizedstructurally and biochemically (12).

We used disulfide trapping to (i) better elucidate how a PIF-tide binds to and activates PDK1, (ii) identify new small moleculefragments targeting the PIF pocket, and (iii) begin to characterizetheir allosteric mechanism. Most remarkably, some of the disul-fide-trapped fragments activate and others inhibit the kinase.Biochemical and X-ray structural studies provided insight intothe binding modes of allosteric modulators and protein confor-mational changes that may propagate their allosteric effects.These studies provide a framework for systematically assessingfunctional and structural effects of small molecules directed atspecific allosteric sites on protein kinases.

ResultsTrapping of Peptides at the PIF Pocket on PDK1. Our first goal wasto use disulfide trapping to better characterize the binding of PIF-tides to PDK1. Because no crystal structures of PIFtide/PDK1complexes have been reported, we have generated an energy-minimized homology model based upon the known structureof a chimeric Akt construct, in which the endogenous C-terminalHM was replaced by the HM of the kinase PRK2 (13) (Fig. S1A).

To test our model, we prepared a panel of recombinant PDK1Cys mutants in which residues at six positions around the PIFpocket were mutated individually to a cysteine (K115C, I119C,V124C, R131C, T148C, and Q150C). All of these Cys mutantsdisplayed comparable catalytic activity, but were somewhat lessactive than the wild-type enzyme (by approximately 20–50%;Fig. S2). We next synthesized a panel of 11 PIFtide variantsderived from the HM of PRK2 in which a disulfide-cappedcysteine replaced residues along the sequence (Fig. 1A). Weallowed each of the six Cys-mutant kinases (at 1 μM) to form a

Author contributions: J.D.S., D.W.W., C.L.M., M.P.J., and J.A.W. designed research; J.D.S.and C.L.M. performed research; J.D.S. and M.A.B. contributed new reagents/analytictools; J.D.S., D.W.W., and C.L.M. analyzed data; and J.D.S. and J.A.W. wrote the paper.

The authors declare no conflict of interest.

Data deposition: Structural coordinates of PDK1 T148C bound to allosteric modulatorshave been deposited in the Protein Data Bank, www.pdb.org (PDB ID codes 3ORZ,3OTU, and 3ORX).1To whom correspondence should be addressed. E-mail: jim.wells@ucsf.edu.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1102376108/-/DCSupplemental.

6056–6061 ∣ PNAS ∣ April 12, 2011 ∣ vol. 108 ∣ no. 15 www.pnas.org/cgi/doi/10.1073/pnas.1102376108

disulfide adduct with each of the 11 Cys-PIFtides (at 100 μM).The concentration of β-mercaptoethanol (β-ME) was varied(100 μM or 1 mM) to adjust for differences in intrinsic reactivitytoward disulfides of the introduced cysteines on PDK1, as gaugedby promiscuity in binding to the library of Cys-PIFtides. The levelof conjugation was determined by mass spectrometry.

The Cys-PIFtides showed pronounced selectivity among thesix PDK1 Cys mutants (Fig. 1B). Furthermore, the extent ofconjugation tracked well with the predicted proximity betweenthe cysteines on the protein and the peptide according to ourcalculated model for the PIFtide/PDK1 complex. For example,PDK1K115C conjugated extensively to the Cys at position 3 inPIFtideQ3C, and the extent of conjugation to other PIFtidestapered off with distance from residue 115 in PDK1. Similareffects were seen with conjugation to Cys at four other positions:119, 124, 131, and 148. One of the PDK1 mutants, PDK1Q150C,was significantly more promiscuous in its reactivity althoughthis mutant did show strongest conjugation to the residue on thePIFtide (D10C) nearest to it in our model.

We tested if the conjugation data could be used to furtherrefine our initial homology model. We performed a conforma-tional search using a conservative restraint of 7.0 Å on theCβ–Cβ distances between PIFtide residues and residues in PDK1that showed the greatest conjugation in disulfide-trapping experi-ments. We followed up with unrestrained molecular dynamics(MD) simulations with explicit aqueous solvent. The PIFtideremained bound to the pocket throughout the simulations anddid not deviate greatly in conformation relative to the initial con-strained model. By contrast, models derived from MD simula-tions on an initially unrestrained complex showed significantlygreater heterogeneity and sampled conformations very differentfrom that of the starting structure (Fig. S1 and Movie S1).

We next assessed the effects of disulfide-trapped PIFtides onthe catalytic activity of their associated PDK1 Cys mutants(Fig. 2). These assays were performed with peptides and PDK1mutants for which we could achieve stoichiometric conjugationat the lowest β-ME concentration tested (100 μM). Kinase activ-ities (initial rates) were measured with a standard assay using

a peptide substrate taken from the activation loop of Akt andradiolabeled ATP.

All of the PDK1 Cys mutants tested had at least one Cys-PIFtide that when conjugated increased kinase activity relativeto unconjugated PDK1 (Fig. 2A). The catalytic activity of themost activated complex, PIFtideD10C∕PDK1Q150C (5.2-fold acti-vation), even exceeded that which we measured for the noncova-lent complex between the parent PIFtide and wild-type PDK1(3.9-fold activation at saturation). Interestingly, although severalPIFtide disulfides could be conjugated stoichiometrically toPDK1Q150C (Fig. 1B), PIFtideD10C was clearly more effective thanthe others at activating this kinase (Fig. 2B). The average errors inactivation or inhibition potency did not exceed 15% of the valuesshown (see Experimental Procedures).

Trapping of Small Molecules at the PIF Pocket. We next evaluatedwhether small molecule disulfides would mimic effects seenfor the larger Cys-PIFtides (molecular mass of approximately2,000 Da). Thus, we screened each of the six PDK1 Cys mutantsagainst a library of 480 disulfide-bearing fragments (molecularmass 200–500) that was synthesized in the Small MoleculeDiscovery Center at University of California, San Francisco.These compounds represent a diversity of scaffolds common indrugs and drug candidates (e.g., heterocycles, piperazines, etc.)and most obey the “Rule of 3,” which was established to judgefragments as suitable starting points for drug discovery efforts(mass <300 Da, predicted cLogP ≤3, number of hydrogen-bonddonors/acceptors ≤3) (14). Preliminary experiments revealedthat the disulfide fragments reacted much more readily withthe PDK1 Cys mutants than did the Cys-PIFtides; at 100 μMβ-ME and 100 μM compound, most small molecule disulfidesreacted quantitatively with every Cys mutant. Therefore, inscreening disulfide fragments for binding, we increased the assaystringency by raising the concentration of β-ME to 10 mM.

Under our assay conditions, the hit rate for disulfide fragmentbinding varied from 1.3% to 13.8% for the six PDK1 Cys mutants(using 20% conjugation to define a hit). The least hit-rich PDK1mutant, PDK1I119C, trapped only six compounds, whereas themost hit-rich mutant, PDK1K115C, trapped 66 compounds. Inter-estingly, the mutants that were the most promiscuous in bindingto small molecules (K115C, R131C, and T148C) were the mostselective in binding to Cys-PIFtides (Fig. 1B). About 3% of thedisulfide fragments in the library (15 out of 480) were trappedto >50% conjugation to one and only one of the six PDK1 Cysmutants. The wild-type PDK1 catalytic domain has four endogen-ous cysteines, which were retained in our mutant constructs.Nevertheless, none of the fragment disulfide hits conjugated towild-type PDK1, indicating specific binding to the Cys incorpo-rated at the PIF pocket.

Whereas Cys-PIFtides activated PDK1 mutants, the disulfidefragment hits comprised both activators and inhibitors, as wellas several inert ligands that neither activated nor inhibited theenzyme (Fig. 3 A and B and Fig. S3). Every PDK1 Cys mutantproduced all three outcomes that varied depending on the frag-

R131C

1

11

I119C

1

11

V124C

1

11

Q150C1

11

% Conjugation: 0% 100%

PIFtide 15-mer : EEQEMFRDFDYIADW 511

K115C

1

11

T148C1

11

AB

Fig. 1. Trapping of Cys-PIFtides to PIF pocket on PDK1. (A) Sequence ofparent PIFtide 15-mer and Cys-scanned region. (B) Conjugation strength ofeach PDK1 Cys mutant to each Cys position in Cys-PIFtides, presented as aheat map on a homology model of the PIFtide/PDK1 complex.

I119C

V124C R131C

K115C

Q150C

T148C1.2

1.6

3.3 1.3

1.20.8 2.0

5.2

E

EM

F RD

E

Q D

F Y

AQ150C

E

EM

F RD

E

Q D

F Y

B FoldActivity

1.0

5.2

Fig. 2. Modulation of PDK1 mutants by disulfide-bound PIFtides. (A) Effectsof nearest-neighbor Cys–Cys cross-links on PDK1 activity. Values shown arefold catalytic activities (relative to a DMSO-treated control) of the indicatedPDK1 mutant conjugated to various Cys-PIFtides. (B) Effects of nearby anddistant cross-links on activity of PDK1Q150C.

Sadowsky et al. PNAS ∣ April 12, 2011 ∣ vol. 108 ∣ no. 15 ∣ 6057

BIOCH

EMISTR

Y

ment bound. Some mutants (e.g., K115C and Q150C) seemed totrap inhibitors or activators preferentially, although the samplesize underlying these trends is small. The mutants showing thegreatest range in activity on stoichiometrically trapping smallmolecule disulfides were PDK1R131C and PDK1T148C where activ-ity varied about 14-fold from the most inhibited to the mostactivated complex. The effects of compound concentration onpercent conjugation (Fig. 3C) and kinase activity (Fig. 3D) werecompared for the most activating and inhibiting fragmentstargeting PDK1T148C—2A2 and 1F8, respectively. The concentra-tions required for 50% conjugation to this mutant (DR50 values)were 5.6 μM for 2A2 and 4.6 μM for 1F8 (at 1 mM β-ME). Thesevalues tracked closely the concentrations required for half-max-imal effect on catalytic activity (EC50 values)—7.1 and 7.2 μMfor 2A2 and 1F8, respectively. None of the most activating orinhibiting disulfide fragments had a significant impact on theactivity of wild-type PDK1 and all were approximately equallyeffective toward mutants in the absence and presence of thedetergent CHAPS, which is known to disperse small moleculeaggregates that can lead to spurious enzyme activation or inhibi-tion (15–17) (Fig. 3B and Fig. S4). Overall, these results indicatethat specific disulfide trapping of compounds at the PIF pocketis critical for the functional effects observed.

Optimizing a Disulfide Fragment Activator. To assess relationshipsbetween structure and activity of disulfide activators, we synthe-sized and tested 30 analogs of 2A2 in which the N-chlorophe-nylpiperazine moiety was replaced with various N-substitutedpiperazines. Compounds were then assayed for binding to andactivation of PDK1T148C (Table S1). Only N-phenyl and N-benzylpiperazines conjugated detectably to this mutant, whereas similarN-acyl and N-benzoyl derivatives did not (at 100 μM compound,10 mM β-ME). Substitution of the aryl ring within these deriva-tives significantly affected both conjugation strength and activa-tion potency and these properties did not seem to be correlated.The best compound of the series (“JS30”; Fig. 3B) activatedPDK1T148C by 6.3-fold, representing a significant improvementover 2A2 (3.9-fold).

Biochemical Characterization of Allosteric PDK1 Modulators. To ex-plore the mechanism behind activation and inhibition of PDK1by PIF pocket ligands, we first wanted to assess whether bindingat the PIF pocket is coupled to substrate binding at the activesite. The weak affinity of the Akt-derived substrate peptide forPDK1 (Km approximately 10 mM) precluded measurement ofMichaelis–Menten parameters (8). Thus, we chose to assessinterplay between the ATP pocket and the PIF pocket using dis-ulfide trapping. Dose-response curves were measured for bindingof the best disulfide fragment activator (JS30) and the best inhi-bitor (1F8) to PDK1T148C in the presence and in the absence ofa saturating concentration of staurosporine (100 μM), a tightlybinding, ATP-competitive inhibitor of PDK1 (Fig. 4A). Thesecurves showed only minimal effects of staurosporine on theefficiency of disulfide-trapping for either JS30 or 1F8.

We next probed whether the allosteric PDK1 modulatorsaffected local or global protein stability. Park and Marqusee haveshown that susceptibility of a protein/ligand complex to proteo-lysis by a nonspecific protease (e.g., thermolysin) is inverselyproportional to the energetic stability of the complex (18). Wecompared thermolysin-catalyzed degradation of PDK1T148C

alone and bound stoichiometrically to one of our best allostericsmall molecule modulators (JS30 or 1F8) or an inert allostericligand (1H9) as visualized by SDS-PAGE (Fig. 4B). All of theseligands stabilized PDK1T148C to degradation; the half-life of theprotein was 1.9 min in the absence of these compounds and6.7–9.5 min in their presence, representing 3.5- to 5-fold stabili-zation. The half-life of wild-type PDK1 (2.5 min) was, by contrast,insignificantly affected by the disulfide fragments (Fig. S5).

We also performed hydrogen-deuterium (HD) exchangeexperiments to assess stabilization of PDK1T148C bound to allos-teric modulators (Fig. S6). Upon stabilization, one expects hydro-gen-bonding interactions involving backbone amide protons tobecome stronger, limiting exchange with deuterium from D2O(19). We exposed unliganded PDK1T148C and PDK1T148C boundstoichiometrically to JS30, 1F8, or 1H9 in H2O to D2O andmonitored deuterium incorporation by mass spectrometry. Themass of the protein alone shifted upward by 76 Da upon exposureto D2O for 1 min, whereas the mass shift was þ53, þ52, and

A

C

B

D2A2 / T148C1F8 / T148C

log ([fragment], µM)-2 -1 0 1 2

0

50

100

% C

on

jug

atio

n

Disulfide fragments

R131C Q150CT148C

0

100

200

300

400

500

ytivit

cA

%

1E6

6H9

1F8

2A2

3H3

5B2

K115C V124C

0

100

200

300

400

500yti

vitc

A%

4D2

2G3 3B91B6

2C7 5B2

C911I

1H9(94 ± 3.3%)

1F8(32 ± 2.2%)

%A

cti

vit

y

log ([fragment], µM)-2 -1 0 1 2

0

100

200

300

400

500 2A2 / T148C2A2 / wild-type1F8 / T148C1F8 / wild-type

2A2(390 ± 10%)

NN

Cl

SS

O

O

O

NH

SSN

SS

OHN

JS30(630 ± 15%)

NN

SS

O

Fig. 3. Activation and inhibition of PDK1 by small molecule fragment dis-ulfides. (A) Activity data (relative to DMSO control) for each PDK1 mutantconjugated to disulfide fragments. (B) Structures of PIF pocket activators2A2 and JS30, inert ligand 1H9, and inhibitor 1F8 for PDK1T148C and theireffects on the activity of PDK1T148C in CHAPS-containing buffer. (C and D)Dose-response curves measuring (C) conjugation or (D) kinase activity forthe best activator, 2A2, and inhibitor, 1F8, against PDK1T148C and comparisonto the effect on wild-type PDK1. Error values calculated from duplicate mea-surements (see SI Text).

DMSOJS301F81H9

0 10 20 30 40 501000

2000

3000

4000

5000

6000

7000

Time, min

Ban

dIn

tens

ity

-0.5 0.0 0.5 1.0 1.5 2.0 2.5

0

20

40

60

80

100

JS30+DMSOJS30+STS1F8+DMSO1F8+STS

log ([fragment], µM)

% C

on

jug

atio

n

A B

Fig. 4. Biochemical characterization of the PDK1T148C activator, JS30, and in-hibitor, 1F8. (A) Disulfide trapping in the presence (filled symbols) or absence(open symbols) of staurosporine (STS). (B) Proteolytic stability assays forPDK1T148C conjugated to JS30, 1F8, or the inert disulfide fragment 1H9.

6058 ∣ www.pnas.org/cgi/doi/10.1073/pnas.1102376108 Sadowsky et al.

þ51 Da for PDK1T148C bound to JS30, 1F8, or 1H9, respectively.Thus, 23–25 amides, or approximately 30% of the total number ofexchangeable amides, were protected from exchange in thebound versus unbound protein.

Structural Characterization of Allosteric PDK1 Modulators. To gaindeeper insight into the binding mode and possible mechanismbehind allosteric modulation of PDK1 by PIF pocket ligands,we solved high-resolution X-ray crystal structures of three com-plexes, two between PDK1T148C and a disulfide fragment activa-tor, 2A2 or JS30, and one between this mutant and a disulfideinhibitor, 1F8 (to 2.0, 2.1, and 2.2 Å resolution, respectively;Table S2). Both activated complexes were crystallized in the pre-sence of the known ATP-competitive ligand bisindolylmaleimide-II (BIM-II), whereas the inhibited complex was crystallized in theabsence of an ATP-site ligand.

Continuous electron density could be traced in all threedisulfide-bound PDK1T148C structures from Cys148 to the ap-pended fragment, 2A2, JS30, or 1F8, allowing unambiguousdetermination of the position of these ligands in the PIF pocket(Fig. 5 A–C). All three fragments target the same general area inthe PIF pocket of PDK1, filling a hydrophobic cavity lined by theside chains of residues Lys115, Ile118, and Ile119 on helix B;Val124, Val127, Thr128, and Arg131 on helix C; and Gln150,Leu155, and Phe157 on β-strand 4. However, the shape of thiscavity differs in the three structures due primarily to movementsof the B and C helices (see below).

In all three structures, PDK1T148C adopts an active-like confor-mation overall, as characterized by specific interactions amongconserved catalytic residues (Fig. 5D and Fig. S7). For example,salt-bridge interactions are formed betweenGlu130 on the C helixand Lys111, residues that are critical for positioning ATP forphosphoryl transfer. Salt-bridge interactions are also apparentbetween phosphorylated Ser241 (pS241) on the activation loopand two key Arg residues, Arg204 of the so-called HRD motifand Arg129 on the C helix, which are responsible for positioningthe activation loop to accept a protein/peptide substrate (20).

Despite the similarities among the inhibited and activatedPDK1T148C structures, several differences are also apparent

(Fig. 5D and Fig. S7). For example, the conformation of the ac-tivation segment (residues 223–252) varies. In the two activatedstructures, this segment is largely disordered and displays poorelectron density between residues 232 and 240. A poorly-definedactivation loop was also observed for one of the two conformersfound in the asymmetric unit of the inhibited 1F8 complex,whereas in the other conformer, the activation segment couldbe almost completely refined, forming a short α-helical turn be-tween residues 231 and 236. The positions of the B and C helicesare also closer to the activation loop and ATP-binding site in the2A2 and JS30 complexes relative to the inhibited 1F8 complex.Residues on the C helix adopt different conformations in thethree structures. For example, in the most activated complex(with JS30) the Tyr126 hydroxyl makes putative H-bonding inter-actions with the side chains of Asp223 (in the DFG motif) andGlu130 and with the backbone amide of Gly225 (in the DFGmotif). In the 2A2 complex, the Tyr126 hydroxyl is positionedslightly further away from these residues and makes a hydrogenbond to a water molecule that instead makes the above contactswith the DFG motif and Glu130. In the inhibited 1F8 complex,the Tyr126 side chain adopts two conformations in the crystal, onein the direction of the active site and one forming an H-bondinginteraction with pSer241, but both too far from the DFG motifor Glu130 to form contacts with either.

Wemutated Tyr126 in PDK1T148C to Phe to investigate the roleof this residue in activation/inhibition by PIF pocket disulfidecompounds (Fig. S8). The resulting Y126F/T148C double mutanthad approximately twofold lower kinase activity relative to theT148C single mutant. The double-mutant was less well activatedby JS30 than was PDK1T148C (5.2- vs. 6.0-fold activation, respec-tively), but was inhibited slightly better by 1F8 relative to thesingle mutant (3.8-fold vs. 3.1-fold inhibition).

DiscussionUnlike conventional screening approaches, disulfide trapping iswell suited for hypothesis-driven ligand discovery efforts of thetype described here in which one asks whether targeting a specificprotein site with small molecules or peptides can produce adesired functional and/or structural effect. Here, we have usedthe technology to gain a better understanding of how PIF pep-tides interact with and allosterically activate PDK1 and to probethe PIF site with small molecule fragments.

Molecular modeling simulations provided a structural hypoth-esis for how PIFtides bind to PDK1 and we subsequently useddisulfide trapping in solution to test this hypothesis. Taken to-gether, our conjugation data strongly support a binding orienta-tion of the PRK2-derived PIFtide in the PIF pocket on PDK1that is similar to that adopted by hydrophobic motifs of AGCkinases bound intramolecularly to their own analogous (HM)pocket (7). Comparison of the catalytic activities of differentcross-linked peptide/PDK1 pairs showed that the extent of PDK1activation is sensitive to PIFtide positioning. For example, severalCys-PIFtide/PDK1 mutant cross-links could be formed betweenresidues that were nearby in our model, but these were nonequi-valent in terms of PDK1 activation. That the most activatedcomplex, PDK1Q150C∕PIFtideD10C, was more active than thecorresponding wild-type (noncovalent) PIFtide/PDK1 complexsuggests that the full activation potential of wild-type PDK1 hasnot been realized in experiments with isolated PIFtides. It ispossible that the entire PIF-containing protein (i.e., PRK2) isrequired for inducing maximal activation of wild-type PDK1.

Screening of a small library of 480 disulfide fragments yieldeda number of ligands with diverse structures and functionalimpacts on PDK1. Each Cys site on PDK1 trapped a distinctpanel of disulfide fragments that were selective for each mutantover other mutants or the wild-type enzyme, suggesting specificbinding modes in each case. The best disulfide fragment activa-tors from our initial screen were nearly as potent as the best

I119

K115

V124 T128

I118

L155F157

V127

1F8/T148CQ150

R131

I119

K115

V124 T128

I118

L155F157

V127

JS30/T148CQ150

R131

A B

C D

I119

K115

V124 T128

I118

L155F157

V127

2A2/T148CQ150

R131

αB

αC

Y126D223

E130

BIM-II

pS241

K111

Fig. 5. Crystal structures of PDK1T148C conjugated to allosteric disulfide frag-ments. Close-up views of the PIF pocket in complex with activators (A) 2A2 or(B) JS30 or (C) inhibitor 1F8, showing residues on the B helix (pink), C helix(yellow), and β-strand 4 (cyan) that form the interaction surface for theseligands. (D) Effects of small molecule modulator binding at the PIF pocketon positioning of the B and C helices (arrows), Tyr126, and other PDK1 activesite residues. Structures shown in D are various mutants of PDK1 bound to PIFpocket ligands JS30 (pink), PS48 (orange) (12), 2A2 (yellow), nothing(i.e., no ligand; gray) (22), or 1F8 (dark/light blue).

Sadowsky et al. PNAS ∣ April 12, 2011 ∣ vol. 108 ∣ no. 15 ∣ 6059

BIOCH

EMISTR

Y

Cys-PIFtide activators; chemical analogs produced even morepotent fragment activators (reaching 6.3-fold activation).

Perhaps most surprising was the fact that the fragments, unlikePIFtides, included both PDK1 activators and inhibitors. Bothbinding and extent of activation or inhibition were highly depen-dent on chemical structure of the bound fragment and positioningof the fragment in the PIF pocket. The activity of PDK1T148C incomplex with small fragments was variable over the largest range(approximately 19-fold), from the most inhibited (in complexwith 1F8) to the most activated (in complex with JS30). Althoughthe fragments are only about 1∕10th the size of PIFtides, theireffects on PDK1 activity were more extreme (i.e., greater activa-tion and inhibition). It is possible that the activation of PDK1observed by PIFtides represents the net sum of interactionsbetween a number of subsites, some of which can be activatingand some inhibiting. Perhaps the small fragments can occupyindividual subsites and thereby generate more potent activat-ing or inhibiting effects on PDK1. Previous disulfide-trappingexperiments on the C5a receptor, a G-protein coupled receptor(GPCR), produced both agonists and antagonists at the same site(4, 5). It is intriguing that a single allosteric site on a kinase, thePIF pocket on PDK1, can similarly generate a continuum ofagonistic and antagonistic effects, by analogy to single ligandbinding sites in GPCRs.

Biochemical data and high-resolution crystal structures foractivating and inhibiting fragments bound stoichiometricallyto PDK1T148C showed that these ligands stabilize the proteinglobally and induce conformational changes that may accountfor their effects on activity. In crystallizing our most strongly in-hibited complex, 1F8/T148C, in the absence of an ATP-site ligandand the most strongly activated complex, JS30/T148C, in thepresence of an ATP-site ligand (BIM-II), we expected to observewidely divergent PDK1 conformations. Surprisingly, the struc-tures of the 1F8∕PDK1T148C and JS30∕PDK1T148C complexeswere quite similar, with all key catalytic residues (e.g., Lys111,Glu130, and Asp223 in the ATP-binding cleft; Arg204 of theHRD motif; and Arg129 on the C helix) poised in canonically“active” conformations (20). However, clear changes were ob-served in the conformation of the activation loop and the posi-tions of the B and C helices, which move approximately 5 Å awayfrom the ATP-binding site in the most inhibited versus the mostactivated complex (Movie S2). Additionally, in both activatedcomplexes, the side chain of Tyr126 in the C helix is directedinward toward the ATP-binding site. In complex with JS30, theTyr126 hydroxyl on PDK1T148C forms a hydrogen bond with theside chains of Asp223 in the DFG motif and Glu130 on the Chelix, residues that are strictly conserved and critically importantin all kinases for positioning ATP during catalysis (21). This typeof interaction has not, to our knowledge, been observed in otherkinase structures. In all PDK1 structures solved to date, Tyr126is positioned outward, forming an H bond to pSer241. In PKAand PKB, the residue corresponding to Tyr126 in PDK1 is a Hisresidue, which forms an electrostatic interaction with an analo-gous phosphorylated residue on the activation loop. It is possiblethat the interaction of Tyr126 with the DFGAsp and/or Glu130 inthe JS30∕PDK1T148C complex plays a role in allosteric activation.Indeed, mutation of Tyr126 to Phe had a modest, but significantnegative impact on activation by JS30. Interestingly, this mutationalso makes inhibition of PDK1T148C by 1F8 slightlymore effective,suggesting that Tyr126 may also destabilize the inhibited state ofthe kinase. Prior to phosphoryl transfer in protein kinases, aMg2þ ion bridges the Asp of the DFGmotif and the gamma phos-phate group of bound ATP (20). It seems unlikely, therefore, thatTyr126 in PDK1 would be able to form a close interaction with theDFG Asp of PDK1 in the Mg2þ∕ATP-bound state. However, it ispossible that Tyr126 assists in release of ADP following phosphor-yl transfer, a step which is rate limiting for many kinases.

Comparisons among the biochemical and structural datareported here for disulfide-bound modulators of PDK1 andelsewhere for PDK1, alone and in complex with a known smallmolecule activator that binds to the PIF pocket, PS48 (12),provide additional clues regarding the mechanism(s) behindallosteric modulation of this kinase. For example, PS48 wasreported to stabilize PDK1, but was paradoxically competitivewith ATP binding (12). We also observe kinase stabilization byPIF pocket-directed disulfide fragments, but do not observesignificant communication with the ATP site by PIF pocket acti-vators, suggesting that this feature is not necessary in all casesfor activation. Our disulfide fragments also bind to the PIF pock-et in slightly different ways relative to PS48. Like PIFtides, PS48possesses two aromatic groups that become buried in the hydro-phobic PIF pocket and a carboxylate that forms an electrostaticinteraction with Arg131 on the C helix (12). Only the hydropho-bic components are reproduced in the disulfide activators pre-sented here, yet activation is observed. The carboxylate/Arg131contact between PIFtides/PS48 and PDK1 may therefore beimportant for binding to the PIF pocket, but may not be impor-tant for activation. Finally, a comparison of the structures here ofPDK1T148C bound to inhibitory and activating disulfide fragmentsand those previously reported of PDK1 absent a PIF pocketligand (22) and in complex with PS48 (12) reveal a compellingcorrelation between C-helix position and kinase activity (Fig. 5Dand Movie S2). The positioning of the C helix relative to the ATPsite for PDK1 bound to various PIF pocket ligands follows thetrend, from nearest to farthest, JS30, PS48, 2A2, none (i.e., boundonly to an ATP-site ligand), and 1F8, which tracks well withthe activities of the corresponding complexes relative to theunliganded proteins: 630%, 400% (12), 390%, 100%, and 32%,respectively. This correlation supports previous hypotheses impli-cating the C helix in allosteric activation of PDK1 (12), but showsadditionally that the C helix may also be positioned to inhibit thekinase. We cannot, however, discount possible effects of crystalpacking differences in the various structures as causes for thestructural changes we describe.

Many kinases possess a functional allosteric pocket analogousto the PIF pocket on PDK1 that should be targetable using theapproach we have described. Within the AGC kinase family,PKA, PKB (Akt), PKC isozymes, RSK, R6K, SGK, PRK1/2,and GRK all have a C-terminal HM segment that docks to apocket analogous to the PIF pocket and cis activates these kinases(7). In Src, a nonreceptor tyrosine kinase, the analogous pocket isoccupied by residues on the SH2-kinase domain linker N-term-inal to the catalytic domain, which stabilize an inactive conforma-tion (23). Other kinases are regulated by interactions of separateprotein domains with a PIF-like pocket (e.g., TPX2/Aurora A,cyclin/CDKs, and the EGFR/EGFR dimer) (24–26). Based onour success with PDK1, we expect that disulfide trapping willallow us to study and validate such sites in an efficient and sys-tematic fashion while simultaneously providing chemical startingpoints for developing bioactive kinase modulators.

Experimental ProceduresPeptide Synthesis. Peptides were synthesized using standardprocedures (see SI Text).

Fragment Disulfide Synthesis. Two equivalents of a commerciallyavailable amine or carboxylic acid (Sigma-Aldrich) were reactedwith one equivalent of either a succinimidyl ester form of theappropriate carboxylic acid (4,4′-dithiodibutyric acid or 3,3′-dithiopropionic acid) or cystamine and diisopropylcarbodiimide,respectively, in a mixture of dimethylformamide and diisopropy-lethylamine (∼10∶1) to generate a disulfide-linked dimer. Thedimer was then reacted with either cystamine or N,N,N′,N′-tetra-methylcystamine in the presence of a catalytic amount of cystea-mine to give the final asymmetric disulfide-capped fragment,

6060 ∣ www.pnas.org/cgi/doi/10.1073/pnas.1102376108 Sadowsky et al.

which was purified by HPLC. The identities and purities ofall reaction products were verified by liquid chromatography(LC)/MS.

Plasmid Construction and Protein Preparation. The catalytic domainof PDK1 (residues 51–359) was cloned into a pFastBacHTavector and expressed in Sf21 insect cells. All mutants were gen-erated by Quikchange mutagenesis (Stratagene) using standardprotocols (see SI Text).

Disulfide-Trapping Conjugation Assays. Immediately prior to disul-fide-trapping experiments, PDK1 was exchanged into 25 mM TrispH 7.5 using a Nap-5 size-exclusion column (GE Amersham)and fresh β-ME (sealed vials from Thermo Scientific) was addedto the desired concentration (see Results). Proteins wereincubated with disulfide compounds in 96-well plates at roomtemperature for 1 h prior to mass spectrometric analysis usinga LCT-Premier LC/electrospray ionization-MS instrument(Waters). Protein masses were deconvoluted using the Max-Entalgorithm within the MassLynx software. Conjugation strengthswere measured by comparing peak areas for the conjugatedversus unconjugated protein.

Kinase Activity Assays. A radioactivity-based kinase assay usingas substrates [γ-32P]-ATP and a peptide derived from the activa-tion loop of Akt (KTFAGTPEYLAPEVRR) was used to monitoreffects of disulfide compounds on PDK1 activity (see SI Text).

Crystallography. Stoichiometric disulfide/PDK1 conjugates wereprepared immediately prior to crystallization by incubatingpartially purified protein with an excess of disulfide compoundin the presence of β-ME until conjugation was complete (asindicated by LC/MS). The conjugates were purified by size-exclu-sion chromatography and concentrated to 8.4, 8.4, or 16 mg∕mLfor the 2A2, JS30, and 1F8 complexes, respectively. To the JS30/

T148C and 2A2/T148C complexes were added BIM-II at anapproximate twofold molar excess. All protein complexes werecrystallized by hanging drop vapor diffusion by mixing the proteinwith reservoir solution. Orange, hexagonal rod-shaped crystalsof PDK1T148C∕2A2∕BIM-II and PDK1T148C∕JS30∕BIM-II wereobtained at 20 °C using the following reservoir solutions: for2A2, 0.1 M sodium citrate pH 5.4, 0.1 M potassium sodiumtartrate, 1.8 M ammonium sulfate; and for JS30, 0.1 M sodiumcitrate pH 5.6, 0.2 M NDSB-211, 1.7 M ammonium sulfate.Square rod-shaped crystals of PDK1T148C∕1F8 were obtainedat 4 °C using the following reservoir solution: 0.1 M Bis-TrispH 6.5, 0.2 M lithium sulfate, 25% ðwt∕volÞ PEG 3350. Crystalswere immersed briefly in cryoprotectant—for 2A2 and 1F8 com-plexes, a mixture of Paratone-N and paraffin oil, or for JS30,0.1 M sodium citrate pH 5.6, 1.7 M ammonium sulfate, 0.2 MNDSB-211, 20% glycerol—and flash-frozen in liquid nitrogen.Diffraction data were collected at the Advanced Light Source,Berkeley, California (Beamline 8.3.1) and scaled withHKL2000. All structures were solved by molecular replacement(in CCP4) using, as a search model, the ATP-bound structure ofPDK1 (Protein Data Bank ID 1H1W) (27). Structures weresubsequently refined iteratively using WinCoot (28) and Phenix(29) (Table S2).

ACKNOWLEDGMENTS. We acknowledge Adam Renslo and PunithaVedantham in the Small Molecule Discovery Center at University ofCalifornia, San Francisco (UCSF) for guiding preparation of the fragmentdisulfide library. We thank J.J. Miranda (UCSF) for instruction in proteinexpression. We thank Susan Taylor (University of California at San Diego)and Adrian Whitty (Boston University) for helpful comments and criticism.This work was supported by grants from the National Institutes ofHealth (R01AI070292 to J.A.W.), the Sandler Foundation (J.A.W.), andpostdoctoral fellowship awards from the California Tobacco RelatedDisease Research Program (110385 to J.D.S.) and the National Institutes ofHealth (F32CA119641 to D.W.W.). The Advanced Light Source is supportedby the director of the Office of Basic Energy Sciences, in the US Departmentof Energy under Contract DE-AC02-05CH11231.

1. Erlanson DA, et al. (2000) Site-directed ligand discovery. Proc Natl Acad Sci USA97:9367–9372.

2. Erlanson DA, Wells JA, Braisted AC (2004) Tethering: Fragment-based drug discovery.Annu Rev Biophys Biomol Struct 33:199–223.

3. Hardy JA, Lam J, Nguyen JT, O’Brien T, Wells JA (2004) Discovery of an allosteric sitein the caspases. Proc Natl Acad Sci USA 101:12461–12466.

4. Buck E, Wells JA (2005) Disulfide trapping to localize small-molecule agonists andantagonists for a G protein-coupled receptor. Proc Natl Acad Sci USA 102:2719–2724.

5. Buck E, Bourne H,Wells JA (2005) Site-specific disulfide capture of agonist and antago-nist peptides on the C5a receptor. J Biol Chem 280:4009–4012.

6. Scheer JM, Romanowski MJ, Wells JA (2006) A common allosteric site and mechanismin caspases. Proc Natl Acad Sci USA 103:7595–7600.

7. Gold MG, Barford D, Komander D (2006) Lining the pockets of kinases and phospha-tases. Curr Opin Struct Biol 16:693–701.

8. Biondi RM, et al. (2000) Identification of a pocket in the PDK1 kinase domain thatinteracts with PIF and the C-terminal residues of PKA. EMBO J 19:979–988.

9. Frodin M, Jensen CJ, Merienne K, Gammeltoft S (2000) A phosphoserine-regulateddocking site in the protein kinase RSK2 that recruits and activates PDK1. EMBO J19:2924–2934.

10. Biondi RM, Kieloch A, Currie RA, Deak M, Alessi DR (2001) The PIF-binding pocket inPDK1 is essential for activation of S6K and SGK, but not PKB. EMBO J 20:4380–4390.

11. Frodin M, et al. (2002) A phosphoserine/threonine-binding pocket in AGC kinasesand PDK1 mediates activation by hydrophobic motif phosphorylation. EMBO J21:5396–5407.

12. Hindie V, et al. (2009) Structure and allosteric effects of low-molecular-weightactivators on the protein kinase PDK1. Nat Chem Biol 5:758–764.

13. Yang J, et al. (2002) Crystal structure of an activated Akt/protein kinase B ternarycomplex with GSK3-peptide and AMP-PNP. Nat Struct Biol 9:940–944.

14. Congreve M, Marshall F (2010) The impact of GPCR structures on pharmacology andstructure-based drug design. Br J Pharmacol 159:986–996.

15. McGovern SL, Caselli E, Grigorieff N, Shoichet BK (2002) A common mechanismunderlying promiscuous inhibitors from virtual and high-throughput screening.J Med Chem 45:1712–1722.

16. Ryan AJ, Gray NM, Lowe PN, Chung CW (2003) Effect of detergent on “promiscuous”inhibitors. J Med Chem 46:3448–3451.

17. Goode DR, Totten RK, Heeres JT, Hergenrother PJ (2008) Identification of promiscuoussmall molecule activators in high-throughput enzyme activation screens. J Med Chem51:2346–2349.

18. Park C, Marqusee S (2005) Pulse proteolysis: A simple method for quantitativedetermination of protein stability and ligand binding. Nat Methods 2:207–212.

19. Hoofnagle AN, Resing KA, Ahn NG (2003) Protein analysis by hydrogen exchangemass spectrometry. Annu Rev Biophys Biomol Struct 32:1–25.

20. Adams JA (2001) Kinetic and catalytic mechanisms of protein kinases. Chem Rev101:2271–2290.

21. Nolen B, Taylor S, Ghosh G (2004) Regulation of protein kinases: Controlling activitythrough activation segment conformation. Mol Cell 15:661–675.

22. Komander D, et al. (2004) Interactions of LY333531 and other bisindolyl maleimideinhibitors with PDK1. Structure 12:215–226.

23. Boggon TJ, Eck MJ (2004) Structure and regulation of Src family kinases. Oncogene23:7918–7927.

24. Bayliss R, Sardon T, Vernos I, Conti E (2003) Structural basis of Aurora-A activationby TPX2 at the mitotic spindle. Mol Cell 12:851–862.

25. Jeffrey PD, et al. (1995) Mechanism of CDK activation revealed by the structure of acyclinA-CDK2 complex. Nature 376:313–320.

26. Zhang X, Gureasko J, Shen K, Cole PA, Kuriyan J (2006) An allosteric mechanismfor activation of the kinase domain of epidermal growth factor receptor. Cell125:1137–1149.

27. Biondi RM, et al. (2002) High resolution crystal structure of the human PDK1 catalyticdomain defines the regulatory phosphopeptide docking site. EMBO J 21:4219–4228.

28. Emsley P, Lohkamp B, Scott WG, Cowtan K (2010) Features and development of Coot.Acta Crystallogr D Biol Crystallogr 66(Pt 4):486–501.

29. Adams PD, et al. (2010) PHENIX: A comprehensive Python-based system for macromo-

lecular structure solution. Acta Crystallogr D Biol Crystallogr 66(Pt 2):213–221.

Sadowsky et al. PNAS ∣ April 12, 2011 ∣ vol. 108 ∣ no. 15 ∣ 6061

BIOCH

EMISTR

Y