3 OCTOBER 2014 bull VOL 346 ISSUE 6205 55SCIENCE sciencemagorg

INTRODUCTION Understanding drug

mechanism poses the daunting challenge

of determining the affinity of the drug for

all potential targets Drug target engage-

ment can be assessed by means of a cel-

lular thermal shift assay (CETSA) based on

ligand-induced changes in protein thermal

stability We combined the CETSA method

with quantitative mass spectrometry to

study the effect of drugs on the thermal

profile of a cellular proteome comprising

more than 7000 proteins The approach

enabled the monitoring of drug targets

and downstream effectors

RATIONALE We devised a method for the

thermal profiling of cellular proteomes

Cells were cultured with or without drugs

and heated to different temperatures so

as to induce protein denaturation and re-

maining soluble proteins were extracted

with buffer At each temperature soluble

proteins were quantified by means of high-

Tracking cancer drugs in living cells by thermal profiling of the proteome

PROTEOMICS

Mikhail M Savitski1 Friedrich B M Reinhard1 Holger Franken1 Thilo Werner1

Maria Faumllth Savitski1 Dirk Eberhard1 Daniel Martinez Molina2 Rozbeh Jafari2

Rebecca Bakszt Dovega2 Susan Klaeger34 Bernhard Kuster34 Paumlr Nordlund25

Marcus Bantscheff1 Gerard Drewes1

RESEARCH ARTICLE SUMMARY

Tracking drugs in living cells Drugs alter the thermal stability of proteins directly through compound bind-

ing or indirectly through changes in overall protein state Thermal proteome prof ling determines melting

curves for thousands of proteins and tracks drug action in cells

resolution mass spectrometry yielding

denaturation curves This allowed deter-

mination of thermal stability and the iden-

tification of ligand-induced shifts To rank

binding affinities among multiple targets

we determined stability profiles across a

range of compound concentrations at a

defined temperature Comparison of the

thermal profiles obtained after drug treat-

ment of intact cells versus cell extract al-

lowed us to distinguish effects induced

by ligand binding from those induced by

downstream modifications

RESULTS We performed thermal pro-

teome profiling (TPP) on human K562 cells

by heating intact cells or cell extracts and

observed marked differences in melting

properties between the two settings with

a trend toward increased protein stability

in cell extract Adenosine triphosphatase

(ATP)ndashbinding proteins showed a signifi-

cant trend toward increased stability in

intact cells suggesting stabilization by the

endogenous ligand This was confirmed

with the addition of ATP to cell extract

which resulted in increased stability for

this protein group The ability of TPP to

identify target binding was validated by

using the broad-specificity inhibitors

staurosporine and

GSK3182571 which in-

duced shifts in the

melting temperatures

of many kinase tar-

gets and also affected

the thermal profiles of

other proteins including regulatory sub-

units of kinase complexes We identified

the heme biosynthesis enzyme ferrochela-

tase (FECH) as an off-target of several ki-

nase inhibitors and showed that the drug

vemurafenib reaches full target occupancy

of its cognate target BRAF and the off-tar-

get FECH within a narrow concentration

window FECH deficiency is genetically

linked to protoporphyria suggesting that

the photosensitivity induced by vemu-

rafenib and other drugs is mediated by

FECH Drug treatment of live cells affected

not only direct target proteins but also

downstream effectors The ABL inhibitor

dasatinib induced thermal shifts in several

proteins downstream of BCR-ABL includ-

ing CRKL and at concentrations in good

agreement with the effect on cell growth

CONCLUSION Thermal profiling of cel-

lular proteomes enables the differential

assessment of protein ligand binding and

other protein modifications providing an

unbiased measure of drug-

target occupancy for multi-

ple targets and facilitating

the identification of mark-

ers for drug efficacy and

toxicity

1Cellzome GmbH Molecular Discovery Research GlaxoSmithKline Meyerhofstrasse 1 Heidelberg Germany2Division of Biophysics Department of Medical Biochemistry and Biophysics Karolinska Institutet Stockholm Sweden3Department of Proteomics and Bioanalytics Technische Universitaumlt Muumlnchen Emil Erlenmeyer Forum 5 Freising Germany4German Cancer Consortium German Cancer Research Center Heidelberg Germany5Centre for Biomedical Structural Biology Nanyang Technological University SingaporeCorresponding author E-mail mikhailmsavitskigskcom (MMS) marcusxbantscheffgskcom (MB) gerardcdrewesgskcom (GD)Cite this article as M M Savitski et al Science 346 1255784 (2014) DOI 101126science1255784

Read the full article at httpdxdoiorg101126science1255784

ON OUR WEB SITE

Drug

Vehicle

( ) =1 minus

1 +minus( minus )

+

mz

Intensity

t (degC)

degC

degCt (degC)

t (degC)

Frac

tio

n n

on

-den

atu

red

Frac

tio

n n

on

-den

atu

red

Frac

tio

n n

on

-den

atu

red

Direct (of-) targets

Indirect targets

Unafected proteins

Published by AAAS

on

Apr

il 15

201

5w

ww

sci

ence

mag

org

Dow

nloa

ded

from

o

n A

pril

15 2

015

ww

ws

cien

cem

ago

rgD

ownl

oade

d fr

om

on

Apr

il 15

201

5w

ww

sci

ence

mag

org

Dow

nloa

ded

from

o

n A

pril

15 2

015

ww

ws

cien

cem

ago

rgD

ownl

oade

d fr

om

on

Apr

il 15

201

5w

ww

sci

ence

mag

org

Dow

nloa

ded

from

o

n A

pril

15 2

015

ww

ws

cien

cem

ago

rgD

ownl

oade

d fr

om

on

Apr

il 15

201

5w

ww

sci

ence

mag

org

Dow

nloa

ded

from

o

n A

pril

15 2

015

ww

ws

cien

cem

ago

rgD

ownl

oade

d fr

om

on

Apr

il 15

201

5w

ww

sci

ence

mag

org

Dow

nloa

ded

from

o

n A

pril

15 2

015

ww

ws

cien

cem

ago

rgD

ownl

oade

d fr

om

on

Apr

il 15

201

5w

ww

sci

ence

mag

org

Dow

nloa

ded

from

RESEARCH ARTICLE

PROTEOMICS

Tracking cancer drugs in living cellsby thermal profiling of the proteomeMikhail M Savitski1dagger Friedrich B M Reinhard1dagger Holger Franken1 Thilo Werner1

Maria Faumllth Savitski1 Dirk Eberhard1 Daniel Martinez Molina2 Rozbeh Jafari2

Rebecca Bakszt Dovega2 Susan Klaeger34 Bernhard Kuster34 Paumlr Nordlund25

Marcus Bantscheff1 Gerard Drewes1

The thermal stability of proteins can be used to assess ligand binding in living cellsWe havegeneralized this concept by determining the thermal profiles of more than 7000 proteins inhuman cells by means of mass spectrometry Monitoring the effects of small-molecule ligandson the profiles delineated more than 50 targets for the kinase inhibitor staurosporineWeidentified the heme biosynthesis enzyme ferrochelatase as a target of kinase inhibitors andsuggest that its inhibition causes the phototoxicity observed with vemurafenib and alectinibThermal shifts were also observed for downstream effectors of drug treatment In live cellsdasatinib induced shifts inBCR-ABLpathway proteins includingCRKCRKLThermal proteomeprofiling provides an unbiasedmeasure of drug-target engagement and facilitates identificationof markers for drug efficacy and toxicity

The complete determination of the proteomicstate of a cell referred to as the proteotypelinks genotype and the cellular phenotypeThus the proteotype of the target cell ortissue represents an important context for

the study of drug action However the task ofaccurately describing a proteotype remains daunt-ing because of the inherent complexity of theproteome comprising in excess of 10000 geneproducts expressed at levels differing over sixor more orders of magnitude many of whichoccur in different splice isoforms andwith differ-ent posttranslational modifications at differentsubcellular localizations (1) Current expressionproteomicsmethods describe proteotypes as listsof proteins and semiquantitative expression lev-els (2) In addition there are methods for thecell-wide assessment of some but not all post-translational modifications (3 4) Studies focusedon the mechanism of bioactive compounds suchas small-molecule drugs or peptides frequentlyrely on affinity-based enrichment strategies toidentify prospective binding partners from a cellextract (5ndash7) These approaches can be combinedfor the differential study of cellular phenotypesmdashfor instance the status of signaling pathways orfor biomarker discovery (8 9) However unbiased

approaches for the cell-wide assessment of pro-tein state and protein function are not availableChanges in the thermal stability of proteins are

frequently used to study ligand binding (10 11)and with the recent development of the cellularthermal shift assay (CETSA) can now be observedin living cells (12) enabling the monitoring oftarget engagement which is a key parameter indrug discovery (13) We extended this approachto address two challenges critical in drug dis-covery target and off-target identification anddiscovery of molecular biomarkers for drug effi-cacy By combining theCETSAmethodwithmulti-plexed quantitative mass spectrometry (MS) weestablished the proteome-wide determination ofprotein thermal stability in intact cells as an in-dependent and complementary strategy for thecharacterization of cellular proteotypes Our ap-proach termed ldquothermal proteome profilingrdquoincludes but is not limited to the monitoring ofprotein-ligand interactions directly in cells or tis-sues because in theory any modification of aprotein can affect its thermal stability (14) Herewe demonstrate that monitoring thermal stabil-ity across cellular proteomes in different statessuch as under drug treatment enables the identi-fication of direct physical interaction partners anddownstream effectors asmarkers of target engage-ment and drug efficacy

Proteome-wide profiling of proteinthermal stability

Tomonitor the thermal stability of proteins across10 different temperatures we used the recentlydevelopedneutron-encoded isobaricmass taggingreagents (TMT10) (15) in conjunction with high-resolution MS This allowed acquisition of fullmelting curves for a large proportion of expressedsoluble proteins in a single liquid chromatographyndash

MS (LC-MS)MS experiment In a typical exper-iment cells were cultured under differential con-ditions such as drug treatment (Fig 1) For eachcondition the cells were divided into 10 aliquotseach of which was briefly heated to a differenttemperature followedby extractionwith phosphate-buffered saline (PBS) Around their intrinsic melt-ing temperature proteins in the cell denature andsubsequently aggregate (16) resulting in their grad-ual disappearance from the PBS-extracted sampleswith increasing temperatures (12 17) This proto-col is only suitable for the soluble fraction of theproteome because membrane proteins are notsolubilized under these conditions After extrac-tion each sample was trypsinated and labeledwith a different isotope-coded isobaric mass tag(15) and the 10 samples from each conditionweremixed and analyzed by means of LC-MSMS Thereporter ion intensities acquired in the MSMSfragment spectra were used to fit a curve andcalculate a specific melting temperature (Tm) foreach protein which was then compared betweenthe vehicle-treated and drug-treated samples Theresulting curves that display the increase in pro-tein aggregationwith temperature inmany casesdirectly reflect the underlying unfolding or ldquomelt-ingrdquo event For many proteins a good correlationbetween the thermofluor unfolding assay andheat-induced precipitation in solution or in cellsby CETSA has been demonstrated (11 17) sug-gesting that properly shaped curves represent realunfolding events Hence ligand-induced curveshifts indicate a change in the thermal stabilityof the protein rather than a change in the aggre-gation properties In a variation of the cell-basedprotocol thermal stability can also be assessed ina cell extract This alternative strategy avoids theseparate extraction of each cell sample potentiallyallowsmore controlled conditions (12) and in con-junction with cell treatment experiments enablesdistinguishing Tm shifts induced by ligand bindingfrom those induced by downstreammodifications

The thermal profile of a cellular proteome

We acquired quantitative thermal stability datafor 5299 proteins across 10 different tempera-tures from the human K562 chronic myeloid leu-kemia cell line This data set could be regardedas the first description of the melting proteome(ldquomeltomerdquo) of a human cell Thermal profileswere acquired in two experimental settings inwhich either intact cells or cell extracts wereheated In both settings we noted a weak butsignificant anticorrelation of the thermal stabil-ity with molecular weight because smaller pro-teins tend to bemore stable (fig S1) A comparisonof the thermal stability across a set of 3204proteins robustly quantified in both intact cellsand cell extract revealed marked differences inmelting properties between cells and extract(Fig 2 and table S1) Hierarchical cluster analysisof the temperature-dependent relative protein con-centrations in the heated cell samples revealeda group of proteins that exhibited an increasein concentration at ~50degC followed by a pro-nounced decrease at ~56degC and another groupof proteins with a concentration increase at ~63degC

RESEARCH

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-1

1Cellzome GmbH Molecular Discovery ResearchGlaxoSmithKline Meyerhofstrasse 1 Heidelberg Germany2Division of Biophysics Department of Medical Biochemistryand Biophysics Karolinska Institutet Stockholm Sweden3Department of Proteomics and Bioanalytics TechnischeUniversitaumlt Muumlnchen Emil Erlenmeyer Forum 5 FreisingGermany 4German Cancer Consortium German CancerResearch Center Heidelberg Germany 5Centre forBiomedical Structural Biology Nanyang TechnologicalUniversity SingaporeCorresponding author E-mail mikhailmsavitskigskcom(MMS) marcusxbantscheffgskcom (MB) gerardcdrewesgskcom (GD) daggerThese authors contributed equally to this work

The concentration increases observed at thesetemperatures indicate increased solubilization atthe initial temperatures followed by aggregationat higher temperatures Gene ontology analysisindicated that the proteins in these clusters arereleased from disintegrating organelles (suchas mitochondria) or large protein assemblies(such as ribosomes) (Fig 2A and fig S2) Whencell extract rather than intact cells was heatedwe did not observe any substantial increases inprotein concentration by heating likely becauseof the disruption of cellular structures duringcell lysis Melting points were determined forproteins passing curve-fitting quality criteria asdescribed in the supplementary materials ma-terials and methods (Fig 2B) On average pro-teins showed 27degC higher Tm values in cell extractas compared with intact cells Although this dif-ference is significant (SEM = 006degC P lt 001t test) the correlation of melting points acquiredin cell extracts and intact cell experiments (R2 =03) is considerably lower than the correlation ofthe respective replicate experiments (intact cellR2 = 093 cell extract R2 = 089) suggesting thatthe disruption of the cellular context heteroge-neously affects protein stability (Fig 2B) Theglobal trend of increased protein stability in cellextract suggests an effect on protein precipita-tion due to lower protein concentrations in cellextract as compared with intact cells This con-trasts with the hypothesis that molecular crowd-

ing in the cytosol should exert a stabilizing effectwhich was based on the observation that phos-phoglycerate kinase exhibits increased stabilityinside the cell (18) This is in line with our datafor this protein but could also be explained by astabilizing effect of the endogenous cosubstrateadenosine 5prime-triphosphate (ATP) (fig S3) Wespeculated that proteins that are bound to aligand inside the cell might show an increase inthermal stability contrary to the general trendbecause the dissociation of protein-ligand com-plexes during extraction would result in theirdestabilization This hypothesiswas substantiatedby an analysis of the Tm values of a set of 440proteins annotated asATPbinders which showeda significant trend (P lt 001 t test) toward in-creased stability in intact cells when comparedwith all other proteins (table S2) This was ex-perimentally confirmed with the determinationof thermal shifts induced by the addition ofMgATP to a K562 cell extract at approximatelyphysiological concentrations (2mM) (19) In thesethermal profiles we detected 213 proteins anno-tated as ATP binders that in two independentreplicates showed the predicted trend toward in-creased stability upon addition of ATP whereasall other proteins detected in these experimentsdid not show this trend (fig S4 and table S3)Further analysis of the thermal proteome pro-files revealed that proteins categorized as DNAbinders were enriched in the set with the lowest

melting points suggesting that these proteinswere extracted without their DNA ligand In-deed the DNA-binding properties of proteinscan be assessed by adding DNA fragments to thecell extract This is exemplified by p53 the globaltranscription regulator for stress Wild-type p53was stabilized upon addition of its cognate ef-fector DNA whereas the p53 R273Hmutant thatdoes not bind these effector sequences (20) wasnot stabilized (fig S5) These results demonstratethe potential of thermal proteome profiling forthe large-scale analysis of proteome-ligand inter-actions including endogenous ligands such ascofactors or metabolites

Monitoring of drug effects on thermalproteome profiles

The analysis of drugmechanism of action by pro-filing both targeted and off-target protein bind-ing will likely represent a major application ofthermal proteome profiling As a proof of prin-ciple we studied kinase inhibitors because kinasesare an important target class with key regulatoryfunctions in cellular pathways An important ques-tion in drug target identification is the prevalenceof false-positive and false-negative identificationsTo investigate the reliability of thermal proteomeprofiling we selected two structurally divergentpromiscuous kinase inhibitors with a known spec-trum of targets staurosporine and GSK3182571which is a close analog of CTx-0294885 (21) Profiles

1255784-2 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 1 Quantitative proteome-wide profiling of protein thermal stability under differentialconditions Cells were cultured under differential conditions such as drug treatment In analternative method the cells were extracted first and the extracts were treated with drug (ldquo1rdquo)

For each condition the cell or cell-extract sample was divided into 10 aliquots (ldquo2rdquo) Aliquots were subjected to heating at the indicated temperatures Samples ofintact cells were subsequently subjected to extraction with PBS (ldquo3rdquo) After digestion with trypsin each sample was labeled with a different TMT10 isotope tag(ldquo4rdquo) Subsequently all samples fromeach conditionweremixed (ldquo4rdquo) and analyzed bymeans of LC-MSMS (ldquo5rdquo)The obtained reporter ion intensitieswere usedto fit a melting curve and calculate the melting temperature Tm of each protein separately for the two conditions (ldquo6rdquo)

RESEARCH | RESEARCH ARTICLE

of K562 cell extract treated with staurosporine orvehicle were generated as two independentreplicates (Fig 3) Of all proteins 92 detectedyielded curves with sufficiently steep slopes toallow the robust identification of ligand-inducedTm shifts However a shallow slope of the melt-ing curve indicated lowermelting point reprodu-cibility (Fig 3A) Whereas most affected proteinsshowed positive shifts consistent with ligand-

induced stabilization a few proteins includingmembers of the protein kinase C family exhib-ited negative shifts whichmay result fromdesta-bilization In total the thermal profiles comprised175 protein kinases of which 51 displayed Tmshifts that passed our significance criteria (Fig3B fig S6 and table S4) A recent comprehensiveanalysis of staurosporine targets identified bymeans of chemoproteomics ldquokinobeadsrdquo profil-

ing (22) detected 229 kinases of which 92 werealso present in both biological replicates of thethermal profiles experiments The limited degreeof overlap may be due to the fact that kinobeadswill fail to identify kinases that do not bind to theimmobilized ligands whereas thermal profilingwill miss proteins owing to insufficient abun-dance andor solubility or the absence of a sig-nificant ligand effect Analysis of the 92 kinasescommon to both data sets revealed that kino-beads identified 66 kinases with a staurosporinemedian inhibitory concentration (IC50) below10mMof which 34 showed a significant Tm shift andan additional 15 showed a reproducible shift ofgreater than 1degC which could become significantwith additional replicates For some proteins weobserved shallow or irregular curves (for exam-ple BMP2K AAK1 in fig S6 G and J respectively)that may reflect multiple transitions possibly be-cause of independent folding domains and aredifficult to fit with our current methods For theremainder of 17 kinases identified as targets bykinobeads we observed very small (lt1degC) or no Tmshifts (Fig 3C) These targets may represent falsenegatives We also observed thermal shifts for pro-teins other than kinases Staurosporine inducedapparent stabilization of coproporphyrinogen-III oxidase and ferrochelatase (FECH) two out ofthe eight enzymes in the heme biosynthesispathway for all of which melting curves wereobtained (Fig 3D and fig S7) Ligand-bindingalso affected the thermal stability of the regu-latory components of some protein complexesmdashfor instance kinase complexes containing cyclins(fig S6 C and J) We further explored this in thecontext of the protein kinase A (PKA) complex(23) Inhibition by staurosporine appeared tostabilize the catalytic subunit but destabilize theregulatory subunit In contrast the addition ofadenosine 3prime5prime-monophosphate (cAMP) whichcauses dissociation of the regulatory subunit fromthe catalytic subunit appeared to stabilize theregulatory subunit but destabilize the catalyticsubunit presumably because of the release of theregulatory subunit (Fig 3E)Comparison of staurosporine Tm shift data

with kinobeads-derived potencies for multipletargets (Fig 3C) showed no significant correla-tion between the affinity for staurosporine andthemagnitude of the Tm shift This is not surpris-ing because the extent of thermal stabilization byligands with a similar binding mode likely de-pends on the stability of the unliganded proteinTo investigate the correlation between ligand af-finity and apparent thermal stabilization we con-sidered a structurally divergent promiscuouskinase inhibitor GSK3182571 (Fig 4 and table S5)We identified 13 targets common to both inhib-itors with an affinity below 1 mM and a less than10-fold difference in affinity For these targetsboth compounds showed similar Tm shifts (R2 =09) (Fig 4A) suggesting that for a given proteinthe Tm shift at saturating ligand concentrationsis dependent on the intrinsic affinity of the ligandDepending on the number of replicate experi-

ments ligand-induced Tm shifts as small as 1degCcan be monitored However to resolve small

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-3

Fig 2 The thermal proteome profile of a human K562 cell (A) Heat map representation of thethermal stability of 3204 soluble proteins in intact cells (left) and cell extract (right) For each protein itsrelative concentration (fold-change) at the indicated temperature compared with the lowest temperature(37degC for intact cells and 40degC for the cell extract) is shown Protein fold-changes from the experiment onintact cells were clusteredTo allow better comparison proteins heated in the cell extract were plotted inthe same order as the proteins heated in intact cells (B) Reproducibility of thermal proteome profilesassessed with replicate experiments on intact cells R2 = 093 (left) cell extract R2 = 089 (middle) anddirect comparison of proteomes from intact cells and cell-extract experiments R2 = 030 (right) Mostproteins showed greater thermal stability (higher Tm values) in cell extract as compared with intact cells(C) ATP-binding proteins show a trend toward increased stability in intact cells as compared with cellextractThe plots showdensity distributions of proteinTmvalues determined in intact cells or in cell extractand the density distribution of the difference between both settings A set of 440 proteins annotated asATP binders show a trend toward increased stability in intact cells The difference in the means of thedistributions was significant (P lt 001 t test)

RESEARCH | RESEARCH ARTICLE

changes in ligand-induced thermostability andto enable the ranking of binding affinities amongmultiple targets we tailored the isothermal dose-response (ITDR) protocol to a defined set of pro-teins (12) In contrast to the full temperatureprofiles in which a compound is assayed at asingle typically saturating concentration across10 temperatures an ITDR profile is generated ata defined temperature over a range of compoundconcentrations To generate affinity data forGSK3182571 we acquired ITDR profiles (tableS6) which showed excellent reproducibility (figS8) and were in good agreement with data fromkinobeads competition-binding experiments bothfor proteins stabilized and proteins destabilizedby ligand binding (Fig 4B and table S7)

Because of its unbiased nature thermal pro-teome profiling can detect unexpected off-targetsof drug compounds The heme biosynthesis path-way has not been related to adverse drug effectsbut is linked to a number of genetic disordersincluding erythropoietic protoporphyria which iscaused by a deficiency in FECH and results inhigh tissue levels of protoporphyrins (24) Wefound two enzymes in the heme pathway inter-actingwith staurosporine and thus askedwhetherdrug interactions with this pathway might resultin adverse effects such as skin or liver disease (25)The melanoma drug vemurafenib frequentlycauses severe photosensitivity concomitant withincreased levels of protoporphyrins (26) ITDRprofiling of cells treatedwith vemurafenib showed

that the main proteins affected were its knowntarget BRAF [negative logarithm of the medianeffective concentration (pEC50) = 61] and FECH(pEC50 = 53) (Fig 5A and table S8) These con-centrations are below the typical steady-state ex-posure for this drug (27) and indicate a less than10-fold window of target occupancy between thedrugrsquos efficacy target and FECH Alectinib asecond-generation anaplastic lymphoma kinase(ALK) inhibitor for the treatment of nonndashsmall-cell lung cancer was also reported to cause photo-sensitivity in contrast to the first-generation drugcrizotinib (28 29) ITDR profiling revealed thatalectinib more potently affected FECH than didvemurafenib whereas crizotinib had no effect(Fig 5B) These results strongly suggest that the

1255784-4 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 3 Differential profiling of drug effects on the thermal proteomeprofile Staurosporine treatment of cell extract yields reproducible thermalshifts allowing robust target identification (A) Cell extract was treated withvehicle or staurosporine Both experiments were performed as two fullyindependent replicates A flat slope of the melting curve relates to lowermelting point reproducibility Melting point differences from the two vehicleexperiments are plotted against the shallowest melting curve slope observedin the four vehiclestaurosporine data sets Proteins with an absolute slopebelow 006 are plotted in gray The histogram shows the distribution ofproteins versus slope values Of all proteins 92 yielded curves with suf-ficiently large slopes (B) Scatter plot of Tm shifts calculated from the two

replicates of the staurosporine versus vehicle treatment experimentStaurosporine-induced Tm shifts that passed the significance criteria areshown in red Proteins with flat slopesmdashas shown in gray in (A)mdashare againshown in gray (C) Comparison between Tm shifts and pIC50 values forstaurosporine as reported from kinobeads data (22) (D) Examples ofmeltingcurves for PKN1 FECH AURKA and SLK with and without staurosporinetreatment (E) The catalytic subunit of PKA is stabilized by staurosporinewhereas the regulatory subunit is destabilized Addition of cAMP followed byWestern blot detection revealed destabilization of the catalytic subunit andstabilization of the regulatory subunit Error bars indicate the SEM from n = 4experiments

RESEARCH | RESEARCH ARTICLE

photosensitivity induced by these drugs is me-diated by FECH and demonstrates that thermalproteome profiling can serve as a stand-alone tech-nique for obtaining quantitative affinity data fortarget-ligand interactions in a cell-based setting

Drug treatment induces Tm shifts fordownstream effector proteins

Any modification of a protein can in principleaffect its thermal stability Therefore compari-son of Tm shifts in cell extract where ligand bind-ing but no downstream effects occur with Tmshifts in intact cells where active signaling takesplace might reveal effector proteins downstreamof the target As a model system we used K562

cells which are transformed by an oncogenic fu-sion protein the BCR-ABL kinase (30) Overallthe samples from intact cells yielded a similarprotein coverage as that of samples from cellextracts (tables S9 and S10) Cultured cells weretreated with vehicle or with the ABL inhibitordasatinib a marketed drug against chronic mye-logenous leukemia (CML) (31) at two concen-trations 05 and 5 mM and the samples wereprofiled (Fig 6A and table S9) Theminimum Tmshift for each protein was calculated from a pairof replicate experiments Four proteins displayedsignificant melting point differences in both the05 and 5 mM dasatinibndashtreated cell samples Wedid not observe a Tm shift for the direct target

BCR-ABL suggesting that this protein is notstabilized by dasatinib binding However sub-stantial and significant Tm shifts were observedfor the adaptor proteinCRKLand thephosphataseSHIP2 (Fig 6B and fig S9) which is also known asINPPL1 These proteins are not likely binding thedrug directly and consistent with this no Tmshift was observed for either protein in cell extracts(table S10) In addition the BCR-ABLndashinteractingproteinCRKshowedapronounced change inmeltingcurve slope in the dasatinib-treated cells (fig S9)CRK and CRKL aremembers of a proto-oncogenefamily that bind to tyrosine-phosphorylated pro-teins and are known downstream effectors ofBCR-ABL signaling (32) CRKLhas been previouslyproposed as a treatment response biomarker (33)Comparison of the CRKL melting curves in K562cells with those in Jurkat cells (table S11) whichare not transformed by tyrosine kinase signalingrevealed that dasatinib treatment of K562 cellsreverted the thermal stability of CRKL to the lev-el observed in Jurkat cells (Fig 6C) and the sameeffect was observed for CRK (fig S10) The Jurkatcell profiles also revealed that ABL1 kinase ap-peared to be more thermostable compared withthe BCR-ABL fusion protein in K562 cells suggest-ing that thermal profiling could be used to identifyprotein fusions which are difficult to identify byexpression proteomics (Fig 6C) In order to eval-uate the drug concentration dependence of theTm-shifted effector proteins we performed ITDRprofiles at three different temperatures (Fig 6Dfig S11 and table S12) The concentrations elicit-ing half-maximal response of the CRKL markerin the ITDR profiles recorded at three differenttemperatureswere between 15 and 32 nMwhichis in good agreement with the known potency ofdasatinib for the inhibition of cell growth (34) In-depth analysis of the 50degC ITDR data revealedadditional markers that either showed small Tmshifts or in the case of the known downstreameffector CRK a pronounced change in the melt-ing curve slope indicating that the ITDRprofiles canoffer greater sensitivity than the full-temperatureprofiles (fig S11) It is tempting to speculate thatthe most sensitive biomarkers are those thatexhibit large or even stoichiometric changes inprotein state which should be robustly detectedby their Tm shift This contrasts with the use of a

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-5

Fig 4 Structurally unrelated kinase inhibitors with sim-ilar affinities induce Tm shifts of comparable magnitudebut assessment of ligand affinity requires ITDR mea-surements (A) Staurosporine Tm shift data are comparedwith data for a structurally unrelated promiscuous kinaseinhibitor GSK3182571 Comparison of Tm shifts induced bystaurosporine or GSK3182571 determined in cell extractProtein targets shown had a pIC50 gt 6 both for staurosporineand GSK3182571 as determined bymeans of kinobeads anddid not differ in affinity between the two compounds bymorethan one order of magnitude (B) Good agreement betweenGSK3182571 pEC50 values determined by means of ITDR with pIC50 values determined with kinobeads(Top) Proteins stabilized by GSK3182571 treatment (Bottom) Proteins destabilized by GSK3182571treatment

Fig 5 The clinical kinase drugsvemurafenib and alectinib which can causephototoxicity as a side effect induce Tmshifts in the heme biosynthesis enzymeFECH (A) Concentration-dependent targetoccupancy of vemurafenib and its cognatetarget BRAF compared with the off-targetFECH ITDR profiling was performed at 55degCwith vemurafenib-treated K562 cells andshowed concentration-dependent thermalstabilization of FECH and BRAF The data arenormalized so that the quantity of solubletarget at the highest compoundconcentration which reflects maximumthermal stabilization is set to 1 and fixed forcurve-fitting (B) ITDR performed at 55degC with K562 cells treated with vemurafenib alectinib or crizotinib over a range of concentrations Alectinib displays amore potent effect on FECH as compared with that by vemurafenib whereas crizotinib a drug not known to cause photosensitivity has no effect

RESEARCH | RESEARCH ARTICLE

posttranslational modification as a biomarkerwhere it is usually difficult to assess the stoi-chiometry of the modificationIn general terms the thermal stability of a

protein may be defined by its mutational statusposttranslationalmodifications and bound ligandssuch as other proteins cofactors metabolites ordrugs Fusionproteins splice variants andposttrans-lational modifications are typically undersampledand therefore not comprehensively detected inMS-basedproteomicsHence the thermal proteomeprofile of a human cell can provide a generalview of the proteomic state or proteotypeThe future scope and applicability of thermal

proteome profilingwill substantially benefit fromcontinued advances in the sensitivity and accu-racyofMS-basedproteomics Futuredevelopmentsshould include the development of protocols formembrane proteins and the application to tis-sues from animal studies or clinical biopsies

Materials and methods

Reagents and cell culture

Reagents andmedia were purchased from Sigma-Aldrich (St Louis MO) unless otherwise notedPBS was prepared by using 137 mM NaCl 27mM KCl 10 mM Na2HPO4 and 2 mM KH2PO4

to buffer at pH 74 and supplemented withcomplete EDTA-free protease inhibitor cocktail(one tablet per 25 ml) (Roche Diagnostics Basel

Switzerland) Staurosporine was from Calbiochem(Millipore Billerica MA) vemurafenib andcrizotinib from Selleck (Houston TX) alectinibfromMedchemExpress (Monmouth Junction NJ)and dasatinib from Toronto Research (TorontoCanada) The synthesis of GSK3182571 is de-scribed in the supplementary methods DNAfragments PG1 (5prime-AGC TTA GAC ATG CCT AGACAT GCC TA -3prime) and PG2 (5prime-AGC TTA GGC ATGTCT AGG CAT GTC TA -3prime) were from Life Tech-nologies (Carlsbad CA) and dissolved in 100 mMTris (pH75) 1 mM NaCl and 10mM EDTA K562Jurkat E61 and A549 cells were from ATCC K562and A549 cells were cultured in RPMI mediumcontaining 10 fetal calf serum (FCS) and JurkatE61 cells in RPMI1640 supplemented with45 gl glucose 10 mM HEPES 1 mM sodiumpyruvate and 10 FCS HT-29 (ATTC no HTB-38)were maintained in Dulbeccorsquos Modified EagleMedium Culture media were supplemented with03 gL L-glutamine and 10 fetal bovine serum(FBS) (Gibco Carlsbad CA) 100 unitsmL peni-cillin and 100 unitsmL streptomycin The cellswere expanded to amaximumof 2times 106 cellsml incase of K562 and 107 cellsml in case of Jurkat E61

Preparation of cell extract forCETSA with Western blot read-out

A549 andHT-29 cells were harvested andwashedwith PBS The cells were diluted in KB buffer

(25mMTris-HCl supplemented with 5mMbeta-glycerophosphate 2 mM DTT 01 mM Na3VO410 mM MgCl2 and complete protease inhibitorcocktail) The cell suspensionswere freeze-thawedthree times by using liquid nitrogen The solublefraction was separated from the cell debris bycentrifugation at 20000 g for 20 min at 4degC FortheCETSAmelting curve experiments cell extractswere diluted with KB buffer and divided into twoaliquots with one aliquot being treated withligandDNA fragments depending on the targetand the other aliquot treated with the solvent ofthe ligand (control) After 10 min incubation atroom temperature the respective cell extractswere heated individually at different temper-atures for 3 min in a thermal cycler [AppliedBiosystems (Foster City CA)Life Technologies]followed by cooling for 3 min at room temper-ature The heated cell extracts were centrifugedat 20000 g for 20min at 4degC in order to separatethe soluble fractions from precipitates The su-pernatantswere transferred to new02mLmicro-tubes and analyzed by means of Western blotanalysis NuPage Bis-Tris 4-12 polyacrylamidegels with MES SDS running buffer (Life Tech-nologies) were used for separation of proteins inthe samples Proteins were transferred to nitro-cellulose membranes by using the iBlot system(Life Technologies) Primary antibodies anti-PKARegulatory I-a (sc-136231) anti-PKA Catalytic-a

1255784-6 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 6 Treatment of K562 cells with dasatinib induces Tm shifts fordownstream effector proteins in the BCR-ABL pathway (A) Cell-wideassessment of Tm shifts induced by dasatinib (05 and 5 mM) in K562 cellsThe minimum Tm shift for each protein was calculated from a pair of fullreplicate experiments Four proteins showing significant melting pointdifferences both in the 05 and 5 mM dasatinib data sets are marked in red(B) Effect of dasatinib on the melting curves for the effector protein CRKL

determined in two biological replicates in intact cells (top) and cell extracts(bottom) (C) The melting curve for CRKL in Jurkat cells closely matchesthe curve obtained in dasatinib-treated K562 cells (top)The ABL1 protein inJurkat cells is much more thermostable compared with the BCR-ABL fusionprotein in K562 cells (D) ITDR curves for CRKL in dasatinib-treated intactcells Experiments were performed at 47degC (green curve) 50degC (purplecurve) and 56degC (blue curve)

RESEARCH | RESEARCH ARTICLE

(sc-28315) anti-p53WT and R273H (sc-126) sec-ondary goat anti-mouse HRP-IgG (sc-2055) andbovine anti-goat HRP-IgG (sc-2352) antibodies(Santa Cruz Biotechnology Dallas TX) were usedfor immunoblotting Chemiluminescence inten-sities were detected and quantified by using aChemiDocXRS+ imaging systemwith Image Labsoftware (Bio-Rad Hercules CA) Data were ex-pressed as means T SEM The data (band inten-sities) from multiple runs (n ge 3) were plottedwith Graphpad Prism software

Preparation of cell extract for thermalproteome profiling

Suspension cultures of K562 cells (15 to 2 times 106

cellsml) or of Jurkat E61 cells (4 times 106 cellsml)were centrifuged at 340 g for 2 min at 4degC Thecells were resuspended in 50 ml PBS After a sec-ond centrifugation step as above the cells wereresuspended in 10 ml ice-cold PBS and centri-fuged again at 340 g for 2 min at 4degC The cellswere resuspended in 15 ml of ice-cold PBS andsnap-frozen in liquidnitrogen The tubewasplacedinto a water bath at 23degC until ~60 of the con-tent was thawed and then transferred on ice un-til the entire contentwas thawed This freezethawcycle was repeated twice before the samplesweresubjected to ultracentrifugation (20 min at 4degCand 100000 g) The protein concentration of thesupernatant was determined by Bradford assay(Bio-Rad) and aliquots were snap-frozen in liquidnitrogen for use in subsequent thermal shiftassays

Thermal profiling using cell extract

A solution of the compound in dimethyl sulfoxide(DMSO) or DMSO alone as vehicle (Table 1) wasadded to the cell extract to 1 final DMSO con-centration The extract was then incubated for10 min at 23degC divided into 10 aliquots of 100 mland transferred into 02-ml polymerase chain re-action (PCR) tubes One each of the compoundand of the vehicle containing samples was heatedin parallel for 3 min to the respective tempera-ture followed by a 3-min incubation time at roomtemperature Subsequently the extract was cen-trifuged at 100000 g for 20 min at 4degC The su-pernatant was fractionated by means of SDS gel

electrophoresis and subjected to sample prepa-ration for MS analysis For ITDR experimentsGSK3182571 was tested as a nine-point serial dilu-tion starting at 100 mM (resulting in concentra-tions of 100 333 111 370 1235 0412 0137 0046and 0015 mM) including one vehicle control inan isothermal dose-response experiment at 53degCfollowing the procedure outlined above for ex-periments in cell extract

Thermal profiling using intact cells

To 24 ml of a suspension of K562 cells (at a den-sity of 15 106 cellsml) or Jurkat E61cells (at adensity of 4 106 cellsml) in a T-flask a volumeof either 120 ml of a solution of the referencecompound dissolved in DMSO or an equivalentamount of DMSO (vehicle) was added Incubationof cellswith compound and vehiclewas conductedin parallel for the duration of 1 hour at 37degC and5CO2 Cells were pelleted at 340 g and 4degC for2min resuspended in 20ml of ice cold PBS andcentrifuged again as indicated above This wash-ing step was repeated once and the cells werecarefully resuspended in 1200 ml PBS 100 ml ofthis resulting cell suspension was transferredinto 02-ml PCR tubes and subjected to a centrifu-gation step at 325 g for 2 min at 4degC Using apipette 80 ml of the PBS supernatant was re-moved By gently tapping the tubes the cells wereresuspended and one each of the compound andof the vehicle containing tubes was heated inparallel in a PCR machine for 3 min to the re-spective temperature (37degC to 67degC) followed bya 3-min incubation time at room temperatureThereafter the cells were snap-frozen in liquidnitrogen for 1 min The cells were thawed brieflyin a water bath at 25degC transferred on ice andresuspended by using a pipette This freeze-thawcycle was repeated once After an addition of fur-ther 30 ml of PBS and resuspension the entirecontent was centrifuged at 100000 g for 20 minat 4degC After centrifugation 30 ml of the super-natant were transferred into a new tube Proteinconcentration of the 37degC and 41degC samples weredetermined and used to normalize loading ofSDS gels Proteins in the supernatant were dena-tured by using SDS sample buffer partially sepa-rated by means of SDS gel electrophoresis and

subjected to sample preparation for MS analysisFor ITDR experiments vemurafenib alectiniband crizotinib were used as a nine-point serialdilution starting at 30 mM (concentrations were30 12 48 192 077 031 012 005 and 002 mM)including one vehicle control at 55degC followingthe procedure outlined aboveDasatinibwas testedas an eight-point serial dilution starting at 1 mM(concentrations usedwere 1 033 0111 0037 00120004 00013 and 000046 mM) including onedasatinib control at 10 mM and one vehicle con-trol on intact cells at 47degC 50degC and 56degC follow-ing the procedure outlined above for experimentson intact cells (Table 1)

Kinobeads assays

Competition binding assays were performed asdescribed previously by using a modified beadmatrix (5 22) Briefly 1 ml (5 mg protein) cellextract was pre-incubatedwith test compound orvehicle for 45 min at 4degC followed by incubationwith kinobeads (35 ml beads per sample) for1 hour at 4degC The beads were washed with lysisbuffer and eluted with 50 ml 2times SDS sample buf-fer GSK3182571was tested at 10 25 0625 015625003906 000977 and 000244 mM

Sample preparation for MS

Gel lanes were cut into three slices covering theentire separation range (~2 cm) and subjectedto in-gel digestion (5) Peptide samples were la-beled with 10-plex TMT (TMT10 Thermo FisherScientific Waltham MA) reagents enabling rel-ative quantification of a broad range of 10 tem-perature points in a single experiment (15 22)For kinobeads and ITDR experiments either10 or 8 TMT labels were used depending on thecompound concentration range chosen Thelabeling reaction was performed in 40 mMtriethylammoniumbicarbonate pH 853 at 22degCand quenched with glycine Labeled peptide ex-tracts were combined to a single sample per ex-periment and as indicated in the supplementarymaterials for most samples additional fraction-ation was performed by using reversed-phasechromatography at pH 12 [1mmXbridge column(Waters Milford MA)] as previously described(table S13) (35)

LC-MSMS analysis

Samples were dried in vacuo and resuspended in005 trifluoroacetic acid in water Of the sam-ple 50 was injected into an Ultimate3000nanoRLSC (Dionex Sunnyvale CA) coupled to aQ Exactive (Thermo Fisher Scientific) Peptideswere separated on custom 50 cm times 100 mM (ID)reversed-phase columns (Reprosil) at 40degC Gra-dient elutionwas performed from 2 acetonitrileto 40acetonitrile in 01 formic acid over 2 hoursSamples were online injected into Q-Exactive massspectrometers operating with a data-dependenttop 10methodMS spectrawere acquired by using70000 resolution and an ion target of 3E6Higherenergy collisional dissociation (HCD) scans wereperformed with 35 NCE at 35000 resolu-tion (at mz 200) and the ion target settingswas set to 2E5 so as to avoid coalescence (15) The

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-7

Table 1 Compoundsused in thermal profilingDashes indicate that compoundwas not tested in intact cells

CompoundConcentration used in

cellular extractexperiments

Concentration(s) inintact cell experiments

Cell line

MgATP 2 mM mdash K562staurosporine 20 mM mdash K562cAMP 1 mM mdash K562GSK3182571 20 mM mdash K562dasatinib 5 mM 05 mM 5 mM K562p53 WT

PG1 100 mM mdash A549PG2 100 mM mdash A549

p53 R273HPG1 100 mM mdash HT-29PG2 100 mM mdash HT-29

RESEARCH | RESEARCH ARTICLE

instruments were operated with Tune 22 or 23and Xcalibur 27 or 3063

Peptide and protein identification

Mascot 24 (Matrix Science Boston MA) wasused for protein identification by using a 10 partspermillionmass tolerance for peptide precursorsand 20 mD (HCD) mass tolerance for fragmentions Carbamidomethylation of cysteine residuesand TMTmodification of lysine residueswere setas fixedmodifications andmethionine oxidationand N-terminal acetylation of proteins and TMTmodification of peptide N-termini were set asvariable modifications The search database con-sisted of a customized version of the Interna-tional Protein Index protein sequence databasecombined with a decoy version of this databasecreated by using a script supplied by Matrix Sci-ence Unless stated otherwise we accepted proteinidentifications as follows (i) For single-spectrumto sequence assignments we required this assign-ment to be the bestmatch and aminimumMascotscore of 31 and a 10times difference of this assign-ment over the next best assignment Based onthese criteria the decoy search results indicatedlt1 false discovery rate (FDR) (ii) For multiplespectrum to sequence assignments and using thesame parameters the decoy search results indi-cate lt01 FDR All identified proteinswere quan-tified FDR for quantified proteins was lt1

Peptide and protein quantification

Reporter ion intensities were read from raw dataandmultiplied with ion accumulation times (theunit is milliseconds) so as to yield ameasure propor-tional to thenumberof ions thismeasure is referredto as ion area (36) Spectra matching to peptideswere filtered according to the following criteriamascot ion score gt15 signal-to-background of theprecursor ion gt4 and signal-to-interference gt05(37) Fold-changes were corrected for isotope pu-rity as described and adjusted for interferencecaused by co-eluting nearly isobaric peaks as esti-mated by the signal-to-interferencemeasure (38) Pro-tein quantification was derived from individualspectra matching to distinct peptides by using asum-based bootstrap algorithm 95 confidenceintervals were calculated for all protein fold-changesthatwere quantifiedwithmore than three spectra (36)

Curve fitting normalization andestimation of slope and melting pointdifferences for CETSA experiments

Melting curves

For all CETSA experiments fold-changes werecalculated by using the lowest temperature con-dition as the reference Relative fold changes as afunction of temperature followeda sigmoidal trendwhich was fitted with the following equationderived from chemical denaturation theory (39)(enwikipediaorgwikiEquilibrium_unfolding)using R (wwwr-projectorg)

f ethT THORN frac14 1 minus plateau

1thorn eminusaTminusbeth THORN thorn plateau

Where T is the temperature and a b and plateauare constants The value of f(T) at the lowest

temperature Tmin was fixed to one The meltingpoint of a protein is defined as the temperatureTm at which half of the protein amount has beendenatured

f(Tm) = 05

Additionally the slope of the melting curve isdefined as the value of the first derivative of f(T)at the point T = Tinfl

slope = f prime(Tinfl)

where Tinfl is the inflection point of the curvemeaning that the second derivative of f (T) equalszero at T = Tinfl

f primeprime(Tinfl) = 0

Normalization

The vehicle and compound-treated experimentswere normalized in the following waySelection of proteins for normalization (i) The

set of proteins identified in both experimentstermed ldquojointPrdquo was determined (ii) For eachexperiment (vehicle or compound) all proteinsfrom jointP fulfilling the following criteria werecollected the fold-changes at the seventh highesttemperature point comparedwith the lowest tem-perature point were between 04 and 06 andat the 9th and 10th highest temperature pointscompared th the lowest temperature pointwere below 03 and 02 respectively (iii) Thebigger of these two protein subsets was thenused for normalization (typically more than200 proteins qualified) This protein set is re-ferred to as ldquonormPrdquoCalculation of correction factors (i) The me-

dian fold-change values over all proteins in normPwere calculated for each of the 10 fold-changesThis was done separately for both experimentsand for each experiment a melting curve wasfitted The curve that was fitted with the best R2

was used to normalize both experiments (ii) Foreach experiment the correction factors werecalculated by using the median fold-changes ofthe proteins in normP For each median fold-change a correction factor by which that fold-change would have to be multiplied with in orderto coincidewith the best-fitting curvewas calculatedNormalization The two sets of correction fac-

tors were applied to all the protein fold-changesin the respective experimentsThis heuristic normalization procedure covers

three important aspects (i) The same subset ofproteins is used to determine normalization co-efficients in both experiments thus circumvent-ing any potential bias caused by proteins onlyidentified in one experiment (ii) The subselectionof late-melting proteins (seventh point between04 and 06) ensures that the normalization ofhigh-temperature points will be more accurate(iii) Last the narrow requirements on the 7th9th and 10th points will ensure that a narrowrange of melting profiles is used for calculatingthe points from which the normalization curvewill be derived This is important because if vastlydifferent melting profiles are combined into asingle one the above equation will not necessar-ily fit well A normalization curve example as well

as melting curves of several proteins before andafter normalization are illustrated in fig S12

Estimation of slope and meltingpoint differences

After normalization melting curves were fittedfor all proteins identified in the vehicle- andcompound-treated experiments Melting pointdifferences and slope differences were calculatedfor proteins whose curves passed the followingtwo requirements (i) both fitted curves for thevehicle and compound treated condition had anR2 of greater than 08 and (ii) the vehicle curvehad a plateau of less than 03Typically close to 80 of proteins that were

identified in both the vehicle- and compound-treated experiment passed these criteria Theslope of the curves had the biggest impact onthe reproducibility of melting point differencesof proteins This observation can be rationalizedin the following way If the curve has a shallowslope then small changes with respect to the ref-erence pointmdashin our case the lowest temperaturemdashwill have a strong impact on where the curvecrosses the 05 fold-change level (melting point)Consequently small variations in the quantitativemeasurement of theMS signals corresponding tothe lowest temperature will lead to strong shiftsin the inferred melting point of the curve Thismeans that it is difficult to get high precisionmeasurements ofmelting point differenceswhencomparing two curves with shallow slopes If thecurve however has a steep slope then the melt-ing point measurement will not be substantiallyaffected by small variations of the quantitativemeasurements of the lowest temperature nor ofany other temperature points Taking this intoaccount we used the following strategy for sig-nificance estimation of protein melting point dif-ferences between vehicle- and compound-treatedexperiments which is conceptually very similarto a broadly used strategy for inferring significantprotein fold-changes (40) The calculated meltingpoint differences of proteins were ordered in des-cending slope order (those with the shallowestslope come first) in which the slope for eachprotein is the lowest (steepest) slope calculatedfrom the two curve fits performed for the proteinin the vehicle- and compound-treated conditionThe proteins were divided into bins The bins areconstructed starting fromproteinswith the high-est (shallowest) slope Each bin consists of at least300 proteins Once each bin has been completedthe remaining number of proteins is counted ifthis number is below 300 the remaining pro-teins are added to the last completed bin Thisdata qualityndashdependent binning strategy is analo-gous to the procedure described in Cox et al withthe only difference being that proteins are sortedby slope instead of abundance (40) The statisticalsignificance of differences in protein meltingpoints was calculated by estimating the left- andright-sided robust standard deviation of the dis-tribution of the measurements using the 158750 and 8413 percentiles and calculating the Pvalues for all measurements for a specific binexactly as previously described (40) Subsequently

1255784-8 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

RESEARCH | RESEARCH ARTICLE

The concentration increases observed at thesetemperatures indicate increased solubilization atthe initial temperatures followed by aggregationat higher temperatures Gene ontology analysisindicated that the proteins in these clusters arereleased from disintegrating organelles (suchas mitochondria) or large protein assemblies(such as ribosomes) (Fig 2A and fig S2) Whencell extract rather than intact cells was heatedwe did not observe any substantial increases inprotein concentration by heating likely becauseof the disruption of cellular structures duringcell lysis Melting points were determined forproteins passing curve-fitting quality criteria asdescribed in the supplementary materials ma-terials and methods (Fig 2B) On average pro-teins showed 27degC higher Tm values in cell extractas compared with intact cells Although this dif-ference is significant (SEM = 006degC P lt 001t test) the correlation of melting points acquiredin cell extracts and intact cell experiments (R2 =03) is considerably lower than the correlation ofthe respective replicate experiments (intact cellR2 = 093 cell extract R2 = 089) suggesting thatthe disruption of the cellular context heteroge-neously affects protein stability (Fig 2B) Theglobal trend of increased protein stability in cellextract suggests an effect on protein precipita-tion due to lower protein concentrations in cellextract as compared with intact cells This con-trasts with the hypothesis that molecular crowd-

ing in the cytosol should exert a stabilizing effectwhich was based on the observation that phos-phoglycerate kinase exhibits increased stabilityinside the cell (18) This is in line with our datafor this protein but could also be explained by astabilizing effect of the endogenous cosubstrateadenosine 5prime-triphosphate (ATP) (fig S3) Wespeculated that proteins that are bound to aligand inside the cell might show an increase inthermal stability contrary to the general trendbecause the dissociation of protein-ligand com-plexes during extraction would result in theirdestabilization This hypothesiswas substantiatedby an analysis of the Tm values of a set of 440proteins annotated asATPbinders which showeda significant trend (P lt 001 t test) toward in-creased stability in intact cells when comparedwith all other proteins (table S2) This was ex-perimentally confirmed with the determinationof thermal shifts induced by the addition ofMgATP to a K562 cell extract at approximatelyphysiological concentrations (2mM) (19) In thesethermal profiles we detected 213 proteins anno-tated as ATP binders that in two independentreplicates showed the predicted trend toward in-creased stability upon addition of ATP whereasall other proteins detected in these experimentsdid not show this trend (fig S4 and table S3)Further analysis of the thermal proteome pro-files revealed that proteins categorized as DNAbinders were enriched in the set with the lowest

melting points suggesting that these proteinswere extracted without their DNA ligand In-deed the DNA-binding properties of proteinscan be assessed by adding DNA fragments to thecell extract This is exemplified by p53 the globaltranscription regulator for stress Wild-type p53was stabilized upon addition of its cognate ef-fector DNA whereas the p53 R273Hmutant thatdoes not bind these effector sequences (20) wasnot stabilized (fig S5) These results demonstratethe potential of thermal proteome profiling forthe large-scale analysis of proteome-ligand inter-actions including endogenous ligands such ascofactors or metabolites

Monitoring of drug effects on thermalproteome profiles

The analysis of drugmechanism of action by pro-filing both targeted and off-target protein bind-ing will likely represent a major application ofthermal proteome profiling As a proof of prin-ciple we studied kinase inhibitors because kinasesare an important target class with key regulatoryfunctions in cellular pathways An important ques-tion in drug target identification is the prevalenceof false-positive and false-negative identificationsTo investigate the reliability of thermal proteomeprofiling we selected two structurally divergentpromiscuous kinase inhibitors with a known spec-trum of targets staurosporine and GSK3182571which is a close analog of CTx-0294885 (21) Profiles

1255784-2 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 1 Quantitative proteome-wide profiling of protein thermal stability under differentialconditions Cells were cultured under differential conditions such as drug treatment In analternative method the cells were extracted first and the extracts were treated with drug (ldquo1rdquo)

For each condition the cell or cell-extract sample was divided into 10 aliquots (ldquo2rdquo) Aliquots were subjected to heating at the indicated temperatures Samples ofintact cells were subsequently subjected to extraction with PBS (ldquo3rdquo) After digestion with trypsin each sample was labeled with a different TMT10 isotope tag(ldquo4rdquo) Subsequently all samples fromeach conditionweremixed (ldquo4rdquo) and analyzed bymeans of LC-MSMS (ldquo5rdquo)The obtained reporter ion intensitieswere usedto fit a melting curve and calculate the melting temperature Tm of each protein separately for the two conditions (ldquo6rdquo)

RESEARCH | RESEARCH ARTICLE

of K562 cell extract treated with staurosporine orvehicle were generated as two independentreplicates (Fig 3) Of all proteins 92 detectedyielded curves with sufficiently steep slopes toallow the robust identification of ligand-inducedTm shifts However a shallow slope of the melt-ing curve indicated lowermelting point reprodu-cibility (Fig 3A) Whereas most affected proteinsshowed positive shifts consistent with ligand-

induced stabilization a few proteins includingmembers of the protein kinase C family exhib-ited negative shifts whichmay result fromdesta-bilization In total the thermal profiles comprised175 protein kinases of which 51 displayed Tmshifts that passed our significance criteria (Fig3B fig S6 and table S4) A recent comprehensiveanalysis of staurosporine targets identified bymeans of chemoproteomics ldquokinobeadsrdquo profil-

ing (22) detected 229 kinases of which 92 werealso present in both biological replicates of thethermal profiles experiments The limited degreeof overlap may be due to the fact that kinobeadswill fail to identify kinases that do not bind to theimmobilized ligands whereas thermal profilingwill miss proteins owing to insufficient abun-dance andor solubility or the absence of a sig-nificant ligand effect Analysis of the 92 kinasescommon to both data sets revealed that kino-beads identified 66 kinases with a staurosporinemedian inhibitory concentration (IC50) below10mMof which 34 showed a significant Tm shift andan additional 15 showed a reproducible shift ofgreater than 1degC which could become significantwith additional replicates For some proteins weobserved shallow or irregular curves (for exam-ple BMP2K AAK1 in fig S6 G and J respectively)that may reflect multiple transitions possibly be-cause of independent folding domains and aredifficult to fit with our current methods For theremainder of 17 kinases identified as targets bykinobeads we observed very small (lt1degC) or no Tmshifts (Fig 3C) These targets may represent falsenegatives We also observed thermal shifts for pro-teins other than kinases Staurosporine inducedapparent stabilization of coproporphyrinogen-III oxidase and ferrochelatase (FECH) two out ofthe eight enzymes in the heme biosynthesispathway for all of which melting curves wereobtained (Fig 3D and fig S7) Ligand-bindingalso affected the thermal stability of the regu-latory components of some protein complexesmdashfor instance kinase complexes containing cyclins(fig S6 C and J) We further explored this in thecontext of the protein kinase A (PKA) complex(23) Inhibition by staurosporine appeared tostabilize the catalytic subunit but destabilize theregulatory subunit In contrast the addition ofadenosine 3prime5prime-monophosphate (cAMP) whichcauses dissociation of the regulatory subunit fromthe catalytic subunit appeared to stabilize theregulatory subunit but destabilize the catalyticsubunit presumably because of the release of theregulatory subunit (Fig 3E)Comparison of staurosporine Tm shift data

with kinobeads-derived potencies for multipletargets (Fig 3C) showed no significant correla-tion between the affinity for staurosporine andthemagnitude of the Tm shift This is not surpris-ing because the extent of thermal stabilization byligands with a similar binding mode likely de-pends on the stability of the unliganded proteinTo investigate the correlation between ligand af-finity and apparent thermal stabilization we con-sidered a structurally divergent promiscuouskinase inhibitor GSK3182571 (Fig 4 and table S5)We identified 13 targets common to both inhib-itors with an affinity below 1 mM and a less than10-fold difference in affinity For these targetsboth compounds showed similar Tm shifts (R2 =09) (Fig 4A) suggesting that for a given proteinthe Tm shift at saturating ligand concentrationsis dependent on the intrinsic affinity of the ligandDepending on the number of replicate experi-

ments ligand-induced Tm shifts as small as 1degCcan be monitored However to resolve small

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-3

Fig 2 The thermal proteome profile of a human K562 cell (A) Heat map representation of thethermal stability of 3204 soluble proteins in intact cells (left) and cell extract (right) For each protein itsrelative concentration (fold-change) at the indicated temperature compared with the lowest temperature(37degC for intact cells and 40degC for the cell extract) is shown Protein fold-changes from the experiment onintact cells were clusteredTo allow better comparison proteins heated in the cell extract were plotted inthe same order as the proteins heated in intact cells (B) Reproducibility of thermal proteome profilesassessed with replicate experiments on intact cells R2 = 093 (left) cell extract R2 = 089 (middle) anddirect comparison of proteomes from intact cells and cell-extract experiments R2 = 030 (right) Mostproteins showed greater thermal stability (higher Tm values) in cell extract as compared with intact cells(C) ATP-binding proteins show a trend toward increased stability in intact cells as compared with cellextractThe plots showdensity distributions of proteinTmvalues determined in intact cells or in cell extractand the density distribution of the difference between both settings A set of 440 proteins annotated asATP binders show a trend toward increased stability in intact cells The difference in the means of thedistributions was significant (P lt 001 t test)

RESEARCH | RESEARCH ARTICLE

changes in ligand-induced thermostability andto enable the ranking of binding affinities amongmultiple targets we tailored the isothermal dose-response (ITDR) protocol to a defined set of pro-teins (12) In contrast to the full temperatureprofiles in which a compound is assayed at asingle typically saturating concentration across10 temperatures an ITDR profile is generated ata defined temperature over a range of compoundconcentrations To generate affinity data forGSK3182571 we acquired ITDR profiles (tableS6) which showed excellent reproducibility (figS8) and were in good agreement with data fromkinobeads competition-binding experiments bothfor proteins stabilized and proteins destabilizedby ligand binding (Fig 4B and table S7)

Because of its unbiased nature thermal pro-teome profiling can detect unexpected off-targetsof drug compounds The heme biosynthesis path-way has not been related to adverse drug effectsbut is linked to a number of genetic disordersincluding erythropoietic protoporphyria which iscaused by a deficiency in FECH and results inhigh tissue levels of protoporphyrins (24) Wefound two enzymes in the heme pathway inter-actingwith staurosporine and thus askedwhetherdrug interactions with this pathway might resultin adverse effects such as skin or liver disease (25)The melanoma drug vemurafenib frequentlycauses severe photosensitivity concomitant withincreased levels of protoporphyrins (26) ITDRprofiling of cells treatedwith vemurafenib showed

that the main proteins affected were its knowntarget BRAF [negative logarithm of the medianeffective concentration (pEC50) = 61] and FECH(pEC50 = 53) (Fig 5A and table S8) These con-centrations are below the typical steady-state ex-posure for this drug (27) and indicate a less than10-fold window of target occupancy between thedrugrsquos efficacy target and FECH Alectinib asecond-generation anaplastic lymphoma kinase(ALK) inhibitor for the treatment of nonndashsmall-cell lung cancer was also reported to cause photo-sensitivity in contrast to the first-generation drugcrizotinib (28 29) ITDR profiling revealed thatalectinib more potently affected FECH than didvemurafenib whereas crizotinib had no effect(Fig 5B) These results strongly suggest that the

1255784-4 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 3 Differential profiling of drug effects on the thermal proteomeprofile Staurosporine treatment of cell extract yields reproducible thermalshifts allowing robust target identification (A) Cell extract was treated withvehicle or staurosporine Both experiments were performed as two fullyindependent replicates A flat slope of the melting curve relates to lowermelting point reproducibility Melting point differences from the two vehicleexperiments are plotted against the shallowest melting curve slope observedin the four vehiclestaurosporine data sets Proteins with an absolute slopebelow 006 are plotted in gray The histogram shows the distribution ofproteins versus slope values Of all proteins 92 yielded curves with suf-ficiently large slopes (B) Scatter plot of Tm shifts calculated from the two

replicates of the staurosporine versus vehicle treatment experimentStaurosporine-induced Tm shifts that passed the significance criteria areshown in red Proteins with flat slopesmdashas shown in gray in (A)mdashare againshown in gray (C) Comparison between Tm shifts and pIC50 values forstaurosporine as reported from kinobeads data (22) (D) Examples ofmeltingcurves for PKN1 FECH AURKA and SLK with and without staurosporinetreatment (E) The catalytic subunit of PKA is stabilized by staurosporinewhereas the regulatory subunit is destabilized Addition of cAMP followed byWestern blot detection revealed destabilization of the catalytic subunit andstabilization of the regulatory subunit Error bars indicate the SEM from n = 4experiments

RESEARCH | RESEARCH ARTICLE

photosensitivity induced by these drugs is me-diated by FECH and demonstrates that thermalproteome profiling can serve as a stand-alone tech-nique for obtaining quantitative affinity data fortarget-ligand interactions in a cell-based setting

Drug treatment induces Tm shifts fordownstream effector proteins

Any modification of a protein can in principleaffect its thermal stability Therefore compari-son of Tm shifts in cell extract where ligand bind-ing but no downstream effects occur with Tmshifts in intact cells where active signaling takesplace might reveal effector proteins downstreamof the target As a model system we used K562

cells which are transformed by an oncogenic fu-sion protein the BCR-ABL kinase (30) Overallthe samples from intact cells yielded a similarprotein coverage as that of samples from cellextracts (tables S9 and S10) Cultured cells weretreated with vehicle or with the ABL inhibitordasatinib a marketed drug against chronic mye-logenous leukemia (CML) (31) at two concen-trations 05 and 5 mM and the samples wereprofiled (Fig 6A and table S9) Theminimum Tmshift for each protein was calculated from a pairof replicate experiments Four proteins displayedsignificant melting point differences in both the05 and 5 mM dasatinibndashtreated cell samples Wedid not observe a Tm shift for the direct target

BCR-ABL suggesting that this protein is notstabilized by dasatinib binding However sub-stantial and significant Tm shifts were observedfor the adaptor proteinCRKLand thephosphataseSHIP2 (Fig 6B and fig S9) which is also known asINPPL1 These proteins are not likely binding thedrug directly and consistent with this no Tmshift was observed for either protein in cell extracts(table S10) In addition the BCR-ABLndashinteractingproteinCRKshowedapronounced change inmeltingcurve slope in the dasatinib-treated cells (fig S9)CRK and CRKL aremembers of a proto-oncogenefamily that bind to tyrosine-phosphorylated pro-teins and are known downstream effectors ofBCR-ABL signaling (32) CRKLhas been previouslyproposed as a treatment response biomarker (33)Comparison of the CRKL melting curves in K562cells with those in Jurkat cells (table S11) whichare not transformed by tyrosine kinase signalingrevealed that dasatinib treatment of K562 cellsreverted the thermal stability of CRKL to the lev-el observed in Jurkat cells (Fig 6C) and the sameeffect was observed for CRK (fig S10) The Jurkatcell profiles also revealed that ABL1 kinase ap-peared to be more thermostable compared withthe BCR-ABL fusion protein in K562 cells suggest-ing that thermal profiling could be used to identifyprotein fusions which are difficult to identify byexpression proteomics (Fig 6C) In order to eval-uate the drug concentration dependence of theTm-shifted effector proteins we performed ITDRprofiles at three different temperatures (Fig 6Dfig S11 and table S12) The concentrations elicit-ing half-maximal response of the CRKL markerin the ITDR profiles recorded at three differenttemperatureswere between 15 and 32 nMwhichis in good agreement with the known potency ofdasatinib for the inhibition of cell growth (34) In-depth analysis of the 50degC ITDR data revealedadditional markers that either showed small Tmshifts or in the case of the known downstreameffector CRK a pronounced change in the melt-ing curve slope indicating that the ITDRprofiles canoffer greater sensitivity than the full-temperatureprofiles (fig S11) It is tempting to speculate thatthe most sensitive biomarkers are those thatexhibit large or even stoichiometric changes inprotein state which should be robustly detectedby their Tm shift This contrasts with the use of a

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-5

Fig 4 Structurally unrelated kinase inhibitors with sim-ilar affinities induce Tm shifts of comparable magnitudebut assessment of ligand affinity requires ITDR mea-surements (A) Staurosporine Tm shift data are comparedwith data for a structurally unrelated promiscuous kinaseinhibitor GSK3182571 Comparison of Tm shifts induced bystaurosporine or GSK3182571 determined in cell extractProtein targets shown had a pIC50 gt 6 both for staurosporineand GSK3182571 as determined bymeans of kinobeads anddid not differ in affinity between the two compounds bymorethan one order of magnitude (B) Good agreement betweenGSK3182571 pEC50 values determined by means of ITDR with pIC50 values determined with kinobeads(Top) Proteins stabilized by GSK3182571 treatment (Bottom) Proteins destabilized by GSK3182571treatment

Fig 5 The clinical kinase drugsvemurafenib and alectinib which can causephototoxicity as a side effect induce Tmshifts in the heme biosynthesis enzymeFECH (A) Concentration-dependent targetoccupancy of vemurafenib and its cognatetarget BRAF compared with the off-targetFECH ITDR profiling was performed at 55degCwith vemurafenib-treated K562 cells andshowed concentration-dependent thermalstabilization of FECH and BRAF The data arenormalized so that the quantity of solubletarget at the highest compoundconcentration which reflects maximumthermal stabilization is set to 1 and fixed forcurve-fitting (B) ITDR performed at 55degC with K562 cells treated with vemurafenib alectinib or crizotinib over a range of concentrations Alectinib displays amore potent effect on FECH as compared with that by vemurafenib whereas crizotinib a drug not known to cause photosensitivity has no effect

RESEARCH | RESEARCH ARTICLE

posttranslational modification as a biomarkerwhere it is usually difficult to assess the stoi-chiometry of the modificationIn general terms the thermal stability of a

protein may be defined by its mutational statusposttranslationalmodifications and bound ligandssuch as other proteins cofactors metabolites ordrugs Fusionproteins splice variants andposttrans-lational modifications are typically undersampledand therefore not comprehensively detected inMS-basedproteomicsHence the thermal proteomeprofile of a human cell can provide a generalview of the proteomic state or proteotypeThe future scope and applicability of thermal

proteome profilingwill substantially benefit fromcontinued advances in the sensitivity and accu-racyofMS-basedproteomics Futuredevelopmentsshould include the development of protocols formembrane proteins and the application to tis-sues from animal studies or clinical biopsies

Materials and methods

Reagents and cell culture

Reagents andmedia were purchased from Sigma-Aldrich (St Louis MO) unless otherwise notedPBS was prepared by using 137 mM NaCl 27mM KCl 10 mM Na2HPO4 and 2 mM KH2PO4

to buffer at pH 74 and supplemented withcomplete EDTA-free protease inhibitor cocktail(one tablet per 25 ml) (Roche Diagnostics Basel

Switzerland) Staurosporine was from Calbiochem(Millipore Billerica MA) vemurafenib andcrizotinib from Selleck (Houston TX) alectinibfromMedchemExpress (Monmouth Junction NJ)and dasatinib from Toronto Research (TorontoCanada) The synthesis of GSK3182571 is de-scribed in the supplementary methods DNAfragments PG1 (5prime-AGC TTA GAC ATG CCT AGACAT GCC TA -3prime) and PG2 (5prime-AGC TTA GGC ATGTCT AGG CAT GTC TA -3prime) were from Life Tech-nologies (Carlsbad CA) and dissolved in 100 mMTris (pH75) 1 mM NaCl and 10mM EDTA K562Jurkat E61 and A549 cells were from ATCC K562and A549 cells were cultured in RPMI mediumcontaining 10 fetal calf serum (FCS) and JurkatE61 cells in RPMI1640 supplemented with45 gl glucose 10 mM HEPES 1 mM sodiumpyruvate and 10 FCS HT-29 (ATTC no HTB-38)were maintained in Dulbeccorsquos Modified EagleMedium Culture media were supplemented with03 gL L-glutamine and 10 fetal bovine serum(FBS) (Gibco Carlsbad CA) 100 unitsmL peni-cillin and 100 unitsmL streptomycin The cellswere expanded to amaximumof 2times 106 cellsml incase of K562 and 107 cellsml in case of Jurkat E61

Preparation of cell extract forCETSA with Western blot read-out

A549 andHT-29 cells were harvested andwashedwith PBS The cells were diluted in KB buffer

(25mMTris-HCl supplemented with 5mMbeta-glycerophosphate 2 mM DTT 01 mM Na3VO410 mM MgCl2 and complete protease inhibitorcocktail) The cell suspensionswere freeze-thawedthree times by using liquid nitrogen The solublefraction was separated from the cell debris bycentrifugation at 20000 g for 20 min at 4degC FortheCETSAmelting curve experiments cell extractswere diluted with KB buffer and divided into twoaliquots with one aliquot being treated withligandDNA fragments depending on the targetand the other aliquot treated with the solvent ofthe ligand (control) After 10 min incubation atroom temperature the respective cell extractswere heated individually at different temper-atures for 3 min in a thermal cycler [AppliedBiosystems (Foster City CA)Life Technologies]followed by cooling for 3 min at room temper-ature The heated cell extracts were centrifugedat 20000 g for 20min at 4degC in order to separatethe soluble fractions from precipitates The su-pernatantswere transferred to new02mLmicro-tubes and analyzed by means of Western blotanalysis NuPage Bis-Tris 4-12 polyacrylamidegels with MES SDS running buffer (Life Tech-nologies) were used for separation of proteins inthe samples Proteins were transferred to nitro-cellulose membranes by using the iBlot system(Life Technologies) Primary antibodies anti-PKARegulatory I-a (sc-136231) anti-PKA Catalytic-a

1255784-6 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 6 Treatment of K562 cells with dasatinib induces Tm shifts fordownstream effector proteins in the BCR-ABL pathway (A) Cell-wideassessment of Tm shifts induced by dasatinib (05 and 5 mM) in K562 cellsThe minimum Tm shift for each protein was calculated from a pair of fullreplicate experiments Four proteins showing significant melting pointdifferences both in the 05 and 5 mM dasatinib data sets are marked in red(B) Effect of dasatinib on the melting curves for the effector protein CRKL

determined in two biological replicates in intact cells (top) and cell extracts(bottom) (C) The melting curve for CRKL in Jurkat cells closely matchesthe curve obtained in dasatinib-treated K562 cells (top)The ABL1 protein inJurkat cells is much more thermostable compared with the BCR-ABL fusionprotein in K562 cells (D) ITDR curves for CRKL in dasatinib-treated intactcells Experiments were performed at 47degC (green curve) 50degC (purplecurve) and 56degC (blue curve)

RESEARCH | RESEARCH ARTICLE

(sc-28315) anti-p53WT and R273H (sc-126) sec-ondary goat anti-mouse HRP-IgG (sc-2055) andbovine anti-goat HRP-IgG (sc-2352) antibodies(Santa Cruz Biotechnology Dallas TX) were usedfor immunoblotting Chemiluminescence inten-sities were detected and quantified by using aChemiDocXRS+ imaging systemwith Image Labsoftware (Bio-Rad Hercules CA) Data were ex-pressed as means T SEM The data (band inten-sities) from multiple runs (n ge 3) were plottedwith Graphpad Prism software

Preparation of cell extract for thermalproteome profiling

Suspension cultures of K562 cells (15 to 2 times 106

cellsml) or of Jurkat E61 cells (4 times 106 cellsml)were centrifuged at 340 g for 2 min at 4degC Thecells were resuspended in 50 ml PBS After a sec-ond centrifugation step as above the cells wereresuspended in 10 ml ice-cold PBS and centri-fuged again at 340 g for 2 min at 4degC The cellswere resuspended in 15 ml of ice-cold PBS andsnap-frozen in liquidnitrogen The tubewasplacedinto a water bath at 23degC until ~60 of the con-tent was thawed and then transferred on ice un-til the entire contentwas thawed This freezethawcycle was repeated twice before the samplesweresubjected to ultracentrifugation (20 min at 4degCand 100000 g) The protein concentration of thesupernatant was determined by Bradford assay(Bio-Rad) and aliquots were snap-frozen in liquidnitrogen for use in subsequent thermal shiftassays

Thermal profiling using cell extract

A solution of the compound in dimethyl sulfoxide(DMSO) or DMSO alone as vehicle (Table 1) wasadded to the cell extract to 1 final DMSO con-centration The extract was then incubated for10 min at 23degC divided into 10 aliquots of 100 mland transferred into 02-ml polymerase chain re-action (PCR) tubes One each of the compoundand of the vehicle containing samples was heatedin parallel for 3 min to the respective tempera-ture followed by a 3-min incubation time at roomtemperature Subsequently the extract was cen-trifuged at 100000 g for 20 min at 4degC The su-pernatant was fractionated by means of SDS gel

electrophoresis and subjected to sample prepa-ration for MS analysis For ITDR experimentsGSK3182571 was tested as a nine-point serial dilu-tion starting at 100 mM (resulting in concentra-tions of 100 333 111 370 1235 0412 0137 0046and 0015 mM) including one vehicle control inan isothermal dose-response experiment at 53degCfollowing the procedure outlined above for ex-periments in cell extract

Thermal profiling using intact cells

To 24 ml of a suspension of K562 cells (at a den-sity of 15 106 cellsml) or Jurkat E61cells (at adensity of 4 106 cellsml) in a T-flask a volumeof either 120 ml of a solution of the referencecompound dissolved in DMSO or an equivalentamount of DMSO (vehicle) was added Incubationof cellswith compound and vehiclewas conductedin parallel for the duration of 1 hour at 37degC and5CO2 Cells were pelleted at 340 g and 4degC for2min resuspended in 20ml of ice cold PBS andcentrifuged again as indicated above This wash-ing step was repeated once and the cells werecarefully resuspended in 1200 ml PBS 100 ml ofthis resulting cell suspension was transferredinto 02-ml PCR tubes and subjected to a centrifu-gation step at 325 g for 2 min at 4degC Using apipette 80 ml of the PBS supernatant was re-moved By gently tapping the tubes the cells wereresuspended and one each of the compound andof the vehicle containing tubes was heated inparallel in a PCR machine for 3 min to the re-spective temperature (37degC to 67degC) followed bya 3-min incubation time at room temperatureThereafter the cells were snap-frozen in liquidnitrogen for 1 min The cells were thawed brieflyin a water bath at 25degC transferred on ice andresuspended by using a pipette This freeze-thawcycle was repeated once After an addition of fur-ther 30 ml of PBS and resuspension the entirecontent was centrifuged at 100000 g for 20 minat 4degC After centrifugation 30 ml of the super-natant were transferred into a new tube Proteinconcentration of the 37degC and 41degC samples weredetermined and used to normalize loading ofSDS gels Proteins in the supernatant were dena-tured by using SDS sample buffer partially sepa-rated by means of SDS gel electrophoresis and

subjected to sample preparation for MS analysisFor ITDR experiments vemurafenib alectiniband crizotinib were used as a nine-point serialdilution starting at 30 mM (concentrations were30 12 48 192 077 031 012 005 and 002 mM)including one vehicle control at 55degC followingthe procedure outlined aboveDasatinibwas testedas an eight-point serial dilution starting at 1 mM(concentrations usedwere 1 033 0111 0037 00120004 00013 and 000046 mM) including onedasatinib control at 10 mM and one vehicle con-trol on intact cells at 47degC 50degC and 56degC follow-ing the procedure outlined above for experimentson intact cells (Table 1)

Kinobeads assays

Competition binding assays were performed asdescribed previously by using a modified beadmatrix (5 22) Briefly 1 ml (5 mg protein) cellextract was pre-incubatedwith test compound orvehicle for 45 min at 4degC followed by incubationwith kinobeads (35 ml beads per sample) for1 hour at 4degC The beads were washed with lysisbuffer and eluted with 50 ml 2times SDS sample buf-fer GSK3182571was tested at 10 25 0625 015625003906 000977 and 000244 mM

Sample preparation for MS

Gel lanes were cut into three slices covering theentire separation range (~2 cm) and subjectedto in-gel digestion (5) Peptide samples were la-beled with 10-plex TMT (TMT10 Thermo FisherScientific Waltham MA) reagents enabling rel-ative quantification of a broad range of 10 tem-perature points in a single experiment (15 22)For kinobeads and ITDR experiments either10 or 8 TMT labels were used depending on thecompound concentration range chosen Thelabeling reaction was performed in 40 mMtriethylammoniumbicarbonate pH 853 at 22degCand quenched with glycine Labeled peptide ex-tracts were combined to a single sample per ex-periment and as indicated in the supplementarymaterials for most samples additional fraction-ation was performed by using reversed-phasechromatography at pH 12 [1mmXbridge column(Waters Milford MA)] as previously described(table S13) (35)

LC-MSMS analysis

Samples were dried in vacuo and resuspended in005 trifluoroacetic acid in water Of the sam-ple 50 was injected into an Ultimate3000nanoRLSC (Dionex Sunnyvale CA) coupled to aQ Exactive (Thermo Fisher Scientific) Peptideswere separated on custom 50 cm times 100 mM (ID)reversed-phase columns (Reprosil) at 40degC Gra-dient elutionwas performed from 2 acetonitrileto 40acetonitrile in 01 formic acid over 2 hoursSamples were online injected into Q-Exactive massspectrometers operating with a data-dependenttop 10methodMS spectrawere acquired by using70000 resolution and an ion target of 3E6Higherenergy collisional dissociation (HCD) scans wereperformed with 35 NCE at 35000 resolu-tion (at mz 200) and the ion target settingswas set to 2E5 so as to avoid coalescence (15) The

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-7

Table 1 Compoundsused in thermal profilingDashes indicate that compoundwas not tested in intact cells

CompoundConcentration used in

cellular extractexperiments

Concentration(s) inintact cell experiments

Cell line

MgATP 2 mM mdash K562staurosporine 20 mM mdash K562cAMP 1 mM mdash K562GSK3182571 20 mM mdash K562dasatinib 5 mM 05 mM 5 mM K562p53 WT

PG1 100 mM mdash A549PG2 100 mM mdash A549

p53 R273HPG1 100 mM mdash HT-29PG2 100 mM mdash HT-29

RESEARCH | RESEARCH ARTICLE

instruments were operated with Tune 22 or 23and Xcalibur 27 or 3063

Peptide and protein identification

Mascot 24 (Matrix Science Boston MA) wasused for protein identification by using a 10 partspermillionmass tolerance for peptide precursorsand 20 mD (HCD) mass tolerance for fragmentions Carbamidomethylation of cysteine residuesand TMTmodification of lysine residueswere setas fixedmodifications andmethionine oxidationand N-terminal acetylation of proteins and TMTmodification of peptide N-termini were set asvariable modifications The search database con-sisted of a customized version of the Interna-tional Protein Index protein sequence databasecombined with a decoy version of this databasecreated by using a script supplied by Matrix Sci-ence Unless stated otherwise we accepted proteinidentifications as follows (i) For single-spectrumto sequence assignments we required this assign-ment to be the bestmatch and aminimumMascotscore of 31 and a 10times difference of this assign-ment over the next best assignment Based onthese criteria the decoy search results indicatedlt1 false discovery rate (FDR) (ii) For multiplespectrum to sequence assignments and using thesame parameters the decoy search results indi-cate lt01 FDR All identified proteinswere quan-tified FDR for quantified proteins was lt1

Peptide and protein quantification

Reporter ion intensities were read from raw dataandmultiplied with ion accumulation times (theunit is milliseconds) so as to yield ameasure propor-tional to thenumberof ions thismeasure is referredto as ion area (36) Spectra matching to peptideswere filtered according to the following criteriamascot ion score gt15 signal-to-background of theprecursor ion gt4 and signal-to-interference gt05(37) Fold-changes were corrected for isotope pu-rity as described and adjusted for interferencecaused by co-eluting nearly isobaric peaks as esti-mated by the signal-to-interferencemeasure (38) Pro-tein quantification was derived from individualspectra matching to distinct peptides by using asum-based bootstrap algorithm 95 confidenceintervals were calculated for all protein fold-changesthatwere quantifiedwithmore than three spectra (36)

Curve fitting normalization andestimation of slope and melting pointdifferences for CETSA experiments

Melting curves

For all CETSA experiments fold-changes werecalculated by using the lowest temperature con-dition as the reference Relative fold changes as afunction of temperature followeda sigmoidal trendwhich was fitted with the following equationderived from chemical denaturation theory (39)(enwikipediaorgwikiEquilibrium_unfolding)using R (wwwr-projectorg)

f ethT THORN frac14 1 minus plateau

1thorn eminusaTminusbeth THORN thorn plateau

Where T is the temperature and a b and plateauare constants The value of f(T) at the lowest

temperature Tmin was fixed to one The meltingpoint of a protein is defined as the temperatureTm at which half of the protein amount has beendenatured

f(Tm) = 05

Additionally the slope of the melting curve isdefined as the value of the first derivative of f(T)at the point T = Tinfl

slope = f prime(Tinfl)

where Tinfl is the inflection point of the curvemeaning that the second derivative of f (T) equalszero at T = Tinfl

f primeprime(Tinfl) = 0

Normalization

The vehicle and compound-treated experimentswere normalized in the following waySelection of proteins for normalization (i) The

set of proteins identified in both experimentstermed ldquojointPrdquo was determined (ii) For eachexperiment (vehicle or compound) all proteinsfrom jointP fulfilling the following criteria werecollected the fold-changes at the seventh highesttemperature point comparedwith the lowest tem-perature point were between 04 and 06 andat the 9th and 10th highest temperature pointscompared th the lowest temperature pointwere below 03 and 02 respectively (iii) Thebigger of these two protein subsets was thenused for normalization (typically more than200 proteins qualified) This protein set is re-ferred to as ldquonormPrdquoCalculation of correction factors (i) The me-

dian fold-change values over all proteins in normPwere calculated for each of the 10 fold-changesThis was done separately for both experimentsand for each experiment a melting curve wasfitted The curve that was fitted with the best R2

was used to normalize both experiments (ii) Foreach experiment the correction factors werecalculated by using the median fold-changes ofthe proteins in normP For each median fold-change a correction factor by which that fold-change would have to be multiplied with in orderto coincidewith the best-fitting curvewas calculatedNormalization The two sets of correction fac-

tors were applied to all the protein fold-changesin the respective experimentsThis heuristic normalization procedure covers

three important aspects (i) The same subset ofproteins is used to determine normalization co-efficients in both experiments thus circumvent-ing any potential bias caused by proteins onlyidentified in one experiment (ii) The subselectionof late-melting proteins (seventh point between04 and 06) ensures that the normalization ofhigh-temperature points will be more accurate(iii) Last the narrow requirements on the 7th9th and 10th points will ensure that a narrowrange of melting profiles is used for calculatingthe points from which the normalization curvewill be derived This is important because if vastlydifferent melting profiles are combined into asingle one the above equation will not necessar-ily fit well A normalization curve example as well

as melting curves of several proteins before andafter normalization are illustrated in fig S12

Estimation of slope and meltingpoint differences

After normalization melting curves were fittedfor all proteins identified in the vehicle- andcompound-treated experiments Melting pointdifferences and slope differences were calculatedfor proteins whose curves passed the followingtwo requirements (i) both fitted curves for thevehicle and compound treated condition had anR2 of greater than 08 and (ii) the vehicle curvehad a plateau of less than 03Typically close to 80 of proteins that were

identified in both the vehicle- and compound-treated experiment passed these criteria Theslope of the curves had the biggest impact onthe reproducibility of melting point differencesof proteins This observation can be rationalizedin the following way If the curve has a shallowslope then small changes with respect to the ref-erence pointmdashin our case the lowest temperaturemdashwill have a strong impact on where the curvecrosses the 05 fold-change level (melting point)Consequently small variations in the quantitativemeasurement of theMS signals corresponding tothe lowest temperature will lead to strong shiftsin the inferred melting point of the curve Thismeans that it is difficult to get high precisionmeasurements ofmelting point differenceswhencomparing two curves with shallow slopes If thecurve however has a steep slope then the melt-ing point measurement will not be substantiallyaffected by small variations of the quantitativemeasurements of the lowest temperature nor ofany other temperature points Taking this intoaccount we used the following strategy for sig-nificance estimation of protein melting point dif-ferences between vehicle- and compound-treatedexperiments which is conceptually very similarto a broadly used strategy for inferring significantprotein fold-changes (40) The calculated meltingpoint differences of proteins were ordered in des-cending slope order (those with the shallowestslope come first) in which the slope for eachprotein is the lowest (steepest) slope calculatedfrom the two curve fits performed for the proteinin the vehicle- and compound-treated conditionThe proteins were divided into bins The bins areconstructed starting fromproteinswith the high-est (shallowest) slope Each bin consists of at least300 proteins Once each bin has been completedthe remaining number of proteins is counted ifthis number is below 300 the remaining pro-teins are added to the last completed bin Thisdata qualityndashdependent binning strategy is analo-gous to the procedure described in Cox et al withthe only difference being that proteins are sortedby slope instead of abundance (40) The statisticalsignificance of differences in protein meltingpoints was calculated by estimating the left- andright-sided robust standard deviation of the dis-tribution of the measurements using the 158750 and 8413 percentiles and calculating the Pvalues for all measurements for a specific binexactly as previously described (40) Subsequently

1255784-8 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

RESEARCH | RESEARCH ARTICLE

of K562 cell extract treated with staurosporine orvehicle were generated as two independentreplicates (Fig 3) Of all proteins 92 detectedyielded curves with sufficiently steep slopes toallow the robust identification of ligand-inducedTm shifts However a shallow slope of the melt-ing curve indicated lowermelting point reprodu-cibility (Fig 3A) Whereas most affected proteinsshowed positive shifts consistent with ligand-

induced stabilization a few proteins includingmembers of the protein kinase C family exhib-ited negative shifts whichmay result fromdesta-bilization In total the thermal profiles comprised175 protein kinases of which 51 displayed Tmshifts that passed our significance criteria (Fig3B fig S6 and table S4) A recent comprehensiveanalysis of staurosporine targets identified bymeans of chemoproteomics ldquokinobeadsrdquo profil-

ing (22) detected 229 kinases of which 92 werealso present in both biological replicates of thethermal profiles experiments The limited degreeof overlap may be due to the fact that kinobeadswill fail to identify kinases that do not bind to theimmobilized ligands whereas thermal profilingwill miss proteins owing to insufficient abun-dance andor solubility or the absence of a sig-nificant ligand effect Analysis of the 92 kinasescommon to both data sets revealed that kino-beads identified 66 kinases with a staurosporinemedian inhibitory concentration (IC50) below10mMof which 34 showed a significant Tm shift andan additional 15 showed a reproducible shift ofgreater than 1degC which could become significantwith additional replicates For some proteins weobserved shallow or irregular curves (for exam-ple BMP2K AAK1 in fig S6 G and J respectively)that may reflect multiple transitions possibly be-cause of independent folding domains and aredifficult to fit with our current methods For theremainder of 17 kinases identified as targets bykinobeads we observed very small (lt1degC) or no Tmshifts (Fig 3C) These targets may represent falsenegatives We also observed thermal shifts for pro-teins other than kinases Staurosporine inducedapparent stabilization of coproporphyrinogen-III oxidase and ferrochelatase (FECH) two out ofthe eight enzymes in the heme biosynthesispathway for all of which melting curves wereobtained (Fig 3D and fig S7) Ligand-bindingalso affected the thermal stability of the regu-latory components of some protein complexesmdashfor instance kinase complexes containing cyclins(fig S6 C and J) We further explored this in thecontext of the protein kinase A (PKA) complex(23) Inhibition by staurosporine appeared tostabilize the catalytic subunit but destabilize theregulatory subunit In contrast the addition ofadenosine 3prime5prime-monophosphate (cAMP) whichcauses dissociation of the regulatory subunit fromthe catalytic subunit appeared to stabilize theregulatory subunit but destabilize the catalyticsubunit presumably because of the release of theregulatory subunit (Fig 3E)Comparison of staurosporine Tm shift data

with kinobeads-derived potencies for multipletargets (Fig 3C) showed no significant correla-tion between the affinity for staurosporine andthemagnitude of the Tm shift This is not surpris-ing because the extent of thermal stabilization byligands with a similar binding mode likely de-pends on the stability of the unliganded proteinTo investigate the correlation between ligand af-finity and apparent thermal stabilization we con-sidered a structurally divergent promiscuouskinase inhibitor GSK3182571 (Fig 4 and table S5)We identified 13 targets common to both inhib-itors with an affinity below 1 mM and a less than10-fold difference in affinity For these targetsboth compounds showed similar Tm shifts (R2 =09) (Fig 4A) suggesting that for a given proteinthe Tm shift at saturating ligand concentrationsis dependent on the intrinsic affinity of the ligandDepending on the number of replicate experi-

ments ligand-induced Tm shifts as small as 1degCcan be monitored However to resolve small

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-3

Fig 2 The thermal proteome profile of a human K562 cell (A) Heat map representation of thethermal stability of 3204 soluble proteins in intact cells (left) and cell extract (right) For each protein itsrelative concentration (fold-change) at the indicated temperature compared with the lowest temperature(37degC for intact cells and 40degC for the cell extract) is shown Protein fold-changes from the experiment onintact cells were clusteredTo allow better comparison proteins heated in the cell extract were plotted inthe same order as the proteins heated in intact cells (B) Reproducibility of thermal proteome profilesassessed with replicate experiments on intact cells R2 = 093 (left) cell extract R2 = 089 (middle) anddirect comparison of proteomes from intact cells and cell-extract experiments R2 = 030 (right) Mostproteins showed greater thermal stability (higher Tm values) in cell extract as compared with intact cells(C) ATP-binding proteins show a trend toward increased stability in intact cells as compared with cellextractThe plots showdensity distributions of proteinTmvalues determined in intact cells or in cell extractand the density distribution of the difference between both settings A set of 440 proteins annotated asATP binders show a trend toward increased stability in intact cells The difference in the means of thedistributions was significant (P lt 001 t test)

RESEARCH | RESEARCH ARTICLE

changes in ligand-induced thermostability andto enable the ranking of binding affinities amongmultiple targets we tailored the isothermal dose-response (ITDR) protocol to a defined set of pro-teins (12) In contrast to the full temperatureprofiles in which a compound is assayed at asingle typically saturating concentration across10 temperatures an ITDR profile is generated ata defined temperature over a range of compoundconcentrations To generate affinity data forGSK3182571 we acquired ITDR profiles (tableS6) which showed excellent reproducibility (figS8) and were in good agreement with data fromkinobeads competition-binding experiments bothfor proteins stabilized and proteins destabilizedby ligand binding (Fig 4B and table S7)

Because of its unbiased nature thermal pro-teome profiling can detect unexpected off-targetsof drug compounds The heme biosynthesis path-way has not been related to adverse drug effectsbut is linked to a number of genetic disordersincluding erythropoietic protoporphyria which iscaused by a deficiency in FECH and results inhigh tissue levels of protoporphyrins (24) Wefound two enzymes in the heme pathway inter-actingwith staurosporine and thus askedwhetherdrug interactions with this pathway might resultin adverse effects such as skin or liver disease (25)The melanoma drug vemurafenib frequentlycauses severe photosensitivity concomitant withincreased levels of protoporphyrins (26) ITDRprofiling of cells treatedwith vemurafenib showed

that the main proteins affected were its knowntarget BRAF [negative logarithm of the medianeffective concentration (pEC50) = 61] and FECH(pEC50 = 53) (Fig 5A and table S8) These con-centrations are below the typical steady-state ex-posure for this drug (27) and indicate a less than10-fold window of target occupancy between thedrugrsquos efficacy target and FECH Alectinib asecond-generation anaplastic lymphoma kinase(ALK) inhibitor for the treatment of nonndashsmall-cell lung cancer was also reported to cause photo-sensitivity in contrast to the first-generation drugcrizotinib (28 29) ITDR profiling revealed thatalectinib more potently affected FECH than didvemurafenib whereas crizotinib had no effect(Fig 5B) These results strongly suggest that the

1255784-4 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 3 Differential profiling of drug effects on the thermal proteomeprofile Staurosporine treatment of cell extract yields reproducible thermalshifts allowing robust target identification (A) Cell extract was treated withvehicle or staurosporine Both experiments were performed as two fullyindependent replicates A flat slope of the melting curve relates to lowermelting point reproducibility Melting point differences from the two vehicleexperiments are plotted against the shallowest melting curve slope observedin the four vehiclestaurosporine data sets Proteins with an absolute slopebelow 006 are plotted in gray The histogram shows the distribution ofproteins versus slope values Of all proteins 92 yielded curves with suf-ficiently large slopes (B) Scatter plot of Tm shifts calculated from the two

replicates of the staurosporine versus vehicle treatment experimentStaurosporine-induced Tm shifts that passed the significance criteria areshown in red Proteins with flat slopesmdashas shown in gray in (A)mdashare againshown in gray (C) Comparison between Tm shifts and pIC50 values forstaurosporine as reported from kinobeads data (22) (D) Examples ofmeltingcurves for PKN1 FECH AURKA and SLK with and without staurosporinetreatment (E) The catalytic subunit of PKA is stabilized by staurosporinewhereas the regulatory subunit is destabilized Addition of cAMP followed byWestern blot detection revealed destabilization of the catalytic subunit andstabilization of the regulatory subunit Error bars indicate the SEM from n = 4experiments

RESEARCH | RESEARCH ARTICLE

photosensitivity induced by these drugs is me-diated by FECH and demonstrates that thermalproteome profiling can serve as a stand-alone tech-nique for obtaining quantitative affinity data fortarget-ligand interactions in a cell-based setting

Drug treatment induces Tm shifts fordownstream effector proteins

Any modification of a protein can in principleaffect its thermal stability Therefore compari-son of Tm shifts in cell extract where ligand bind-ing but no downstream effects occur with Tmshifts in intact cells where active signaling takesplace might reveal effector proteins downstreamof the target As a model system we used K562

cells which are transformed by an oncogenic fu-sion protein the BCR-ABL kinase (30) Overallthe samples from intact cells yielded a similarprotein coverage as that of samples from cellextracts (tables S9 and S10) Cultured cells weretreated with vehicle or with the ABL inhibitordasatinib a marketed drug against chronic mye-logenous leukemia (CML) (31) at two concen-trations 05 and 5 mM and the samples wereprofiled (Fig 6A and table S9) Theminimum Tmshift for each protein was calculated from a pairof replicate experiments Four proteins displayedsignificant melting point differences in both the05 and 5 mM dasatinibndashtreated cell samples Wedid not observe a Tm shift for the direct target

BCR-ABL suggesting that this protein is notstabilized by dasatinib binding However sub-stantial and significant Tm shifts were observedfor the adaptor proteinCRKLand thephosphataseSHIP2 (Fig 6B and fig S9) which is also known asINPPL1 These proteins are not likely binding thedrug directly and consistent with this no Tmshift was observed for either protein in cell extracts(table S10) In addition the BCR-ABLndashinteractingproteinCRKshowedapronounced change inmeltingcurve slope in the dasatinib-treated cells (fig S9)CRK and CRKL aremembers of a proto-oncogenefamily that bind to tyrosine-phosphorylated pro-teins and are known downstream effectors ofBCR-ABL signaling (32) CRKLhas been previouslyproposed as a treatment response biomarker (33)Comparison of the CRKL melting curves in K562cells with those in Jurkat cells (table S11) whichare not transformed by tyrosine kinase signalingrevealed that dasatinib treatment of K562 cellsreverted the thermal stability of CRKL to the lev-el observed in Jurkat cells (Fig 6C) and the sameeffect was observed for CRK (fig S10) The Jurkatcell profiles also revealed that ABL1 kinase ap-peared to be more thermostable compared withthe BCR-ABL fusion protein in K562 cells suggest-ing that thermal profiling could be used to identifyprotein fusions which are difficult to identify byexpression proteomics (Fig 6C) In order to eval-uate the drug concentration dependence of theTm-shifted effector proteins we performed ITDRprofiles at three different temperatures (Fig 6Dfig S11 and table S12) The concentrations elicit-ing half-maximal response of the CRKL markerin the ITDR profiles recorded at three differenttemperatureswere between 15 and 32 nMwhichis in good agreement with the known potency ofdasatinib for the inhibition of cell growth (34) In-depth analysis of the 50degC ITDR data revealedadditional markers that either showed small Tmshifts or in the case of the known downstreameffector CRK a pronounced change in the melt-ing curve slope indicating that the ITDRprofiles canoffer greater sensitivity than the full-temperatureprofiles (fig S11) It is tempting to speculate thatthe most sensitive biomarkers are those thatexhibit large or even stoichiometric changes inprotein state which should be robustly detectedby their Tm shift This contrasts with the use of a

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-5

Fig 4 Structurally unrelated kinase inhibitors with sim-ilar affinities induce Tm shifts of comparable magnitudebut assessment of ligand affinity requires ITDR mea-surements (A) Staurosporine Tm shift data are comparedwith data for a structurally unrelated promiscuous kinaseinhibitor GSK3182571 Comparison of Tm shifts induced bystaurosporine or GSK3182571 determined in cell extractProtein targets shown had a pIC50 gt 6 both for staurosporineand GSK3182571 as determined bymeans of kinobeads anddid not differ in affinity between the two compounds bymorethan one order of magnitude (B) Good agreement betweenGSK3182571 pEC50 values determined by means of ITDR with pIC50 values determined with kinobeads(Top) Proteins stabilized by GSK3182571 treatment (Bottom) Proteins destabilized by GSK3182571treatment

Fig 5 The clinical kinase drugsvemurafenib and alectinib which can causephototoxicity as a side effect induce Tmshifts in the heme biosynthesis enzymeFECH (A) Concentration-dependent targetoccupancy of vemurafenib and its cognatetarget BRAF compared with the off-targetFECH ITDR profiling was performed at 55degCwith vemurafenib-treated K562 cells andshowed concentration-dependent thermalstabilization of FECH and BRAF The data arenormalized so that the quantity of solubletarget at the highest compoundconcentration which reflects maximumthermal stabilization is set to 1 and fixed forcurve-fitting (B) ITDR performed at 55degC with K562 cells treated with vemurafenib alectinib or crizotinib over a range of concentrations Alectinib displays amore potent effect on FECH as compared with that by vemurafenib whereas crizotinib a drug not known to cause photosensitivity has no effect

RESEARCH | RESEARCH ARTICLE

posttranslational modification as a biomarkerwhere it is usually difficult to assess the stoi-chiometry of the modificationIn general terms the thermal stability of a

protein may be defined by its mutational statusposttranslationalmodifications and bound ligandssuch as other proteins cofactors metabolites ordrugs Fusionproteins splice variants andposttrans-lational modifications are typically undersampledand therefore not comprehensively detected inMS-basedproteomicsHence the thermal proteomeprofile of a human cell can provide a generalview of the proteomic state or proteotypeThe future scope and applicability of thermal

proteome profilingwill substantially benefit fromcontinued advances in the sensitivity and accu-racyofMS-basedproteomics Futuredevelopmentsshould include the development of protocols formembrane proteins and the application to tis-sues from animal studies or clinical biopsies

Materials and methods

Reagents and cell culture

Reagents andmedia were purchased from Sigma-Aldrich (St Louis MO) unless otherwise notedPBS was prepared by using 137 mM NaCl 27mM KCl 10 mM Na2HPO4 and 2 mM KH2PO4

to buffer at pH 74 and supplemented withcomplete EDTA-free protease inhibitor cocktail(one tablet per 25 ml) (Roche Diagnostics Basel

Switzerland) Staurosporine was from Calbiochem(Millipore Billerica MA) vemurafenib andcrizotinib from Selleck (Houston TX) alectinibfromMedchemExpress (Monmouth Junction NJ)and dasatinib from Toronto Research (TorontoCanada) The synthesis of GSK3182571 is de-scribed in the supplementary methods DNAfragments PG1 (5prime-AGC TTA GAC ATG CCT AGACAT GCC TA -3prime) and PG2 (5prime-AGC TTA GGC ATGTCT AGG CAT GTC TA -3prime) were from Life Tech-nologies (Carlsbad CA) and dissolved in 100 mMTris (pH75) 1 mM NaCl and 10mM EDTA K562Jurkat E61 and A549 cells were from ATCC K562and A549 cells were cultured in RPMI mediumcontaining 10 fetal calf serum (FCS) and JurkatE61 cells in RPMI1640 supplemented with45 gl glucose 10 mM HEPES 1 mM sodiumpyruvate and 10 FCS HT-29 (ATTC no HTB-38)were maintained in Dulbeccorsquos Modified EagleMedium Culture media were supplemented with03 gL L-glutamine and 10 fetal bovine serum(FBS) (Gibco Carlsbad CA) 100 unitsmL peni-cillin and 100 unitsmL streptomycin The cellswere expanded to amaximumof 2times 106 cellsml incase of K562 and 107 cellsml in case of Jurkat E61

Preparation of cell extract forCETSA with Western blot read-out

A549 andHT-29 cells were harvested andwashedwith PBS The cells were diluted in KB buffer

(25mMTris-HCl supplemented with 5mMbeta-glycerophosphate 2 mM DTT 01 mM Na3VO410 mM MgCl2 and complete protease inhibitorcocktail) The cell suspensionswere freeze-thawedthree times by using liquid nitrogen The solublefraction was separated from the cell debris bycentrifugation at 20000 g for 20 min at 4degC FortheCETSAmelting curve experiments cell extractswere diluted with KB buffer and divided into twoaliquots with one aliquot being treated withligandDNA fragments depending on the targetand the other aliquot treated with the solvent ofthe ligand (control) After 10 min incubation atroom temperature the respective cell extractswere heated individually at different temper-atures for 3 min in a thermal cycler [AppliedBiosystems (Foster City CA)Life Technologies]followed by cooling for 3 min at room temper-ature The heated cell extracts were centrifugedat 20000 g for 20min at 4degC in order to separatethe soluble fractions from precipitates The su-pernatantswere transferred to new02mLmicro-tubes and analyzed by means of Western blotanalysis NuPage Bis-Tris 4-12 polyacrylamidegels with MES SDS running buffer (Life Tech-nologies) were used for separation of proteins inthe samples Proteins were transferred to nitro-cellulose membranes by using the iBlot system(Life Technologies) Primary antibodies anti-PKARegulatory I-a (sc-136231) anti-PKA Catalytic-a

1255784-6 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 6 Treatment of K562 cells with dasatinib induces Tm shifts fordownstream effector proteins in the BCR-ABL pathway (A) Cell-wideassessment of Tm shifts induced by dasatinib (05 and 5 mM) in K562 cellsThe minimum Tm shift for each protein was calculated from a pair of fullreplicate experiments Four proteins showing significant melting pointdifferences both in the 05 and 5 mM dasatinib data sets are marked in red(B) Effect of dasatinib on the melting curves for the effector protein CRKL

determined in two biological replicates in intact cells (top) and cell extracts(bottom) (C) The melting curve for CRKL in Jurkat cells closely matchesthe curve obtained in dasatinib-treated K562 cells (top)The ABL1 protein inJurkat cells is much more thermostable compared with the BCR-ABL fusionprotein in K562 cells (D) ITDR curves for CRKL in dasatinib-treated intactcells Experiments were performed at 47degC (green curve) 50degC (purplecurve) and 56degC (blue curve)

RESEARCH | RESEARCH ARTICLE

(sc-28315) anti-p53WT and R273H (sc-126) sec-ondary goat anti-mouse HRP-IgG (sc-2055) andbovine anti-goat HRP-IgG (sc-2352) antibodies(Santa Cruz Biotechnology Dallas TX) were usedfor immunoblotting Chemiluminescence inten-sities were detected and quantified by using aChemiDocXRS+ imaging systemwith Image Labsoftware (Bio-Rad Hercules CA) Data were ex-pressed as means T SEM The data (band inten-sities) from multiple runs (n ge 3) were plottedwith Graphpad Prism software

Preparation of cell extract for thermalproteome profiling

Suspension cultures of K562 cells (15 to 2 times 106

cellsml) or of Jurkat E61 cells (4 times 106 cellsml)were centrifuged at 340 g for 2 min at 4degC Thecells were resuspended in 50 ml PBS After a sec-ond centrifugation step as above the cells wereresuspended in 10 ml ice-cold PBS and centri-fuged again at 340 g for 2 min at 4degC The cellswere resuspended in 15 ml of ice-cold PBS andsnap-frozen in liquidnitrogen The tubewasplacedinto a water bath at 23degC until ~60 of the con-tent was thawed and then transferred on ice un-til the entire contentwas thawed This freezethawcycle was repeated twice before the samplesweresubjected to ultracentrifugation (20 min at 4degCand 100000 g) The protein concentration of thesupernatant was determined by Bradford assay(Bio-Rad) and aliquots were snap-frozen in liquidnitrogen for use in subsequent thermal shiftassays

Thermal profiling using cell extract

A solution of the compound in dimethyl sulfoxide(DMSO) or DMSO alone as vehicle (Table 1) wasadded to the cell extract to 1 final DMSO con-centration The extract was then incubated for10 min at 23degC divided into 10 aliquots of 100 mland transferred into 02-ml polymerase chain re-action (PCR) tubes One each of the compoundand of the vehicle containing samples was heatedin parallel for 3 min to the respective tempera-ture followed by a 3-min incubation time at roomtemperature Subsequently the extract was cen-trifuged at 100000 g for 20 min at 4degC The su-pernatant was fractionated by means of SDS gel

electrophoresis and subjected to sample prepa-ration for MS analysis For ITDR experimentsGSK3182571 was tested as a nine-point serial dilu-tion starting at 100 mM (resulting in concentra-tions of 100 333 111 370 1235 0412 0137 0046and 0015 mM) including one vehicle control inan isothermal dose-response experiment at 53degCfollowing the procedure outlined above for ex-periments in cell extract

Thermal profiling using intact cells

To 24 ml of a suspension of K562 cells (at a den-sity of 15 106 cellsml) or Jurkat E61cells (at adensity of 4 106 cellsml) in a T-flask a volumeof either 120 ml of a solution of the referencecompound dissolved in DMSO or an equivalentamount of DMSO (vehicle) was added Incubationof cellswith compound and vehiclewas conductedin parallel for the duration of 1 hour at 37degC and5CO2 Cells were pelleted at 340 g and 4degC for2min resuspended in 20ml of ice cold PBS andcentrifuged again as indicated above This wash-ing step was repeated once and the cells werecarefully resuspended in 1200 ml PBS 100 ml ofthis resulting cell suspension was transferredinto 02-ml PCR tubes and subjected to a centrifu-gation step at 325 g for 2 min at 4degC Using apipette 80 ml of the PBS supernatant was re-moved By gently tapping the tubes the cells wereresuspended and one each of the compound andof the vehicle containing tubes was heated inparallel in a PCR machine for 3 min to the re-spective temperature (37degC to 67degC) followed bya 3-min incubation time at room temperatureThereafter the cells were snap-frozen in liquidnitrogen for 1 min The cells were thawed brieflyin a water bath at 25degC transferred on ice andresuspended by using a pipette This freeze-thawcycle was repeated once After an addition of fur-ther 30 ml of PBS and resuspension the entirecontent was centrifuged at 100000 g for 20 minat 4degC After centrifugation 30 ml of the super-natant were transferred into a new tube Proteinconcentration of the 37degC and 41degC samples weredetermined and used to normalize loading ofSDS gels Proteins in the supernatant were dena-tured by using SDS sample buffer partially sepa-rated by means of SDS gel electrophoresis and

subjected to sample preparation for MS analysisFor ITDR experiments vemurafenib alectiniband crizotinib were used as a nine-point serialdilution starting at 30 mM (concentrations were30 12 48 192 077 031 012 005 and 002 mM)including one vehicle control at 55degC followingthe procedure outlined aboveDasatinibwas testedas an eight-point serial dilution starting at 1 mM(concentrations usedwere 1 033 0111 0037 00120004 00013 and 000046 mM) including onedasatinib control at 10 mM and one vehicle con-trol on intact cells at 47degC 50degC and 56degC follow-ing the procedure outlined above for experimentson intact cells (Table 1)

Kinobeads assays

Competition binding assays were performed asdescribed previously by using a modified beadmatrix (5 22) Briefly 1 ml (5 mg protein) cellextract was pre-incubatedwith test compound orvehicle for 45 min at 4degC followed by incubationwith kinobeads (35 ml beads per sample) for1 hour at 4degC The beads were washed with lysisbuffer and eluted with 50 ml 2times SDS sample buf-fer GSK3182571was tested at 10 25 0625 015625003906 000977 and 000244 mM

Sample preparation for MS

Gel lanes were cut into three slices covering theentire separation range (~2 cm) and subjectedto in-gel digestion (5) Peptide samples were la-beled with 10-plex TMT (TMT10 Thermo FisherScientific Waltham MA) reagents enabling rel-ative quantification of a broad range of 10 tem-perature points in a single experiment (15 22)For kinobeads and ITDR experiments either10 or 8 TMT labels were used depending on thecompound concentration range chosen Thelabeling reaction was performed in 40 mMtriethylammoniumbicarbonate pH 853 at 22degCand quenched with glycine Labeled peptide ex-tracts were combined to a single sample per ex-periment and as indicated in the supplementarymaterials for most samples additional fraction-ation was performed by using reversed-phasechromatography at pH 12 [1mmXbridge column(Waters Milford MA)] as previously described(table S13) (35)

LC-MSMS analysis

Samples were dried in vacuo and resuspended in005 trifluoroacetic acid in water Of the sam-ple 50 was injected into an Ultimate3000nanoRLSC (Dionex Sunnyvale CA) coupled to aQ Exactive (Thermo Fisher Scientific) Peptideswere separated on custom 50 cm times 100 mM (ID)reversed-phase columns (Reprosil) at 40degC Gra-dient elutionwas performed from 2 acetonitrileto 40acetonitrile in 01 formic acid over 2 hoursSamples were online injected into Q-Exactive massspectrometers operating with a data-dependenttop 10methodMS spectrawere acquired by using70000 resolution and an ion target of 3E6Higherenergy collisional dissociation (HCD) scans wereperformed with 35 NCE at 35000 resolu-tion (at mz 200) and the ion target settingswas set to 2E5 so as to avoid coalescence (15) The

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-7

Table 1 Compoundsused in thermal profilingDashes indicate that compoundwas not tested in intact cells

CompoundConcentration used in

cellular extractexperiments

Concentration(s) inintact cell experiments

Cell line

MgATP 2 mM mdash K562staurosporine 20 mM mdash K562cAMP 1 mM mdash K562GSK3182571 20 mM mdash K562dasatinib 5 mM 05 mM 5 mM K562p53 WT

PG1 100 mM mdash A549PG2 100 mM mdash A549

p53 R273HPG1 100 mM mdash HT-29PG2 100 mM mdash HT-29

RESEARCH | RESEARCH ARTICLE

instruments were operated with Tune 22 or 23and Xcalibur 27 or 3063

Peptide and protein identification

Mascot 24 (Matrix Science Boston MA) wasused for protein identification by using a 10 partspermillionmass tolerance for peptide precursorsand 20 mD (HCD) mass tolerance for fragmentions Carbamidomethylation of cysteine residuesand TMTmodification of lysine residueswere setas fixedmodifications andmethionine oxidationand N-terminal acetylation of proteins and TMTmodification of peptide N-termini were set asvariable modifications The search database con-sisted of a customized version of the Interna-tional Protein Index protein sequence databasecombined with a decoy version of this databasecreated by using a script supplied by Matrix Sci-ence Unless stated otherwise we accepted proteinidentifications as follows (i) For single-spectrumto sequence assignments we required this assign-ment to be the bestmatch and aminimumMascotscore of 31 and a 10times difference of this assign-ment over the next best assignment Based onthese criteria the decoy search results indicatedlt1 false discovery rate (FDR) (ii) For multiplespectrum to sequence assignments and using thesame parameters the decoy search results indi-cate lt01 FDR All identified proteinswere quan-tified FDR for quantified proteins was lt1

Peptide and protein quantification

Reporter ion intensities were read from raw dataandmultiplied with ion accumulation times (theunit is milliseconds) so as to yield ameasure propor-tional to thenumberof ions thismeasure is referredto as ion area (36) Spectra matching to peptideswere filtered according to the following criteriamascot ion score gt15 signal-to-background of theprecursor ion gt4 and signal-to-interference gt05(37) Fold-changes were corrected for isotope pu-rity as described and adjusted for interferencecaused by co-eluting nearly isobaric peaks as esti-mated by the signal-to-interferencemeasure (38) Pro-tein quantification was derived from individualspectra matching to distinct peptides by using asum-based bootstrap algorithm 95 confidenceintervals were calculated for all protein fold-changesthatwere quantifiedwithmore than three spectra (36)

Curve fitting normalization andestimation of slope and melting pointdifferences for CETSA experiments

Melting curves

For all CETSA experiments fold-changes werecalculated by using the lowest temperature con-dition as the reference Relative fold changes as afunction of temperature followeda sigmoidal trendwhich was fitted with the following equationderived from chemical denaturation theory (39)(enwikipediaorgwikiEquilibrium_unfolding)using R (wwwr-projectorg)

f ethT THORN frac14 1 minus plateau

1thorn eminusaTminusbeth THORN thorn plateau

Where T is the temperature and a b and plateauare constants The value of f(T) at the lowest

temperature Tmin was fixed to one The meltingpoint of a protein is defined as the temperatureTm at which half of the protein amount has beendenatured

f(Tm) = 05

Additionally the slope of the melting curve isdefined as the value of the first derivative of f(T)at the point T = Tinfl

slope = f prime(Tinfl)

where Tinfl is the inflection point of the curvemeaning that the second derivative of f (T) equalszero at T = Tinfl

f primeprime(Tinfl) = 0

Normalization

The vehicle and compound-treated experimentswere normalized in the following waySelection of proteins for normalization (i) The

set of proteins identified in both experimentstermed ldquojointPrdquo was determined (ii) For eachexperiment (vehicle or compound) all proteinsfrom jointP fulfilling the following criteria werecollected the fold-changes at the seventh highesttemperature point comparedwith the lowest tem-perature point were between 04 and 06 andat the 9th and 10th highest temperature pointscompared th the lowest temperature pointwere below 03 and 02 respectively (iii) Thebigger of these two protein subsets was thenused for normalization (typically more than200 proteins qualified) This protein set is re-ferred to as ldquonormPrdquoCalculation of correction factors (i) The me-

dian fold-change values over all proteins in normPwere calculated for each of the 10 fold-changesThis was done separately for both experimentsand for each experiment a melting curve wasfitted The curve that was fitted with the best R2

was used to normalize both experiments (ii) Foreach experiment the correction factors werecalculated by using the median fold-changes ofthe proteins in normP For each median fold-change a correction factor by which that fold-change would have to be multiplied with in orderto coincidewith the best-fitting curvewas calculatedNormalization The two sets of correction fac-

tors were applied to all the protein fold-changesin the respective experimentsThis heuristic normalization procedure covers

three important aspects (i) The same subset ofproteins is used to determine normalization co-efficients in both experiments thus circumvent-ing any potential bias caused by proteins onlyidentified in one experiment (ii) The subselectionof late-melting proteins (seventh point between04 and 06) ensures that the normalization ofhigh-temperature points will be more accurate(iii) Last the narrow requirements on the 7th9th and 10th points will ensure that a narrowrange of melting profiles is used for calculatingthe points from which the normalization curvewill be derived This is important because if vastlydifferent melting profiles are combined into asingle one the above equation will not necessar-ily fit well A normalization curve example as well

as melting curves of several proteins before andafter normalization are illustrated in fig S12

Estimation of slope and meltingpoint differences

After normalization melting curves were fittedfor all proteins identified in the vehicle- andcompound-treated experiments Melting pointdifferences and slope differences were calculatedfor proteins whose curves passed the followingtwo requirements (i) both fitted curves for thevehicle and compound treated condition had anR2 of greater than 08 and (ii) the vehicle curvehad a plateau of less than 03Typically close to 80 of proteins that were

identified in both the vehicle- and compound-treated experiment passed these criteria Theslope of the curves had the biggest impact onthe reproducibility of melting point differencesof proteins This observation can be rationalizedin the following way If the curve has a shallowslope then small changes with respect to the ref-erence pointmdashin our case the lowest temperaturemdashwill have a strong impact on where the curvecrosses the 05 fold-change level (melting point)Consequently small variations in the quantitativemeasurement of theMS signals corresponding tothe lowest temperature will lead to strong shiftsin the inferred melting point of the curve Thismeans that it is difficult to get high precisionmeasurements ofmelting point differenceswhencomparing two curves with shallow slopes If thecurve however has a steep slope then the melt-ing point measurement will not be substantiallyaffected by small variations of the quantitativemeasurements of the lowest temperature nor ofany other temperature points Taking this intoaccount we used the following strategy for sig-nificance estimation of protein melting point dif-ferences between vehicle- and compound-treatedexperiments which is conceptually very similarto a broadly used strategy for inferring significantprotein fold-changes (40) The calculated meltingpoint differences of proteins were ordered in des-cending slope order (those with the shallowestslope come first) in which the slope for eachprotein is the lowest (steepest) slope calculatedfrom the two curve fits performed for the proteinin the vehicle- and compound-treated conditionThe proteins were divided into bins The bins areconstructed starting fromproteinswith the high-est (shallowest) slope Each bin consists of at least300 proteins Once each bin has been completedthe remaining number of proteins is counted ifthis number is below 300 the remaining pro-teins are added to the last completed bin Thisdata qualityndashdependent binning strategy is analo-gous to the procedure described in Cox et al withthe only difference being that proteins are sortedby slope instead of abundance (40) The statisticalsignificance of differences in protein meltingpoints was calculated by estimating the left- andright-sided robust standard deviation of the dis-tribution of the measurements using the 158750 and 8413 percentiles and calculating the Pvalues for all measurements for a specific binexactly as previously described (40) Subsequently

1255784-8 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

RESEARCH | RESEARCH ARTICLE

changes in ligand-induced thermostability andto enable the ranking of binding affinities amongmultiple targets we tailored the isothermal dose-response (ITDR) protocol to a defined set of pro-teins (12) In contrast to the full temperatureprofiles in which a compound is assayed at asingle typically saturating concentration across10 temperatures an ITDR profile is generated ata defined temperature over a range of compoundconcentrations To generate affinity data forGSK3182571 we acquired ITDR profiles (tableS6) which showed excellent reproducibility (figS8) and were in good agreement with data fromkinobeads competition-binding experiments bothfor proteins stabilized and proteins destabilizedby ligand binding (Fig 4B and table S7)

Because of its unbiased nature thermal pro-teome profiling can detect unexpected off-targetsof drug compounds The heme biosynthesis path-way has not been related to adverse drug effectsbut is linked to a number of genetic disordersincluding erythropoietic protoporphyria which iscaused by a deficiency in FECH and results inhigh tissue levels of protoporphyrins (24) Wefound two enzymes in the heme pathway inter-actingwith staurosporine and thus askedwhetherdrug interactions with this pathway might resultin adverse effects such as skin or liver disease (25)The melanoma drug vemurafenib frequentlycauses severe photosensitivity concomitant withincreased levels of protoporphyrins (26) ITDRprofiling of cells treatedwith vemurafenib showed

that the main proteins affected were its knowntarget BRAF [negative logarithm of the medianeffective concentration (pEC50) = 61] and FECH(pEC50 = 53) (Fig 5A and table S8) These con-centrations are below the typical steady-state ex-posure for this drug (27) and indicate a less than10-fold window of target occupancy between thedrugrsquos efficacy target and FECH Alectinib asecond-generation anaplastic lymphoma kinase(ALK) inhibitor for the treatment of nonndashsmall-cell lung cancer was also reported to cause photo-sensitivity in contrast to the first-generation drugcrizotinib (28 29) ITDR profiling revealed thatalectinib more potently affected FECH than didvemurafenib whereas crizotinib had no effect(Fig 5B) These results strongly suggest that the

1255784-4 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 3 Differential profiling of drug effects on the thermal proteomeprofile Staurosporine treatment of cell extract yields reproducible thermalshifts allowing robust target identification (A) Cell extract was treated withvehicle or staurosporine Both experiments were performed as two fullyindependent replicates A flat slope of the melting curve relates to lowermelting point reproducibility Melting point differences from the two vehicleexperiments are plotted against the shallowest melting curve slope observedin the four vehiclestaurosporine data sets Proteins with an absolute slopebelow 006 are plotted in gray The histogram shows the distribution ofproteins versus slope values Of all proteins 92 yielded curves with suf-ficiently large slopes (B) Scatter plot of Tm shifts calculated from the two

replicates of the staurosporine versus vehicle treatment experimentStaurosporine-induced Tm shifts that passed the significance criteria areshown in red Proteins with flat slopesmdashas shown in gray in (A)mdashare againshown in gray (C) Comparison between Tm shifts and pIC50 values forstaurosporine as reported from kinobeads data (22) (D) Examples ofmeltingcurves for PKN1 FECH AURKA and SLK with and without staurosporinetreatment (E) The catalytic subunit of PKA is stabilized by staurosporinewhereas the regulatory subunit is destabilized Addition of cAMP followed byWestern blot detection revealed destabilization of the catalytic subunit andstabilization of the regulatory subunit Error bars indicate the SEM from n = 4experiments

RESEARCH | RESEARCH ARTICLE

photosensitivity induced by these drugs is me-diated by FECH and demonstrates that thermalproteome profiling can serve as a stand-alone tech-nique for obtaining quantitative affinity data fortarget-ligand interactions in a cell-based setting

Drug treatment induces Tm shifts fordownstream effector proteins

Any modification of a protein can in principleaffect its thermal stability Therefore compari-son of Tm shifts in cell extract where ligand bind-ing but no downstream effects occur with Tmshifts in intact cells where active signaling takesplace might reveal effector proteins downstreamof the target As a model system we used K562

cells which are transformed by an oncogenic fu-sion protein the BCR-ABL kinase (30) Overallthe samples from intact cells yielded a similarprotein coverage as that of samples from cellextracts (tables S9 and S10) Cultured cells weretreated with vehicle or with the ABL inhibitordasatinib a marketed drug against chronic mye-logenous leukemia (CML) (31) at two concen-trations 05 and 5 mM and the samples wereprofiled (Fig 6A and table S9) Theminimum Tmshift for each protein was calculated from a pairof replicate experiments Four proteins displayedsignificant melting point differences in both the05 and 5 mM dasatinibndashtreated cell samples Wedid not observe a Tm shift for the direct target

BCR-ABL suggesting that this protein is notstabilized by dasatinib binding However sub-stantial and significant Tm shifts were observedfor the adaptor proteinCRKLand thephosphataseSHIP2 (Fig 6B and fig S9) which is also known asINPPL1 These proteins are not likely binding thedrug directly and consistent with this no Tmshift was observed for either protein in cell extracts(table S10) In addition the BCR-ABLndashinteractingproteinCRKshowedapronounced change inmeltingcurve slope in the dasatinib-treated cells (fig S9)CRK and CRKL aremembers of a proto-oncogenefamily that bind to tyrosine-phosphorylated pro-teins and are known downstream effectors ofBCR-ABL signaling (32) CRKLhas been previouslyproposed as a treatment response biomarker (33)Comparison of the CRKL melting curves in K562cells with those in Jurkat cells (table S11) whichare not transformed by tyrosine kinase signalingrevealed that dasatinib treatment of K562 cellsreverted the thermal stability of CRKL to the lev-el observed in Jurkat cells (Fig 6C) and the sameeffect was observed for CRK (fig S10) The Jurkatcell profiles also revealed that ABL1 kinase ap-peared to be more thermostable compared withthe BCR-ABL fusion protein in K562 cells suggest-ing that thermal profiling could be used to identifyprotein fusions which are difficult to identify byexpression proteomics (Fig 6C) In order to eval-uate the drug concentration dependence of theTm-shifted effector proteins we performed ITDRprofiles at three different temperatures (Fig 6Dfig S11 and table S12) The concentrations elicit-ing half-maximal response of the CRKL markerin the ITDR profiles recorded at three differenttemperatureswere between 15 and 32 nMwhichis in good agreement with the known potency ofdasatinib for the inhibition of cell growth (34) In-depth analysis of the 50degC ITDR data revealedadditional markers that either showed small Tmshifts or in the case of the known downstreameffector CRK a pronounced change in the melt-ing curve slope indicating that the ITDRprofiles canoffer greater sensitivity than the full-temperatureprofiles (fig S11) It is tempting to speculate thatthe most sensitive biomarkers are those thatexhibit large or even stoichiometric changes inprotein state which should be robustly detectedby their Tm shift This contrasts with the use of a

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-5

Fig 4 Structurally unrelated kinase inhibitors with sim-ilar affinities induce Tm shifts of comparable magnitudebut assessment of ligand affinity requires ITDR mea-surements (A) Staurosporine Tm shift data are comparedwith data for a structurally unrelated promiscuous kinaseinhibitor GSK3182571 Comparison of Tm shifts induced bystaurosporine or GSK3182571 determined in cell extractProtein targets shown had a pIC50 gt 6 both for staurosporineand GSK3182571 as determined bymeans of kinobeads anddid not differ in affinity between the two compounds bymorethan one order of magnitude (B) Good agreement betweenGSK3182571 pEC50 values determined by means of ITDR with pIC50 values determined with kinobeads(Top) Proteins stabilized by GSK3182571 treatment (Bottom) Proteins destabilized by GSK3182571treatment

Fig 5 The clinical kinase drugsvemurafenib and alectinib which can causephototoxicity as a side effect induce Tmshifts in the heme biosynthesis enzymeFECH (A) Concentration-dependent targetoccupancy of vemurafenib and its cognatetarget BRAF compared with the off-targetFECH ITDR profiling was performed at 55degCwith vemurafenib-treated K562 cells andshowed concentration-dependent thermalstabilization of FECH and BRAF The data arenormalized so that the quantity of solubletarget at the highest compoundconcentration which reflects maximumthermal stabilization is set to 1 and fixed forcurve-fitting (B) ITDR performed at 55degC with K562 cells treated with vemurafenib alectinib or crizotinib over a range of concentrations Alectinib displays amore potent effect on FECH as compared with that by vemurafenib whereas crizotinib a drug not known to cause photosensitivity has no effect

RESEARCH | RESEARCH ARTICLE

posttranslational modification as a biomarkerwhere it is usually difficult to assess the stoi-chiometry of the modificationIn general terms the thermal stability of a

protein may be defined by its mutational statusposttranslationalmodifications and bound ligandssuch as other proteins cofactors metabolites ordrugs Fusionproteins splice variants andposttrans-lational modifications are typically undersampledand therefore not comprehensively detected inMS-basedproteomicsHence the thermal proteomeprofile of a human cell can provide a generalview of the proteomic state or proteotypeThe future scope and applicability of thermal

proteome profilingwill substantially benefit fromcontinued advances in the sensitivity and accu-racyofMS-basedproteomics Futuredevelopmentsshould include the development of protocols formembrane proteins and the application to tis-sues from animal studies or clinical biopsies

Materials and methods

Reagents and cell culture

Reagents andmedia were purchased from Sigma-Aldrich (St Louis MO) unless otherwise notedPBS was prepared by using 137 mM NaCl 27mM KCl 10 mM Na2HPO4 and 2 mM KH2PO4

to buffer at pH 74 and supplemented withcomplete EDTA-free protease inhibitor cocktail(one tablet per 25 ml) (Roche Diagnostics Basel

Switzerland) Staurosporine was from Calbiochem(Millipore Billerica MA) vemurafenib andcrizotinib from Selleck (Houston TX) alectinibfromMedchemExpress (Monmouth Junction NJ)and dasatinib from Toronto Research (TorontoCanada) The synthesis of GSK3182571 is de-scribed in the supplementary methods DNAfragments PG1 (5prime-AGC TTA GAC ATG CCT AGACAT GCC TA -3prime) and PG2 (5prime-AGC TTA GGC ATGTCT AGG CAT GTC TA -3prime) were from Life Tech-nologies (Carlsbad CA) and dissolved in 100 mMTris (pH75) 1 mM NaCl and 10mM EDTA K562Jurkat E61 and A549 cells were from ATCC K562and A549 cells were cultured in RPMI mediumcontaining 10 fetal calf serum (FCS) and JurkatE61 cells in RPMI1640 supplemented with45 gl glucose 10 mM HEPES 1 mM sodiumpyruvate and 10 FCS HT-29 (ATTC no HTB-38)were maintained in Dulbeccorsquos Modified EagleMedium Culture media were supplemented with03 gL L-glutamine and 10 fetal bovine serum(FBS) (Gibco Carlsbad CA) 100 unitsmL peni-cillin and 100 unitsmL streptomycin The cellswere expanded to amaximumof 2times 106 cellsml incase of K562 and 107 cellsml in case of Jurkat E61

Preparation of cell extract forCETSA with Western blot read-out

A549 andHT-29 cells were harvested andwashedwith PBS The cells were diluted in KB buffer

(25mMTris-HCl supplemented with 5mMbeta-glycerophosphate 2 mM DTT 01 mM Na3VO410 mM MgCl2 and complete protease inhibitorcocktail) The cell suspensionswere freeze-thawedthree times by using liquid nitrogen The solublefraction was separated from the cell debris bycentrifugation at 20000 g for 20 min at 4degC FortheCETSAmelting curve experiments cell extractswere diluted with KB buffer and divided into twoaliquots with one aliquot being treated withligandDNA fragments depending on the targetand the other aliquot treated with the solvent ofthe ligand (control) After 10 min incubation atroom temperature the respective cell extractswere heated individually at different temper-atures for 3 min in a thermal cycler [AppliedBiosystems (Foster City CA)Life Technologies]followed by cooling for 3 min at room temper-ature The heated cell extracts were centrifugedat 20000 g for 20min at 4degC in order to separatethe soluble fractions from precipitates The su-pernatantswere transferred to new02mLmicro-tubes and analyzed by means of Western blotanalysis NuPage Bis-Tris 4-12 polyacrylamidegels with MES SDS running buffer (Life Tech-nologies) were used for separation of proteins inthe samples Proteins were transferred to nitro-cellulose membranes by using the iBlot system(Life Technologies) Primary antibodies anti-PKARegulatory I-a (sc-136231) anti-PKA Catalytic-a

1255784-6 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 6 Treatment of K562 cells with dasatinib induces Tm shifts fordownstream effector proteins in the BCR-ABL pathway (A) Cell-wideassessment of Tm shifts induced by dasatinib (05 and 5 mM) in K562 cellsThe minimum Tm shift for each protein was calculated from a pair of fullreplicate experiments Four proteins showing significant melting pointdifferences both in the 05 and 5 mM dasatinib data sets are marked in red(B) Effect of dasatinib on the melting curves for the effector protein CRKL

determined in two biological replicates in intact cells (top) and cell extracts(bottom) (C) The melting curve for CRKL in Jurkat cells closely matchesthe curve obtained in dasatinib-treated K562 cells (top)The ABL1 protein inJurkat cells is much more thermostable compared with the BCR-ABL fusionprotein in K562 cells (D) ITDR curves for CRKL in dasatinib-treated intactcells Experiments were performed at 47degC (green curve) 50degC (purplecurve) and 56degC (blue curve)

RESEARCH | RESEARCH ARTICLE

(sc-28315) anti-p53WT and R273H (sc-126) sec-ondary goat anti-mouse HRP-IgG (sc-2055) andbovine anti-goat HRP-IgG (sc-2352) antibodies(Santa Cruz Biotechnology Dallas TX) were usedfor immunoblotting Chemiluminescence inten-sities were detected and quantified by using aChemiDocXRS+ imaging systemwith Image Labsoftware (Bio-Rad Hercules CA) Data were ex-pressed as means T SEM The data (band inten-sities) from multiple runs (n ge 3) were plottedwith Graphpad Prism software

Preparation of cell extract for thermalproteome profiling

Suspension cultures of K562 cells (15 to 2 times 106

cellsml) or of Jurkat E61 cells (4 times 106 cellsml)were centrifuged at 340 g for 2 min at 4degC Thecells were resuspended in 50 ml PBS After a sec-ond centrifugation step as above the cells wereresuspended in 10 ml ice-cold PBS and centri-fuged again at 340 g for 2 min at 4degC The cellswere resuspended in 15 ml of ice-cold PBS andsnap-frozen in liquidnitrogen The tubewasplacedinto a water bath at 23degC until ~60 of the con-tent was thawed and then transferred on ice un-til the entire contentwas thawed This freezethawcycle was repeated twice before the samplesweresubjected to ultracentrifugation (20 min at 4degCand 100000 g) The protein concentration of thesupernatant was determined by Bradford assay(Bio-Rad) and aliquots were snap-frozen in liquidnitrogen for use in subsequent thermal shiftassays

Thermal profiling using cell extract

A solution of the compound in dimethyl sulfoxide(DMSO) or DMSO alone as vehicle (Table 1) wasadded to the cell extract to 1 final DMSO con-centration The extract was then incubated for10 min at 23degC divided into 10 aliquots of 100 mland transferred into 02-ml polymerase chain re-action (PCR) tubes One each of the compoundand of the vehicle containing samples was heatedin parallel for 3 min to the respective tempera-ture followed by a 3-min incubation time at roomtemperature Subsequently the extract was cen-trifuged at 100000 g for 20 min at 4degC The su-pernatant was fractionated by means of SDS gel

electrophoresis and subjected to sample prepa-ration for MS analysis For ITDR experimentsGSK3182571 was tested as a nine-point serial dilu-tion starting at 100 mM (resulting in concentra-tions of 100 333 111 370 1235 0412 0137 0046and 0015 mM) including one vehicle control inan isothermal dose-response experiment at 53degCfollowing the procedure outlined above for ex-periments in cell extract

Thermal profiling using intact cells

To 24 ml of a suspension of K562 cells (at a den-sity of 15 106 cellsml) or Jurkat E61cells (at adensity of 4 106 cellsml) in a T-flask a volumeof either 120 ml of a solution of the referencecompound dissolved in DMSO or an equivalentamount of DMSO (vehicle) was added Incubationof cellswith compound and vehiclewas conductedin parallel for the duration of 1 hour at 37degC and5CO2 Cells were pelleted at 340 g and 4degC for2min resuspended in 20ml of ice cold PBS andcentrifuged again as indicated above This wash-ing step was repeated once and the cells werecarefully resuspended in 1200 ml PBS 100 ml ofthis resulting cell suspension was transferredinto 02-ml PCR tubes and subjected to a centrifu-gation step at 325 g for 2 min at 4degC Using apipette 80 ml of the PBS supernatant was re-moved By gently tapping the tubes the cells wereresuspended and one each of the compound andof the vehicle containing tubes was heated inparallel in a PCR machine for 3 min to the re-spective temperature (37degC to 67degC) followed bya 3-min incubation time at room temperatureThereafter the cells were snap-frozen in liquidnitrogen for 1 min The cells were thawed brieflyin a water bath at 25degC transferred on ice andresuspended by using a pipette This freeze-thawcycle was repeated once After an addition of fur-ther 30 ml of PBS and resuspension the entirecontent was centrifuged at 100000 g for 20 minat 4degC After centrifugation 30 ml of the super-natant were transferred into a new tube Proteinconcentration of the 37degC and 41degC samples weredetermined and used to normalize loading ofSDS gels Proteins in the supernatant were dena-tured by using SDS sample buffer partially sepa-rated by means of SDS gel electrophoresis and

subjected to sample preparation for MS analysisFor ITDR experiments vemurafenib alectiniband crizotinib were used as a nine-point serialdilution starting at 30 mM (concentrations were30 12 48 192 077 031 012 005 and 002 mM)including one vehicle control at 55degC followingthe procedure outlined aboveDasatinibwas testedas an eight-point serial dilution starting at 1 mM(concentrations usedwere 1 033 0111 0037 00120004 00013 and 000046 mM) including onedasatinib control at 10 mM and one vehicle con-trol on intact cells at 47degC 50degC and 56degC follow-ing the procedure outlined above for experimentson intact cells (Table 1)

Kinobeads assays

Competition binding assays were performed asdescribed previously by using a modified beadmatrix (5 22) Briefly 1 ml (5 mg protein) cellextract was pre-incubatedwith test compound orvehicle for 45 min at 4degC followed by incubationwith kinobeads (35 ml beads per sample) for1 hour at 4degC The beads were washed with lysisbuffer and eluted with 50 ml 2times SDS sample buf-fer GSK3182571was tested at 10 25 0625 015625003906 000977 and 000244 mM

Sample preparation for MS

Gel lanes were cut into three slices covering theentire separation range (~2 cm) and subjectedto in-gel digestion (5) Peptide samples were la-beled with 10-plex TMT (TMT10 Thermo FisherScientific Waltham MA) reagents enabling rel-ative quantification of a broad range of 10 tem-perature points in a single experiment (15 22)For kinobeads and ITDR experiments either10 or 8 TMT labels were used depending on thecompound concentration range chosen Thelabeling reaction was performed in 40 mMtriethylammoniumbicarbonate pH 853 at 22degCand quenched with glycine Labeled peptide ex-tracts were combined to a single sample per ex-periment and as indicated in the supplementarymaterials for most samples additional fraction-ation was performed by using reversed-phasechromatography at pH 12 [1mmXbridge column(Waters Milford MA)] as previously described(table S13) (35)

LC-MSMS analysis

Samples were dried in vacuo and resuspended in005 trifluoroacetic acid in water Of the sam-ple 50 was injected into an Ultimate3000nanoRLSC (Dionex Sunnyvale CA) coupled to aQ Exactive (Thermo Fisher Scientific) Peptideswere separated on custom 50 cm times 100 mM (ID)reversed-phase columns (Reprosil) at 40degC Gra-dient elutionwas performed from 2 acetonitrileto 40acetonitrile in 01 formic acid over 2 hoursSamples were online injected into Q-Exactive massspectrometers operating with a data-dependenttop 10methodMS spectrawere acquired by using70000 resolution and an ion target of 3E6Higherenergy collisional dissociation (HCD) scans wereperformed with 35 NCE at 35000 resolu-tion (at mz 200) and the ion target settingswas set to 2E5 so as to avoid coalescence (15) The

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-7

Table 1 Compoundsused in thermal profilingDashes indicate that compoundwas not tested in intact cells

CompoundConcentration used in

cellular extractexperiments

Concentration(s) inintact cell experiments

Cell line

MgATP 2 mM mdash K562staurosporine 20 mM mdash K562cAMP 1 mM mdash K562GSK3182571 20 mM mdash K562dasatinib 5 mM 05 mM 5 mM K562p53 WT

PG1 100 mM mdash A549PG2 100 mM mdash A549

p53 R273HPG1 100 mM mdash HT-29PG2 100 mM mdash HT-29

RESEARCH | RESEARCH ARTICLE

instruments were operated with Tune 22 or 23and Xcalibur 27 or 3063

Peptide and protein identification

Mascot 24 (Matrix Science Boston MA) wasused for protein identification by using a 10 partspermillionmass tolerance for peptide precursorsand 20 mD (HCD) mass tolerance for fragmentions Carbamidomethylation of cysteine residuesand TMTmodification of lysine residueswere setas fixedmodifications andmethionine oxidationand N-terminal acetylation of proteins and TMTmodification of peptide N-termini were set asvariable modifications The search database con-sisted of a customized version of the Interna-tional Protein Index protein sequence databasecombined with a decoy version of this databasecreated by using a script supplied by Matrix Sci-ence Unless stated otherwise we accepted proteinidentifications as follows (i) For single-spectrumto sequence assignments we required this assign-ment to be the bestmatch and aminimumMascotscore of 31 and a 10times difference of this assign-ment over the next best assignment Based onthese criteria the decoy search results indicatedlt1 false discovery rate (FDR) (ii) For multiplespectrum to sequence assignments and using thesame parameters the decoy search results indi-cate lt01 FDR All identified proteinswere quan-tified FDR for quantified proteins was lt1

Peptide and protein quantification

Reporter ion intensities were read from raw dataandmultiplied with ion accumulation times (theunit is milliseconds) so as to yield ameasure propor-tional to thenumberof ions thismeasure is referredto as ion area (36) Spectra matching to peptideswere filtered according to the following criteriamascot ion score gt15 signal-to-background of theprecursor ion gt4 and signal-to-interference gt05(37) Fold-changes were corrected for isotope pu-rity as described and adjusted for interferencecaused by co-eluting nearly isobaric peaks as esti-mated by the signal-to-interferencemeasure (38) Pro-tein quantification was derived from individualspectra matching to distinct peptides by using asum-based bootstrap algorithm 95 confidenceintervals were calculated for all protein fold-changesthatwere quantifiedwithmore than three spectra (36)

Curve fitting normalization andestimation of slope and melting pointdifferences for CETSA experiments

Melting curves

For all CETSA experiments fold-changes werecalculated by using the lowest temperature con-dition as the reference Relative fold changes as afunction of temperature followeda sigmoidal trendwhich was fitted with the following equationderived from chemical denaturation theory (39)(enwikipediaorgwikiEquilibrium_unfolding)using R (wwwr-projectorg)

f ethT THORN frac14 1 minus plateau

1thorn eminusaTminusbeth THORN thorn plateau

Where T is the temperature and a b and plateauare constants The value of f(T) at the lowest

temperature Tmin was fixed to one The meltingpoint of a protein is defined as the temperatureTm at which half of the protein amount has beendenatured

f(Tm) = 05

Additionally the slope of the melting curve isdefined as the value of the first derivative of f(T)at the point T = Tinfl

slope = f prime(Tinfl)

where Tinfl is the inflection point of the curvemeaning that the second derivative of f (T) equalszero at T = Tinfl

f primeprime(Tinfl) = 0

Normalization

The vehicle and compound-treated experimentswere normalized in the following waySelection of proteins for normalization (i) The

set of proteins identified in both experimentstermed ldquojointPrdquo was determined (ii) For eachexperiment (vehicle or compound) all proteinsfrom jointP fulfilling the following criteria werecollected the fold-changes at the seventh highesttemperature point comparedwith the lowest tem-perature point were between 04 and 06 andat the 9th and 10th highest temperature pointscompared th the lowest temperature pointwere below 03 and 02 respectively (iii) Thebigger of these two protein subsets was thenused for normalization (typically more than200 proteins qualified) This protein set is re-ferred to as ldquonormPrdquoCalculation of correction factors (i) The me-

dian fold-change values over all proteins in normPwere calculated for each of the 10 fold-changesThis was done separately for both experimentsand for each experiment a melting curve wasfitted The curve that was fitted with the best R2

was used to normalize both experiments (ii) Foreach experiment the correction factors werecalculated by using the median fold-changes ofthe proteins in normP For each median fold-change a correction factor by which that fold-change would have to be multiplied with in orderto coincidewith the best-fitting curvewas calculatedNormalization The two sets of correction fac-

tors were applied to all the protein fold-changesin the respective experimentsThis heuristic normalization procedure covers

three important aspects (i) The same subset ofproteins is used to determine normalization co-efficients in both experiments thus circumvent-ing any potential bias caused by proteins onlyidentified in one experiment (ii) The subselectionof late-melting proteins (seventh point between04 and 06) ensures that the normalization ofhigh-temperature points will be more accurate(iii) Last the narrow requirements on the 7th9th and 10th points will ensure that a narrowrange of melting profiles is used for calculatingthe points from which the normalization curvewill be derived This is important because if vastlydifferent melting profiles are combined into asingle one the above equation will not necessar-ily fit well A normalization curve example as well

as melting curves of several proteins before andafter normalization are illustrated in fig S12

Estimation of slope and meltingpoint differences

After normalization melting curves were fittedfor all proteins identified in the vehicle- andcompound-treated experiments Melting pointdifferences and slope differences were calculatedfor proteins whose curves passed the followingtwo requirements (i) both fitted curves for thevehicle and compound treated condition had anR2 of greater than 08 and (ii) the vehicle curvehad a plateau of less than 03Typically close to 80 of proteins that were

identified in both the vehicle- and compound-treated experiment passed these criteria Theslope of the curves had the biggest impact onthe reproducibility of melting point differencesof proteins This observation can be rationalizedin the following way If the curve has a shallowslope then small changes with respect to the ref-erence pointmdashin our case the lowest temperaturemdashwill have a strong impact on where the curvecrosses the 05 fold-change level (melting point)Consequently small variations in the quantitativemeasurement of theMS signals corresponding tothe lowest temperature will lead to strong shiftsin the inferred melting point of the curve Thismeans that it is difficult to get high precisionmeasurements ofmelting point differenceswhencomparing two curves with shallow slopes If thecurve however has a steep slope then the melt-ing point measurement will not be substantiallyaffected by small variations of the quantitativemeasurements of the lowest temperature nor ofany other temperature points Taking this intoaccount we used the following strategy for sig-nificance estimation of protein melting point dif-ferences between vehicle- and compound-treatedexperiments which is conceptually very similarto a broadly used strategy for inferring significantprotein fold-changes (40) The calculated meltingpoint differences of proteins were ordered in des-cending slope order (those with the shallowestslope come first) in which the slope for eachprotein is the lowest (steepest) slope calculatedfrom the two curve fits performed for the proteinin the vehicle- and compound-treated conditionThe proteins were divided into bins The bins areconstructed starting fromproteinswith the high-est (shallowest) slope Each bin consists of at least300 proteins Once each bin has been completedthe remaining number of proteins is counted ifthis number is below 300 the remaining pro-teins are added to the last completed bin Thisdata qualityndashdependent binning strategy is analo-gous to the procedure described in Cox et al withthe only difference being that proteins are sortedby slope instead of abundance (40) The statisticalsignificance of differences in protein meltingpoints was calculated by estimating the left- andright-sided robust standard deviation of the dis-tribution of the measurements using the 158750 and 8413 percentiles and calculating the Pvalues for all measurements for a specific binexactly as previously described (40) Subsequently

1255784-8 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

RESEARCH | RESEARCH ARTICLE

photosensitivity induced by these drugs is me-diated by FECH and demonstrates that thermalproteome profiling can serve as a stand-alone tech-nique for obtaining quantitative affinity data fortarget-ligand interactions in a cell-based setting

Drug treatment induces Tm shifts fordownstream effector proteins

Any modification of a protein can in principleaffect its thermal stability Therefore compari-son of Tm shifts in cell extract where ligand bind-ing but no downstream effects occur with Tmshifts in intact cells where active signaling takesplace might reveal effector proteins downstreamof the target As a model system we used K562

cells which are transformed by an oncogenic fu-sion protein the BCR-ABL kinase (30) Overallthe samples from intact cells yielded a similarprotein coverage as that of samples from cellextracts (tables S9 and S10) Cultured cells weretreated with vehicle or with the ABL inhibitordasatinib a marketed drug against chronic mye-logenous leukemia (CML) (31) at two concen-trations 05 and 5 mM and the samples wereprofiled (Fig 6A and table S9) Theminimum Tmshift for each protein was calculated from a pairof replicate experiments Four proteins displayedsignificant melting point differences in both the05 and 5 mM dasatinibndashtreated cell samples Wedid not observe a Tm shift for the direct target

BCR-ABL suggesting that this protein is notstabilized by dasatinib binding However sub-stantial and significant Tm shifts were observedfor the adaptor proteinCRKLand thephosphataseSHIP2 (Fig 6B and fig S9) which is also known asINPPL1 These proteins are not likely binding thedrug directly and consistent with this no Tmshift was observed for either protein in cell extracts(table S10) In addition the BCR-ABLndashinteractingproteinCRKshowedapronounced change inmeltingcurve slope in the dasatinib-treated cells (fig S9)CRK and CRKL aremembers of a proto-oncogenefamily that bind to tyrosine-phosphorylated pro-teins and are known downstream effectors ofBCR-ABL signaling (32) CRKLhas been previouslyproposed as a treatment response biomarker (33)Comparison of the CRKL melting curves in K562cells with those in Jurkat cells (table S11) whichare not transformed by tyrosine kinase signalingrevealed that dasatinib treatment of K562 cellsreverted the thermal stability of CRKL to the lev-el observed in Jurkat cells (Fig 6C) and the sameeffect was observed for CRK (fig S10) The Jurkatcell profiles also revealed that ABL1 kinase ap-peared to be more thermostable compared withthe BCR-ABL fusion protein in K562 cells suggest-ing that thermal profiling could be used to identifyprotein fusions which are difficult to identify byexpression proteomics (Fig 6C) In order to eval-uate the drug concentration dependence of theTm-shifted effector proteins we performed ITDRprofiles at three different temperatures (Fig 6Dfig S11 and table S12) The concentrations elicit-ing half-maximal response of the CRKL markerin the ITDR profiles recorded at three differenttemperatureswere between 15 and 32 nMwhichis in good agreement with the known potency ofdasatinib for the inhibition of cell growth (34) In-depth analysis of the 50degC ITDR data revealedadditional markers that either showed small Tmshifts or in the case of the known downstreameffector CRK a pronounced change in the melt-ing curve slope indicating that the ITDRprofiles canoffer greater sensitivity than the full-temperatureprofiles (fig S11) It is tempting to speculate thatthe most sensitive biomarkers are those thatexhibit large or even stoichiometric changes inprotein state which should be robustly detectedby their Tm shift This contrasts with the use of a

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-5

Fig 4 Structurally unrelated kinase inhibitors with sim-ilar affinities induce Tm shifts of comparable magnitudebut assessment of ligand affinity requires ITDR mea-surements (A) Staurosporine Tm shift data are comparedwith data for a structurally unrelated promiscuous kinaseinhibitor GSK3182571 Comparison of Tm shifts induced bystaurosporine or GSK3182571 determined in cell extractProtein targets shown had a pIC50 gt 6 both for staurosporineand GSK3182571 as determined bymeans of kinobeads anddid not differ in affinity between the two compounds bymorethan one order of magnitude (B) Good agreement betweenGSK3182571 pEC50 values determined by means of ITDR with pIC50 values determined with kinobeads(Top) Proteins stabilized by GSK3182571 treatment (Bottom) Proteins destabilized by GSK3182571treatment

Fig 5 The clinical kinase drugsvemurafenib and alectinib which can causephototoxicity as a side effect induce Tmshifts in the heme biosynthesis enzymeFECH (A) Concentration-dependent targetoccupancy of vemurafenib and its cognatetarget BRAF compared with the off-targetFECH ITDR profiling was performed at 55degCwith vemurafenib-treated K562 cells andshowed concentration-dependent thermalstabilization of FECH and BRAF The data arenormalized so that the quantity of solubletarget at the highest compoundconcentration which reflects maximumthermal stabilization is set to 1 and fixed forcurve-fitting (B) ITDR performed at 55degC with K562 cells treated with vemurafenib alectinib or crizotinib over a range of concentrations Alectinib displays amore potent effect on FECH as compared with that by vemurafenib whereas crizotinib a drug not known to cause photosensitivity has no effect

RESEARCH | RESEARCH ARTICLE

posttranslational modification as a biomarkerwhere it is usually difficult to assess the stoi-chiometry of the modificationIn general terms the thermal stability of a

protein may be defined by its mutational statusposttranslationalmodifications and bound ligandssuch as other proteins cofactors metabolites ordrugs Fusionproteins splice variants andposttrans-lational modifications are typically undersampledand therefore not comprehensively detected inMS-basedproteomicsHence the thermal proteomeprofile of a human cell can provide a generalview of the proteomic state or proteotypeThe future scope and applicability of thermal

proteome profilingwill substantially benefit fromcontinued advances in the sensitivity and accu-racyofMS-basedproteomics Futuredevelopmentsshould include the development of protocols formembrane proteins and the application to tis-sues from animal studies or clinical biopsies

Materials and methods

Reagents and cell culture

Reagents andmedia were purchased from Sigma-Aldrich (St Louis MO) unless otherwise notedPBS was prepared by using 137 mM NaCl 27mM KCl 10 mM Na2HPO4 and 2 mM KH2PO4

to buffer at pH 74 and supplemented withcomplete EDTA-free protease inhibitor cocktail(one tablet per 25 ml) (Roche Diagnostics Basel

Switzerland) Staurosporine was from Calbiochem(Millipore Billerica MA) vemurafenib andcrizotinib from Selleck (Houston TX) alectinibfromMedchemExpress (Monmouth Junction NJ)and dasatinib from Toronto Research (TorontoCanada) The synthesis of GSK3182571 is de-scribed in the supplementary methods DNAfragments PG1 (5prime-AGC TTA GAC ATG CCT AGACAT GCC TA -3prime) and PG2 (5prime-AGC TTA GGC ATGTCT AGG CAT GTC TA -3prime) were from Life Tech-nologies (Carlsbad CA) and dissolved in 100 mMTris (pH75) 1 mM NaCl and 10mM EDTA K562Jurkat E61 and A549 cells were from ATCC K562and A549 cells were cultured in RPMI mediumcontaining 10 fetal calf serum (FCS) and JurkatE61 cells in RPMI1640 supplemented with45 gl glucose 10 mM HEPES 1 mM sodiumpyruvate and 10 FCS HT-29 (ATTC no HTB-38)were maintained in Dulbeccorsquos Modified EagleMedium Culture media were supplemented with03 gL L-glutamine and 10 fetal bovine serum(FBS) (Gibco Carlsbad CA) 100 unitsmL peni-cillin and 100 unitsmL streptomycin The cellswere expanded to amaximumof 2times 106 cellsml incase of K562 and 107 cellsml in case of Jurkat E61

Preparation of cell extract forCETSA with Western blot read-out

A549 andHT-29 cells were harvested andwashedwith PBS The cells were diluted in KB buffer

(25mMTris-HCl supplemented with 5mMbeta-glycerophosphate 2 mM DTT 01 mM Na3VO410 mM MgCl2 and complete protease inhibitorcocktail) The cell suspensionswere freeze-thawedthree times by using liquid nitrogen The solublefraction was separated from the cell debris bycentrifugation at 20000 g for 20 min at 4degC FortheCETSAmelting curve experiments cell extractswere diluted with KB buffer and divided into twoaliquots with one aliquot being treated withligandDNA fragments depending on the targetand the other aliquot treated with the solvent ofthe ligand (control) After 10 min incubation atroom temperature the respective cell extractswere heated individually at different temper-atures for 3 min in a thermal cycler [AppliedBiosystems (Foster City CA)Life Technologies]followed by cooling for 3 min at room temper-ature The heated cell extracts were centrifugedat 20000 g for 20min at 4degC in order to separatethe soluble fractions from precipitates The su-pernatantswere transferred to new02mLmicro-tubes and analyzed by means of Western blotanalysis NuPage Bis-Tris 4-12 polyacrylamidegels with MES SDS running buffer (Life Tech-nologies) were used for separation of proteins inthe samples Proteins were transferred to nitro-cellulose membranes by using the iBlot system(Life Technologies) Primary antibodies anti-PKARegulatory I-a (sc-136231) anti-PKA Catalytic-a

1255784-6 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 6 Treatment of K562 cells with dasatinib induces Tm shifts fordownstream effector proteins in the BCR-ABL pathway (A) Cell-wideassessment of Tm shifts induced by dasatinib (05 and 5 mM) in K562 cellsThe minimum Tm shift for each protein was calculated from a pair of fullreplicate experiments Four proteins showing significant melting pointdifferences both in the 05 and 5 mM dasatinib data sets are marked in red(B) Effect of dasatinib on the melting curves for the effector protein CRKL

determined in two biological replicates in intact cells (top) and cell extracts(bottom) (C) The melting curve for CRKL in Jurkat cells closely matchesthe curve obtained in dasatinib-treated K562 cells (top)The ABL1 protein inJurkat cells is much more thermostable compared with the BCR-ABL fusionprotein in K562 cells (D) ITDR curves for CRKL in dasatinib-treated intactcells Experiments were performed at 47degC (green curve) 50degC (purplecurve) and 56degC (blue curve)

RESEARCH | RESEARCH ARTICLE

(sc-28315) anti-p53WT and R273H (sc-126) sec-ondary goat anti-mouse HRP-IgG (sc-2055) andbovine anti-goat HRP-IgG (sc-2352) antibodies(Santa Cruz Biotechnology Dallas TX) were usedfor immunoblotting Chemiluminescence inten-sities were detected and quantified by using aChemiDocXRS+ imaging systemwith Image Labsoftware (Bio-Rad Hercules CA) Data were ex-pressed as means T SEM The data (band inten-sities) from multiple runs (n ge 3) were plottedwith Graphpad Prism software

Preparation of cell extract for thermalproteome profiling

Suspension cultures of K562 cells (15 to 2 times 106

cellsml) or of Jurkat E61 cells (4 times 106 cellsml)were centrifuged at 340 g for 2 min at 4degC Thecells were resuspended in 50 ml PBS After a sec-ond centrifugation step as above the cells wereresuspended in 10 ml ice-cold PBS and centri-fuged again at 340 g for 2 min at 4degC The cellswere resuspended in 15 ml of ice-cold PBS andsnap-frozen in liquidnitrogen The tubewasplacedinto a water bath at 23degC until ~60 of the con-tent was thawed and then transferred on ice un-til the entire contentwas thawed This freezethawcycle was repeated twice before the samplesweresubjected to ultracentrifugation (20 min at 4degCand 100000 g) The protein concentration of thesupernatant was determined by Bradford assay(Bio-Rad) and aliquots were snap-frozen in liquidnitrogen for use in subsequent thermal shiftassays

Thermal profiling using cell extract

A solution of the compound in dimethyl sulfoxide(DMSO) or DMSO alone as vehicle (Table 1) wasadded to the cell extract to 1 final DMSO con-centration The extract was then incubated for10 min at 23degC divided into 10 aliquots of 100 mland transferred into 02-ml polymerase chain re-action (PCR) tubes One each of the compoundand of the vehicle containing samples was heatedin parallel for 3 min to the respective tempera-ture followed by a 3-min incubation time at roomtemperature Subsequently the extract was cen-trifuged at 100000 g for 20 min at 4degC The su-pernatant was fractionated by means of SDS gel

electrophoresis and subjected to sample prepa-ration for MS analysis For ITDR experimentsGSK3182571 was tested as a nine-point serial dilu-tion starting at 100 mM (resulting in concentra-tions of 100 333 111 370 1235 0412 0137 0046and 0015 mM) including one vehicle control inan isothermal dose-response experiment at 53degCfollowing the procedure outlined above for ex-periments in cell extract

Thermal profiling using intact cells

To 24 ml of a suspension of K562 cells (at a den-sity of 15 106 cellsml) or Jurkat E61cells (at adensity of 4 106 cellsml) in a T-flask a volumeof either 120 ml of a solution of the referencecompound dissolved in DMSO or an equivalentamount of DMSO (vehicle) was added Incubationof cellswith compound and vehiclewas conductedin parallel for the duration of 1 hour at 37degC and5CO2 Cells were pelleted at 340 g and 4degC for2min resuspended in 20ml of ice cold PBS andcentrifuged again as indicated above This wash-ing step was repeated once and the cells werecarefully resuspended in 1200 ml PBS 100 ml ofthis resulting cell suspension was transferredinto 02-ml PCR tubes and subjected to a centrifu-gation step at 325 g for 2 min at 4degC Using apipette 80 ml of the PBS supernatant was re-moved By gently tapping the tubes the cells wereresuspended and one each of the compound andof the vehicle containing tubes was heated inparallel in a PCR machine for 3 min to the re-spective temperature (37degC to 67degC) followed bya 3-min incubation time at room temperatureThereafter the cells were snap-frozen in liquidnitrogen for 1 min The cells were thawed brieflyin a water bath at 25degC transferred on ice andresuspended by using a pipette This freeze-thawcycle was repeated once After an addition of fur-ther 30 ml of PBS and resuspension the entirecontent was centrifuged at 100000 g for 20 minat 4degC After centrifugation 30 ml of the super-natant were transferred into a new tube Proteinconcentration of the 37degC and 41degC samples weredetermined and used to normalize loading ofSDS gels Proteins in the supernatant were dena-tured by using SDS sample buffer partially sepa-rated by means of SDS gel electrophoresis and

subjected to sample preparation for MS analysisFor ITDR experiments vemurafenib alectiniband crizotinib were used as a nine-point serialdilution starting at 30 mM (concentrations were30 12 48 192 077 031 012 005 and 002 mM)including one vehicle control at 55degC followingthe procedure outlined aboveDasatinibwas testedas an eight-point serial dilution starting at 1 mM(concentrations usedwere 1 033 0111 0037 00120004 00013 and 000046 mM) including onedasatinib control at 10 mM and one vehicle con-trol on intact cells at 47degC 50degC and 56degC follow-ing the procedure outlined above for experimentson intact cells (Table 1)

Kinobeads assays

Competition binding assays were performed asdescribed previously by using a modified beadmatrix (5 22) Briefly 1 ml (5 mg protein) cellextract was pre-incubatedwith test compound orvehicle for 45 min at 4degC followed by incubationwith kinobeads (35 ml beads per sample) for1 hour at 4degC The beads were washed with lysisbuffer and eluted with 50 ml 2times SDS sample buf-fer GSK3182571was tested at 10 25 0625 015625003906 000977 and 000244 mM

Sample preparation for MS

Gel lanes were cut into three slices covering theentire separation range (~2 cm) and subjectedto in-gel digestion (5) Peptide samples were la-beled with 10-plex TMT (TMT10 Thermo FisherScientific Waltham MA) reagents enabling rel-ative quantification of a broad range of 10 tem-perature points in a single experiment (15 22)For kinobeads and ITDR experiments either10 or 8 TMT labels were used depending on thecompound concentration range chosen Thelabeling reaction was performed in 40 mMtriethylammoniumbicarbonate pH 853 at 22degCand quenched with glycine Labeled peptide ex-tracts were combined to a single sample per ex-periment and as indicated in the supplementarymaterials for most samples additional fraction-ation was performed by using reversed-phasechromatography at pH 12 [1mmXbridge column(Waters Milford MA)] as previously described(table S13) (35)

LC-MSMS analysis

Samples were dried in vacuo and resuspended in005 trifluoroacetic acid in water Of the sam-ple 50 was injected into an Ultimate3000nanoRLSC (Dionex Sunnyvale CA) coupled to aQ Exactive (Thermo Fisher Scientific) Peptideswere separated on custom 50 cm times 100 mM (ID)reversed-phase columns (Reprosil) at 40degC Gra-dient elutionwas performed from 2 acetonitrileto 40acetonitrile in 01 formic acid over 2 hoursSamples were online injected into Q-Exactive massspectrometers operating with a data-dependenttop 10methodMS spectrawere acquired by using70000 resolution and an ion target of 3E6Higherenergy collisional dissociation (HCD) scans wereperformed with 35 NCE at 35000 resolu-tion (at mz 200) and the ion target settingswas set to 2E5 so as to avoid coalescence (15) The

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-7

Table 1 Compoundsused in thermal profilingDashes indicate that compoundwas not tested in intact cells

CompoundConcentration used in

cellular extractexperiments

Concentration(s) inintact cell experiments

Cell line

MgATP 2 mM mdash K562staurosporine 20 mM mdash K562cAMP 1 mM mdash K562GSK3182571 20 mM mdash K562dasatinib 5 mM 05 mM 5 mM K562p53 WT

PG1 100 mM mdash A549PG2 100 mM mdash A549

p53 R273HPG1 100 mM mdash HT-29PG2 100 mM mdash HT-29

RESEARCH | RESEARCH ARTICLE

instruments were operated with Tune 22 or 23and Xcalibur 27 or 3063

Peptide and protein identification

Mascot 24 (Matrix Science Boston MA) wasused for protein identification by using a 10 partspermillionmass tolerance for peptide precursorsand 20 mD (HCD) mass tolerance for fragmentions Carbamidomethylation of cysteine residuesand TMTmodification of lysine residueswere setas fixedmodifications andmethionine oxidationand N-terminal acetylation of proteins and TMTmodification of peptide N-termini were set asvariable modifications The search database con-sisted of a customized version of the Interna-tional Protein Index protein sequence databasecombined with a decoy version of this databasecreated by using a script supplied by Matrix Sci-ence Unless stated otherwise we accepted proteinidentifications as follows (i) For single-spectrumto sequence assignments we required this assign-ment to be the bestmatch and aminimumMascotscore of 31 and a 10times difference of this assign-ment over the next best assignment Based onthese criteria the decoy search results indicatedlt1 false discovery rate (FDR) (ii) For multiplespectrum to sequence assignments and using thesame parameters the decoy search results indi-cate lt01 FDR All identified proteinswere quan-tified FDR for quantified proteins was lt1

Peptide and protein quantification

Reporter ion intensities were read from raw dataandmultiplied with ion accumulation times (theunit is milliseconds) so as to yield ameasure propor-tional to thenumberof ions thismeasure is referredto as ion area (36) Spectra matching to peptideswere filtered according to the following criteriamascot ion score gt15 signal-to-background of theprecursor ion gt4 and signal-to-interference gt05(37) Fold-changes were corrected for isotope pu-rity as described and adjusted for interferencecaused by co-eluting nearly isobaric peaks as esti-mated by the signal-to-interferencemeasure (38) Pro-tein quantification was derived from individualspectra matching to distinct peptides by using asum-based bootstrap algorithm 95 confidenceintervals were calculated for all protein fold-changesthatwere quantifiedwithmore than three spectra (36)

Curve fitting normalization andestimation of slope and melting pointdifferences for CETSA experiments

Melting curves

For all CETSA experiments fold-changes werecalculated by using the lowest temperature con-dition as the reference Relative fold changes as afunction of temperature followeda sigmoidal trendwhich was fitted with the following equationderived from chemical denaturation theory (39)(enwikipediaorgwikiEquilibrium_unfolding)using R (wwwr-projectorg)

f ethT THORN frac14 1 minus plateau

1thorn eminusaTminusbeth THORN thorn plateau

Where T is the temperature and a b and plateauare constants The value of f(T) at the lowest

temperature Tmin was fixed to one The meltingpoint of a protein is defined as the temperatureTm at which half of the protein amount has beendenatured

f(Tm) = 05

Additionally the slope of the melting curve isdefined as the value of the first derivative of f(T)at the point T = Tinfl

slope = f prime(Tinfl)

where Tinfl is the inflection point of the curvemeaning that the second derivative of f (T) equalszero at T = Tinfl

f primeprime(Tinfl) = 0

Normalization

The vehicle and compound-treated experimentswere normalized in the following waySelection of proteins for normalization (i) The

set of proteins identified in both experimentstermed ldquojointPrdquo was determined (ii) For eachexperiment (vehicle or compound) all proteinsfrom jointP fulfilling the following criteria werecollected the fold-changes at the seventh highesttemperature point comparedwith the lowest tem-perature point were between 04 and 06 andat the 9th and 10th highest temperature pointscompared th the lowest temperature pointwere below 03 and 02 respectively (iii) Thebigger of these two protein subsets was thenused for normalization (typically more than200 proteins qualified) This protein set is re-ferred to as ldquonormPrdquoCalculation of correction factors (i) The me-

dian fold-change values over all proteins in normPwere calculated for each of the 10 fold-changesThis was done separately for both experimentsand for each experiment a melting curve wasfitted The curve that was fitted with the best R2

was used to normalize both experiments (ii) Foreach experiment the correction factors werecalculated by using the median fold-changes ofthe proteins in normP For each median fold-change a correction factor by which that fold-change would have to be multiplied with in orderto coincidewith the best-fitting curvewas calculatedNormalization The two sets of correction fac-

tors were applied to all the protein fold-changesin the respective experimentsThis heuristic normalization procedure covers

three important aspects (i) The same subset ofproteins is used to determine normalization co-efficients in both experiments thus circumvent-ing any potential bias caused by proteins onlyidentified in one experiment (ii) The subselectionof late-melting proteins (seventh point between04 and 06) ensures that the normalization ofhigh-temperature points will be more accurate(iii) Last the narrow requirements on the 7th9th and 10th points will ensure that a narrowrange of melting profiles is used for calculatingthe points from which the normalization curvewill be derived This is important because if vastlydifferent melting profiles are combined into asingle one the above equation will not necessar-ily fit well A normalization curve example as well

as melting curves of several proteins before andafter normalization are illustrated in fig S12

Estimation of slope and meltingpoint differences

After normalization melting curves were fittedfor all proteins identified in the vehicle- andcompound-treated experiments Melting pointdifferences and slope differences were calculatedfor proteins whose curves passed the followingtwo requirements (i) both fitted curves for thevehicle and compound treated condition had anR2 of greater than 08 and (ii) the vehicle curvehad a plateau of less than 03Typically close to 80 of proteins that were

identified in both the vehicle- and compound-treated experiment passed these criteria Theslope of the curves had the biggest impact onthe reproducibility of melting point differencesof proteins This observation can be rationalizedin the following way If the curve has a shallowslope then small changes with respect to the ref-erence pointmdashin our case the lowest temperaturemdashwill have a strong impact on where the curvecrosses the 05 fold-change level (melting point)Consequently small variations in the quantitativemeasurement of theMS signals corresponding tothe lowest temperature will lead to strong shiftsin the inferred melting point of the curve Thismeans that it is difficult to get high precisionmeasurements ofmelting point differenceswhencomparing two curves with shallow slopes If thecurve however has a steep slope then the melt-ing point measurement will not be substantiallyaffected by small variations of the quantitativemeasurements of the lowest temperature nor ofany other temperature points Taking this intoaccount we used the following strategy for sig-nificance estimation of protein melting point dif-ferences between vehicle- and compound-treatedexperiments which is conceptually very similarto a broadly used strategy for inferring significantprotein fold-changes (40) The calculated meltingpoint differences of proteins were ordered in des-cending slope order (those with the shallowestslope come first) in which the slope for eachprotein is the lowest (steepest) slope calculatedfrom the two curve fits performed for the proteinin the vehicle- and compound-treated conditionThe proteins were divided into bins The bins areconstructed starting fromproteinswith the high-est (shallowest) slope Each bin consists of at least300 proteins Once each bin has been completedthe remaining number of proteins is counted ifthis number is below 300 the remaining pro-teins are added to the last completed bin Thisdata qualityndashdependent binning strategy is analo-gous to the procedure described in Cox et al withthe only difference being that proteins are sortedby slope instead of abundance (40) The statisticalsignificance of differences in protein meltingpoints was calculated by estimating the left- andright-sided robust standard deviation of the dis-tribution of the measurements using the 158750 and 8413 percentiles and calculating the Pvalues for all measurements for a specific binexactly as previously described (40) Subsequently

1255784-8 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

RESEARCH | RESEARCH ARTICLE

posttranslational modification as a biomarkerwhere it is usually difficult to assess the stoi-chiometry of the modificationIn general terms the thermal stability of a

protein may be defined by its mutational statusposttranslationalmodifications and bound ligandssuch as other proteins cofactors metabolites ordrugs Fusionproteins splice variants andposttrans-lational modifications are typically undersampledand therefore not comprehensively detected inMS-basedproteomicsHence the thermal proteomeprofile of a human cell can provide a generalview of the proteomic state or proteotypeThe future scope and applicability of thermal

proteome profilingwill substantially benefit fromcontinued advances in the sensitivity and accu-racyofMS-basedproteomics Futuredevelopmentsshould include the development of protocols formembrane proteins and the application to tis-sues from animal studies or clinical biopsies

Materials and methods

Reagents and cell culture

Reagents andmedia were purchased from Sigma-Aldrich (St Louis MO) unless otherwise notedPBS was prepared by using 137 mM NaCl 27mM KCl 10 mM Na2HPO4 and 2 mM KH2PO4

to buffer at pH 74 and supplemented withcomplete EDTA-free protease inhibitor cocktail(one tablet per 25 ml) (Roche Diagnostics Basel

Switzerland) Staurosporine was from Calbiochem(Millipore Billerica MA) vemurafenib andcrizotinib from Selleck (Houston TX) alectinibfromMedchemExpress (Monmouth Junction NJ)and dasatinib from Toronto Research (TorontoCanada) The synthesis of GSK3182571 is de-scribed in the supplementary methods DNAfragments PG1 (5prime-AGC TTA GAC ATG CCT AGACAT GCC TA -3prime) and PG2 (5prime-AGC TTA GGC ATGTCT AGG CAT GTC TA -3prime) were from Life Tech-nologies (Carlsbad CA) and dissolved in 100 mMTris (pH75) 1 mM NaCl and 10mM EDTA K562Jurkat E61 and A549 cells were from ATCC K562and A549 cells were cultured in RPMI mediumcontaining 10 fetal calf serum (FCS) and JurkatE61 cells in RPMI1640 supplemented with45 gl glucose 10 mM HEPES 1 mM sodiumpyruvate and 10 FCS HT-29 (ATTC no HTB-38)were maintained in Dulbeccorsquos Modified EagleMedium Culture media were supplemented with03 gL L-glutamine and 10 fetal bovine serum(FBS) (Gibco Carlsbad CA) 100 unitsmL peni-cillin and 100 unitsmL streptomycin The cellswere expanded to amaximumof 2times 106 cellsml incase of K562 and 107 cellsml in case of Jurkat E61

Preparation of cell extract forCETSA with Western blot read-out

A549 andHT-29 cells were harvested andwashedwith PBS The cells were diluted in KB buffer

(25mMTris-HCl supplemented with 5mMbeta-glycerophosphate 2 mM DTT 01 mM Na3VO410 mM MgCl2 and complete protease inhibitorcocktail) The cell suspensionswere freeze-thawedthree times by using liquid nitrogen The solublefraction was separated from the cell debris bycentrifugation at 20000 g for 20 min at 4degC FortheCETSAmelting curve experiments cell extractswere diluted with KB buffer and divided into twoaliquots with one aliquot being treated withligandDNA fragments depending on the targetand the other aliquot treated with the solvent ofthe ligand (control) After 10 min incubation atroom temperature the respective cell extractswere heated individually at different temper-atures for 3 min in a thermal cycler [AppliedBiosystems (Foster City CA)Life Technologies]followed by cooling for 3 min at room temper-ature The heated cell extracts were centrifugedat 20000 g for 20min at 4degC in order to separatethe soluble fractions from precipitates The su-pernatantswere transferred to new02mLmicro-tubes and analyzed by means of Western blotanalysis NuPage Bis-Tris 4-12 polyacrylamidegels with MES SDS running buffer (Life Tech-nologies) were used for separation of proteins inthe samples Proteins were transferred to nitro-cellulose membranes by using the iBlot system(Life Technologies) Primary antibodies anti-PKARegulatory I-a (sc-136231) anti-PKA Catalytic-a

1255784-6 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

Fig 6 Treatment of K562 cells with dasatinib induces Tm shifts fordownstream effector proteins in the BCR-ABL pathway (A) Cell-wideassessment of Tm shifts induced by dasatinib (05 and 5 mM) in K562 cellsThe minimum Tm shift for each protein was calculated from a pair of fullreplicate experiments Four proteins showing significant melting pointdifferences both in the 05 and 5 mM dasatinib data sets are marked in red(B) Effect of dasatinib on the melting curves for the effector protein CRKL

determined in two biological replicates in intact cells (top) and cell extracts(bottom) (C) The melting curve for CRKL in Jurkat cells closely matchesthe curve obtained in dasatinib-treated K562 cells (top)The ABL1 protein inJurkat cells is much more thermostable compared with the BCR-ABL fusionprotein in K562 cells (D) ITDR curves for CRKL in dasatinib-treated intactcells Experiments were performed at 47degC (green curve) 50degC (purplecurve) and 56degC (blue curve)

RESEARCH | RESEARCH ARTICLE

(sc-28315) anti-p53WT and R273H (sc-126) sec-ondary goat anti-mouse HRP-IgG (sc-2055) andbovine anti-goat HRP-IgG (sc-2352) antibodies(Santa Cruz Biotechnology Dallas TX) were usedfor immunoblotting Chemiluminescence inten-sities were detected and quantified by using aChemiDocXRS+ imaging systemwith Image Labsoftware (Bio-Rad Hercules CA) Data were ex-pressed as means T SEM The data (band inten-sities) from multiple runs (n ge 3) were plottedwith Graphpad Prism software

Preparation of cell extract for thermalproteome profiling

Suspension cultures of K562 cells (15 to 2 times 106

cellsml) or of Jurkat E61 cells (4 times 106 cellsml)were centrifuged at 340 g for 2 min at 4degC Thecells were resuspended in 50 ml PBS After a sec-ond centrifugation step as above the cells wereresuspended in 10 ml ice-cold PBS and centri-fuged again at 340 g for 2 min at 4degC The cellswere resuspended in 15 ml of ice-cold PBS andsnap-frozen in liquidnitrogen The tubewasplacedinto a water bath at 23degC until ~60 of the con-tent was thawed and then transferred on ice un-til the entire contentwas thawed This freezethawcycle was repeated twice before the samplesweresubjected to ultracentrifugation (20 min at 4degCand 100000 g) The protein concentration of thesupernatant was determined by Bradford assay(Bio-Rad) and aliquots were snap-frozen in liquidnitrogen for use in subsequent thermal shiftassays

Thermal profiling using cell extract

A solution of the compound in dimethyl sulfoxide(DMSO) or DMSO alone as vehicle (Table 1) wasadded to the cell extract to 1 final DMSO con-centration The extract was then incubated for10 min at 23degC divided into 10 aliquots of 100 mland transferred into 02-ml polymerase chain re-action (PCR) tubes One each of the compoundand of the vehicle containing samples was heatedin parallel for 3 min to the respective tempera-ture followed by a 3-min incubation time at roomtemperature Subsequently the extract was cen-trifuged at 100000 g for 20 min at 4degC The su-pernatant was fractionated by means of SDS gel

electrophoresis and subjected to sample prepa-ration for MS analysis For ITDR experimentsGSK3182571 was tested as a nine-point serial dilu-tion starting at 100 mM (resulting in concentra-tions of 100 333 111 370 1235 0412 0137 0046and 0015 mM) including one vehicle control inan isothermal dose-response experiment at 53degCfollowing the procedure outlined above for ex-periments in cell extract

Thermal profiling using intact cells

To 24 ml of a suspension of K562 cells (at a den-sity of 15 106 cellsml) or Jurkat E61cells (at adensity of 4 106 cellsml) in a T-flask a volumeof either 120 ml of a solution of the referencecompound dissolved in DMSO or an equivalentamount of DMSO (vehicle) was added Incubationof cellswith compound and vehiclewas conductedin parallel for the duration of 1 hour at 37degC and5CO2 Cells were pelleted at 340 g and 4degC for2min resuspended in 20ml of ice cold PBS andcentrifuged again as indicated above This wash-ing step was repeated once and the cells werecarefully resuspended in 1200 ml PBS 100 ml ofthis resulting cell suspension was transferredinto 02-ml PCR tubes and subjected to a centrifu-gation step at 325 g for 2 min at 4degC Using apipette 80 ml of the PBS supernatant was re-moved By gently tapping the tubes the cells wereresuspended and one each of the compound andof the vehicle containing tubes was heated inparallel in a PCR machine for 3 min to the re-spective temperature (37degC to 67degC) followed bya 3-min incubation time at room temperatureThereafter the cells were snap-frozen in liquidnitrogen for 1 min The cells were thawed brieflyin a water bath at 25degC transferred on ice andresuspended by using a pipette This freeze-thawcycle was repeated once After an addition of fur-ther 30 ml of PBS and resuspension the entirecontent was centrifuged at 100000 g for 20 minat 4degC After centrifugation 30 ml of the super-natant were transferred into a new tube Proteinconcentration of the 37degC and 41degC samples weredetermined and used to normalize loading ofSDS gels Proteins in the supernatant were dena-tured by using SDS sample buffer partially sepa-rated by means of SDS gel electrophoresis and

subjected to sample preparation for MS analysisFor ITDR experiments vemurafenib alectiniband crizotinib were used as a nine-point serialdilution starting at 30 mM (concentrations were30 12 48 192 077 031 012 005 and 002 mM)including one vehicle control at 55degC followingthe procedure outlined aboveDasatinibwas testedas an eight-point serial dilution starting at 1 mM(concentrations usedwere 1 033 0111 0037 00120004 00013 and 000046 mM) including onedasatinib control at 10 mM and one vehicle con-trol on intact cells at 47degC 50degC and 56degC follow-ing the procedure outlined above for experimentson intact cells (Table 1)

Kinobeads assays

Competition binding assays were performed asdescribed previously by using a modified beadmatrix (5 22) Briefly 1 ml (5 mg protein) cellextract was pre-incubatedwith test compound orvehicle for 45 min at 4degC followed by incubationwith kinobeads (35 ml beads per sample) for1 hour at 4degC The beads were washed with lysisbuffer and eluted with 50 ml 2times SDS sample buf-fer GSK3182571was tested at 10 25 0625 015625003906 000977 and 000244 mM

Sample preparation for MS

Gel lanes were cut into three slices covering theentire separation range (~2 cm) and subjectedto in-gel digestion (5) Peptide samples were la-beled with 10-plex TMT (TMT10 Thermo FisherScientific Waltham MA) reagents enabling rel-ative quantification of a broad range of 10 tem-perature points in a single experiment (15 22)For kinobeads and ITDR experiments either10 or 8 TMT labels were used depending on thecompound concentration range chosen Thelabeling reaction was performed in 40 mMtriethylammoniumbicarbonate pH 853 at 22degCand quenched with glycine Labeled peptide ex-tracts were combined to a single sample per ex-periment and as indicated in the supplementarymaterials for most samples additional fraction-ation was performed by using reversed-phasechromatography at pH 12 [1mmXbridge column(Waters Milford MA)] as previously described(table S13) (35)

LC-MSMS analysis

Samples were dried in vacuo and resuspended in005 trifluoroacetic acid in water Of the sam-ple 50 was injected into an Ultimate3000nanoRLSC (Dionex Sunnyvale CA) coupled to aQ Exactive (Thermo Fisher Scientific) Peptideswere separated on custom 50 cm times 100 mM (ID)reversed-phase columns (Reprosil) at 40degC Gra-dient elutionwas performed from 2 acetonitrileto 40acetonitrile in 01 formic acid over 2 hoursSamples were online injected into Q-Exactive massspectrometers operating with a data-dependenttop 10methodMS spectrawere acquired by using70000 resolution and an ion target of 3E6Higherenergy collisional dissociation (HCD) scans wereperformed with 35 NCE at 35000 resolu-tion (at mz 200) and the ion target settingswas set to 2E5 so as to avoid coalescence (15) The

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-7

Table 1 Compoundsused in thermal profilingDashes indicate that compoundwas not tested in intact cells

CompoundConcentration used in

cellular extractexperiments

Concentration(s) inintact cell experiments

Cell line

MgATP 2 mM mdash K562staurosporine 20 mM mdash K562cAMP 1 mM mdash K562GSK3182571 20 mM mdash K562dasatinib 5 mM 05 mM 5 mM K562p53 WT

PG1 100 mM mdash A549PG2 100 mM mdash A549

p53 R273HPG1 100 mM mdash HT-29PG2 100 mM mdash HT-29

RESEARCH | RESEARCH ARTICLE

instruments were operated with Tune 22 or 23and Xcalibur 27 or 3063

Peptide and protein identification

Mascot 24 (Matrix Science Boston MA) wasused for protein identification by using a 10 partspermillionmass tolerance for peptide precursorsand 20 mD (HCD) mass tolerance for fragmentions Carbamidomethylation of cysteine residuesand TMTmodification of lysine residueswere setas fixedmodifications andmethionine oxidationand N-terminal acetylation of proteins and TMTmodification of peptide N-termini were set asvariable modifications The search database con-sisted of a customized version of the Interna-tional Protein Index protein sequence databasecombined with a decoy version of this databasecreated by using a script supplied by Matrix Sci-ence Unless stated otherwise we accepted proteinidentifications as follows (i) For single-spectrumto sequence assignments we required this assign-ment to be the bestmatch and aminimumMascotscore of 31 and a 10times difference of this assign-ment over the next best assignment Based onthese criteria the decoy search results indicatedlt1 false discovery rate (FDR) (ii) For multiplespectrum to sequence assignments and using thesame parameters the decoy search results indi-cate lt01 FDR All identified proteinswere quan-tified FDR for quantified proteins was lt1

Peptide and protein quantification

Reporter ion intensities were read from raw dataandmultiplied with ion accumulation times (theunit is milliseconds) so as to yield ameasure propor-tional to thenumberof ions thismeasure is referredto as ion area (36) Spectra matching to peptideswere filtered according to the following criteriamascot ion score gt15 signal-to-background of theprecursor ion gt4 and signal-to-interference gt05(37) Fold-changes were corrected for isotope pu-rity as described and adjusted for interferencecaused by co-eluting nearly isobaric peaks as esti-mated by the signal-to-interferencemeasure (38) Pro-tein quantification was derived from individualspectra matching to distinct peptides by using asum-based bootstrap algorithm 95 confidenceintervals were calculated for all protein fold-changesthatwere quantifiedwithmore than three spectra (36)

Curve fitting normalization andestimation of slope and melting pointdifferences for CETSA experiments

Melting curves

For all CETSA experiments fold-changes werecalculated by using the lowest temperature con-dition as the reference Relative fold changes as afunction of temperature followeda sigmoidal trendwhich was fitted with the following equationderived from chemical denaturation theory (39)(enwikipediaorgwikiEquilibrium_unfolding)using R (wwwr-projectorg)

f ethT THORN frac14 1 minus plateau

1thorn eminusaTminusbeth THORN thorn plateau

Where T is the temperature and a b and plateauare constants The value of f(T) at the lowest

temperature Tmin was fixed to one The meltingpoint of a protein is defined as the temperatureTm at which half of the protein amount has beendenatured

f(Tm) = 05

Additionally the slope of the melting curve isdefined as the value of the first derivative of f(T)at the point T = Tinfl

slope = f prime(Tinfl)

where Tinfl is the inflection point of the curvemeaning that the second derivative of f (T) equalszero at T = Tinfl

f primeprime(Tinfl) = 0

Normalization

The vehicle and compound-treated experimentswere normalized in the following waySelection of proteins for normalization (i) The

set of proteins identified in both experimentstermed ldquojointPrdquo was determined (ii) For eachexperiment (vehicle or compound) all proteinsfrom jointP fulfilling the following criteria werecollected the fold-changes at the seventh highesttemperature point comparedwith the lowest tem-perature point were between 04 and 06 andat the 9th and 10th highest temperature pointscompared th the lowest temperature pointwere below 03 and 02 respectively (iii) Thebigger of these two protein subsets was thenused for normalization (typically more than200 proteins qualified) This protein set is re-ferred to as ldquonormPrdquoCalculation of correction factors (i) The me-

dian fold-change values over all proteins in normPwere calculated for each of the 10 fold-changesThis was done separately for both experimentsand for each experiment a melting curve wasfitted The curve that was fitted with the best R2

was used to normalize both experiments (ii) Foreach experiment the correction factors werecalculated by using the median fold-changes ofthe proteins in normP For each median fold-change a correction factor by which that fold-change would have to be multiplied with in orderto coincidewith the best-fitting curvewas calculatedNormalization The two sets of correction fac-

tors were applied to all the protein fold-changesin the respective experimentsThis heuristic normalization procedure covers

three important aspects (i) The same subset ofproteins is used to determine normalization co-efficients in both experiments thus circumvent-ing any potential bias caused by proteins onlyidentified in one experiment (ii) The subselectionof late-melting proteins (seventh point between04 and 06) ensures that the normalization ofhigh-temperature points will be more accurate(iii) Last the narrow requirements on the 7th9th and 10th points will ensure that a narrowrange of melting profiles is used for calculatingthe points from which the normalization curvewill be derived This is important because if vastlydifferent melting profiles are combined into asingle one the above equation will not necessar-ily fit well A normalization curve example as well

as melting curves of several proteins before andafter normalization are illustrated in fig S12

Estimation of slope and meltingpoint differences

After normalization melting curves were fittedfor all proteins identified in the vehicle- andcompound-treated experiments Melting pointdifferences and slope differences were calculatedfor proteins whose curves passed the followingtwo requirements (i) both fitted curves for thevehicle and compound treated condition had anR2 of greater than 08 and (ii) the vehicle curvehad a plateau of less than 03Typically close to 80 of proteins that were

identified in both the vehicle- and compound-treated experiment passed these criteria Theslope of the curves had the biggest impact onthe reproducibility of melting point differencesof proteins This observation can be rationalizedin the following way If the curve has a shallowslope then small changes with respect to the ref-erence pointmdashin our case the lowest temperaturemdashwill have a strong impact on where the curvecrosses the 05 fold-change level (melting point)Consequently small variations in the quantitativemeasurement of theMS signals corresponding tothe lowest temperature will lead to strong shiftsin the inferred melting point of the curve Thismeans that it is difficult to get high precisionmeasurements ofmelting point differenceswhencomparing two curves with shallow slopes If thecurve however has a steep slope then the melt-ing point measurement will not be substantiallyaffected by small variations of the quantitativemeasurements of the lowest temperature nor ofany other temperature points Taking this intoaccount we used the following strategy for sig-nificance estimation of protein melting point dif-ferences between vehicle- and compound-treatedexperiments which is conceptually very similarto a broadly used strategy for inferring significantprotein fold-changes (40) The calculated meltingpoint differences of proteins were ordered in des-cending slope order (those with the shallowestslope come first) in which the slope for eachprotein is the lowest (steepest) slope calculatedfrom the two curve fits performed for the proteinin the vehicle- and compound-treated conditionThe proteins were divided into bins The bins areconstructed starting fromproteinswith the high-est (shallowest) slope Each bin consists of at least300 proteins Once each bin has been completedthe remaining number of proteins is counted ifthis number is below 300 the remaining pro-teins are added to the last completed bin Thisdata qualityndashdependent binning strategy is analo-gous to the procedure described in Cox et al withthe only difference being that proteins are sortedby slope instead of abundance (40) The statisticalsignificance of differences in protein meltingpoints was calculated by estimating the left- andright-sided robust standard deviation of the dis-tribution of the measurements using the 158750 and 8413 percentiles and calculating the Pvalues for all measurements for a specific binexactly as previously described (40) Subsequently

1255784-8 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

RESEARCH | RESEARCH ARTICLE

(sc-28315) anti-p53WT and R273H (sc-126) sec-ondary goat anti-mouse HRP-IgG (sc-2055) andbovine anti-goat HRP-IgG (sc-2352) antibodies(Santa Cruz Biotechnology Dallas TX) were usedfor immunoblotting Chemiluminescence inten-sities were detected and quantified by using aChemiDocXRS+ imaging systemwith Image Labsoftware (Bio-Rad Hercules CA) Data were ex-pressed as means T SEM The data (band inten-sities) from multiple runs (n ge 3) were plottedwith Graphpad Prism software

Preparation of cell extract for thermalproteome profiling

Suspension cultures of K562 cells (15 to 2 times 106

cellsml) or of Jurkat E61 cells (4 times 106 cellsml)were centrifuged at 340 g for 2 min at 4degC Thecells were resuspended in 50 ml PBS After a sec-ond centrifugation step as above the cells wereresuspended in 10 ml ice-cold PBS and centri-fuged again at 340 g for 2 min at 4degC The cellswere resuspended in 15 ml of ice-cold PBS andsnap-frozen in liquidnitrogen The tubewasplacedinto a water bath at 23degC until ~60 of the con-tent was thawed and then transferred on ice un-til the entire contentwas thawed This freezethawcycle was repeated twice before the samplesweresubjected to ultracentrifugation (20 min at 4degCand 100000 g) The protein concentration of thesupernatant was determined by Bradford assay(Bio-Rad) and aliquots were snap-frozen in liquidnitrogen for use in subsequent thermal shiftassays

Thermal profiling using cell extract

A solution of the compound in dimethyl sulfoxide(DMSO) or DMSO alone as vehicle (Table 1) wasadded to the cell extract to 1 final DMSO con-centration The extract was then incubated for10 min at 23degC divided into 10 aliquots of 100 mland transferred into 02-ml polymerase chain re-action (PCR) tubes One each of the compoundand of the vehicle containing samples was heatedin parallel for 3 min to the respective tempera-ture followed by a 3-min incubation time at roomtemperature Subsequently the extract was cen-trifuged at 100000 g for 20 min at 4degC The su-pernatant was fractionated by means of SDS gel

electrophoresis and subjected to sample prepa-ration for MS analysis For ITDR experimentsGSK3182571 was tested as a nine-point serial dilu-tion starting at 100 mM (resulting in concentra-tions of 100 333 111 370 1235 0412 0137 0046and 0015 mM) including one vehicle control inan isothermal dose-response experiment at 53degCfollowing the procedure outlined above for ex-periments in cell extract

Thermal profiling using intact cells

To 24 ml of a suspension of K562 cells (at a den-sity of 15 106 cellsml) or Jurkat E61cells (at adensity of 4 106 cellsml) in a T-flask a volumeof either 120 ml of a solution of the referencecompound dissolved in DMSO or an equivalentamount of DMSO (vehicle) was added Incubationof cellswith compound and vehiclewas conductedin parallel for the duration of 1 hour at 37degC and5CO2 Cells were pelleted at 340 g and 4degC for2min resuspended in 20ml of ice cold PBS andcentrifuged again as indicated above This wash-ing step was repeated once and the cells werecarefully resuspended in 1200 ml PBS 100 ml ofthis resulting cell suspension was transferredinto 02-ml PCR tubes and subjected to a centrifu-gation step at 325 g for 2 min at 4degC Using apipette 80 ml of the PBS supernatant was re-moved By gently tapping the tubes the cells wereresuspended and one each of the compound andof the vehicle containing tubes was heated inparallel in a PCR machine for 3 min to the re-spective temperature (37degC to 67degC) followed bya 3-min incubation time at room temperatureThereafter the cells were snap-frozen in liquidnitrogen for 1 min The cells were thawed brieflyin a water bath at 25degC transferred on ice andresuspended by using a pipette This freeze-thawcycle was repeated once After an addition of fur-ther 30 ml of PBS and resuspension the entirecontent was centrifuged at 100000 g for 20 minat 4degC After centrifugation 30 ml of the super-natant were transferred into a new tube Proteinconcentration of the 37degC and 41degC samples weredetermined and used to normalize loading ofSDS gels Proteins in the supernatant were dena-tured by using SDS sample buffer partially sepa-rated by means of SDS gel electrophoresis and

subjected to sample preparation for MS analysisFor ITDR experiments vemurafenib alectiniband crizotinib were used as a nine-point serialdilution starting at 30 mM (concentrations were30 12 48 192 077 031 012 005 and 002 mM)including one vehicle control at 55degC followingthe procedure outlined aboveDasatinibwas testedas an eight-point serial dilution starting at 1 mM(concentrations usedwere 1 033 0111 0037 00120004 00013 and 000046 mM) including onedasatinib control at 10 mM and one vehicle con-trol on intact cells at 47degC 50degC and 56degC follow-ing the procedure outlined above for experimentson intact cells (Table 1)

Kinobeads assays

Competition binding assays were performed asdescribed previously by using a modified beadmatrix (5 22) Briefly 1 ml (5 mg protein) cellextract was pre-incubatedwith test compound orvehicle for 45 min at 4degC followed by incubationwith kinobeads (35 ml beads per sample) for1 hour at 4degC The beads were washed with lysisbuffer and eluted with 50 ml 2times SDS sample buf-fer GSK3182571was tested at 10 25 0625 015625003906 000977 and 000244 mM

Sample preparation for MS

Gel lanes were cut into three slices covering theentire separation range (~2 cm) and subjectedto in-gel digestion (5) Peptide samples were la-beled with 10-plex TMT (TMT10 Thermo FisherScientific Waltham MA) reagents enabling rel-ative quantification of a broad range of 10 tem-perature points in a single experiment (15 22)For kinobeads and ITDR experiments either10 or 8 TMT labels were used depending on thecompound concentration range chosen Thelabeling reaction was performed in 40 mMtriethylammoniumbicarbonate pH 853 at 22degCand quenched with glycine Labeled peptide ex-tracts were combined to a single sample per ex-periment and as indicated in the supplementarymaterials for most samples additional fraction-ation was performed by using reversed-phasechromatography at pH 12 [1mmXbridge column(Waters Milford MA)] as previously described(table S13) (35)

LC-MSMS analysis

Samples were dried in vacuo and resuspended in005 trifluoroacetic acid in water Of the sam-ple 50 was injected into an Ultimate3000nanoRLSC (Dionex Sunnyvale CA) coupled to aQ Exactive (Thermo Fisher Scientific) Peptideswere separated on custom 50 cm times 100 mM (ID)reversed-phase columns (Reprosil) at 40degC Gra-dient elutionwas performed from 2 acetonitrileto 40acetonitrile in 01 formic acid over 2 hoursSamples were online injected into Q-Exactive massspectrometers operating with a data-dependenttop 10methodMS spectrawere acquired by using70000 resolution and an ion target of 3E6Higherenergy collisional dissociation (HCD) scans wereperformed with 35 NCE at 35000 resolu-tion (at mz 200) and the ion target settingswas set to 2E5 so as to avoid coalescence (15) The

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-7

Table 1 Compoundsused in thermal profilingDashes indicate that compoundwas not tested in intact cells

CompoundConcentration used in

cellular extractexperiments

Concentration(s) inintact cell experiments

Cell line

MgATP 2 mM mdash K562staurosporine 20 mM mdash K562cAMP 1 mM mdash K562GSK3182571 20 mM mdash K562dasatinib 5 mM 05 mM 5 mM K562p53 WT

PG1 100 mM mdash A549PG2 100 mM mdash A549

p53 R273HPG1 100 mM mdash HT-29PG2 100 mM mdash HT-29

RESEARCH | RESEARCH ARTICLE

instruments were operated with Tune 22 or 23and Xcalibur 27 or 3063

Peptide and protein identification

Mascot 24 (Matrix Science Boston MA) wasused for protein identification by using a 10 partspermillionmass tolerance for peptide precursorsand 20 mD (HCD) mass tolerance for fragmentions Carbamidomethylation of cysteine residuesand TMTmodification of lysine residueswere setas fixedmodifications andmethionine oxidationand N-terminal acetylation of proteins and TMTmodification of peptide N-termini were set asvariable modifications The search database con-sisted of a customized version of the Interna-tional Protein Index protein sequence databasecombined with a decoy version of this databasecreated by using a script supplied by Matrix Sci-ence Unless stated otherwise we accepted proteinidentifications as follows (i) For single-spectrumto sequence assignments we required this assign-ment to be the bestmatch and aminimumMascotscore of 31 and a 10times difference of this assign-ment over the next best assignment Based onthese criteria the decoy search results indicatedlt1 false discovery rate (FDR) (ii) For multiplespectrum to sequence assignments and using thesame parameters the decoy search results indi-cate lt01 FDR All identified proteinswere quan-tified FDR for quantified proteins was lt1

Peptide and protein quantification

Reporter ion intensities were read from raw dataandmultiplied with ion accumulation times (theunit is milliseconds) so as to yield ameasure propor-tional to thenumberof ions thismeasure is referredto as ion area (36) Spectra matching to peptideswere filtered according to the following criteriamascot ion score gt15 signal-to-background of theprecursor ion gt4 and signal-to-interference gt05(37) Fold-changes were corrected for isotope pu-rity as described and adjusted for interferencecaused by co-eluting nearly isobaric peaks as esti-mated by the signal-to-interferencemeasure (38) Pro-tein quantification was derived from individualspectra matching to distinct peptides by using asum-based bootstrap algorithm 95 confidenceintervals were calculated for all protein fold-changesthatwere quantifiedwithmore than three spectra (36)

Curve fitting normalization andestimation of slope and melting pointdifferences for CETSA experiments

Melting curves

For all CETSA experiments fold-changes werecalculated by using the lowest temperature con-dition as the reference Relative fold changes as afunction of temperature followeda sigmoidal trendwhich was fitted with the following equationderived from chemical denaturation theory (39)(enwikipediaorgwikiEquilibrium_unfolding)using R (wwwr-projectorg)

f ethT THORN frac14 1 minus plateau

1thorn eminusaTminusbeth THORN thorn plateau

Where T is the temperature and a b and plateauare constants The value of f(T) at the lowest

temperature Tmin was fixed to one The meltingpoint of a protein is defined as the temperatureTm at which half of the protein amount has beendenatured

f(Tm) = 05

Additionally the slope of the melting curve isdefined as the value of the first derivative of f(T)at the point T = Tinfl

slope = f prime(Tinfl)

where Tinfl is the inflection point of the curvemeaning that the second derivative of f (T) equalszero at T = Tinfl

f primeprime(Tinfl) = 0

Normalization

The vehicle and compound-treated experimentswere normalized in the following waySelection of proteins for normalization (i) The

set of proteins identified in both experimentstermed ldquojointPrdquo was determined (ii) For eachexperiment (vehicle or compound) all proteinsfrom jointP fulfilling the following criteria werecollected the fold-changes at the seventh highesttemperature point comparedwith the lowest tem-perature point were between 04 and 06 andat the 9th and 10th highest temperature pointscompared th the lowest temperature pointwere below 03 and 02 respectively (iii) Thebigger of these two protein subsets was thenused for normalization (typically more than200 proteins qualified) This protein set is re-ferred to as ldquonormPrdquoCalculation of correction factors (i) The me-

dian fold-change values over all proteins in normPwere calculated for each of the 10 fold-changesThis was done separately for both experimentsand for each experiment a melting curve wasfitted The curve that was fitted with the best R2

was used to normalize both experiments (ii) Foreach experiment the correction factors werecalculated by using the median fold-changes ofthe proteins in normP For each median fold-change a correction factor by which that fold-change would have to be multiplied with in orderto coincidewith the best-fitting curvewas calculatedNormalization The two sets of correction fac-

tors were applied to all the protein fold-changesin the respective experimentsThis heuristic normalization procedure covers

three important aspects (i) The same subset ofproteins is used to determine normalization co-efficients in both experiments thus circumvent-ing any potential bias caused by proteins onlyidentified in one experiment (ii) The subselectionof late-melting proteins (seventh point between04 and 06) ensures that the normalization ofhigh-temperature points will be more accurate(iii) Last the narrow requirements on the 7th9th and 10th points will ensure that a narrowrange of melting profiles is used for calculatingthe points from which the normalization curvewill be derived This is important because if vastlydifferent melting profiles are combined into asingle one the above equation will not necessar-ily fit well A normalization curve example as well

as melting curves of several proteins before andafter normalization are illustrated in fig S12

Estimation of slope and meltingpoint differences

After normalization melting curves were fittedfor all proteins identified in the vehicle- andcompound-treated experiments Melting pointdifferences and slope differences were calculatedfor proteins whose curves passed the followingtwo requirements (i) both fitted curves for thevehicle and compound treated condition had anR2 of greater than 08 and (ii) the vehicle curvehad a plateau of less than 03Typically close to 80 of proteins that were

identified in both the vehicle- and compound-treated experiment passed these criteria Theslope of the curves had the biggest impact onthe reproducibility of melting point differencesof proteins This observation can be rationalizedin the following way If the curve has a shallowslope then small changes with respect to the ref-erence pointmdashin our case the lowest temperaturemdashwill have a strong impact on where the curvecrosses the 05 fold-change level (melting point)Consequently small variations in the quantitativemeasurement of theMS signals corresponding tothe lowest temperature will lead to strong shiftsin the inferred melting point of the curve Thismeans that it is difficult to get high precisionmeasurements ofmelting point differenceswhencomparing two curves with shallow slopes If thecurve however has a steep slope then the melt-ing point measurement will not be substantiallyaffected by small variations of the quantitativemeasurements of the lowest temperature nor ofany other temperature points Taking this intoaccount we used the following strategy for sig-nificance estimation of protein melting point dif-ferences between vehicle- and compound-treatedexperiments which is conceptually very similarto a broadly used strategy for inferring significantprotein fold-changes (40) The calculated meltingpoint differences of proteins were ordered in des-cending slope order (those with the shallowestslope come first) in which the slope for eachprotein is the lowest (steepest) slope calculatedfrom the two curve fits performed for the proteinin the vehicle- and compound-treated conditionThe proteins were divided into bins The bins areconstructed starting fromproteinswith the high-est (shallowest) slope Each bin consists of at least300 proteins Once each bin has been completedthe remaining number of proteins is counted ifthis number is below 300 the remaining pro-teins are added to the last completed bin Thisdata qualityndashdependent binning strategy is analo-gous to the procedure described in Cox et al withthe only difference being that proteins are sortedby slope instead of abundance (40) The statisticalsignificance of differences in protein meltingpoints was calculated by estimating the left- andright-sided robust standard deviation of the dis-tribution of the measurements using the 158750 and 8413 percentiles and calculating the Pvalues for all measurements for a specific binexactly as previously described (40) Subsequently

1255784-8 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

RESEARCH | RESEARCH ARTICLE

instruments were operated with Tune 22 or 23and Xcalibur 27 or 3063

Peptide and protein identification

Mascot 24 (Matrix Science Boston MA) wasused for protein identification by using a 10 partspermillionmass tolerance for peptide precursorsand 20 mD (HCD) mass tolerance for fragmentions Carbamidomethylation of cysteine residuesand TMTmodification of lysine residueswere setas fixedmodifications andmethionine oxidationand N-terminal acetylation of proteins and TMTmodification of peptide N-termini were set asvariable modifications The search database con-sisted of a customized version of the Interna-tional Protein Index protein sequence databasecombined with a decoy version of this databasecreated by using a script supplied by Matrix Sci-ence Unless stated otherwise we accepted proteinidentifications as follows (i) For single-spectrumto sequence assignments we required this assign-ment to be the bestmatch and aminimumMascotscore of 31 and a 10times difference of this assign-ment over the next best assignment Based onthese criteria the decoy search results indicatedlt1 false discovery rate (FDR) (ii) For multiplespectrum to sequence assignments and using thesame parameters the decoy search results indi-cate lt01 FDR All identified proteinswere quan-tified FDR for quantified proteins was lt1

Peptide and protein quantification

Reporter ion intensities were read from raw dataandmultiplied with ion accumulation times (theunit is milliseconds) so as to yield ameasure propor-tional to thenumberof ions thismeasure is referredto as ion area (36) Spectra matching to peptideswere filtered according to the following criteriamascot ion score gt15 signal-to-background of theprecursor ion gt4 and signal-to-interference gt05(37) Fold-changes were corrected for isotope pu-rity as described and adjusted for interferencecaused by co-eluting nearly isobaric peaks as esti-mated by the signal-to-interferencemeasure (38) Pro-tein quantification was derived from individualspectra matching to distinct peptides by using asum-based bootstrap algorithm 95 confidenceintervals were calculated for all protein fold-changesthatwere quantifiedwithmore than three spectra (36)

Curve fitting normalization andestimation of slope and melting pointdifferences for CETSA experiments

Melting curves

For all CETSA experiments fold-changes werecalculated by using the lowest temperature con-dition as the reference Relative fold changes as afunction of temperature followeda sigmoidal trendwhich was fitted with the following equationderived from chemical denaturation theory (39)(enwikipediaorgwikiEquilibrium_unfolding)using R (wwwr-projectorg)

f ethT THORN frac14 1 minus plateau

1thorn eminusaTminusbeth THORN thorn plateau

Where T is the temperature and a b and plateauare constants The value of f(T) at the lowest

temperature Tmin was fixed to one The meltingpoint of a protein is defined as the temperatureTm at which half of the protein amount has beendenatured

f(Tm) = 05

Additionally the slope of the melting curve isdefined as the value of the first derivative of f(T)at the point T = Tinfl

slope = f prime(Tinfl)

where Tinfl is the inflection point of the curvemeaning that the second derivative of f (T) equalszero at T = Tinfl

f primeprime(Tinfl) = 0

Normalization

The vehicle and compound-treated experimentswere normalized in the following waySelection of proteins for normalization (i) The

set of proteins identified in both experimentstermed ldquojointPrdquo was determined (ii) For eachexperiment (vehicle or compound) all proteinsfrom jointP fulfilling the following criteria werecollected the fold-changes at the seventh highesttemperature point comparedwith the lowest tem-perature point were between 04 and 06 andat the 9th and 10th highest temperature pointscompared th the lowest temperature pointwere below 03 and 02 respectively (iii) Thebigger of these two protein subsets was thenused for normalization (typically more than200 proteins qualified) This protein set is re-ferred to as ldquonormPrdquoCalculation of correction factors (i) The me-

dian fold-change values over all proteins in normPwere calculated for each of the 10 fold-changesThis was done separately for both experimentsand for each experiment a melting curve wasfitted The curve that was fitted with the best R2

was used to normalize both experiments (ii) Foreach experiment the correction factors werecalculated by using the median fold-changes ofthe proteins in normP For each median fold-change a correction factor by which that fold-change would have to be multiplied with in orderto coincidewith the best-fitting curvewas calculatedNormalization The two sets of correction fac-

tors were applied to all the protein fold-changesin the respective experimentsThis heuristic normalization procedure covers

three important aspects (i) The same subset ofproteins is used to determine normalization co-efficients in both experiments thus circumvent-ing any potential bias caused by proteins onlyidentified in one experiment (ii) The subselectionof late-melting proteins (seventh point between04 and 06) ensures that the normalization ofhigh-temperature points will be more accurate(iii) Last the narrow requirements on the 7th9th and 10th points will ensure that a narrowrange of melting profiles is used for calculatingthe points from which the normalization curvewill be derived This is important because if vastlydifferent melting profiles are combined into asingle one the above equation will not necessar-ily fit well A normalization curve example as well

as melting curves of several proteins before andafter normalization are illustrated in fig S12

Estimation of slope and meltingpoint differences

After normalization melting curves were fittedfor all proteins identified in the vehicle- andcompound-treated experiments Melting pointdifferences and slope differences were calculatedfor proteins whose curves passed the followingtwo requirements (i) both fitted curves for thevehicle and compound treated condition had anR2 of greater than 08 and (ii) the vehicle curvehad a plateau of less than 03Typically close to 80 of proteins that were

identified in both the vehicle- and compound-treated experiment passed these criteria Theslope of the curves had the biggest impact onthe reproducibility of melting point differencesof proteins This observation can be rationalizedin the following way If the curve has a shallowslope then small changes with respect to the ref-erence pointmdashin our case the lowest temperaturemdashwill have a strong impact on where the curvecrosses the 05 fold-change level (melting point)Consequently small variations in the quantitativemeasurement of theMS signals corresponding tothe lowest temperature will lead to strong shiftsin the inferred melting point of the curve Thismeans that it is difficult to get high precisionmeasurements ofmelting point differenceswhencomparing two curves with shallow slopes If thecurve however has a steep slope then the melt-ing point measurement will not be substantiallyaffected by small variations of the quantitativemeasurements of the lowest temperature nor ofany other temperature points Taking this intoaccount we used the following strategy for sig-nificance estimation of protein melting point dif-ferences between vehicle- and compound-treatedexperiments which is conceptually very similarto a broadly used strategy for inferring significantprotein fold-changes (40) The calculated meltingpoint differences of proteins were ordered in des-cending slope order (those with the shallowestslope come first) in which the slope for eachprotein is the lowest (steepest) slope calculatedfrom the two curve fits performed for the proteinin the vehicle- and compound-treated conditionThe proteins were divided into bins The bins areconstructed starting fromproteinswith the high-est (shallowest) slope Each bin consists of at least300 proteins Once each bin has been completedthe remaining number of proteins is counted ifthis number is below 300 the remaining pro-teins are added to the last completed bin Thisdata qualityndashdependent binning strategy is analo-gous to the procedure described in Cox et al withthe only difference being that proteins are sortedby slope instead of abundance (40) The statisticalsignificance of differences in protein meltingpoints was calculated by estimating the left- andright-sided robust standard deviation of the dis-tribution of the measurements using the 158750 and 8413 percentiles and calculating the Pvalues for all measurements for a specific binexactly as previously described (40) Subsequently

1255784-8 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

RESEARCH | RESEARCH ARTICLE

an adjustment for multiple hypothesis testingwas performed on the full data set by usingBenjamini-Hochberg (BH) correction (41)In order to select proteins whose thermal sta-

bility is significantly changed by compound treat-ment in two biological replicates (two pairs ofvehicle- and compound-treated experiments) thefollowing rules were used (i) The melting pointdifference between vehicle- and compound-treatedconditions for a protein had a BH-corrected Pvalue of less than 005 in one biological replicateand less than 010 in the other (ii) Both meltingpoint differences were either positive or negativein the two biological replicates (iii) The smallestabsolute melting point difference of the proteinin the two biological replicates was greater thanthe absolutemelting point difference of that sameprotein between the two vehicle experiments (iv)In each biological replicate the steepest slopeof the protein melting curve in the vehicle- andcompound-treated conditions pair was below ndash006Given the stringent requirements on the qual-

ity of the curve fits of the compared proteins weallowed proteins quantified with single peptidesto be part of the analysis

Calculation of pEC50 values fromITDR experiments

For isothermal dose-response experiments thevehicle condition was used as the reference forfold-change calculation For proteins stabilizedby the compound treatment at least a 50 in-crease in protein stabilization at the highest com-pound concentration compared with the vehiclecondition (FChighest concentration gt15) was requiredin order to attempt to calculate a pEC50 Beforefitting the sigmoidal dose response curve (fittedby using GraphPad Prism top and bottom fixedat 1 and 0 respectively variable slope) all fold-change values were transformed as follows

FCtransformed = (FC ndash1)(FChighest concentration ndash 1)

After this transformation the fold-change valueis zero at the vehicle condition and one at thecondition corresponding to the highest drug con-centration In order to ascertain that the proteinstabilizationwas due to compound treatment andnot random fluctuations we additionally requiredthat the fitted sigmoidal curve had an R2 of atleast 08 Because the transformed fold-change atthe highest compound concentration conditionwasfixed at one this concentration was assumed tobe sufficiently high to achieve the maximum ef-fect of the compound-induced protein stabilizationFor proteins destabilized by the compound

treatment at least a 50 increase in protein sta-bilization at the vehicle condition comparedwiththe condition with the highest compound con-centration was required (FChighest concentration lt06666) In this case the fold-changeswere trans-formed as follows

FCtransformed =(FC ndash FChighest concentration)(1 ndash FChighest concentration)

After this transformation the fold-change valueis 1 at the vehicle condition and 0 at the con-

dition corresponding to the highest drug con-centration The sigmoidal curve fits and R2 gt 08criterion were performed and applied in thesame way as above Conversely here it has to beassumed that the transformed fold-change at thehighest compound concentration is sufficientlyhigh in order for the compound-induced proteindestabilization to have reached the maximumeffect Given the potentially narrow fold-changerange we required proteins to be quantified withmore than three spectra in order to considerthem for pEC50 calculations

pIC50 calculations

For kinobeads experiments the vehicle condi-tion was used as the reference for fold-changecalculation Sigmoidal dose-response curves werefitted by using R (wwwr-projectorg) and the drcpackage (wwwbioassaydk) as previously de-scribed (5)

Heat map generation

Only proteins that were quantified with at leasttwo distinct peptides in the intact-cell experimentand additionally were quantified in the secondintact-cell experiment and in the two cell-extractexperiments were included (table S1) The clus-tering was performed in Spotfire by using com-plete linkage with Euclidean distance

GO analysis

The web tool Gorilla (cbl-gorillacstechnionacil)(42) was used to identify enrichedGO terms in (i)selected clusters in the melting proteome heat-map (Fig 2 and fig S2) by using all the proteinsin the whole-cell heatmap as background and (ii)the 200 proteins with the highest or lowest melt-ing temperature in the whole-cells experimentand the cell extract as well as for the 200 proteinswith the biggest positive or negative differencebetween whole cells and cell extract For the sec-ond enrichment analysis all proteins that werepresent in both the whole-cell experiment andthe cell-extract experiment and fulfilled the filter-ing criteria described above were used as back-ground Significant (FDRQ value lt001) GO termsare reported in table S2

REFERENCES AND NOTES

1 M Wilhelm et al Mass-spectrometry-based draft of the humanproteome Nature 509 582ndash587 (2014) doi 101038nature13319 pmid 24870543

2 N A Kulak G Pichler I Paron N Nagaraj M Mann Minimalencapsulated proteomic-sample processing applied to copy-number estimation in eukaryotic cells Nat Methods 11319ndash324 (2014) doi 101038nmeth2834 pmid 24487582

3 J Cox M Mann Quantitative high-resolution proteomics fordata-driven systems biology Annu Rev Biochem 80 273ndash299(2011) doi 101146annurev-biochem-061308-093216pmid 21548781

4 A Bensimon A J Heck R Aebersold Mass spectrometry-based proteomics and network biology Annu Rev Biochem81 379ndash405 (2012) doi 101146annurev-biochem-072909-100424 pmid 22439968

5 M Bantscheff et al Quantitative chemical proteomics revealsmechanisms of action of clinical ABL kinase inhibitors NatBiotechnol 25 1035ndash1044 (2007) doi 101038nbt1328pmid 17721511

6 M Bantscheff et al Chemoproteomics profiling of HDACinhibitors reveals selective targeting of HDAC complexesNat Biotechnol 29 255ndash265 (2011) doi 101038nbt1759pmid 21258344

7 A P Frei et al Direct identification of ligand-receptorinteractions on living cells and tissues Nat Biotechnol 30997ndash1001 (2012) doi 101038nbt2354 pmid 22983091

8 G E Winter et al Systems-pharmacology dissection of a drugsynergy in imatinib-resistant CML Nat Chem Biol 8 905ndash912(2012) doi 101038nchembio1085 pmid 23023260

9 J S Duncan et al Dynamic reprogramming of the kinomein response to targeted MEK inhibition in triple-negativebreast cancer Cell 149 307ndash321 (2012) doi 101016jcell201202053 pmid 22500798

10 C N Pace T McGrath Substrate stabilization of lysozymeto thermal and guanidine hydrochloride denaturationJ Biol Chem 255 3862ndash3865 (1980) pmid 7372654

11 M Vedadi et al Chemical screening methods to identifyligands that promote protein stability protein crystallizationand structure determination Proc Natl Acad Sci USA 10315835ndash15840 (2006) doi 101073pnas0605224103pmid 17035505

12 D Martinez Molina et al Monitoring drug target engagementin cells and tissues using the cellular thermal shift assayScience 341 84ndash87 (2013) doi 101126science1233606pmid 23828940

13 G M Simon M J Niphakis B F Cravatt Determining targetengagement in living systems Nat Chem Biol 9 200ndash205(2013) doi 101038nchembio1211 pmid 23508173

14 J A Fraser et al A novel p53 phosphorylation site withinthe MDM2 ubiquitination signal II A model in whichphosphorylation at SER269 induces a mutant conformation top53 J Biol Chem 285 37773ndash37786 (2010) doi 101074jbcM110143107 pmid 20847049

15 T Werner et al Ion coalescence of neutron encoded TMT10-plex reporter ions Anal Chem 86 3594ndash3601 (2014)doi 101021ac500140s pmid 24579773

16 B I Kurganov Kinetics of protein aggregation Quantitativeestimation of the chaperone-like activity in test-systems basedon suppression of protein aggregation Biochemistry (Mosc)67 409ndash422 (2002) doi 101023A1015277805345pmid 11996654

17 I Asial et al Engineering protein thermostability using ageneric activity-independent biophysical screen inside the cellNat Commun 4 2901 (2013) doi 101038ncomms3901pmid 24352381

18 S Ebbinghaus A Dhar J D McDonald M Gruebele Proteinfolding stability and dynamics imaged in a living cellNat Methods 7 319ndash323 (2010) doi 101038nmeth1435pmid 20190760

19 I Becher et al Affinity profiling of the cellular kinome for thenucleotide cofactors ATP ADP and GTP ACS Chem Biol 8599ndash607 (2013) doi 101021cb3005879 pmid 23215245

20 W S El-Deiry S E Kern J A Pietenpol K W KinzlerB Vogelstein Definition of a consensus binding site for p53Nat Genet 1 45ndash49 (1992) doi 101038ng0492-45pmid 1301998

21 L Zhang et al Characterization of the novel broad-spectrumkinase inhibitor CTx-0294885 as an affinity reagent for massspectrometry-based kinome profiling J Proteome Res 123104ndash3116 (2013) doi 101021pr3008495 pmid 23692254

22 T Werner et al High-resolution enabled TMT 8-plexingAnal Chem 84 7188ndash7194 (2012) doi 101021ac301553xpmid 22881393

23 P Zhang et al Structure and allostery of the PKA RIIbtetrameric holoenzyme Science 335 712ndash716 (2012)doi 101126science1213979 pmid 22323819

24 H A Dailey P N Meissner Erythroid heme biosynthesis andits disorders Cold Spring Harb Perspect Med 3 a011676(2013) doi 101101cshperspecta011676 pmid 23471474

25 S Wahlin P Harper Skin ferrochelatase and photosensitivityin mice and man J Invest Dermatol 130 631ndash633 (2010)doi 101038jid2009246 pmid 19657351

26 P Gelot et al Vemurafenib An unusual UVA-inducedphotosensitivity Exp Dermatol 22 297ndash298 (2013)doi 101111exd12119 pmid 23528218

27 G Bollag et al Clinical efficacy of a RAF inhibitor needs broadtarget blockade in BRAF-mutant melanoma Nature 467596ndash599 (2010) doi 101038nature09454 pmid 20823850

28 C A Perez M Velez L E Raez E S Santos Overcoming theresistance to crizotinib in patients with non-small cell lungcancer harboring EML4ALK translocation Lung Cancer 84110ndash115 (2014) doi 101016jlungcan201402001pmid 24598368

29 S Ou et al paper presented at the European CancerCongress 2013 abstract 44 (2013) available at httpeccamsterdam2013ecco-orgeuScientific-ProgrammeAbstract-searchaspxabstractid=8959

SCIENCE sciencemagorg 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 1255784-9

RESEARCH | RESEARCH ARTICLE

30 Y Ben-Neriah G Q Daley A M Mes-Masson O N WitteD Baltimore The chronic myelogenous leukemia-specific P210protein is the product of the bcrabl hybrid gene Science 233212ndash214 (1986) doi 101126science3460176 pmid 3460176

31 N P Shah et al Overriding imatinib resistance with a novelABL kinase inhibitor Science 305 399ndash401 (2004)doi 101126science1099480 pmid 15256671

32 T Oda et al Crkl is the major tyrosine-phosphorylated proteinin neutrophils from patients with chronic myelogenousleukemia J Biol Chem 269 22925ndash22928 (1994)pmid 8083188

33 A Hamilton et al BCR-ABL activity and its response to drugscan be determined in CD34+ CML stem cells by CrkLphosphorylation status using flow cytometry Leukemia 201035ndash1039 (2006) doi 101038sjleu2404189 pmid 16572205

34 Y Deguchi et al Comparison of imatinib dasatinib nilotiniband INNO-406 in imatinib-resistant cell lines Leuk Res 32980ndash983 (2008) doi 101016jleukres200711008pmid 18191450

35 U Kruse et al Chemoproteomics-based kinome profiling andtarget deconvolution of clinical multi-kinase inhibitors inprimary chronic lymphocytic leukemia cells Leukemia 2589ndash100 (2011) doi 101038leu2010233 pmid 20944678

36 M M Savitski et al Delayed fragmentation and optimizedisolation width settings for improvement of proteinidentification and accuracy of isobaric mass tag quantificationon Orbitrap-type mass spectrometers Anal Chem 838959ndash8967 (2011) doi 101021ac201760x pmid 22017476

37 M M Savitski et al Targeted data acquisition for improvedreproducibility and robustness of proteomic massspectrometry assays J Am Soc Mass Spectrom 211668ndash1679 (2010) doi 101016jjasms201001012pmid 20171116

38 M M Savitski et al Measuring and managing ratiocompression for accurate iTRAQTMT quantificationJ Proteome Res 12 3586ndash3598 (2013) doi 101021pr400098r pmid 23768245

39 J A Schellman The thermodynamics of solvent exchangeBiopolymers 34 1015ndash1026 (1994) doi 101002bip360340805 pmid 8075384

40 J Cox M Mann MaxQuant enables high peptide identificationrates individualized ppb-range mass accuracies andproteome-wide protein quantification Nat Biotechnol 261367ndash1372 (2008) doi 101038nbt1511 pmid 19029910

41 Y Benjamini Y Hochberg Controlling the false discoveryrate A practical and powerful approach to multiple testingJ R Stat Soc B 57 289ndash300 (1995)

42 E Eden R Navon I Steinfeld D Lipson Z Yakhini GOrillaA tool for discovery and visualization of enriched GO termsin ranked gene lists BMC Bioinformatics 10 48 (2009)doi 1011861471-2105-10-48 pmid 19192299

ACKNOWLEDGMENTS

We thank D Thomson for the synthesis of GSK3182571 J Stuhlfauthfor cell culture M Jundt K Kammerer M Kloumls-HudakM Paulmann I Toumlgel and T Rudi for expert technical assistance

and F Weisbrodt for help with the figures We are grateful to M Hannand G Neubauer for discussions and support PN acknowledgesfunding from Karolinska Institute (DPA) the Swedish ResearchCouncil (Vetenskapsraringdet) and the Swedish Cancer Society(Cancerfonden) MMS FBMR MB PN and GD conceivedthe project MMS FBMR TW DE DMM RJ MB and GDdesigned the experiments FBMR TW DE RBD RJ andDMM conducted and supervised experiments MMS FBMRHF TW MFS and MB analyzed proteomics data MMSFBMR SK BK MB and PN contributed to the manuscriptand GD wrote the manuscript PN and DMM are cofoundersand shareholders of Pelago Bioscience BK is a cofounder andshareholder of OmicScouts GmbH PN is the inventor of a patent(approved in the UK and pending in other districts) controlled byPelago Biosciences AB and Evitra Proteoma AB covering the basicCETSA method RH and DM also have associations to PelagoBioscience AB Mass spectrometry data are available for downloadat ProteomicsDB (wwwproteomicsdborg data set identifierPRDB004185)

SUPPLEMENTARY MATERIALS

wwwsciencemagorgcontent34662051255784supplDC1Materials and MethodsFigs S1 to S12Tables S1 to S13

8 May 2014 accepted 2 September 2014101126science1255784

1255784-10 3 OCTOBER 2014 bull VOL 346 ISSUE 6205 sciencemagorg SCIENCE

RESEARCH | RESEARCH ARTICLE

DOI 101126science1255784 (2014)346 Science

et alMikhail M SavitskiproteomeTracking cancer drugs in living cells by thermal profiling of the

This copy is for your personal non-commercial use only

clicking herecolleagues clients or customers by you can order high-quality copies for yourIf you wish to distribute this article to others

herefollowing the guidelines

can be obtained byPermission to republish or repurpose articles or portions of articles

) April 15 2015 wwwsciencemagorg (this information is current as of

The following resources related to this article are available online at

httpwwwsciencemagorgcontent34662051255784fullhtmlversion of this article at

including high-resolution figures can be found in the onlineUpdated information and services

httpwwwsciencemagorgcontentsuppl2014100134662051255784DC1html can be found at Supporting Online Material

httpwwwsciencemagorgcontent34662051255784fullhtmlrelatedfound at

can berelated to this article A list of selected additional articles on the Science Web sites

httpwwwsciencemagorgcontent34662051255784fullhtmlref-list-1 9 of which can be accessed freecites 41 articlesThis article

httpwwwsciencemagorgcontent34662051255784fullhtmlrelated-urls1 articles hosted by HighWire Press seecited by This article has been

httpwwwsciencemagorgcgicollectionbiochemBiochemistry

subject collectionsThis article appears in the following

registered trademark of AAAS is aScience2014 by the American Association for the Advancement of Science all rights reserved The title

CopyrightAmerican Association for the Advancement of Science 1200 New York Avenue NW Washington DC 20005 (print ISSN 0036-8075 online ISSN 1095-9203) is published weekly except the last week in December by theScience

on

Apr

il 15

201

5w

ww

sci

ence

mag

org

Dow

nloa

ded

from