The Drug Sensitivity and Resistance Testing (DSRT) Approach

A phenotypic screening and machine learning platform e ciently identiﬁes triple negative breast cancer-selective and readily druggable targets Prson Gautam 1 Alok Jaiswal 1 Tero Aittokallio 1, 2 Hassan Al Ali 3 Krister Wennerberg 1,4

Identifying e ective oncogenic targets is challenged by the complexity of genetic alterations in 1Institute for Molecular Medicine Finland (FIMM), HiLIFE, University of Helsinki, Finland cancer and their poorly understood relation to cell function and survival. There is a need for meth- Current kinome coverage of kinase inhibitors in TNBC exhibit diverse kinase dependencies MFM-223 is selectively addicted to FGFR2 2Department of Mathematics and Statistics, University of Turku, Finland 3The Miami Project to Cure Paralysis, Peggy and Harold Katz Family Drug Discovery Center, A A Sylvester Comprehensive Cancer Center, and Department of Neurological Surgery and Medicine ods that rapidly and accurately identify “pharmacologically e ective” targets without the require- clinical evaluation TN Kinases MFM-223 CAL-120 MDA-MB-231 TNBC TNBC TNBC TNBC TNBC TNBC HER2+ 100 University of Miami Miller School of Medicine, Miami, FL 33136, USA. non- HER2+ FGFR1 0.97 0.00 0.00 MFM-223 BL1 BL2 M MSL IM LAR ER+, PR+ 50 ment for priori knowledge of complex signaling networks. We developed an approach that uses ma- cancerous FGFR2 56.46 0.00 0.00 CAL-120 25 4 MDA-MB-231 Biotech Research & Innovation Centre (BRIC) and Novo Nordisk Foundation Center HCC1937 CAL-85-1 CAL-120 MDA-MB-231 DU4475 CAL-148 MCF-10A SK-BR-3 BT-474 FGFR3 25.10 0.00 0.00 0 chine learning to relate results from unbiased phenotypic screening of kinase inhibitors to their bio- for Stem Cell Biology (DanStem), University of Copenhagen, Denmark HCC1599 HDQ-P1 BT-549 MDA-MB-436 MFM-223 FGFR4 0.00 0.00 0.00 MAXIS*Bk Clinical status MDA-MB-468 CAL-51 Hs578T MDA-MB-453 score chemical activity data. This process, which we call idTRAX (Identiﬁcation of Drug TaRgets and An- HCC1143 Approved Phase 4 B Erdafitinb AZD4547 Lucitanib Discordance within kinase dependency profiles generated by different ti-targets by Cellular and Molecular Cross-referencing), discovers targets that are pharmacological- 100 100 100 Phase 3 B 100 67 gene silencing approaches and idTRAX ly responsive and readily druggable. Additionally, the identiﬁed targets are not typically overcome TKL Phase 2 33 75 75 75 TK 0 -33 Phase 1 -67 -100 50 50 50 Cell lines by the robustness of signaling networks, because only targets that e ectively induce a phenotype y inhibition MAXIS*BK score A bili t a

25 25 25 0.3 14 3

upon pharmacological engagement are selected. We applied this methodology to triple negative %v i CAL-14 8 STE HCC159 9 MDA-MB-23 1 CAL-5 1 CAL-85- 1 Hs578 T HCC 1 SK-BR- 3 MDA-MB-43 6 HCC193 7 MFM-22 3 DU447 5 MDA-MB-46 8 CAL-12 0 MCF-10 A BT-54 9 HDQ-P 1 MDA-MB-45 3 BT-47 4 0.2

HCC1599 HCC1937 MDA-MB-436 MDA-MB-468 MDA-MB-231 HCC1143 HDQ-P1 MCF-10A BT-549 Hs578T CAL-85-1 BT-474 SK-BR-3 MDA-MB-453 MFM-223 CAL-51 CAL-120 CAL-148 DU4475 0 0 0 breast cancer, which still lacks targeted therapy. We screened 19 breast cancer cell lines with ~500 PRKD2 PRKD3 0.1 1 10 100 1000 0.1 1 10 100 1000 1 10 100 1000 10000 0.1 PRKD1 Concentration (nM) PRKX small-molecule kinase inhibitors with annotated kinase activity data and identified cell-line specific PRKACA 0.0 PRKCQ SGK3 PRKCH -0.1 kinase targets, i.e. kinases whose inhibition causes cell line-specific cytotoxicity or cytostaticity. We PRKCG MFM-223 CAL-120 MDA-MB-231 Spearman’s rho ( ρ) PRKCA C CMGC AKT3 -0.2 AKT2 100 100 100 also found that triple negative breast cancer cell lines exhibit heterogeneous target patterns, indi- AKT1 RPS6KA1 ****

RPS6KA6 io n 75 75 75 siRNA_shRNA RPS6KA4 shRNA_CRISPR siRNA_CRISPR shRNA_idTRAX siRNA_idTRAX CRISPR_idTRAX cating the need of personalized medicine approach to tackle them. On further analysis, we found CK1 RPS6KA5 PRKG2 PRKG1 50 50 50 ROCK2 that the correlation between targets identified by the pharmacological approach and those identi- ROCK1 B CAMK1D =0.2551, p=0.0006 =0.1147, p=0.1263 =0.1077, p=0.1511 CHEK2 viability inhibi t 25 25 25 ρ ρ ρ CAMK2D % 5 5 5 fied by knockdown/knockout-based screens was low, suggesting that gene-silencing approaches PDPK1 BRSK1 BRSK2 0 0 0 MARK1 MARK4 FGFR1 FGFR2 FGFR3 FGFR4 Death FGFR1 FGFR2 FGFR3 FGFR4 Death FGFR1 FGFR2 FGFR3 FGFR4 Death may not be the most ecient at identifying targets for small-molecule drug discovery or repurpos- SIK2 Scramble Scramble Scramble 0 0 0 MARK3 siRNA MELK NUAK1 ing. Our approach provides a platform for rapidly identifying sample-specific drug targets and po- TSSK1B TSSK2 -5 -5 -5 PHKG2 A) Heat map based on MAXIS*Bk score, predicting FGFR2 addiction unique to MFM-223 cell line. Atypical Protein Kinases AGC PHKG1 z-score (siRN A) tentially guiding personalized therapy regimens. GRK7 B) Validated of FGFR addiction with three individuals FGFR KIs. The dose response curves are color coded z-GARP score (shRN A) MAPKAPK3 z-GARP score (shRN A) ABC1 PLK1 respective to cell lines as mentioned in the legend. -10 -10 -10 MAP4K2 -2 -1 0 1 -2 -1 0 1 -10 -5 0 5 STK26 C) Bar plot illustrating the effect of siRNA-based knockdown of four different FGFR isoforms in the three respective Alpha STK3 CERES score (CRISPR) CERES score (CRISPR) z-score (siRNA) STK4 MINK1 cell lines. Brd STK10 The drug sensitivity and resistance testing (DSRT) approach PAK6 ρ=0.1455, p=0.0520 ρ=0.0227, p=0.7629 ρ=0.01629, p=0.8286 CAMK Others PAK5 PDHK PAK3 well established generic SRPK1 0 0 0 PI3K CHEK1 PIKK vulnerabilities ABL1 ABL2 25 25 25 SRC RIO YES1 idTRAX identifies pharmacologically effective targets not FYN TIF1 FGR 50 50 50 Printing of drug plates containing BLK detected by RNAi ~600 oncology drugs in 5 different HCK LCK A 75 75 75 concentrations in a 10,000-fold PIK3CA PIK3CD 100 MFM-223 MAXIS*Bk score MAXIS*Bk score MAXIS*Bk score concentration range. BTK Kinases MFM-223 CAL-148 CAL-120 MDA-MB-231

CAL−51 CAL-148 CAL−51 TXK 50 100 100 100

20 40 AKT1 54.96 45.74 0.00 0.00 TEC CAL-120 ITK 25

30 10 AKT2 44.98 46.98 0.00 0.00 MDA-MB-231 -2 -1 0 1 -10 -5 0 5 -10 -5 0 5 FRK 0 SRMS

DSS CERES score (CRISPR) z-GARP score (shRNA) z-score (siRNA) 0 idTRAX to de-convolve cell line-selective kinase AKT3 20.86 12.93 0.00 0.00 Selective DSS Selective FER MAXIS*Bk FES 10 −10 dependency from drug sensitivities PTK2B score

0 EPHA3

in ib

cin

−3 21 di

cin

ine

ulin

391

502

iclib

sine 128

y atin

y in

bine

ista ene

lone

bic

−

lin

olis

halan

−2 −391

abine

oci

rinad

aenol

Stattic

dib

26458

YM155

cetax

staur

PIK−

AT 101 AT efit

T−406

pimy

ara

BI 2536 (+)J

emetan

BIIB021

Afatinib

GSK−J4 X2

tupi

AVN944

UCN−01

AUY922

Indibulin

MK1775

icam Nut

alrub

Alv

demetan Olapar K

amatinib

AZD8055

Linsitinib SNS−032

Paclitaxel

Dina

AZD2014

Neratinib

046 Do lumetinib Belinostat

Valru − Dact

TAK−901

Volasertib

TGX

AZD7762 Melp

ABT−751

Patupilone

UNC0642

Entinostat

Bleomycin

Ble

Pl Clomif

Vinblastine

Plicamycin

emcit Carfilzomib Floxuridine Pa

erd Vincristine

Trametinib

efametinib

Cytarabine Dapo

Vinorelbine Floxuridine

Bortezomib

Clo

Pravast

SB 743921 Tosedostat

Clofarab Ixabepilone

Mido

−ade

SK2636771

Bortezomib

Ser Dacomitinib

Pevonedistat

Lestaurtinib Se S

o−ade Tubastatin A Tubastatin

Mitomycin C Hydroxyurea

Hydroxyurea

Omacetaxine

Pevone

Panobinostat

Dactinomycin

Methotrexate

Tanes

Methotrexate

Fost

Camptotechin EPHA4

GSK21

PF−

Camptotechin

NVP

PF−00477736

Mercaptopurine

−chloro

−chloro−adenosine

8−c

(5Z)−7−Oxo

mino

8−amino−chloro−adenosine Correlation between kinase dependencies assimilated using different screening approaches (drug, siRNA, shRNA, and CRISPR 8−a EPHB2 EPHB3 inhibitor EPHA2 B Afuresertib Uprosertib Ipatasertib screening). siRNA screening data (z-score), shRNA data (z-GARP scores) and CRISPR data (CERES score) were extracted from Apitolisib EPHB4 FGFR1 PI3K kinase FGFR1 100 100 100 Campbell et al [22], Marcotte et al [23] and from Meyers et al [24] studies, respectively. MAXIS*Bk score represents kinase dependency mTOR FGFR3 Axitinib 100 100 100 Vandetanib Dactolisib FGFR2 CDK JAK profile based on compound screening. Only kinases common to all datasets (n=179) were used in the analysis. RET 75 75 75 RET Incubation for 72 hours Dovitinib Lapatinib Lucitanib 75 75 75 CUDC−102 LTK R KIT A) Dot plot comparing correlation (Spearman rank correlation) between different functional profiling approaches. For each comparison, 100 max 50 50 50 at 37°C, 5% CO Erlotinib ALK Drug response TRK %viability %viability 2 %viability 25 25 25 Danusertib ROS1 Regorafenib AKT only the cell lines that are common between the two data sets were incorporated. Colors indicate individual cell lines. Dasatinib INSR 80 0 0 0 UCN−01 50 50 50 INSRR y inhibition Midostaurin -10 -9 -8 -7 -6 -9 -8 -7 -6 -5 -8 -7 -6 -5 -4 Aurora VEGFR B) Scatter plot comparing scores from different methods in a representative cell line MDA-MB-231. The points correspond to EGFR SRC IGF1R HER2

(%) log[M] log[M] log[M] Imatinib bili t 60 NTRK1 known mutation a B−Raf Saracatinib Perifosine 25 the common kinases between the data sets used in the analysis. Slope hibition Foretinib NTRK3 25 25 RON TAK−901 100 100 100

in MFM223 NTRK2 %v i Pictilisib 40 sorafenib 75 75 75 AXL

Relative in PDGFR MEK AZD1480 MERTK 50 0 50 50 FGFR 0 0 PDK 20 AT9283 TYRO3 A %viability %viability CHK min Rebastinib 25 %viability 25 25 PKC MET Nilotinib MUSK Rmin MGCD−265 DDR2 0.1 1 10 100 1000 1 10 100 1000 10000 1 10 100 1000 10000 FLT 0 0 0 Trametinib 0 FLT3 EphR ABL SYK -9 -8 -7 -6 -5 -9 -8 -7 -6 -5 -9 -8 -7 -6 -5 IC50 Concentration (nM) TIE2 TEK Drug concentration (logM) log[M] log[M] log[M] Lestaurtinib HER2+ EGFR Conclusions Tandutinib ERBB4 cell lines ERBB2 Curve fitting and MET Kinase inhibitors in clinical evaluation are more biased towards tyrosine kinases, our data reveal Drug sensitivity score (DSS) Inhibitor activity data MST1R drug sensitivity score (DSS) calculation TNK2 MFM-223 CAL-120 MDA-MB-231 * TNK1 C opportunities for targeting other types of kinases and calls for improving representation of non-ty- NEK1 NEK9 100 100 100 AURKC rosine kinases in clinical development Cell lines/ Patient samples DAY 3: Measuring cytotoxicity and AURKB

JAK2 io n viability of cells JAK3 75 75 75 CSF1R DAY 3 : Generation of dose response PDGFRB We developed a machine learning algorithm which maps the drug sensitivity based phenotypic curves, calculation of drug sensitivity Machine learning algorithm FLT4 50 50 50 KDR * score (DSS) and identification of most (idTRAX) FLT1 data to kinase activity profile to de-convolve cell line-selective kinase addictions effective/selective drugs. Drug set driver muation RAF1 DAY 1: Dispensing of cells on BRAF viability inhibi t 25 25 25 drug plates. enrichment and meta analysis. PIM1 % in DU4475 PIM3 PIM2 MKNK1 0 0 0 Our approach underscores the heterogeneity of kinase addiction in TNBC known addition in LRRK2 MAP2K1 AKT1 AKT2 AKT3 Death AKT1 AKT2 AKT3 Death AKT1 AKT2 AKT3 Death * MDA-MB-231 (KRAS mutant) CDK5 Scramble Scramble Scramble Kinases CDK1 siRNA CDK2 CDK3 The algorithm provides a platform for identifying selective vulnerabilities and thereby guiding tar- CDK6 * MAPKAPK3 FGFR2 NTRK1 SGK3 RPS6KA1 PRKD2 CHEK2 STK3 MKNK1 PRKCQ PRKCA STK10 EPHA3 EPHB4 PAK5 ABL2 TXK EPHB3 RPS6KA5 RPS6KA4 PRKG1 ROCK2 ROCK1 MARK4 ITK MET MUSK DDR2 TEK FLT1 BRAF PI4KB YES1 FRK MAPK14 MST1R MAPK9 EGFR RAF1 MAPK11 AKT1 MINK1 CDK6 TSSK1B MAP4K2 STK26 ROS1 MELK INSRR TNK1 RPS6KA6 INSR CSNK2A1 CDK5 JAK3 DYRK2 TSSK2 CDK1 NUAK1 CLK2 LTK NEK9 FER HIPK1 GSK3A DYRK1A FGR AURKC IKBKE ERBB4 ERBB2 PIK3CA IGF1R DYRK1B PTK2B HIPK4 PLK1 PIK3CD CSNK1A1 MAPK9 geted therapy regimens References DU4475 MAPK14 A) Heat map based on MAXIS*Bk score, predicting AKT addiction unique to MFM-223 and CAL-148 cell lines. CAL-148 MAPK11 CAL-120 GSK3A B) Validated of AKT addition with three individuals AKT KIs. The dose response curves are color coded respective Pemovska, T. et al. Individualized systems medicine strategy to tailor treatments for patients with chemorefractory acute myeloid leukemia. Cancer Discov. 3, 1416–1429 (2013). CAL-51 GSK3B 1. MFM-223 MAXIS*Bk-score CSNK2A1 to cell lines as mentioned in the legend. MDA-MB-453 DYRK1B SK-BR-3 Chemical screening approach will provide more readily translatable vulnerabilities in comparison DYRK1A C) Bar plot illustrating the effect of siRNA-based knockdown of three different AKT isoforms in the three respective 2. Yadav, B. et al. Quantitative scoring of differential drug sensitivity for individually optimized anticancer therapies. Sci. Rep. 4, 1–10 (2014). BT-474 DYRK2 CAL-85-1 100 67 33 0 -33 -67 -100 * Hs578T HIPK1 cell lines. to loss-of-function screening approach IRAK4 3. Lehmann, et al. ‘Identification of human triple-negative breast cancer subtypes and preclinical models for selection of targeted therapies’, J Clin Invest.,121: 2750-2767 (2011). BT-549 MCF-10A CLK2 HDQ-P1 viability inhibition HIPK4 4. Gautam, P. et al. Identification of selective cytotoxic and synthetic lethal drug responses in triple negative breast cancer cells. Mol. Cancer. 15, 1–16 (2016). Cell lines HCC1143 TTK MDA-MB-231 CHUK MDA-MB-468 CSNK1G1 5. Al-Ali, H. et al. Rational polypharmacology: systematically identification and engaging multiple drug targets to promote axon growth. ACS Chem Biol. 10(8):1939-51 (2015). MDA-MB-436 CSNK1A1 HCC1937 TBK1 In close partnership with: HCC1599 IKBKE FIMM- Institute for Molecular Medicine Finland, PI4KB 6. Campbell, J. et al. Large-scale profiling of kinase dependencies in cancer cell lines. Cell Rep. 14(10): 2490-501 (2016). P.O.Box 20, FI-00014 University of Helsinki, Finland Cell line-specific kinase A) Transcriptomics-based classification of TNBC cell lines (Lehmann et al, 2011) 7. Marcotte, R. et al. Functional genomic landscape of human breast cancer drivers, vulnerabilities, and resistance. Cell. 164(1-2):293-309. (2016). Biomedicum Helsinki 2U, Tukholmankatu 8, 00290 Helsinki, Finland B) Heat map based on MAXIS*BK score. Known cell lines-specific kinase addictions are highlighted 8. Aguirre, A.J. et al. Genomic copy number dictates a gene-independent cell response to CRISPR/Cas9 targeting. Cancer Discov. 6(8):914-29 (2016). Illustration of the kinase dependency deconvolution approach in green boxes. www. mm.