article published online: 22 May 2017 | doi: 10.1038/nchembio.2382

A combinatorial screen of the CLOUD uncovers a synergy targeting the androgen receptor

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

Marco P Licciardello1,9, Anna Ringler1, Patrick Markt1,9, Freya Klepsch1,9, Charles-Hugues Lardeau1,2, Sara Sdelci1, Erika Schirghuber1,2, André C Müller1, Michael Caldera1, Anja Wagner3, Rebecca Herzog3, Thomas Penz1, Michael Schuster1, Bernd Boidol1,2, Gerhard Dürnberger1,9, Yasin Folkvaljon4, Pär Stattin4,5, Vladimir Ivanov6, Jacques Colinge1,9, Christoph Bock1,7,8   , Klaus Kratochwill3   , Jörg Menche1   , Keiryn L Bennett1    & Stefan Kubicek1,2*    Approved drugs are invaluable tools to study biochemical pathways, and further characterization of these compounds may lead to repurposing of single drugs or combinations. Here we describe a collection of 308 small molecules representing the diversity of structures and molecular targets of all FDA-approved chemical entities. The CeMM Library of Unique Drugs (CLOUD) covers prodrugs and active forms at pharmacologically relevant concentrations and is ideally suited for combinatorial studies. We screened pairwise combinations of CLOUD drugs for impairment of cancer cell viability and discovered a synergistic interaction between flutamide and phenprocoumon (PPC). The combination of these drugs modulates the stability of the androgen receptor (AR) and resensitizes AR-mutant prostate cancer cells to flutamide. Mechanistically, we show that the AR is a substrate for g-carboxylation, a post-translational modification inhibited by PPC. Collectively, our data suggest that PPC could be repurposed to tackle resistance to antiandrogens in prostate cancer patients.

D

rug discovery is a challenging process that often spans 10–15 years1. Despite recent technological advances, the number of new molecular entities approved each year still falls below expectations2. Researchers in both industry and academia have tried to bypass the hurdles of drug discovery by repositioning clinical compounds for new therapeutic indications, a process commonly referred to as drug repurposing3. Patient diversity and drug resistance represent substantial obstacles in long-term pharmacological treatment of diseases such as cancer and infections4,5. Multicomponent therapeutics can be effective alternatives in these circumstances6, providing a new strategy for more personalized medicine. Approved drugs, which have been certified to be safe, effective and bioavailable and have well-defined targets and mechanisms of action, are an ideal starting point for the development of new multicomponent regimes and provide a fast track to new clinical applications. These unique features have prompted efforts to catalog and collect all drugs approved for human or veterinary use7,8; however, access to these comprehensive libraries is limited. Moreover, a systematic pairwise combinatorial high-throughput screen (HTS) of all 14,814 molecules currently in the US National Institutes of Health Chemical Genomics Center (NCGC) pharmaceutical collection8 would generate more than 100,000,000 data points, an effort beyond the current capacity of screening facilities. In this study, we report a strategy to computationally derive a set of representative drugs that have been approved by the US Food

and Drug Administration (FDA). Our aim was to generate a collection that optimally captures the biochemical diversity of all clinical compounds while fitting on a single 384-well screening plate. We annotated this collection with reported human peak plasma concentrations to encourage HTS at pharmacologically relevant doses. Moreover, for all substances administered as prodrugs, we included the respective active form. We named this collection of clinical compounds the CeMM Library of Unique Drugs (CLOUD). The nonredundant nature of our library allows systematic investigation of all combinations of CLOUD compounds. In a combinatorial HTS assessing cancer cell viability, we uncovered a synergistic interaction between flutamide and PPC, approved drugs representing antiandrogens and vitamin K antagonists in the CLOUD. We found that the synergy between these two drug classes was conserved in multiple prostate cancer cell lines. Concomitant administration of flutamide and PPC impaired the viability of prostate cells carrying AR mutations associated with resistance to therapy among patients. Our findings suggest that vitamin K antagonists could be repurposed in the clinic to address resistance to antiandrogens mediated by AR mutations.

RESULTS The CLOUD is a representative library of approved drugs

To generate a representative library of clinical compounds, we retrieved all approved products from the Drugs@FDA Database (https://www.accessdata.fda.gov/scripts/cder/daf/index.cfm). This

CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences, Vienna, Austria. 2Christian Doppler Laboratory for Chemical Epigenetics and Anti infectives, CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences, Vienna, Austria. 3Christian Doppler Laboratory for Molecular Stress Research in Peritoneal Dialysis, Department of Pediatrics and Adolescent Medicine, Medical University of Vienna, Vienna, Austria. 4Department of Surgery and Perioperative Sciences, Urology and Andrology, Umeå University Hospital, Umeå, Sweden. 5Department of Surgical Sciences, Uppsala University, Uppsala, Sweden. 6Enamine Ltd., Kiev, Ukraine. 7Department of Laboratory Medicine, Medical University of Vienna, Vienna, Austria. 8Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbru¨cken, Germany. 9Present addresses: Cancer Research UK Cancer Therapeutics Unit, The Institute of Cancer Research, London, UK (M.P.L.); Pharmacy Department, Innsbruck University Hospital, Innsbruck, Austria (P.M.); Chemical Computing Group, Inc., Cologne, Germany (F.K.); Gregor Mendel Institute, Vienna, Austria (G.D.); Faculté de Médecine, Université de Montpellier, and Institut de Recherche en Cancérologie de Montpellier, INSERM U1194, Institut régional du Cancer Montpellier, Montpellier, France (J.C.). *e-mail: [email protected] 1

nature CHEMICAL BIOLOGY | Advance online publication | www.nature.com/naturechemicalbiology

1

article

Nature chemical biology doi: 10.1038/nchembio.2382

resource includes prescription and over-the-counter small molecules and therapeutic biologicals approved for human use, as well as drugs discontinued for reasons other than safety. First, we determined

a

and extracted 2,171 unique active pharmaceutical ingredients responsible for the biological effects of the 26,800 products retrieved from the database (Fig. 1a). To obtain a small molecule collection

b

STEAM

26,800 FDA-approved products

O NH2

2,171 active ingredients

N

H2N

N

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

+

O

NH

O

+

35 active forms of prodrugs

N H

O

HO

O

H N

N H

O

H N H

NH

O NH

S S

NH

HO

S

O

H N

O H

N H

O

NH

O O

O H

OH O

N

O S

N

N

N

O

H

OH

Dihydrofolate reductase inhibitors

O N H

O O

O N H

N

N

Dihydrofolate reductase inhibitors

NH

O

Histone deacetylase inhibitors

Histone deacetylase inhibitors

O

CLOUD

c

NH2

O O

H2N

Clustering 239 unique drugs

O

OH

H2N

N

920 drugs with known target

34 drugs with unknown target

OH

HN

OH

N

STEA M

M STEA

954 systemic small molecules

N H

NH2 N

O

HN

OH

N H

N

N

O

O

O

N

H2N O

H2N

OH O

N

NH2

1,415 unique small molecules

OH

N H

N

N

O

CLOUD

2

DIPS score

1

0

−1

DIPS score > median (58) DIPS score < median (11)

−2

Clusters of STEAM drugs

d

100

STEAM CLOUD

e

30 25

80

Drugs (%)

Drugs (%)

20 60

40

15 10

20

00

,0 00 >1

00

1,0

–1

10

0–

10

10

1–

1

.1

0

0. 1–

4

1

3

–0

2

01

1

Violations of Lipinski's rule of five

0.

0

<0 .0

0

5

Peak plasma concentration (µM)

Figure 1 | The CLOUD. (a) Schematic representation of the filtering and clustering procedure leading to the 308 CLOUD drugs. (b) Examples of STEAM drug clusters and representative CLOUD drugs. The cluster of dihydrofolate reductase inhibitors centers on methotrexate, which was then selected for the CLOUD (top); the two structurally very different histone deacetylase inhibitors were both kept in the CLOUD (bottom). A complete list of clusters and selected CLOUD drugs can be found in Supplementary Data Set 1. (c) Pairwise DIPS scores within 69 STEAM clusters. (d) Violations of Lipinski’s rule of five by STEAM and CLOUD drugs. (e) Peak plasma concentration ranges of all CLOUD drugs. 2

nature chemical biology | Advance online publication | www.nature.com/naturechemicalbiology

article

The CLOUD covers the biochemical space of approved drugs

To ensure that the CLOUD functionally preserved most of the biological activities addressed by STEAM drugs, we analyzed gene expression profiles reported in the Connectivity Map (CMap)9 for compounds belonging to the same class. To eliminate batch effects from CMap data, we used drug-induced gene expression profile similarity (DIPS) scores10, which provide efficient data normalization. Pairwise DIPS scores within most STEAM clusters were higher than that of random drug pairs, indicating a similar influence of co-clustered drugs on cellular responses (Fig. 1c). We then addressed the clustering procedure for potential biases in organisms, molecular targets or physicochemical properties. The 920 STEAM drugs with known molecular targets engage mostly human proteins (Supplementary Fig. 2a). CLOUD drugs showed a similar distribution, with a minor reduction in bacteria-specific compounds owing to the high number of structurally similar antibiotics acting on dihydropteroate synthase, penicillinbinding proteins (PBP) and the ribosome. A molecular target classification of STEAM drugs showed that the majority of approved small molecules engage G protein–coupled receptors (Supplementary Fig. 2b), as reported11. Again, CLOUD compounds showed a similar

a

b

6

0.5 Deviation

2 0 –2

0.0 –0.5

–4 –6

Flutamide PPC

1.0

Flutamide PPC

4 Log standard score

–1.0 –1

–0.5

c

0 0.5 Deviation

Synergy

1

PPC (µM)

Antagonism

PPC (µM)

11 .6 15 .4 20 .6 27 .4 36 .6 48 .8 65 .0 86 .7

11 .6 15 .4 20 .6 27 .4 36 .6 48 .8 65 .0 86 .7 4.9

4.9

6.5

6.5 Flutamide (µM)

suitable for HTS, we discarded all macromolecules, narrowing our set down to 1,928 small molecules. We also removed all salt fragments, which left 1,415 unique molecular entities. We then discarded all FDA-approved molecules that exert their biological effects through mechanisms other than protein–ligand interaction, are not used to treat diseases or are found only in topical products. This filtering scheme produced a collection of 954 systemically active small molecules, which we called STEAM (Supplementary Results, Supplementary Data Set 1). To condense the STEAM into the CLOUD, we appended biological activities to all drugs with known molecular target and grouped them into classes (Supplementary Data Set 1). Each class is made up of drugs targeting the same protein family with the same mechanism of action. Compounds belonging to the same class are often structurally very similar owing to ‘me too’ drugs. While minor structural differences might change important pharmacological parameters in vivo, we reasoned that molecules that are structurally very similar and belong to the same activity class would behave redundantly in most screening assays. Therefore, we clustered all drugs within a specific group according to chemical structure and selected molecules at cluster centers. For example, the clustering algorithm selected methotrexate out of four structurally similar dihydrofolate reductase inhibitors (Fig. 1b). In contrast, vorinostat and romidepsin are distinct small molecule inhibitors of histone deacetylases. In this and similar cases we kept multiple compounds for the same target (Fig. 1b). Moreover, certain drug classes cover a broader range of therapeutic activities: serotonin receptor agonists can be used to treat anxiety, migraine and disorders of gastrointestinal motility, depending on the specific receptor subtype they engage. For such classes the clustering parameters were relaxed to allow for the selection of structurally related compounds covering a broader range of therapeutic indications. Structural clustering of 176 classes resulted in 239 representative drugs optimized for chemical diversity and coverage of biological activities. To enable detection of compound activity in both biochemical and cellular assays, we searched the literature for prodrugs among these 239 molecules. We identified 35 prodrugs, for which the corresponding active forms were added to our screening collection (Supplementary Table 1). Finally, the 34 STEAM drugs with unknown molecular target were included to generate the CLOUD collection of 308 small molecules (Supplementary Data Set 1 and Supplementary Table 1). The CLOUD covers more drug classes than do other commercially available libraries containing FDA-approved drugs and exclusively provides active forms of prodrugs (Supplementary Fig. 1).

Flutamide (µM)

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

Nature chemical biology doi: 10.1038/nchembio.2382

8.7 11.6 15.5 20.6

8.7 11.6 15.5 20.6

27.5

27.5

36.7

36.7 0

Fractional inhibition

1

–1

Deviation (Differential volume = 23.9)

1

Figure 2 | A combinatorial HTS of the CLOUD uncovers the synergy between flutamide and PPC. (a) Butterfly plot summarizing screen results. Dots represent combinations of two CLOUD drugs. The deviation from the Bliss score and the log of the standard score of the deviation are plotted. Synergies with a deviation >0.7 (red) and antagonisms with a deviation below −0.5 (blue) are indicated. (b) Synergies and antagonisms validated in the counter-screen plotted according to deviation from the Bliss score. The top 20 synergies (red) and antagonisms (blue) are indicated. (c) Dose-response matrix of KBM7 cells treated with flutamide and PPC at the indicated concentrations for 3 d. The average fractional inhibition of viability of 2 biological replicates (left) and deviations from Bliss score (right) are shown. The differential volume is the sum of all deviations.

pattern. Furthermore, orally bioavailable drugs usually adhere to ‘Lipinski’s rule of five’12, and approximately 80% of CLOUD and STEAM drugs did not violate this rule (Fig. 1d). Moreover, when we analyzed individually the physicochemical properties indicated in Lipinski’s rule of five and the numbers of rotatable bonds, we observed no significant differences in distribution between CLOUD and STEAM drugs (Supplementary Fig. 3). This confirmed that the 239 CLOUD drugs with known molecular targets not only reflect the target distribution but also cover the chemical space of STEAM compounds. Approved drugs have been extensively annotated with pharmacological data, including human peak plasma concentrations. While biochemical and cellular HTS commonly apply compounds in the micromolar range, human peak plasma concentrations span several orders of magnitude (Fig. 1e). To foster clinically relevant discoveries, we prepared stocks of CLOUD drugs enabling screens in the range of their respective plasma concentrations. Furthermore, we arranged all CLOUD compounds in a single 384-well plate with the aim of creating an easily accessible reference library for repurposing and annotating clinical drugs.

Flutamide and PPC interact synergistically

The nonredundant nature of the CLOUD enables effective combinatorial screenings exploring new applications of approved drugs. In one such screen, we investigated the effect of pairwise combinations of CLOUD compounds on the viability of KBM7 cells, a near-haploid human chronic myeloid leukemia (CML) cell line

nature CHEMICAL BIOLOGY | Advance online publication | www.nature.com/naturechemicalbiology

3

Nature chemical biology doi: 10.1038/nchembio.2382 a contribution from basal AR signaling activation due to the presence of endogenous ligands, we repeated the dose-response treatment in steroid-deprived conditions, confirming the sensitivity of LNCaP cells to the combination (Supplementary Fig. 10). Next, we a

4

PC-3 cells PPC (µM) 11 .6 15 .4 20 .6 27 .4 36 .6 48 .8 65 .0 86 .7

11 .6 15 .4 20 .6 27 .4 36 .6 48 .8 65 .0 86 .7

4.9

4.9 6.5

Flutamide (µM)

Flutamide (µM)

6.5 8.7 11.6 15.5 20.6

8.7 11.6 15.5 20.6

27.5

27.5

36.7

36.7 0

Fractional inhibition

c

1

0

11 .6 15 .4 20 .6 27 .4 36 .6 48 .8 65 .0 86 .7

15 .4 20 .5 27 .4 36 .5 48 .6 64 .9 86 .5 11 5. 3

10.1

Bicalutamide (µM)

7.6

6.5 8.7

15.5 20.6 27.5

1

LNCaP cells PPC (µM)

4.9

11.6

36.7

13.4 17.9 23.9 31.8 42.5 56.6

0

e

Fractional inhibition

d

LNCaP cells Warfarin (µM)

Fractional inhibition

1

DMSO

104

0

103

102

102

101

101

101

102

103

104

PPC

104

100 100 104

103

103

102

102

101

101

100 100

101

102

Fractional inhibition

103

104

100 100

1

Flutamide

104

103

100 100

AR and vitamin K antagonists synergize in prostate cancer To evaluate the translational potential of the CLOUD and the synergistic interaction uncovered in our combinatorial screen, we assessed the effect of the combination of flutamide and PPC on various prostate cancer cell lines. We performed dose-response matrix measurements with LNCaP cells, an AR-dependent prostate cancer cell line carrying a T877A substitution in the AR23. LNCaP cells showed a marked sensitivity to the combination of flutamide and PPC (Fig. 3a and Supplementary Fig. 8a), whereas PC-3 cells, which express almost undetectable amounts of AR24, were only mildly affected at very high concentrations (Fig. 3b and Supplementary Fig. 8b). Flutamide and PPC also synergized to impair the viability of 22Rv1 cells, which harbor a duplication of exon 3 in the AR gene as well as a H874Y mutation in the AR C-terminal domain (Supplementary Fig. 9a). The two approved drugs synergized also in VCaP and LAPC4 cells, which express wild-type AR, albeit to a lower extent (Supplementary Fig. 9b,c). In contrast, DU145 cells, which express very low levels of AR, were not affected by the combination, similarly to PC-3 cells (Supplementary Fig. 9d). To rule out

b

LNCaP cells PPC (µM)

Flutamide (µM)

that enables the rapid downstream functional characterization of drug targets and mechanisms of action13,14 (Supplementary Table 2 and Supplementary Data Set 2). Using the ABL inhibitor dasatinib as a positive control and measuring cellular ATP levels as a surrogate for cell viability, we tested 40,160 pairwise combinations and achieved robust screening results with an average Z′ factor of 0.4. In line with most drugs acting by mechanisms other than cytotoxicity, the majority of compounds did not affect the viability of KBM7 cells when used alone or in combination (Supplementary Fig. 4). We analyzed combinations of CLOUD compounds for potential synergy or antagonism according to the Bliss independence model15 (Fig. 2a). To corroborate the data, we also evaluated our results using the highest single agent (HSA) model16 and observed good correlation with the Bliss independence approach (Supplementary Fig. 5a). As most CLOUD drugs did not affect the viability of KBM7 cells even at the highest concentration tested, we could calculate Loewe scores17 for only 1,327 combinations. Predictions of synergy or antagonism according to the Chou–Talalay extension of Loewe additivity18 correlated with the Bliss independence model (Supplementary Fig. 5b). We further refined our results following recently published criteria based on a revision of the Saariselkä agreement19 (Supplementary Fig. 5c and Supplementary Data Set 3). On the basis of the results from the Bliss independence analysis, we selected 254 synergistic and antagonistic hit combinations (Fig. 2a and Supplementary Table 2) for counter-screening and further investigated the top 20 synergies and antagonisms in dose-response matrices (Fig. 2b and Supplementary Table 3). The synergistic interaction between flutamide, a nonsteroidal DHT-competitive AR antagonist approved for the treatment of prostate cancer, and PPC, a coumarin anticoagulant used for the treatment of thrombosis, stood out as the most compelling hit of the screen (Supplementary Fig. 6 and Fig. 2c). KBM7 cells enable rapid generation of human gene knockouts by insertional mutagenesis20,21. As flutamide targets the AR, we hypothesized that our synergistic interaction would rely on the presence of the nuclear receptor. We therefore addressed the effect of the drug combination on a KBM7 clone carrying a gene trap in the AR gene (ARGT KBM7). We first validated the loss of AR expression using RT-qPCR and western blotting (Supplementary Fig. 7a,b). Compared to wild-type KBM7 cells, ARGT KBM7 cells showed increased resistance to the combination of flutamide and PPC (Supplementary Fig. 7c). A closer inspection at the position of the gene-trap locus revealed an AR variant (AR45)22 downstream of the inserted cassette (Supplementary Fig. 7d). The expression of this isoform and residual transcripts of the full-length AR in ARGT KBM7 cells (Supplementary Fig. 7e) might explain the incomplete resistance to the combination.

PI

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

article

101

102

103

104

Flutamide + PPC

101

102

103

104

Annexin V

Figure 3 | The combination of antiandrogens and vitamin K antagonists induces apoptosis of LNCaP prostate cancer cells. (a,b) Dose-response matrices for LNCaP (a) and PC-3 cells (b) treated with flutamide and PPC at the indicated concentrations for 3 d. (c,d) Dose-response matrices for LNCaP cells treated with flutamide and warfarin (c) or bicalutamide and PPC (d) at the indicated concentrations for 3 d. Data are average fractional inhibition of viability of 2 biological replicates (the corresponding deviations and differential volumes are illustrated in Supplementary Figs. 8 and 11). (e) Propidium iodide–annexin V staining of LNCaP cells treated with DMSO, 15 μM flutamide, 35 μM PPC or the combination for 3 d. Only the combination of the two drugs induces cellular apoptosis. Representative plots of 2 biological replicates are shown.

nature chemical biology | Advance online publication | www.nature.com/naturechemicalbiology

article

Nature chemical biology doi: 10.1038/nchembio.2382 a

40

DMSO

PPC

Flutamide

Flutamide + PPC

b

Low

AR

β-tubulin

High

FIRE intensity

Merge

DMSO

20

AR_FIRE

10

DAPI

Relative mRNA level

30

8

6 AR_FIRE

2

c

– –

– +

+ –

+ +

130 kDa

Flutamide PPC WB: AR PPC

100 kDa

d

+ – –

– + +

130 kDa

+ + +

Bortezomib Flutamide PPC WB: AR

100 kDa

2

4

6

f

Flutamide + PPC

DMSO 0

DAPI

WB: GAPDH

35 kDa

AR_FIRE

– – –

DAPI

WB: GAPDH

35 kDa

e

KLK2

AR_FIRE

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

TMPRSS2

KLK3

AR

Flutamide + PPC

0

DAPI

Flutamide

4

8

0

10

100 kDa

100 kDa

35 kDa

35 kDa

2

4

6

8

10

– – –

Hours WB: AR

130 kDa 100 kDa

WB: GAPDH 35 kDa

– + 30 30 5 5

– 15 10

+ 15 10

Vitamin K Flutamide PPC WB: AR WB: GAPDH

Figure 4 | The combination of flutamide and PPC induces AR degradation. (a) RT-qPCR analysis of expression of AR and canonical targets of AR signaling in LNCaP cells cultured in steroid-deprived medium and treated with 15 μM flutamide, 35 μM PPC or the combination for 24 h. Data are normalized to actin expression, and DMSO treatment is set to 1. Error bars, s.d. of 3 biological replicates. (b) Immunofluorescence analysis of LNCaP cells treated with 15 μM flutamide, 35 μM PPC or the combination for 24 h. Scale bar, 20 μm. (c) Western blotting (WB) of LNCaP cells treated as in b. GAPDH was used as a loading control. (d) Western blotting of LNCaP cells treated with 30 nM bortezomib, 15 μM flutamide or 35 μM PPC as indicated. GAPDH was used as a loading control. (e) Western blotting of LNCaP cells treated with DMSO (left) or the combination (right). After 24 h, cycloheximide was added to the medium at a final concentration of 2.5 μg/ml, and samples were harvested at the indicated time points. GAPDH was used as a loading control. (f) Western blotting of LNCaP cells treated with flutamide and PPC at the indicated concentrations for 24 h ± 10 mg/ml vitamin K. GAPDH was used as a loading control. Full scans of western blotting experiments are shown in Supplementary Figure 16.

asked whether the synergistic compounds could be exchanged for other members of the drug classes they represent in the CLOUD. PPC was chosen as the representative vitamin K antagonist, a drug class that also includes warfarin. Of note, treatment of LNCaP cells with the combination of flutamide and warfarin resulted in a synergistic toxicity similar to that observed with PPC (Fig. 3c and Supplementary Fig. 11a). Flutamide is the representative androgen receptor antagonist in the CLOUD; other STEAM drugs in this class include bicalutamide and nilutamide. When we tested these compounds in combination with PPC in LNCaP cells, we observed strong synergy with bicalutamide (Fig. 3d and Supplementary Fig. 11b). Enzalutamide and nilutamide also showed a synergistic interaction with PPC, albeit to a lower extent compared to flutamide (Supplementary Fig. 11c,d). We then addressed whether PPC exerted its effect on the viability of LNCaP cells via inhibition of vitamin K epoxide reductase complex subunit 1 (VKORC1), the canonical mechanism of action

described for coumarin anticoagulants. Vitamin K functions as the cofactor of γ-glutamyl carboxylase (GGCX), which catalyzes γ-carboxylation of glutamic acid residues in proteins involved in the coagulation cascade. VKORC1 converts oxidized vitamin K back to the reduced form sustaining the cellular pool of the cofactor25. The addition of vitamin K partially rescued the viability of LNCaP cells treated with flutamide and PPC in combination, with limited vitamin K solubility and uptake probably preventing complete antagonism (Supplementary Fig. 12a). Moreover, knockdown of GGCX increased the toxicity of flutamide in LNCaP cells but did not further decrease viability when flutamide was used in combination with PPC (Supplementary Fig. 12b,c), supporting inhibition of protein γ-carboxylation as part of the mechanism underlying the synergistic interaction. Propidium iodide–annexin V staining followed by flow cytometry analysis of LNCaP cells showed that only the concomitant administration of flutamide and PPC induced apoptosis in the AR-dependent

nature CHEMICAL BIOLOGY | Advance online publication | www.nature.com/naturechemicalbiology

5

Nature chemical biology doi: 10.1038/nchembio.2382

prostate cancer cells (Fig. 3e). Global gene expression profiling of LNCaP cells revealed that after 6 h, flutamide and PPC changed the expression of only a limited number of genes when used individually. However, the administration of the drugs together affected 1,411 transcripts (Supplementary Fig. 13a). Gene set enrichment analysis26 identified apoptosis-associated terms to be significantly (false discovery rate q-value 0.002) upregulated when flutamide and PPC were used in combination (Supplementary Fig. 13b). Overall, these data corroborate the concept behind the CLOUD and show that combinations of different antiandrogens and vitamin K antagonists synergistically lead to prostate cancer cell death.

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

60 °C

72 °C

60 °C

130 kDa 100 kDa

DMSO

130 kDa 100 kDa

130 kDa 100 kDa

PPC

130 kDa 100 kDa

72 °C

DMSO PPC Treated lysates WB: AR

Pretreated cells WB: AR

b

pI 3.0

250 kDa 150 kDa 100 kDa 75 kDa

4.0

5.0

6.0

6.5 7.0 8.0

9.0

10.0

50 kDa

Flutamide and PPC induce AR degradation

To further dissect the mechanism of action behind this synergistic interaction, we analyzed the effect of the combination on AR signaling. While flutamide normally acts as an AR antagonist, mutations in the AR ligand binding domain can switch flutamide activity to an agonistic one and have been described as a common resistance mechanism in prostate cancer27,28. RT-qPCR experiments confirmed that flutamide behaves as an AR agonist in LNCaP cells harboring an AR T877A variant, as the drug increased the expression of AR signaling canonical targets such as KLK3, TMPRSS2 and KLK2 after 24 h (Fig. 4a). In contrast, PPC alone did not affect the expression of these genes. Notably, the concomitant administration of flutamide and PPC restrained the expression of KLK3 and KLK2 (to levels similar to those observed with vehicle treatment) and decreased AR expression in LNCaP cells (Fig. 4a). As PPC alone did not interfere with the induction of AR signaling promoted by the potent synthetic agonist R1881 (Supplementary Fig. 14), we concluded that the vitamin K antagonist does not abolish binding of AR ligands to the receptor. The decrease in AR mRNA upon treatment of LNCaP cells with the combination of flutamide and PPC could reflect a lower transcription rate or decreased mRNA stability. We assessed the turnover of AR mRNA in the presence of a transcription inhibitor over 10 h and found no difference between LNCaP cells treated with DMSO and those treated with the drug combination (Supplementary Fig. 15). In addition to transcripts, AR protein levels were reduced upon co-administration of flutamide and PPC, as assessed by immunofluorescence and western blotting (Fig. 4b,c and Supplementary Fig. 16). Notably, AR mRNA and protein decreased after 8 h (Supplementary Fig. 17). Treatment of LNCaP cells with the proteasome inhibitor bortezomib reduced AR protein levels, as described29, and partly restored the levels of AR upon coadministration with flutamide and PPC (Fig. 4d). In addition, turnover measurements revealed decreased AR half-life in LNCaP cells treated with the combination (Fig. 4e). Vitamin K partly rescued AR depletion in cells treated with flutamide and PPC and increased viability of LNCaP cells in a dose-dependent manner (Fig. 4f and Supplementary Fig. 18). These results suggest that the concomitant administration of flutamide and PPC to LNCaP cells leads to AR degradation that occurs, at least in part, via the proteasome and by means of GGCX inhibition. As AR regulates its own expression30, transcriptional changes probably contribute to the overall decrease in protein abundance. Loss of AR has been described as sufficient to cause apoptosis of LNCaP cells31,32. In line with those findings, we were unable to derive viable LNCaP clones with stable AR knockdown or knockout. Short-term AR knockdown via small hairpin RNA (shRNA) in LNCaP cells showed induction of apoptosis after 120 h (Supplementary Fig. 19). The slower kinetics, as compared to the drug combination treatment, correlated with delayed loss of AR in the knockdown experiment.

The AR is post-translationally modified by g-carboxylation

PPC inhibits γ-carboxylation of proteins involved in blood coagulation. We hypothesized that this post-translational modification 6

a

37 kDa 25 kDa 20 kDa 15 kDa

DMSO

c

PPC

d

Vitamin K (µg/ml) 0 5 20

IP Gla

250 kDa

0

130 kDa

130 kDa 100 kDa

35

67.5

100

PPC (µM) WB: AR

100 kDa Input

WB: Gla

70 kDa 55 kDa

0

35

67.5

100

130 kDa

PPC (µM) WB: AR

100 kDa

35 kDa WB: GAPDH

35 kDa

e

2+

[M+2H] 548.3752

100

Relative abundance (%)

article

ME ac

cbx-me

VQL G L G R

me

50

b2+ 375.3732

y1+

y5+ 529.2697

189.1471 y2+ y3+ y4+ 246.1866 359.3984 416.3701 b3+ 474.2753

200

300

400

500

617.3347

y7+ 756.4996

b7+ 885.3997

b5+ b4+ 715.3505 b6+ 602.2208 y6+ 772.3926 657.4696

600 700 m/z

800

900

1,000 1,100

Figure 5 | The AR is post-translationally modified by g-glutamyl carboxylation. (a) CETSA followed by western blotting (WB) of LNCaP cells treated with DMSO or 35 μM PPC for 2 d (left) and LNCaP cell lysates treated with DMSO or 100 μM PPC for 30 min (right). (b) 2D difference gel electrophoresis (DIGE) of AR immunoprecipitated from LNCaP cells treated with DMSO (green) or PPC (red). (c) Western blotting for global γ-carboxylation in LNCaP cells treated with increasing concentrations of vitamin K. (d) Immunoprecipitation (IP) of γ-carboxylated AR from LNCaP lysates treated with increasing amounts of PPC. Full scans of western blots and 2D gel experiments are shown in Supplementary Figure 20. (e) MS/MS spectrum of the methylated γ-carboxylated AR(1–9) peptide derived from LNCaP cells treated with vitamin K.

could also occur on the AR and induce global conformational changes affecting protein stability and ligand binding. First, we evaluated the effect of PPC on the stability of the AR using a cellular thermal shift assay (CETSA)33, which confirmed that PPC affected the thermal stability of the AR

nature chemical biology | Advance online publication | www.nature.com/naturechemicalbiology

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

Nature chemical biology doi: 10.1038/nchembio.2382 (Fig. 5a and Supplementary Fig. 20). However, this shift in stability was observed only when we cultured LNCaP cells in the presence of PPC for 2 d before the experiment. In the absence of such pretreatment, the addition of PPC to the cell lysate did not stabilize the AR (Fig. 5a), suggesting there is not a direct engagement of the receptor by PPC. We then assessed whether PPC treatment could otherwise modify the AR and analyzed the migration pattern of the protein by means of 2D gel electrophoresis (Fig. 5b). We observed nonoverlapping spots when comparing AR immunoprecipitated from LNCaP cells pretreated with DMSO or PPC. Overall, AR from cells treated with PPC migrated at a higher isoelectric point, consistent with decreased negative charge. Both the CETSA and 2D gel electrophoresis results thus suggest post-translational modification changes on the AR induced by PPC. Next, we assessed global γ-carboxylation to determine whether LNCaP cells are capable of γ-carboxylating protein substrates. Immunoblotting with an antibody recognizing γ-carboxylation in a sequence-independent context revealed several background bands over a wide range of molecular weights (Fig. 5c). However, addition of vitamin K increased the abundance of only a few bands, the most prominent of which corresponded to the molecular weight of the AR (Fig. 5c). Notably, the same antibody immunoprecipitated the AR from LNCaP cell lysates, and PPC treatment reduced the amount of immunoprecipitated protein in a dose-dependent manner (Fig. 5d). Finally, to identify potential AR γ-carboxylation sites, we immunoprecipitated the receptor from LNCaP cells treated with vitamin K. We prepared the samples using procedures previously described for γ-carboxylated substrates34 and analyzed them by liquid chromatography–mass spectrometry. We detected a total of 19 AR-derived peptides and observed γ-carboxylation on E2 and E93 (Fig. 5e). These results show that the AR is γ-carboxylated and that this post-translational modification is inhibited by PPC.

DISCUSSION

In an attempt to physically obtain a complete set of all FDA-approved drugs for screening purposes, we realized that no commercially available compound collection covers all drug classes and that combinatorial screenings of even relatively small libraries would overload the infrastructure of most screening platforms. After an extensive literature search, we generated the CLOUD through cheminformatic selection of a set of FDA-approved drugs representing the entire target and chemical space of all clinical compounds. In contrast to existing collections, all CLOUD drugs fit on a single 384-well plate, are screened in the range of their human plasma concentration and include active forms of compounds clinically administered as prodrugs. At the CeMM chemical screening platform, we now use the CLOUD as a reference set in the development and optimization of all high-content and highthroughput assays. The reductionist approach behind the CLOUD inevitably leads to the omission of additional targets and potential polypharmacological features of some approved drugs. We have appended such additional information, based on a recent study35, to the description of the library (Supplementary Data Set 1). The discovery of the synergistic interaction between flutamide and PPC in a combinatorial HTS of the CLOUD, and the conservation of this synergy among other members of the drug classes these compounds represent, validate the concept underlying our nonredundant collection of FDA-approved small molecules. We identified the drug synergy in an AR-positive CML cell line and followed up in prostate cancer cells, as AR signaling has been shown to have a crucial role in the development of the disease36. Current prostate cancer pharmacological treatments aim at reducing androgen levels or inhibiting downstream signaling with AR antagonists37,38. However, even though patients tend to respond to antiandrogens initially, they inevitably develop resistance39,40.

article

Drug combinations can circumvent or delay the development of resistance mechanisms, and the use of already approved drugs provides different pharmacological and clinical benefits. We observed synergy between flutamide and PPC at clinically relevant concentrations. While the maximum synergy occurs at concentrations higher than those observed in patients, LNCaP cell viability is clearly affected at the peak plasma concentrations of PPC and flutamide. Furthermore, periodic dosing and long-term treatment of patients is likely to result in higher cumulative exposure than that observed in 72 h experiments where single drug doses are administered to cancer cells in vitro. Our data therefore predict that patients receiving vitamin K antagonists are less likely to develop antiandrogen resistance and should therefore experience longer progression-free survival on these drugs. We retrospectively tested this hypothesis using records from the Prostate Cancer Database Sweden (PCBaSe)41. In brief, PCBaSe is a population-based cohort of more than 100,000 men with prostate cancer including data from the National Prostate Cancer Register, the Prescribed Drug Registry and the Cause of Death Registry. There was a nonsignificant decrease in risk of prostate cancer death in 25 men treated with the combination for an extended time (Supplementary Fig. 21). Because the combination of these drugs in clinical practice is rare, larger cohorts are needed to provide final evidence for the clinical effects of the combination of antiandrogens with vitamin K antagonists. Notably, similar attempts have tried to support evidence that warfarin use decreases the risk of prostate cancer mortality, but clinical results were inconclusive42. Ultimately, a thoroughly controlled prospective study together with rigorous patient stratification would be desirable to comprehensively address the clinical benefits of combining antiandrogens with vitamin K antagonists in the context of prostate cancer. Mechanistically, we observed a reduction in AR protein stability and abundance upon treatment of LNCaP cells with the combination of flutamide and PPC. The downregulation of this important nuclear receptor is very likely to be the main apoptotic trigger in AR-dependent LNCaP cells. We also report that γ-carboxylation, which has previously been described only for a restricted group of proteins involved in the coagulation cascade, also occurs on at least two glutamic acid residues of the AR N-terminal domain. One of these AR modification sites was recently described by another group43. We show that γ-carboxylation alters the thermal stability and isoelectric point of the AR. Our findings support a model where PPC precludes γ-carboxylation, and binding of flutamide to the uncarboxylated receptor results in a conformational change that induces protein degradation. Currently, we can only speculate on the conformational changes the AR undergoes before destabilization. The presence of γ-carboxylation sites at the AR N terminus might be relevant for interactions with the ligand binding domain, which are known to be critical for AR activation. These interactions could be further influenced by mutations in the ligand binding domain. Indeed, flutamide and PPC show greater fitness impairment in LNCaP and 22Rv1 cells harboring mutations in the C terminus of the AR. Of note, resistance mutations in the ligand binding domain have been described to affect AR stability in response to different drugs44,45. In summary, our results define a role for γ-carboxylation in AR signaling and suggest that vitamin K antagonists could be clinically repurposed for the formulation of more specific prostate cancer treatments. Received 30 November 2015; accepted 2 March 2017; published online 22 May 2017

Methods

Methods, including statements of data availability and any associated accession codes and references, are available in the online version of the paper.

nature CHEMICAL BIOLOGY | Advance online publication | www.nature.com/naturechemicalbiology

7

article

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

References

Nature chemical biology doi: 10.1038/nchembio.2382

1. Paul, S.M. et al. How to improve R&D productivity: the pharmaceutical industry’s grand challenge. Nat. Rev. Drug Discov. 9, 203–214 (2010). 2. Scannell, J.W., Blanckley, A., Boldon, H. & Warrington, B. Diagnosing the decline in pharmaceutical R&D efficiency. Nat. Rev. Drug Discov. 11, 191–200 (2012). 3. Ashburn, T.T. & Thor, K.B. Drug repositioning: identifying and developing new uses for existing drugs. Nat. Rev. Drug Discov. 3, 673–683 (2004). 4. Clavel, F. & Hance, A.J. HIV drug resistance. N. Engl. J. Med. 350, 1023–1035 (2004). 5. Holohan, C., Van Schaeybroeck, S., Longley, D.B. & Johnston, P.G. Cancer drug resistance: an evolving paradigm. Nat. Rev. Cancer 13, 714–726 (2013). 6. Bock, C. & Lengauer, T. Managing drug resistance in cancer: lessons from HIV therapy. Nat. Rev. Cancer 12, 494–501 (2012). 7. Chong, C.R., Chen, X., Shi, L., Liu, J.O. & Sullivan, D.J. Jr. A clinical drug library screen identifies astemizole as an antimalarial agent. Nat. Chem. Biol. 2, 415–416 (2006). 8. Huang, R. et al. The NCGC pharmaceutical collection: a comprehensive resource of clinically approved drugs enabling repurposing and chemical genomics. Sci. Transl. Med. 3, 80ps16 (2011). 9. Lamb, J. et al. The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease. Science 313, 1929–1935 (2006). 10. Iskar, M. et al. Drug-induced regulation of target expression. PLoS Comput. Biol. 6, e1000925 (2010). 11. Overington, J.P., Al-Lazikani, B. & Hopkins, A.L. How many drug targets are there? Nat. Rev. Drug Discov. 5, 993–996 (2006). 12. Lipinski, C.A., Lombardo, F., Dominy, B.W. & Feeney, P.J. Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings. Adv. Drug Deliv. Rev. 46, 3–26 (2001). 13. Reiling, J.H. et al. A haploid genetic screen identifies the major facilitator domain containing 2A (MFSD2A) transporter as a key mediator in the response to tunicamycin. Proc. Natl. Acad. Sci. USA 108, 11756–11765 (2011). 14. Winter, G.E. et al. The solute carrier SLC35F2 enables YM155-mediated DNA damage toxicity. Nat. Chem. Biol. 10, 768–773 (2014). 15. Bliss, C.I. The toxicity of poisons applied jointly. Ann. Appl. Biol. 26, 585–615 (1939). 16. Berenbaum, M.C. What is synergy? Pharmacol. Rev. 41, 93–141 (1989). 17. Loewe, S. Die quantitativen Probleme der Pharmakologie. Ergeb. Physiol. 27, 48–187 (1928). 18. Chou, T.C. & Talalay, P. Quantitative analysis of dose-effect relationships: the combined effects of multiple drugs or enzyme inhibitors. Adv. Enzyme Regul. 22, 27–55 (1984). 19. Tang, J., Wennerberg, K. & Aittokallio, T. What is synergy? The Saariselkä agreement revisited. Front. Pharmacol. 6, 181 (2015). 20. Bürckstümmer, T. et al. A reversible gene trap collection empowers haploid genetics in human cells. Nat. Methods 10, 965–971 (2013). 21. Carette, J.E. et al. Haploid genetic screens in human cells identify host factors used by pathogens. Science 326, 1231–1235 (2009). 22. Dehm, S.M. & Tindall, D.J. Alternatively spliced androgen receptor variants. Endocr. Relat. Cancer 18, R183–R196 (2011). 23. Veldscholte, J. et al. The androgen receptor in LNCaP cells contains a mutation in the ligand binding domain which affects steroid binding characteristics and response to anti-androgens. J. Steroid Biochem. Mol. Biol. 41, 665–669 (1992). 24. Alimirah, F., Chen, J., Basrawala, Z., Xin, H. & Choubey, D. DU-145 and PC-3 human prostate cancer cell lines express androgen receptor: implications for the androgen receptor functions and regulation. FEBS Lett. 580, 2294–2300 (2006). 25. Suttie, J.W. Vitamin K-dependent carboxylase. Annu. Rev. Biochem. 54, 459–477 (1985). 26. Subramanian, A. et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc. Natl. Acad. Sci. USA 102, 15545–15550 (2005). 27. Feldman, B.J. & Feldman, D. The development of androgen-independent prostate cancer. Nat. Rev. Cancer 1, 34–45 (2001). 28. Gaddipati, J.P. et al. Frequent detection of codon 877 mutation in the androgen receptor gene in advanced prostate cancers. Cancer Res. 54, 2861–2864 (1994). 29. Bhattacharyya, R.S., Krishnan, A.V., Swami, S. & Feldman, D. Fulvestrant (ICI 182,780) down-regulates androgen receptor expression and diminishes androgenic responses in LNCaP human prostate cancer cells. Mol. Cancer Ther. 5, 1539–1549 (2006). 30. Grad, J.M., Lyons, L.S., Robins, D.M. & Burnstein, K.L. The androgen receptor (AR) amino-terminus imposes androgen-specific regulation of AR gene expression via an exonic enhancer. Endocrinology 142, 1107–1116 (2001).

8

31. Eder, I.E. et al. Inhibition of LncaP prostate cancer cells by means of androgen receptor antisense oligonucleotides. Cancer Gene Ther. 7, 997–1007 (2000). 32. Liao, X., Tang, S., Thrasher, J.B., Griebling, T.L. & Li, B. Small-interfering RNA-induced androgen receptor silencing leads to apoptotic cell death in prostate cancer. Mol. Cancer Ther. 4, 505–515 (2005). 33. Martinez Molina, D. et al. Monitoring drug target engagement in cells and tissues using the cellular thermal shift assay. Science 341, 84–87 (2013). 34. Hallgren, K.W., Zhang, D., Kinter, M., Willard, B. & Berkner, K.L. Methylation of γ-carboxylated Glu (Gla) allows detection by liquid chromatography-mass spectrometry and the identification of Gla residues in the γ-glutamyl carboxylase. J. Proteome Res. 12, 2365–2374 (2013). 35. Santos, R. et al. A comprehensive map of molecular drug targets. Nat. Rev. Drug Discov. 16, 19–34 (2017). 36. Taplin, M.-E. Drug insight: role of the androgen receptor in the development and progression of prostate cancer. Nat. Clin. Pract. Oncol. 4, 236–244 (2007). 37. de Bono, J.S. et al. Abiraterone and increased survival in metastatic prostate cancer. N. Engl. J. Med. 364, 1995–2005 (2011). 38. Tran, C. et al. Development of a second-generation antiandrogen for treatment of advanced prostate cancer. Science 324, 787–790 (2009). 39. Chen, C.D. et al. Molecular determinants of resistance to antiandrogen therapy. Nat. Med. 10, 33–39 (2004). 40. Balbas, M.D. et al. Overcoming mutation-based resistance to anti-androgens with rational drug design. eLife 2, e00499 (2013). 41. Van Hemelrijck, M. et al. Cohort Profile: the National Prostate Cancer Register of Sweden and Prostate Cancer data Base Sweden 2.0. Int. J. Epidemiol. 42, 956–967 (2013). 42. Tagalakis, V., Tamim, H., Blostein, M., Hanley, J.A. & Kahn, S.R. Risk of prostate cancer death in long-term users of warfarin: a population-based case-control study. Cancer Causes Control 24, 1079–1085 (2013). 43. Tew, B.Y. et al. Vitamin K epoxide reductase regulation of androgen receptor activity. Oncotarget 8, 13818–13831 (2017). 44. Itsumi, M. et al. PMA induces androgen receptor downregulation and cellular apoptosis in prostate cancer cells. J. Mol. Endocrinol. 53, 31–41 (2014). 45. Yu, Z. et al. Galeterone prevents androgen receptor binding to chromatin and enhances degradation of mutant androgen receptor. Clin. Cancer Res. 20, 4075–4085 (2014).

Acknowledgments

We thank M. Iskar (EMBL) and P. Bork (EMBL) for providing us with DIPS scores, R. Schüle (Albert-Ludwigs-University Freiburg, Germany) for LAPC4 cells, and G. Winter (CeMM) and G. Superti-Furga (CeMM) for thoughtful discussions and initializing synergy screenings at CeMM. S.K. acknowledges support by a Marie Curie Career Integration Grant, the Austrian Federal Ministry of Science, Research and Economy and the National Foundation for Research, Technology, and Development and the Austrian Science Fund (FWF): F4701-B20.

Author contributions

P.M., F.K. and S.K. designed and assembled the CLOUD. M.P.L., F.K., C.-H.L. and S.K. designed and performed the screen of the CLOUD. M.P.L., F.K., M.C. and J.M. analyzed the data from the screen. M.P.L. and A.R. designed and performed viability, RT-qPCR, western blotting, immunoprecipitation and immunofluorescence experiments. S.S. performed immunofluorescence experiments. A.R. and E.S. designed and performed gene expression and immunoprecipitation experiments. A.R. and B.B. designed and performed knockdown experiments. A.C.M. and K.L.B. designed and performed proteomics experiments. A.W., R.H. and K.K. performed and analyzed 2D gel electrophoresis experiments. T.P., M.S. and C.B. performed and analyzed RNA-seq experiments. G.D. and J.C. performed DIPS score analyses. Y.F. and P.S. analyzed patient data in PCBaSe. V.I. synthesized, provided and quality controlled chemicals. M.P.L. and S.K. wrote the manuscript.

Competing financial interests

The authors declare competing financial interests: details accompany the online version of the paper.

Additional information

Any supplementary information, chemical compound information and source data are available in the online version of the paper. Reprints and permissions information is available online at http://www.nature.com/reprints/index.html. Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Correspondence and requests for materials should be addressed to S.K.

nature chemical biology | Advance online publication | www.nature.com/naturechemicalbiology

ONLINE METHODS

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

Cell culture. KBM7 cells were grown in Isocove’s modified Dulbecco’s medium supplemented with 10% FBS (Thermo Fisher Scientific) and 1% penicillin–streptomycin (Sigma-Aldrich). ARGT KBM7 cells were obtained from Haplogen. LNCaP (ATCC, CRL-1740), PC-3 (ATCC, CRL-1435), 22Rv1 (ATCC, CRL-2505) and DU145 (ATCC, HTB-81) cells were grown in RPMI1640 supplemented with 10% FBS and 1% penicillin–streptomycin. VCaP (ATCC, CRL-2876) cells were grown in DMEM supplemented with 10% FBS and 1% penicillin–streptomycin. LAPC4 cells (kind gift from R.S.) were grown in Iscove’s modified Dulbecco’s medium supplemented with 10% FBS and 1% penicillin–streptomycin. For steroid-deprived experiments, LNCaP cells were cultured in RPMI-1640 supplemented with 10% charcoal stripped FBS (Thermo Fisher Scientific) and 1% penicillin–streptomycin. Generation of the STEAM. The complete list of all FDA-approved products was downloaded from the Drugs@FDA database (accessed 18 January 2011), and the names of all active ingredients were extracted. Information on biological activity, Chemical Abstracts Service registry numbers, canonical smiles, and molecular weights was retrieved mainly from the DrugBank (https://www. drugbank.ca/) or the Therapeutic Target Database (http://bidd.nus.edu.sg/ group/cjttd/). Other sources included Clarke’s Analysis of Drugs and Poisons, the KEGG drug database, the Handbook of Clinical DrugData, and Martindale. Active ingredients were filtered as described in the text to obtain 954 systemically active small molecules. Generation of the CLOUD. Structural clustering of the 920 STEAM drugs with known targets was performed using a Pipeline Pilot protocol (Accelrys). Structures were represented as extended connectivity fingerprints (ECFP), and the Tanimoto coefficient was applied as a measurement of structural distance. For cluster formation, the Tanimoto dissimilarity was set to 0.85. If this threshold could not cover most of the therapeutic activities within a drug class, the threshold was lowered in a stepwise procedure. The script was used to identify the cluster centers within each of the 176 drug classes. All cluster centers were kept, providing a final set of 239 structurally unique CLOUD drugs. To these, we added 34 STEAM drugs with unknown targets and the active forms of 35 prodrugs. Compounds were purchased mainly from Enamine Ltd, Toronto Research Chemical and Sigma-Aldrich. Controlled substances and unstable or unavailable compounds are indicated in Supplementary Data Set 1 and Supplementary Table 1. Physicochemical property calculations. Molecular weight, logP, the number of rotatable bonds, hydrogen bond acceptors or donors and violations of the Lipinski’s rule of five were calculated using the chemistry components of the software Pipeline Pilot (Accelrys). KBM7 cell viability screen. CLOUD drugs and combinations thereof were transferred on 384-well plates using an acoustic liquid handler (Echo, Labcyte) and 5,000 cells per well were dispensed on top of the drugs using a dispenser (Thermo Fisher Scientific) for a total of 50 μl/well. Viability was measured after 72 h using CellTiter-Glo (Promega) in a multilabel plate reader (EnVision, PekinElmer). Signal was then normalized to negative (DMSO) and positive (10 μM dasatinib) controls and set between 0 and 100. Noisy compounds, defined according to median absolute deviation (MAD), were excluded from the analysis together with their corresponding combinations. Drug combinations were analyzed according to the Bliss independence model15. Briefly, the effect of the combination of drug A and drug B can be predicted to be C = A + B − A × B, where A and B are the effects of the single drugs expressed as fractional inhibition of viability between 0 and 1. A deviation of the experimental value from the Bliss prediction was calculated. Positive deviations denote synergies, while negative deviations indicate antagonisms. Top hits were selected setting thresholds for deviation (>0.7 for synergies and less than −0.5 for antagonisms) and Z score (>1). Furthermore, we also calculated scores for drug combinations according to the highest single agent (HSA)16, the Loewe additivity model17,18 and criteria based on a revisited Saariselkä agreement19 as already described. In all, 254 hits were selected and tested again in a counter-screen. The top 20 validated synergies and antagonisms were further validated in 4 × doi:10.1038/nchembio.2382

4-concentration dose-response matrices and analyzed in a similar manner. Here, synergisms and antagonisms were measured calculating differential volumes representing the sum of all deviation values within a specific matrix. Differential volumes >1 were taken as synergies, and differential volumes less than −1 were considered antagonisms. Reverse transcription quantitative polymerase chain reaction. RT-qPCR was performed as described46. Primers are listed in Supplementary Table 4. Western blotting. Cells were lysed in RIPA buffer supplemented with a cocktail of protease inhibitors (Roche). Lysates were resolved by SDS–PAGE. AR-specific (5153) and GAPDH-specific (5174) antibodies were purchased from Cell Signaling Technology. αTUB-specific (ab7291) and GGCX-specific (ab197982) antibodies were purchased from Abcam. Anti-γ-carboxyglutamyl antibody (3570) was purchased from Sekisui Diagnostics. Immunofluorescence. Cells were grown on coverslips, fixed with methanol at −20 °C for 24 h, washed once with PBS, permeabilized in solution-B (3% BSA in PBS, 0.1% Triton X-100, 0.02% azide) for 30 min and incubated with primary antibodies in solution-B for 30 min at room temperature. βTUB-specific antibody (T4026, dilution 1:1,000) was purchased from Sigma-Aldrich, and AR-specific antibody (5153, dilution 1:500) was purchased from Cell Signaling Technologies. Primary antibodies were detected using Alexa Fluor 488 goat anti-rabbit IgG and Alexa Fluor 555 goat anti-mouse IgG (Thermo Fisher Scientific) diluted in solution-B. DNA was stained with DAPI (Thermo Fisher Scientific, 0.01 mg/ml). Images were taken using a Leica DMI6000 B inverted confocal system and a 63× 1.30 ACS Apo lens, and edited using Leica LAS AF software (Leica Microsystems) and Fiji (ImageJ). Apoptosis assay. In all, 50,000 cells/well were seeded in 6-well plates and treated with DMSO, 15 μM flutamide, 35 μM PPC or flutamide and PPC in combination after 24 h. After 3 d, cells were labeled with Alexa Fluor 647 annexin V–specific antibody (BioLegend) and propidium iodide and analyzed using flow cytometry (BD FACSCalibur, BD Biosciences). Caspase3/7 activity was measure by Caspase-Glo 3/7 Assay (Promega) according to the manufacturer’s protocol. Cellular thermal shift assay. LNCaP cells were treated with DMSO or 35 μM PPC for 2 d and then harvested maintaining a constant concentration of PPC in all subsequent handling buffers. Otherwise, lysates were directly treated with the drug as described33. Immunoprecipitation. Cells were lysed in immunoprecipitation (IP) buffer (50 mM Tris, pH 8, 150 mM NaCl, 1% IGEPAL) supplemented with a cocktail of protease inhibitors (Roche). Immunoprecipitations were performed at 4 °C and using Dynabeads Protein G (Thermo Fisher Scientific). Lysates were precleared with beads for 30 min, incubated with antibodies overnight (1 μg antibody was used for 50 μg protein) and immunoprecipitated for 2 h. After washing, immunoprecipitated proteins were eluted either by boiling in Laemmli buffer or with 50 mM glycine, pH 2.8. Gene expression profiling. LNCaP cells were treated with DMSO, 15 μM flutamide, 35 μM PPC or the combination for 6 h. Cells were lysed and RNA isolated using the RNeasy Mini Kit (Qiagen) according to the manufacturer’s protocol. The amount of total RNA was quantified using a Qubit 2.0 Fluorometric Quantitation system (Thermo Fisher Scientific), and the RNA integrity number (RIN) was determined using an Experion Automated Electrophoresis System (Bio-Rad). RNA-seq libraries were prepared with TruSeq Stranded mRNA LT sample preparation kit (Illumina) using Sciclone and Zephyr liquid handling robotics (PerkinElmer). Library amounts were quantified using the Qubit 2.0 Fluorometric Quantitation system, and the size distribution was assessed using an Experion Automated Electrophoresis System. Libraries were pooled and sequenced on the Illumina HiSeq platform using the 50-bp single-read configuration. Reads were aligned with TopHat (v2.0.4) with the–no-noveljuncs–no-novel-indels options47. Gene expression was calculated as reads per nature CHEMICAL BIOLOGY

kb per millions of reads (RPKMs) using RPKM_count.py from the RSeQC package48, and the NCBI RNA reference sequences collection (RefSeq) was downloaded from UCSC (http://genome.ucsc.edu/).

© 2017 Nature America, Inc., part of Springer Nature. All rights reserved.

2D gel electrophoresis. IP lysates eluted with glycine buffer were used for 2D difference gel electrophoresis (DIGE). Buffer was exchanged to DIGE sample buffer (30 mM Tris, pH 8.5, 7 M urea, 2 M thiourea, 4% 3-[(3-cholamidopropyl) dimethylammonio]-1-propanesulfonate (CHAPS), 1 mM EDTA) using ultrafiltration columns (Millipore), and pH was adjusted to 8.5 as previously described49. Proteins were labeled with 200 pmol cyanine dyes (Refraction-2D, NH DyeAGNOSTICS) according to the manufacturer’s protocol. Rehydration and isoelectric focusing on IPG strips (ReadyStrip IPG Strips, 11 cm, pH 3–10, nonlinear (Bio-Rad)) was followed by SDS–PAGE on gradient gels (Criterion, Any kDa, Bio-Rad) and detection with a Typhoon fluorescence scanner (GE Healthcare) at the recommended excitation and emission wavelengths for G-Dye200 and G-Dye300. Images were analyzed using Delta2D 4.5 software (Decodon). Knockdown experiments. Lentiviral vectors harboring shRNA sequences in a pLKO.1 backbone were obtained from Sigma-Aldrich. AR_shRNA1: CCGGCACCAATGTCAACTCCAGGATCTCGAGATCCT GGAGTTGACATTGGTGTTTTT AR_shRNA2: CGGCACCAATGTCAACTCCAGGATCTCGAGATCCTG GAGTTGACATTGGTGTTTTTG

nature chemical biology

GGCX: CCGGCCTTTCTTAGACGCCAACAAACTCGAGTTTGTTGGC GTCTAAGAAAGGTTTTTG Liquid chromatography–mass spectrometry. Tryptically digested AR was analyzed on a nano-high-performance liquid chromatography system (Agilent Technologies) coupled to a linear trap quadrupole Orbitrap Velos mass spectrometer (Thermo Fisher Scientific) essentially as described50. Data availability. RNA-seq data are available in Gene Expression Omnibus (GEO) under accession number GSE94783.

46. Licciardello, M.P. et al. NOTCH1 activation in breast cancer confers sensitivity to inhibition of SUMOylation. Oncogene 29, 319 (2014). 47. Kim, D. et al. TopHat2: accurate alignment of transcriptomes in the presence of insertions, deletions and gene fusions. Genome Biol. 14, R36 (2013). 48. Wang, L., Wang, S. & Li, W. RSeQC: quality control of RNA-seq experiments. Bioinformatics 28, 2184–2185 (2012). 49. Rudashevskaya, E.L. et al. A method to resolve the composition of heterogeneous affinity-purified protein complexes assembled around a common protein by chemical cross-linking, gel electrophoresis and mass spectrometry. Nat. Protoc. 8, 75–97 (2013). 50. Huber, M.L. et al. abFASP-MS: affinity-based filter-aided sample preparation mass spectrometry for quantitative analysis of chemically labeled protein complexes. J. Proteome Res. 13, 1147–1155 (2014).

doi:10.1038/nchembio.2382