Transitions to catalytically inactive conformationsin EGFR kinaseYibing Shana,1, Anton Arkhipova, Eric T. Kima, Albert C. Pana, and David E. Shawa,b,1

aD. E. Shaw Research, New York, NY 10036; and bCenter for Computational Biology and Bioinformatics, Columbia University, New York, NY 10032

Edited† by José N. Onuchic, Rice University, Houston, TX, and approved March 5, 2013 (received for review November 30, 2012)

The epidermal growth factor receptor (EGFR) is a key protein incellular signaling, and its kinase domain (EGFR kinase) is an intenselypursued target of small-molecule drugs. Although both catalyticallyactive and inactive conformations of EGFR kinase have been re-solved crystallographically, experimental characterization of thetransitions between these conformations remains difficult. Usingunbiased, all-atom molecular dynamics simulations, we observedEGFR kinase spontaneously transition from the active to the so-called “Src-like inactive” conformation by way of two sets of inter-mediate conformations: One corresponds to a previously identifiedlocally disordered state and the other to previously undescribed“extended” conformations,markedby theopeningof theATP-bindingsite between the two lobes of the kinase domain.We also simulatedthe protonation-dependent transition of EGFR kinase to another[“Asp-Phe-Gly-out” (“DFG-out”)] inactive conformation and ob-served similar intermediate conformations. A key element observedin the simulated transitions is local unfolding, or “cracking,” whichsupports a prediction of energy landscape theory. We used hydro-gen–deuterium (H/D) exchange measurements to corroborate oursimulations and found that the simulated intermediate conforma-tions correlate better with the H/D exchange data than existingactive or inactive EGFR kinase crystal structures. The intermediateconformations revealed by our simulations of the transition processdiffer significantly from the existing crystal structures and may pro-vide unique possibilities for structure-based drug discovery.

molten globule intermediate state | hinge unfolding | high-entropytransition state | activation loop conformations

Epidermal growth factor receptor (EGFR, also known asErbB1/Her1) kinase, the cytoplasmic kinase domain of the

cell-surface receptor EGFR (1), is a widely studied protein do-main and the target of a number of small-molecule drugs used totreat a variety of cancers (2). Crystal structures reveal that theactive state of EGFR kinase shares characteristics common toalmost all active protein kinases (3). Two inactive conformationsof EGFR kinase have also been obtained from crystal structures(4, 5), one resembling the “Src-like inactive” state and one the“Asp-Phe-Gly-out” (“DFG-out”) state, each of which is observedin many other protein kinases (6).Despite these structural findings, the transition pathways be-

tween the active and inactive states of EGFR kinase remain un-clear. Likemany other protein domains, EGFR kinase presumablyadopts a multitude of intermediate conformations; identifyingsuch conformations is crucial to a comprehensive, molecular-levelunderstanding of the protein’s dynamic behavior in solution. It isalso possible that such intermediate conformations could be cap-tured and stabilized by suitable small molecules, offering uniqueopportunities for drug discovery. Intermediate conformations ofEGFR kinase, however, have not been directly observed experi-mentally. Molecular dynamics (MD) simulation has increasinglybeen used to characterize protein conformational changes andtransition pathways, including those of other protein kinases (e.g.,refs. 7–13). Here we report unbiased, all-atom simulations ofEGFR kinase transitioning from its active state to the Src-like (4)and DFG-out (14, 15) inactive states through multiple intermedi-ate conformations.

EGFR kinase, like other protein kinase domains, consists of anN-terminal and a C-terminal lobe (the N and C lobes), betweenwhich the ATP-binding site is located. The N lobe consists of fiveβ-strands (β1–β5) and the αC helix, whereas the C lobe is pre-dominantly helical, with an “activation loop” that is highly con-formationally variable. In the active state of EGFR kinase (3, 5),the αC helix is adjacent to the ATP-binding site (the “αC-in”conformation), and the catalytically important Asp831 residue ofthe conserved “DFG” motif is found within that site (“DFG-in”).The activation loop maintains a β9 strand and an overall confor-mation compatible with substrate binding (Fig. 1A). In the Src-likeinactive state of EGFR kinase, the αC helix is located farther fromthe ATP-binding site than it is in the active state (an “αC-out”conformation). Meanwhile, the residues of the β9 strand forma two-turn helix packed with the αC helix (Fig. 1A). The DFG-outinactive conformation differs from the active conformation in thatit features a “flipped” DFG motif, in which Phe832, instead ofAsp831, is found in the ATP-binding site. In addition to thesepreviously identified conformations, a recent study (16) founda locally disordered state, in which the αC helix is partially disor-dered and placed “out,” whereas the activation loop remains in anactive-like conformation with an intact β9 strand.The simulations reported here show transitions between the

active state and both the Src-like inactive and the DFG-out in-active states of EGFR kinase. The previously identified locallydisordered state appears as an intermediate in both pathwaysand is followed by highly extended conformations (Fig. 2). Wealso performed hydrogen–deuterium (H/D) exchange experi-ments, the results of which correlate well with the solvent ex-posure of the backbone amide groups in the locally disorderedintermediate state observed in our simulations, thus supportingour computational findings. The simulations also reveal thatthese intermediate conformations are accompanied by localunfolding, or “cracking,” at the so-called “hinge” region; thisobservation supports the prediction of energy landscape theorythat conformational changes in proteins may require local un-folding in their transition pathways (17). Moreover, our simu-lations show the extraordinary conformational flexibility of theactivation loop, reflecting the unique role of this segment in theconformational dynamics of protein kinases in general.

ResultsIn the present study, using unbiased, all-atomMD, we simulated thetransition of EGFR kinase from an active conformation to each ofthe two inactive conformations captured by X-ray crystallography.

Author contributions: Y.S. and D.E.S. designed research; A.A. and E.T.K. performed re-search; Y.S. and A.C.P. analyzed data; and Y.S. and D.E.S. wrote the paper.

The authors declare no conflict of interest.†This Direct Submission article had a prearranged editor.

Freely available online through the PNAS open access option.

See Commentary on page 7114.1To whom correspondence may be addressed. E-mail: David.Shaw@DEShawResearch.comor Yibing.Shan@DEShawResearch.com.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1220843110/-/DCSupplemental.

7270–7275 | PNAS | April 30, 2013 | vol. 110 | no. 18 www.pnas.org/cgi/doi/10.1073/pnas.1220843110

Dow

nloa

ded

by g

uest

on

Janu

ary

14, 2

021

Transition to the Src-Like Inactive Conformation. We performedthree separate simulations of EGFR kinase starting from the ac-tive structure. In two of them we observed a spontaneoustransition to the Src-like inactive conformation. (In the thirdsimulation we observed a transition to the DFG-out inactiveconformation, as is discussed in later sections.) One transition wascompleted 7 μs into a 23-μs simulation (Fig. 2D), and the other wascompleted 8.7 μs into a 12-μs simulation (Fig. S1). In each case,once the EGFR kinase arrived at the inactive conformation, itremained there for the remainder of the simulation. In this sec-tion, our analysis focuses on the transition to the Src-like inactiveconformation in the former simulation, which is qualitativelysimilar to that observed in the latter.In our simulations, the active state of the EGFR monomer

proved quite unstable, a result predicted by earlier findings in-dicating that EGFR kinase is activated only in an asymmetricdimer (5). The Src-like inactive conformations reached in oursimulations are in good agreement with the crystal structure ofinactive EGFR kinase, notably in terms of the position of the αChelix and the conversion of the active conformation’s activationloop β9 strand into the two-turn helix characteristic of the Src-like inactive conformation (Fig. 1B). This is a significant con-formational change, involving an ∼12-Å backbone root-mean-square deviation (RMSD) rearrangement of the residues of thetwo-turn helix and an ∼6-Å RMSD displacement of the αC helix.The inactive conformations we obtained are consistent with

the crystal structure not only in backbone arrangement, but alsoin the details of key interactions. In the active conformation ofmany kinases, for example, Glu738 is paired with Lys721 in a saltbridge; in the Src-like inactive conformation, Lys721 is pairedwith Lys836 of the two-turn helix instead. This swap of saltbridges, which crystal structures show to be common in the de-activation of many kinases, was accurately captured by our sim-ulations. The Lys721–Lys836 salt bridge formed and remainedstable once the EGFR kinase reached the Src-like inactiveconformation (Fig. 2D).To further show that the Src-like inactive conformations

reached by our simulations are consistent with their crystalstructure counterpart, we simulated EGFR kinase from the in-active structure [Protein Data Bank (PDB) ID code 2GS7] for10 μs. We found that the conformational space sampled by thissimulation overlapped well with the space sampled after the sim-ulation starting from the active conformation reached the Src-likeinactive conformation (Fig. 2F).

Intermediate Conformations. The N and C lobes are connected bya hinge region, C-terminal to β5, which includes a short loop, an

αD helix, and a part of the αE helix (Fig. 2). Our simulationsshowed that in the transition to the Src-like inactive conformation,the αC helix readily departed from its initial active conformationwithout altering the activation loop, resulting in an intermediateconformational state (Fig. 2 and Fig. S2B) that resembles thepreviously identified locally disordered state of EGFR kinase (16).This state ended ∼3.6 μs into the simulation and was followed bya hinge-like separation of the N and C lobes, yielding substantiallyextended conformations marked by their increased radii of gyra-tion and solvent-accessible surface areas (SASAs).The additional space between the two lobes allowed a rear-

rangement of the activation loop of ∼20 Å RMSD and ultimatelyallowed the formation of the two-turn helix at∼6.2 μs (Fig. 2). Thekinase then contracted, with a decreasing radius of gyration (Fig.2D), and eventually settled at the Src-like inactive conformation.Notably, the extended conformations differ from the active con-formation (11 Å RMSD; Fig. 2F) more than the ensuing Src-likeinactive conformation does (8 Å RMSD), indicating the difficultyin inferring intermediate transitional conformations from knownstarting and ending structures of a conformational change.

Cracking in the Intermediate Conformation. Conformational dy-namics are of critical importance to the function of proteins (18) asdynamic entities (19). Conformational transitions from an “open”state to a “closed” state (20) are common to many proteins.Miyashita et al. proposed that local protein unfolding, or cracking,is a common characteristic of these transitions in crossing free-energy barriers (17). They argued that cracking may lower the free-energy barrier of the conformational change and that the entropicgain resulting from the greater number of accessible conformationsmay overcome the enthalpic loss due to the disruption of fa-vorable contacts that results from unfolding. Various modelswere subsequently developed to demonstrate the cracking mech-anism (20, 21), but cracking has only rarely been observed in fullyatomistic simulations of protein conformational changes withoutusing artificial biasing forces (22).In our simulations of EGFR kinase, we observed that the ex-

tended conformations were accompanied by spontaneous crackingat the hinge region. During cracking, the region’s helical struc-tures, particularly those of the αD and αE helices, tended to dis-solve, leading to a decreased number of helical residues (Fig. 2B).Below, we show that similar cracking at this region was also ob-served in EGFR kinase’s transition to the DFG-out inactiveconformation.It has been proposed that cracking facilitates conformational

changes in proteins by raising conformational entropy, therebylowering the free energy of the transition state (17, 23). We thusmeasured conformational fluctuations, which are indicative ofconformational entropy, in the EGFR kinase simulations. Wefound a surge of conformational fluctuations in the extendedconformations, especially near the activation loop and in the hingeregion (Fig. 2E). This is broadly consistent with a crackingmechanism, although the precise location of the transition state isunclear. The coupling of the activation loop and the hinge regionin conformational fluctuation may be understood structurally:Cracking at the hinge region leads to a separation of the N and Clobes, making additional volume accessible to the activation loop,which is located between the two lobes. At ∼7 μs, after the com-pletion of EGFR kinase’s transition to the Src-like inactive con-formation and the accompanying reduction in its radius of gyration,the high conformational entropy subsided and the cracked helicesbegan to recover (Fig. 2 D and E).

Corroboration from Hydrogen–Deuterium Exchange Experiments.Seeking experimental corroboration of the intermediate con-formations of EGFR kinase generated by our simulations, weanalyzed our simulation trajectory in light of the data from H/Dexchange experiments. For each snapshot of the 27-μs simulation,

αC helix

2-turn helix β9 sheetP loop

hinge

Src-like inactive crystal structureActive crystal structure(Starting simulation conformation)

Src-like inactive crystal structureSimulation conformation

αC helix

2-turn helix

activationloop

activationloop

A B

Fig. 1. Comparison of key structural elements of simulation and crystalstructures of EGFR kinase. (A) Superposition of the crystal structures of theactive and the Src-like inactive conformations (PDB ID codes 2ITP and 2GS7).(B) Superposition of the conformation reached in our simulation with theSrc-like inactive crystal structure.

Shan et al. PNAS | April 30, 2013 | vol. 110 | no. 18 | 7271

BIOPH

YSICSAND

COMPU

TATIONALBIOLO

GY

SEECO

MMEN

TARY

Dow

nloa

ded

by g

uest

on

Janu

ary

14, 2

021

we calculated the SASA of each residue’s backbone amide groupsand calculated their correlation coefficient, R2, to the H/D ex-change profile of EGFR kinase. (The amide SASA profiles wereaveraged over 1 μs before they were correlated with the H/D ex-change data.) We observed a peak correlation (R2 = 0.56) between2.6 and 3.6 μs of simulation time (Fig. 3), the period during whichthe EGFR kinase adopted the locally disordered conformations(Fig. 2). This correlation of the simulation-generated intermediateconformations to the H/D exchange profile was higher than that ofthe crystallographically determined active conformation (R2= 0.48)or the Src-like inactive conformation (R2= 0.35), lending additionalexperimental support to the simulations.Proteins in solution exist in multiple conformations; active and

Src-like inactive conformations of EGFR kinase presumably co-exist, collectively determining overall H/D exchange. A combina-tion of the amide SASA profiles of these conformations, optimizedto reflect their Boltzmann weights, could thus potentially correlatewith the H/D exchange data better than either conformation in-dividually. Our analysis, however, showed that such a combinationyielded a maximum correlation of R2 = 0.49—a negligible im-provement from that of the active conformation alone (Fig. S3)and lower than the correlation of the intermediate conformation(R2 = 0.56). The relatively poor correlation of the weightedcombination of the active and the Src-like inactive crystal struc-tures to theH/D exchange profile is consistent with the importanceof the locally disordered state in the conformational equilibrium ofEGFR kinase. Analysis based on the slow binding of lapatinibsuggested that the locally disordered state may be prevalent formonomeric EGFR kinases, whereas the Boltzmann weight of theSrc-like inactive conformation may be low (16). Although thisdisordered state was relatively transient in the simulations

reported here (Fig. 2), this appears to be an artifact of the forcefield we used, which may not be sufficiently accurate for a reliableestimate of the Boltzmman weight of a conformational state. Therelatively poor correlation may also be explained, in part, by thefact that the active conformation of EGFR kinase is viable only indimers (5), but the effect of dimerization on H/D exchange is notaccounted for in the analysis.

Src-Like Inactive Conformation as an Intermediate Conformation ofthe DFG Flip. In a third simulation of EGFR kinase starting from itsactive conformation, we protonated the Asp residue of the DFGmotif after ∼17 μs. At ∼32 μs we observed the completion ofa transition to the DFG-out inactive conformation (Fig. 4). In Ablkinase, MD simulations combined with experimental work havepreviously shown that protonation of the Asp residue promotesa DFG flip and favors the DFG-out conformation (24) and thatthe DFG flip is coupled with large-scale motion of the αC helix(25). Overall conservation of both sequence and structure in theregions adjacent to the DFG motif strongly suggests that theprotonation dependence of the DFG flip is conserved in manyprotein tyrosine kinases, including EGFR kinase. In the presentsimulation, the DFG flip finished only after Asp831 was pro-tonated (Methods). Our estimation of pKa (26), based on the(heavy-atom) conformations, shows a significant increase in thepKa of Asp831 accompanying the DFG flip in the simulation (Fig.S4); this observation is consistent with the notion that theDFG flipis protonation dependent.An increasing number of crystal structures of EGFR kinase in

the DFG-out conformation have been resolved (14, 15), suggest-ing that this conformation may be common for the kinase in so-lution. A generalized Born estimation of conformational free

Locally disordered

disordered αC helix

A

D

αE helix

hinge region

Active αC helix

activation loop αD helix

αE helix

C lobe

N lobe

2-turn helix

Src-like inactive

αC helix

cracked hinge

αC helix

Extended

activation loop

αD helix C

Hinge cracking

RMSD w. r. t. active (Å)

RM

SD

w. r

. t. i

nact

ive

(Å)

Active Structure

Src-like inactiveStructure

Transition by MD (23 μs)MD from the inactive (10 μs)

LocallyDisordered

Extended

Src-like inactive

0 5 10 15 20 25Time (μs)

15

20

25

30

Hel

ix c

ount

Hinge

Val762-Thr879Cα-Cα (Å)

20

25

30

35

40

19

20

21

22

Rg (

Å)

0

5

10

15

20R

MS

D (Å

) w. r. t. Src-like inactive w. r. t. active

Cracking

F

0

5

10

15

20

E762-K860

E762-K745

Dis

tanc

e (Å

)

673723

773823

873923

Act

ivat

ion

loop

Hin

geαG

hel

ix

Time (μs)

Residue no.

1.0

1.4

1.8

2.2

2.6

Time (μs)0 5 10 15 20

E

αC helix

hinge helix

Val762

B

Activation loop

Active conformationExtended conformation

Thr879

Aver

age

RM

SF

(Å)

Residue-specific RMSF (Å)

0 2 4 6 8 10 12 140

2

4

6

8

10

12

14

1.5

3.0

0.0

4.5

6.0

7.5

9.0

≥10

0 5 10 15 20

Fig. 2. Transition to the Src-like inactive state. (A)Cartoon representation of the transition pathway.The extended conformation features an openinterlobe arrangement and cracking at the hingeregion. (B) Extended conformation (blue) comparedwith the initial active conformation (red). The αChelix and the activation loop are highlighted. Thedistance between the Cα atoms of Val762 andThr879, which is used as an indicator of the pro-tein’s degree of extension, is marked. (C) Close-upof the hinge region, with (blue) and without (red)cracking. (D) Transition to the Src-like inactive con-formation measured by the switch of salt bridges,RMSD, radius of gyration, and the number of helicalresidues in the hinge region (residues 795–831). Thelocally disordered state is highlighted by a yellowbackground and the extended conformation by alight green background. (Bottom) Dissolution of he-lical structure in the hinge region indicates that theseintermediate conformations were accompanied bycracking. (E ) Average and residue-specific root-mean-square fluctuation (RMSF) as functions of timein the simulation. The RMSF was calculated with arunning time window of 1 μs with respect to theaverage conformation in that time window. Onlythe Cα atoms of each residue were used in the cal-culation. As highlighted by the green backgroundand orange rectangles, the extended conformationspreceding the transition to the Src-like inactive con-formation were marked by high RMSF, indicative ofhigh conformational entropy, particularly at theregions of the hinge and the activation loop. (F )Transitions projected onto a 2D space of RMSD withrespect to the active and Src-like inactive crystalstructures, respectively. RMSD is calculated using the Cα atoms of the αC and the two-turn helices (residues 756–769 and 857–863). Each dot representsa snapshot of the trajectory; the interval between the snapshots shown is 25 ns. The light blue region includes conformations within 8.2 Å RMSD of both activeand Src-like inactive crystal structures. As shown, the transition (red dots, pink arrows) passed through two sets of intermediate conformations, which wereidentified by visual inspection. The simulation starting from the inactive conformation is also shown (purple dots).

7272 | www.pnas.org/cgi/doi/10.1073/pnas.1220843110 Shan et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

14, 2

021

energy (27) (Fig. S4) suggests that the free energy of the DFG-outconformation is comparable to that of the DFG-in and thattransitions between these conformations involve crossing a free-energy barrier.The simulation also showed substantial motion of the activa-

tion loop accompanying the DFG flip. Starting from the activeconformation, the activation loop underwent a dramatic rear-rangement of more than 20 Å RMSD, arriving at a so-called“autoinhibitory conformation,” previously seen in IRK kinase(28), in which the activation loop is folded onto the kinase

substrate-binding site (Fig. 5B). Notably, in our simulation, theSrc-like inactive conformation was an intermediate in the DFGflip (Fig. 4), confirming an earlier prediction by Levinson et al.based upon their analysis of Abl crystal structures (25). At thecompletion of the DFG flip, the activation loop returned to thestarting conformation by way of a reverse conformational ar-rangement of ∼20 Å RMSD, during which the β9 strand wasrecovered (Fig. 5A; event 3 in Fig. 4C, Bottom). The endingconformation of the present simulation is highly consistent withthe DFG-out crystal structure (PDB ID code 2RF9) of EGFRkinase, which also maintains an intact β9 strand.As in the transition to the Src-like inactive conformations, we

observed cracking at the hinge region in the rearrangement ofthe activation loop during the DFG flip. The cracking allowedsubstantial movement of the N and the C lobes with respect toone another, which is required for the DFG flip (Fig. 4). Anal-ogously to the transition toward the Src-like inactive conforma-tion, the completion of the DFG flip was followed by an overallcontraction of the protein (Fig. 4C) and a reconstitution of thehelical structure at the hinge region.

Conformational Variability of the Activation Loop. Activation loopsin protein kinases are known for their central role in kinaseregulation and in the binding of kinase drugs. Modeling theconformation of the activation loop is thus highly desirable, butsuch characterization is challenging due to the region’s confor-mational variability, which renders it unresolved in many kinasecrystal structures. Broadly, four conformations of the activationloop (Fig. 5), which may differ from each other by as much as20 Å RMSD, are seen in available kinase structures: active, Src-like inactive, substrate competitive, and detached. Although thelatter two conformations have not, to our knowledge, beencaptured in EGFR kinase crystal structures, we observed themin our simulations.The activation loop’s active conformation, which includes the

β9 strand, is shared by almost all protein kinases, whereas itsinactive conformations vary a great deal among different kinases.The Src-like inactive conformation is seen in many kinases, in-cluding Src, Abl, CDK2, and EGFR. The substrate-competitiveconformation, which features an activation loop positioned to-ward the hinge and a shielded substrate-binding site, is adoptedby IRK, Abl, Src, and p38 MAP kinases, among others. In thedetached conformation, the activation loop, especially its shortαEF helix, is detached from the main body of the kinase. Thisconformation is typically seen in kinase dimers in which the

R2

R2 = 0.56Time: 2.4–3.4 μs

200 806040 100

Am

ide

SA

SA

(Å2 )

Å( AS

AS edi

mA

2 )

Tota

l SA

SA

(Å2 )

EAPNQALLRI LKETEFKK I KVLGSG

LDYVREHKDNI GSQYLLNWCVQI AKGMNYLEDRRLVHRDLAARNVLVKTPQH

I SSI LEKGERLPQPPI CT I DVYMI MVKCWMI DADSRPKFREL I I EFSKMARD

AFGTVYKGLWI PEGEKVK I PVA KELREATSPKANKEI LDEAYVMASVDNPHV

VK I TDFGLAKLLGAEEKE HAEGGKVPI KWMALESI LHRI YTHQSDVWSYGVT

PQRYLVI QGDERMHLPSPTDSNFYRALMDEEDMDDV

CRLLGI CLTSTVQL I TQLMPFGCL

VWELMTFGSKPYDGI PASE

700 720 740 760 780

068008 840 880

950930 970 990 1010

820

910

pH = 8

Iβ1 β2 β3 β4 β5αC

αEαD Activation segmentY

Catalytic loop

Hinge

900

(30 sec)

Simulation time (μs)

H/D exchange (%)Hydrogen–deuterium exchange of EGFR kinase

see panel B Amide SASA w. r. t.H/D exchange rate

αF

αG αH αI

QGAMG

VDADEYLI PQQG

>30%>50%>70%

>10%

>90%

500

400

300

20014000

18000

17000

16000

15000

0.25

0.35

0.45

0.55

0 5 10 15 20 25

0.0

3.5

3.0

2.5

2.0

1.5

1.0

0.5

A

B C

Fig. 3. Corroboration from H/D exchange experiments. (A) Total SASA ofEGFR kinase and its amide groups as functions of the simulation time. Alsoshown is the correlation of the SASA of the amide groups with the measuredH/D exchange rates (pH 8.0). The amide SASA profile was calculated for eachsnapshot in the simulation and averaged over a time window of 1 μs beforecomparison with the H/D exchange data. Note that the peak of the corre-lation (highlighted) occurs for the locally disordered state (Fig. 2D). (B)Correlation of the amide SASA and the measured H/D exchange rates at thehighlighted simulation time. Each dot represents a peptide fragment ofEGFR kinase (C) for which the H/D exchange rate is measured. Note thata residue may belong to multiple overlapping fragments due to the non-specific nature of the proteolysis step in the H/D experiment, resulting inmore dots on the correlation graph than the total number of fragmentsshown in C. The amide SASA of a fragment is averaged over its constituentresidues. (C) An H/D exchange profile of EGFR kinase measured at pH 8.0.

Conformation 1 DFG-in (start)2 Intermediate3 Intermediate4 DFG-out (end)

αC helix

DFG motif

A

αC helix

DFG motif

B

Hinge

Src-like inactive conformation in DFG flip

2-turnhelix

Time (μs)

C

19

20

21

20

25

30

2468

10

0 10 20 30 40 50 605

1525 w. r. t. active

w. r. t. Src-like inactive

Rg (

Å)

α c

ount

R

MS

D (Å

)R

MS

D (Å

)

DFGDFGProtonationProtonation

Cracking at high Rg

1

23

Fig. 4. Transition to the DFG-out inactive conformation. (A) Snapshots of the DFG flip, at 0, 13.42, 17.35, and 52.13 μs of the MD simulation, respectively. Thesnapshot at 0 μs is identical to the crystal structure of the active conformation (PDB ID code 2ITP). Note the motion of the αC helix involved in the DFG flip,measured by radius of gyration, number of helical residues in the hinge region, and RMSD with respect to crystal structures. (B) Snapshot taken at 23.32 μs, inwhich the Src-like inactive conformation serves as an intermediate of the DFG flip. (C) DFG flip measured by radius of gyration, RMSDs, and other metrics. Therunning average with windows of 100 ns (green) is also shown in the second panel from the top. The extended conformation (with light blue background) isaccompanied by cracking. Three conformational events of the activation loop are marked: 1, visiting the Src-like inactive conformation as an intermediate; 2,visiting the substrate-competitive conformation, which is separated from the initial active conformation by as much as 20 Å RMSD; and 3, return to the initialconformation. The RMSD calculation is performed using the Cα atoms of the two-turn helix (residues 857–863).

Shan et al. PNAS | April 30, 2013 | vol. 110 | no. 18 | 7273

BIOPH

YSICSAND

COMPU

TATIONALBIOLO

GY

SEECO

MMEN

TARY

Dow

nloa

ded

by g

uest

on

Janu

ary

14, 2

021

activation loops engage in domain swapping, which may underlietranskinase autophosphorylation (29, 30).Our simulations have captured multiple conformational changes

of the activation loop in EGFR kinase, some of which have not yetbeen seen in crystal structures. We observed the activation loop inthe substrate-competitive conformation (Fig. 5B), for instance, asan intermediate during the DFG flip. The detached conformationwas observed from 0.3 to 0.8 μs during the 12-μs simulation, inwhich the activation loop eventually settled into the Src-like in-active conformation. In the detached conformation, the αEF helixnear the activation loop’s C terminus was dislodged, and the acti-vation loop itself assumed an extended conformation away fromthe main body of the kinase (Fig. 5C). Although transient andhighly flexible in our simulation, this conformation could poten-tially be stabilized by a dimerization partner.The activation loop also visited other conformations in the

simulations. In one such conformation, for example, the middlesection of the loop formed a four-turn helix, reminiscent of a helixin the activation loop of MPSK1 kinase (PDB ID code 2BUJ).This conformation (Fig. 5D), which lasted ∼3 μs (17.5–20.5 μs) inthe 23-μs simulation, is intriguing because the helix puts Tyr845,a well-known phosphorylation site of EGFR kinase, in an entirelyexposed position along with a group of glutamate residues (Glu865,Glu866, Glu868, and Glu872). The resemblance of this arrange-ment to poly(Glu4Tyr) peptide, a favorable peptide substrate oftyrosine kinases, raises the possibility that this helical conformationof the activation loop might be involved in the phosphorylationof Tyr845.

DiscussionIn the present study, conformational transitions associated with thedeactivation of EGFR kinase were simulated in atomic detail usingunbiased MD. The fact that both Src-like and DFG-out inactiveconformations of EGFRkinase were generated without using priorcrystallographic information demonstrates the increasing potentialof MD simulations in the prediction of protein conformations.

The simulations suggest a set of highly extended conformationsin the transition pathway. These extended conformations arecrucial to the overall conformational changes of EGFR kinase:They separate the N and C lobes, thus making room for a rear-rangement of the activation loop. Perhaps surprisingly, by manymeasurements (including backbone RMSD), the starting, activeconformation of EGFR kinase differs more from the intermediateconformations than from the ending conformations (Fig. 2F). Thishighlights the underlying difficulty in inferring intermediate con-formations from ending conformations using methods such asbiased MD simulations (SI Methods and Fig. S5). The inter-mediate conformations are difficult to characterize experimentallybecause of their low Boltzmann weights. They are, however, po-tentially important in the regulation and function of protein kinases,for instance by exposing the activation loop for phosphorylation,a process known to be central to kinase regulation.Protein kinase activation loops are extremely conformationally

flexible, a trait reflected in many crystal structures. This flexibilityis associated with the activation loop’s role as a critical elementin protein kinase regulation: Its phosphorylation stabilizes thekinase’s conformation and enables robust protein kinase activity(6). Our simulations illustrated the EGFR kinase activationloop’s extraordinary flexibility, beyond that exhibited in existingcrystal structures (Fig. 5). Moreover, our simulations showedthat, even in the process of a seemingly local conformationalchange like the DFG flip, the activation loop routinely under-goes conformational changes of as much as 20 Å RMSD.Protein conformational changes are sometimes seen as con-

certed motions of structurally well-maintained subunits, while theprotein as a whole remains structurally intact in a compact andnative-like state in the process. Recently, however, it has beenproposed that local partial unfolding (cracking) may occur duringprotein conformational changes, reducing the energy barriers tothe protein’s global motions (17). In our simulations, cracking wasindeed observed in the vicinity of the hinge region linking theN and C lobes; this led to highly extended conformations, whichthe kinase visited while interconverting between its active and Src-like inactive conformations. These simulations are among the fewexamples thus far of spontaneous cracking being directly observedin unbiased atomistic simulations (22). Importantly, the simu-lations suggest that a protein may need to significantly deviatefrom its relatively compact native-like conformation in the processof a conformational change. It is also worth noting that, in oursimulations, cracking was observed only on the microsecondtimescale. Our results suggest that to determine whether crackingaccompanies a protein conformational change, simulations of thistimescale may sometimes be needed even when biasing forcesare applied.In previous work (16), we showed that EGFR kinase has

a locally disordered state and that cancer mutations in EGFRkinase suppress this disordered state, leading to elevated levelsof EGFR dimerization and activation. In the present study, thedisordered conformations were observed in the transition fromthe active to the Src-like inactive conformation (Fig. 2). More-over, these conformations were found to correlate well with theH/D exchange profile of EGFR kinase, consistent with a signifi-cant Boltzmann weight for the locally disordered state, as waspreviously inferred from the binding kinetics of lapatinib (16). Itshould be noted that, although the disordered conformations inour present simulations closely resemble their counterparts fromthe previous simulations overall, the helicity of the αC-helixsegment appeared more robust in the present simulations (Fig.S2), presumably due to subtle differences in the force fields weused. [We used the CHARMM22* force field (31) in the presentstudy and an Amber force field (32, 33) in the previous one.]Although the force field used here appears to exaggerate thestability of the Src-like and DFG-out inactive states relative tothe locally disordered state, and thus likely accelerates the rate of

C lobe

N lobeαC helix

Activation loop

C lobe

N lobe αC helix

Activation loop

C lobe

N lobe

αC helix

C lobe

N lobe

αC helix

2-turn helix2-turn helix

Tyr845

activation-loophelix formation

detachedactivation loop

Compared toPDB 2RF9

Compared to PDB 1OPJ

A B

C D

Fig. 5. Conformations of the activation loop found in the simulations. (A)Active-like arrangement of the activation loop coupled with a DFG-outconformation. This conformation, with the β9 strand intact, resembles theMig6-bound structure of EGFR kinase (red). (B) Substrate-competitive con-formation, compared with the structure of imatinib-bound Abl kinase (red).This (DFG-out) conformation is an intermediate of the simulated DFG flip.(C) Detached activation loop with a dislodged αEF helix. This is a transientconformation at 6.12 μs into the simulation starting from the Src-likeinactive structure. A similar conformation has been observed in crystalstructures of other protein kinases [e.g., in the crystal structure of domain-swapped OSR1 (PDB ID code 3DAK)]. (D) Conformation in which the acti-vation loop forms a helix that exposes the phosphorylation site Tyr845. Thetwo-turn helix is intact in this conformation.

7274 | www.pnas.org/cgi/doi/10.1073/pnas.1220843110 Shan et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

14, 2

021

transitions to these states, this does allow simulation of the con-formational changes in a reasonable time.The catalytic cycle of a protein kinase involves micro- to milli-

second conformational changes accompanying nucleotide bindingor release (34). It has previously been proposed that such con-formational changes may involve DFG flips in response to localelectrostatic changes in the catalytic cycle (24). We speculate thatnormal tyrosine kinase catalysis also requires access to the DFGflip’s highly extended conformations. This conjecture is in partinspired by the autoinhibition of Src and Abl kinases by SH2 andSH3 domains that bind at the “back” of the kinase domain (35–37). Because the binding to SH2 and SH3 obstructs neither ATPnor substrate binding directly, how the catalytic domain is inhibi-ted remains somewhat puzzling. On the basis of the present sim-ulations, we suggest that the inhibition could be realized by theSH2 and SH3 domains serving as a rigid clamp (38) that restrainsthe kinase domain and hinders the movement of its N and C lobes.In a similar fashion, the FGF receptor and ZAP70 kinases may beautoinhibited (39, 40) by so-called “molecular brakes” near thehinge region, which restrain the hinge region and ultimately thetwo kinase lobes.

MethodsWe performed all-atom molecular dynamics simulations of EGFR kinase, withwater represented explicitly, using the CHARMM22* forcefield (31) and three-point (TIP3P) water model (41, 42) on Anton (43), a special-purpose machine.The cubic simulated systems of 82.5 Å per sidewere solvated inwater with 0.15M NaCl; they contained 51,000 atoms. The simulations were initiated from theactive structure of EGFR kinase (PDB ID code 2ITP).

The simulation of a DFG flip consisted of two consecutive simulations: onestarting from the active structure of EGFR kinase and the other starting fromthe end conformation of the first simulation. In the second simulation,Asp831 was protonated.

The simulations were performed in the constant pressure and temperature(NPT) ensemble with T = 310 K, P = 1 bar, and Berendsen’s coupling scheme(44) with one temperature group. Water molecules and all bond lengths tohydrogen atoms were constrained using M-SHAKE (45). Van der Waals andshort-range electrostatic interactions were cut off at 12.5 Å. Long-rangeelectrostatic interactions were calculated using the k-space Gaussian SplitEwald method (46) with a 64 × 64 × 64 mesh. The simulation time step was 2.5fs; the r-RESPA integration method was used, with long-range electrostaticsevaluated every 5 fs (47). No artificial (biasing) forces were applied. A detaileddescription of the H/D exchange experiment is available elsewhere (16).

ACKNOWLEDGMENTS. The authors thank Michael Eastwood for discussionson cracking, Stefano Piana for discussions on force fields, and Mollie Kirkand Berkman Frank for editorial assistance.

1. Yarden Y, Sliwkowski MX (2001) Untangling the ErbB signalling network. Nat Rev MolCell Biol 2(2):127–137.

2. Lurje G, Lenz H-J (2009) EGFR signaling and drug discovery. Oncology 77(6):400–410.3. Stamos J, Sliwkowski MX, Eigenbrot C (2002) Structure of the epidermal growth

factor receptor kinase domain alone and in complex with a 4-anilinoquinazoline in-hibitor. J Biol Chem 277(48):46265–46272.

4. Wood ER, et al. (2004) A unique structure for epidermal growth factor receptorbound to GW572016 (Lapatinib): Relationships among protein conformation, in-hibitor off-rate, and receptor activity in tumor cells. Cancer Res 64(18):6652–6659.

5. Zhang XW, Gureasko J, Shen K, Cole PA, Kuriyan J (2006) An allosteric mechanism foractivation of the kinase domain of epidermal growth factor receptor. Cell 125(6):1137–1149.

6. Huse M, Kuriyan J (2002) The conformational plasticity of protein kinases. Cell 109(3):275–282.

7. Huang H, Ozkirimli E, Post CB (2009) Comparison of three perturbation moleculardynamics methods for modeling conformational transitions. J Chem Theory Comput5(5):1301–1314.

8. Lovera S, et al. (2012) The different flexibility of c-Src and c-Abl kinases regulates theaccessibility of a druggable inactive conformation. J Am Chem Soc 134(5):2496–2499.

9. Lu B, Wong CF, McCammon JA (2005) Release of ADP from the catalytic subunit ofprotein kinase A: A molecular dynamics simulation study. Protein Sci 14(1):159–168.

10. Ozkirimli E, Post CB (2006) Src kinase activation: A switched electrostatic network.Protein Sci 15(5):1051–1062.

11. Wan S, Wright DW, Coveney PV (2012) Mechanism of drug efficacy within the EGFreceptor revealed by microsecond molecular dynamics simulation. Mol Cancer Ther11(11):2394–2400.

12. Yang S, Banavali NK, Roux B (2009) Mapping the conformational transition in Srcactivation by cumulating the information from multiple molecular dynamics trajec-tories. Proc Natl Acad Sci USA 106(10):3776–3781.

13. Frembgen-Kesner T, Elcock AH (2006) Computational sampling of a cryptic drugbinding site in a protein receptor: Explicit solvent molecular dynamics and inhibitordocking to p38 MAP kinase. J Mol Biol 359(1):202–214.

14. Gajiwala KS, et al. (2013) Insights into the aberrant activity of mutant EGFR kinasedomain and drug recognition. Structure 21(2):209–219.

15. Zhang X, et al. (2007) Inhibition of the EGF receptor by binding of MIG6 to an acti-vating kinase domain interface. Nature 450(7170):741–744.

16. Shan Y, et al. (2012) Oncogenic mutations counteract intrinsic disorder in the EGFRkinase and promote receptor dimerization. Cell 149(4):860–870.

17. Miyashita O, Onuchic JN, Wolynes PG (2003) Nonlinear elasticity, proteinquakes, andthe energy landscapes of functional transitions in proteins. Proc Natl Acad Sci USA100(22):12570–12575.

18. Wolf-Watz M, et al. (2004) Linkage between dynamics and catalysis in a thermophilic-mesophilic enzyme pair. Nat Struct Mol Biol 11(10):945–949.

19. Frauenfelder H, Sligar SG, Wolynes PG (1991) The energy landscapes and motions ofproteins. Science 254(5038):1598–1603.

20. Hyeon C, Jennings PA, Adams JA, Onuchic JN (2009) Ligand-induced global transitionsin the catalytic domain of protein kinase A. Proc Natl Acad Sci USA 106(9):3023–3028.

21. Whitford PC, Miyashita O, Levy Y, Onuchic JN (2007) Conformational transitions ofadenylate kinase: Switching by cracking. J Mol Biol 366(5):1661–1671.

22. Brokaw JB, Chu JW (2010) On the roles of substrate binding and hinge unfolding inconformational changes of adenylate kinase. Biophys J 99(10):3420–3429.

23. Whitford PC, Sanbonmatsu KY, Onuchic JN (2012) Biomolecular dynamics: Order-disorder transitions and energy landscapes. Rep Prog Phys 75(7):076601.

24. Shan Y, et al. (2009) A conserved protonation-dependent switch controls drugbinding in the Abl kinase. Proc Natl Acad Sci USA 106(1):139–144.

25. Levinson NM, et al. (2006) A Src-like inactive conformation in the abl tyrosine kinasedomain. PLoS Biol 4(5):e144.

26. Gordon JC, et al. (2005) H++: A server for estimating pKas and adding missing hy-drogens to macromolecules. Nucleic Acids Res 33(Web Server issue):W368–W371.

27. Onufriev A, Bashford D, Case DA (2000) Modification of the generalized Born modelsuitable for macromolecules. J Phys Chem B 104(15):3712–3720.

28. Hubbard SR, Wei L, Ellis L, Hendrickson WA (1994) Crystal structure of the tyrosinekinase domain of the human insulin receptor. Nature 372(6508):746–754.

29. Pike ACW, et al. (2008) Activation segment dimerization: A mechanism for kinaseautophosphorylation of non-consensus sites. EMBO J 27(4):704–714.

30. Li Y, Palmer AG (2009) Domain swapping in the kinase superfamily: OSR1 joins themix. Protein Sci 18(4):678–681.

31. Piana S, Lindorff-Larsen K, Shaw DE (2011) How robust are protein folding simu-lations with respect to force field parameterization? Biophys J 100(9):L47–L49.

32. Hornak V, et al. (2006) Comparison of multiple Amber force fields and developmentof improved protein backbone parameters. Proteins 65(3):712–725.

33. Lindorff-Larsen K, et al. (2010) Improved side-chain torsion potentials for the Amberff99SB protein force field. Proteins 78(8):1950–1958.

34. Adams JA (2001) Kinetic and catalytic mechanisms of protein kinases. Chem Rev101(8):2271–2290.

35. Nagar B, et al. (2003) Structural basis for the autoinhibition of c-Abl tyrosine kinase.Cell 112(6):859–871.

36. Sicheri F, Moarefi I, Kuriyan J (1997) Crystal structure of the Src family tyrosine kinaseHck. Nature 385(6617):602–609.

37. Xu W, Harrison SC, Eck MJ (1997) Three-dimensional structure of the tyrosine kinasec-Src. Nature 385(6617):595–602.

38. Young MA, Gonfloni S, Superti-Furga G, Roux B, Kuriyan J (2001) Dynamic couplingbetween the SH2 and SH3 domains of c-Src and Hck underlies their inactivation byC-terminal tyrosine phosphorylation. Cell 105(1):115–126.

39. Chen H, et al. (2007) A molecular brake in the kinase hinge region regulates theactivity of receptor tyrosine kinases. Mol Cell 27(5):717–730.

40. Deindl S, et al. (2007) Structural basis for the inhibition of tyrosine kinase activity ofZAP-70. Cell 129(4):735–746.

41. Jorgensen WL, Chandrasekhar J, Madura JD, Impey RW, Klein ML (1983) Comparisonof simple potential functions for simulating liquid water. J Chem Phys 79(2):926–935.

42. MacKerell AD, Jr., et al. (1998) All-atom empirical potential for molecular modelingand dynamics studies of proteins. J Phys Chem B 102(18):3586–3616.

43. Shaw DE, et al. (2009) Millisecond-scale molecular dynamics simulations on Anton.Proceedings of the Conference on High Performance Computing, Networking, Stor-age and Analysis (SC09) (Association for Computing Machinery, New York), 10.1145/1654059.1654099.

44. Berendsen HJC, Postma JPM, DiNola A, Haak JR (1984) Molecular dynamics withcoupling to an external bath. J Chem Phys 81(8):3684–3690.

45. Krautler V, van Gunsteren WF, Hunenberger PH (2001) A fast SHAKE algorithm tosolve distance constraint equations for small molecules in molecular dynamics simu-lations. J Comput Chem 22(5):501–508.

46. Shan Y, Klepeis JL, Eastwood MP, Dror RO, Shaw DE (2005) Gaussian split Ewald: Afast Ewald mesh method for molecular simulation. J Chem Phys 122(5):54101.

47. Tuckerman M, Berne BJ, Martyna GJ (1992) Reversible multiple time scale moleculardynamics. J Chem Phys 97(3):1990–2001.

Shan et al. PNAS | April 30, 2013 | vol. 110 | no. 18 | 7275

BIOPH

YSICSAND

COMPU

TATIONALBIOLO

GY

SEECO

MMEN

TARY

Dow

nloa

ded

by g

uest

on

Janu

ary

14, 2

021