Machine-Learning and Chemicogenomics Approach Defi Nes and Predicts Cross-Talk of Hippo and MAPK Pathways

Published OnlineFirst November 18, 2020; DOI: 10.1158/2159-8290.CD-20-0706

reseArch ArtIcle

Machine-Learning and Chemicogenomics Approach Defi nes and Predicts Cross-Talk of Hippo and MAPK Pathways

Trang H. Pham 1 , Thijs J. Hagenbeek 1 , Ho-June Lee 1 , Jason Li 2 , Christopher M. Rose 3 , Eva Lin 1 , Mamie Yu 1 , Scott E. Martin 1 , Robert Piskol 2 , Jennifer A. Lacap 4 , Deepak Sampath 4 , Victoria C. Pham 3 , Zora Modrusan 5 , Jennie R. Lill 3 , Christiaan Klijn 2 , Shiva Malek 1 , Matthew T. Chang 2 , and Anwesha Dey 1

ABstrAct Hippo pathway dysregulation occurs in multiple cancers through genetic and nongenetic alterations, resulting in translocation of YAP to the nucleus and activation of the TEAD family of transcription factors. Unlike other oncogenic pathways such as RAS, defi ning tumors that are Hippo pathway–dependent is far more complex due to the lack of hotspot genetic alterations. Here, we developed a machine-learning framework to identify a robust, cancer type–agnostic gene expression signature to quantitate Hippo pathway activity and cross-talk as well as predict YAP/TEAD dependency across cancers. Further, through chemical genetic interaction screens and multiomics analyses, we discover a direct interaction between MAPK signaling and TEAD stability such that knockdown of YAP combined with MEK inhibition results in robust inhibition of tumor cell growth in Hippo dysregulated tumors. This multifaceted approach underscores how computational models combined with experimental studies can inform precision medicine approaches including predictive diagnostics and combination strategies.

SIGNIFICANCE: An integrated chemicogenomics strategy was developed to identify a lineage- independent signature for the Hippo pathway in cancers. Evaluating transcriptional profi les using a machine-learning method led to identifi cation of a relationship between YAP/TAZ dependency and MAPK pathway activity. The results help to nominate potential combination therapies with Hippo pathway inhibition.

1Department of Discovery Oncology, Genentech, Inc., South San Francisco, Corresponding Authors: Anwesha Dey, Genentech Inc., 1 DNA Way, MS California. 2 Department of Bioinformatics, Genentech, Inc., South San Fran- 41-1a, South San Francisco, CA 94080. Phone: 650-678-8953; Fax: cisco, California. 3Department of Microchemistry, Proteomics, and Lipidomics, 650-225-6443; E-mail: [email protected] ; and Matthew T. Chang, matthew. Genentech, Inc., South San Francisco, California. 4 Department of Translational [email protected] 5 Oncology, Genentech, Inc., South San Francisco, California. Department of Cancer Discov 2021;11:778–93 Molecular Biology, Genentech, Inc., South San Francisco, California. doi: 10.1158/2159-8290.CD-20-0706 Note: Supplementary data for this article are available at Cancer Discovery Online (http://cancerdiscovery.aacrjournals.org/). ©2020 American Association for Cancer Research. T.H. Pham, T.J. Hagenbeek, and H.-J. Lee contributed equally to this work.

778 | CANCER DISCOVERY March 2021 AACRJournals.org

Introduction (4–6). These led to the discovery of the conserved Hippo pathway core components consisting of serine/threonine kinases One challenge of cancer precision medicine is the hetero- named Mammalian STE20-like 1/2 (MST1/2) with adaptor geneity of genetic and nongenetic alterations that result in protein SAV1 that directly phosphorylate the large tumor aberrant pathway signaling. Recurrent mutations and genetic suppressors (LATS1/2). Together with the kinase activator alterations have been identified in many oncogenic signaling MOB1, LATS1/2 can phosphorylate the two major down- pathways, including MAPK and PI3K (1, 2), whereas other stream coactivators YAP (YAP1) and TAZ (WWTR1; Fig. 1A, signaling pathways such as Hippo lack canonical hotspot inset). When the pathway is deregulated, unphosphorylated mutations. Yet dysregulation in Hippo pathway signaling is YAP and TAZ are translocated to the nucleus and activate known to drive oncogenesis across numerous cancer types. downstream target gene expression by binding to TEAD The Hippo pathway is emerging as the target of drug dis- family transcription factors (TF; refs. 7–13; Fig. 1A, inset). covery efforts, but it lacks hotspot mutations; identifying Widespread dysregulation of the Hippo pathway components relevant Hippo pathway–dependent patient population(s) has been observed in multiple human cancer types including and biomarker(s) of response is a prerequisite for precision glioma and breast, liver, lung, prostate, colorectal, and gas- medicine in tumors that leverage this pathway. tric cancers (14–17). Furthermore, tumors with dysregulated The Hippo pathway controls multiple cellular functions Hippo components are not only insensitive to the intrinsic that drive oncogenesis, including proliferation, cell fate deter- cellular death barriers (3, 18) but also resistant to chemo and mination, and cell survival. Perturbation of the pathway has molecular targeted therapies (19–21). been shown to trigger tumorigenesis in mice (3). The pathway Extensive studies have established the importance of the is evolutionarily conserved across diverse species and was first Hippo pathway in biology and cancers. As drug develop- identified inDrosophila melanogaster through multiple genetic ment interest in targeting the pathway continues to grow screens for gene mutations that cause overgrowth phenotype (22–25), one key clinical challenge is to identify patient

March 2021 CANCER DISCOVERY | 779

RESEARCH ARTICLE Pham et al.

MST1/2 A SAV1 B P Nucleus Cytoplasm TCGA cervical squamous 14 LATS1/2 12 MOB1

count) 10 P 2 P P 8 (log YAP/TAZ P YAP/TAZ 6 P TEAD 4 14-3-3 mRNA expression YAP1 MST1 4% 12 STK3 1% NF2 2% Cervical squam. LATS1 6% LATS2 3% 10 YAP1 12% Most frequent n = 50 n = 140 YAP1 amplification 8 LATS1 alteration Ovarian LATS2 alteration TCGA head and neck squamous MST1 alteration NF2 alteration 14 6 Head and neck squam. 12

count) 10 2 Urothelial bladder 8 Esophageal (log 4 6 4 Sarcoma mRNA expression YAP1 Melanoma Lung adeno. MST1 2% Glioblastoma STK3 1%

amplified) % of samples ( YAP1 2 Uterine Uterine sarc. NF2 2% Lung squam. Stomach adeno Renal (clear cell) AML Colon LATS1 2% Liver Renal (papillary) DLBC Mesothelioma LATS2 2% 0 Thymoma uveal mel Cholangiocarcinoma YAP1 6% 02468101231 n = 65 n = 429 Genetic alteration % of samples altered (other) Amplification Deep deletion Tr uncating mutation

Inframe mutationMissense mutation

C siNTC Upregulated siYAP/TAZ Downregulated nregula dow ted SF268 383 D on ge m ne m s o ( 221 C n Detroit 562 7/7 (2%) =

3 , ES-2 6/7 (9%) 7 2

5/7 (10%) ,

RVH-421 6

4/7 (11%) % HCC1954 3/7 (14%) ) SKG-II 2/7 (20%) OVCAR-8 1/7 (34%) 040 80 02,000 4,000 01234 YAP1 mRNA # of differentially CTGF (RPKM) expressed genes (log fold change) 2 F Significance score (−log10 P value) 0 1230 123 E Unclustered Genes not (n = 135) broadly expressed Cluster 1 (n = 4) (n = 220) Genes with inconsistent changes in expression Cluster 2 (n = 2,994) (n = 145)

Common Cluster 3 downregulated Genes without consistent WGCNA (n = 80) changes in expression genes Cluster 4 (n = 3,721) upon siYAP1/TAZ KD (n = 621) (n = 41) Detroit 562 (HNSCC) PA-TU-8902 (Panc)

Figure 1. Frequent YAP1 amplification associates with upregulation ofYAP1 RNA expression and aberrant pathway signaling. A, Frequency of YAP1 copy-number amplifications y( axis) compared with the most frequent alteration in core Hippo pathway members other than YAP1 (x axis) across TCGA cohort. Each point is colored based on the most frequently mutated core Hippo pathway member. B, Pattern of mutations in core Hippo pathway members is mutually exclusive across cancers, and shown here are cervical squamous and head/neck squamous. Top of each oncoprint is YAP1 mRNA expression where YAP1 overexpression was found predominantly in YAP1-amplified samples. C, RNA-seq data from siYAP1+WWTR1 vs. siNTC of seven different cell lines carrying YAP1 amplifications, resulting in broad gene expression changes including significant downregulation ofCTGF in all cell lines. D, Forty- six percent of significantly downregulated genes were identified in at least three cell lines uponYAP1/WWTR1 knockdown. E, Schematic of weighted gene co-expression network analysis (WGCNA) that defines four gene clusters associated withYAP1 /WWTR1 knockdown. Common downregulated genes defined as significantly downregulated in at least three of seven cell lines. F, ATAC-seq analysis identifies that Cluster 2 genes are mostly associated with loss of chromatin accessibility upon YAP1/WWTR1 knockdown in Detroit 562 and PA-TU-8902 cells.

780 | CANCER DISCOVERY March 2021 AACRJournals.org

Machine-Learning Approach Predicts Hippo Pathway Dependency RESEARCH ARTICLE populations that would benefit from such a therapy. Previ- broadly expressed across all tissues in addition to being sig- ous studies on the Hippo pathway have either defined broad nificantly and consistently downregulated (in at least three genetic alterations in pathway component(s) or focused on of seven cell lines) upon YAP1/WWTR1 knockdown (Fig. individual cancer types or cancer cell lines. This experimen- 1D). This identified four distinct gene clusters of coexpressed tal strategy has established the role of the Hippo pathway genes and one cluster of noncorrelated genes (Fig. 1E). Inter- in cancers; however, regulation of Hippo pathway signaling estingly, we noted that many of the canonical Hippo path- can be highly complex with many linked signaling inputs way–regulated genes (e.g., CTGF, CYR61, etc.) were found from the orthogonal pathways (26). Here, we employ an within gene Cluster 2, suggesting that Cluster 2 may be most integrated experimental–computational strategy to identify proximal to Hippo pathway signaling. Among the 145 genes a lineage-independent signature for the Hippo pathway in in Cluster 2, 86% (n = 124) have not been reported in previ- cancers. By evaluating transcriptional profiles, we observed a ous YAP/TAZ gene signatures (Supplementary Fig. S1C) in relationship between YAP/TAZ dependency and MAPK path- which we orthogonally validate several genes using RT-PCR way activity, leading us to nominate potential combination (Supplementary Fig. S1D). To further validate Cluster 2, we therapies with Hippo pathway inhibition. leveraged recent systematic CRISPR and RNAi dependency screens (30). Although these datasets utilize only single-gene Results knockout, nevertheless, we performed gene-wide regression analysis with overlapping cell lines to assess whether the new Computational Pan-Cancer Approach to Quantify gene set is associated with a given gene knockout/knockdown. Hippo Pathway Dysregulation Among the most significant gene dependencies, this anal- To understand the role of the Hippo pathway in human ysis confirmed many Hippo pathway effectors included cancers, we first examined the pathway alternations using WWTR1, YAP1, and TEAD1 (Supplementary Fig. S1E and data from The Cancer Genome Atlas (TCGA; ref. 27). Although Supplementary Table S1). We then performed an unbiased genetic alterations in the Hippo pathway are infrequent (1%– analysis of somatic genetic predictors of Cluster 2 single- 15% across individual cancer types), YAP1 amplifications are sample gene set enrichment analysis scores in the TCGA among the most frequent alterations pan-cancer in the Hippo pan-cancer cohort (Supplementary Fig. S1F). The most pathway (Fig. 1A) and most frequently observed in patients significant results includedNF2 loss-of-function mutations with cervical and head and neck squamous cell cancer. As and homozygous deletions (Supplementary Fig. S1F and expected, genetic YAP1 amplifications, but no other Hippo S1G), consistent with Hippo pathway regulation. Further- pathway alterations, were exclusively associated with YAP1 more, we performed RNA-seq on three independent, NF2- RNA overexpression in multiple cancer types (Fig. 1B). Fur- null (Hippo pathway altered) cell lines [i.e., GOS-3 (glioma), thermore, genetic YAP1 amplifications, along with alterations MDA-MB-231 (triple-negative breast cancer), and MS751 in other Hippo pathway members, were mutually exclusive (cervical)] after YAP1/WWTR1 knockdown. Consistent with across samples of patients with cancer (Fig. 1B), suggesting the original seven YAP1-amplified cell lines, we observed these low-frequency mutations may function similarly to similar numbers of overlapping, significantly downregulated deregulate the Hippo pathway. genes in each of the three independent NF2-null cell lines As YAP is a transcriptional coactivator, the most frequently (Supplementary Fig. S1H). altered regulator of the Hippo pathway, and previously associ- Last, we performed Assay for Transposase-Accessible Chro- ated with treatment resistance (19, 20), we hypothesized that matin using sequencing (ATAC-seq) on Detroit 562 and its oncogenic potential must be mediated by its downstream PA-TU-8902, a pancreatic adenocarcinoma cancer cell line transcriptional target genes. We aimed to develop a first-prin- with TEAD4 amplification, and confirmed that Cluster 2 ciple approach to map a lineage-independent transcriptional genes were most strongly associated with loss of chromatin signature for Hippo deregulation. We first identified seven accessibility upon YAP1/WWTR1 knockdown (Fig. 1F; Sup- cell lines originating from different tissues, but all carry YAP1 plementary Fig. S1I and S1J). Taken together, Cluster 2 genes amplification (copy number, 6.29 ± 1.50) with markedYAP1 included the most well-known canonical Hippo pathway mRNA overexpression (Fig. 1C). We performed knockdown marker genes, mostly correlated with previously reported of YAP1 and its paralog WWTR1 and then performed RNA Hippo pathway activity genes, and were associated with loss sequencing (RNA-seq) on the parental and YAP1/WWTR1 of chromatin accessibility upon knockdown. knockdown lines. YAP1/WWTR1 knockdown resulted in broad transcriptomic deregulation in all cell lines (Fig. 1C). Machine-Learning Approach to Predict Although CTGF expression (a canonical Hippo pathway tar- YAP/WWTR1 Dependency and MAPK get gene) was significantly decreased in all cell lines (Fig. 1C), Pathway Combination there was no clear association between CTGF expression and Aberrant Hippo pathway signaling has been known to magnitude of global gene expression changes to a cell line’s drive oncogenesis in several cancer types, many of which sensitivity to YAP1/WWTR1 knockdown (Supplementary Fig. lack a known Hippo pathway genetic alteration. To better S1A and S1B). Taken together, this suggested that YAP/TAZ identify potential Hippo pathway–dependent populations, dependency may be more complex, which necessitates expand- we sought to predict YAP/TAZ dependency using the cluster ing beyond a single marker of pathway activity to capture the genes we have identified here. We identified and performed pathway dependency over different cell lineages. RNA-seq on a broader set of 42 cancer cell lines exhibiting a We performed an unbiased weighted correlation network spectrum of Hippo pathway activity. Next, we assessed each analysis (28, 29) among a consensus set of genes that were cell line’s sensitivity to YAP1/WWTR1 knockdown to train a

March 2021 CANCER DISCOVERY | 781

RESEARCH ARTICLE Pham et al. machine-learning (ML) computational model to predict YAP/ ency, we noted that cell lines with the largest magnitude of TAZ dependency given a cell line’s parental cluster gene decrease in Cluster 4 genes were also those that were most expression profile. The ensemble-based algorithm learned a dependent on YAP1/WWTR1 knockdown (Fig. 2G). Taken combination of gene expression values to predict the change together, this suggests that additional suppression of MAPK in viability after YAP1/WWTR1 knockdown (Fig. 2A). Cluster pathway may serve to further enhance therapeutic efficacy of 2 score was the most correlated to predicted dependency, a Hippo pathway inhibitor. further supporting that Cluster 2 is most proximal to aberrant Hippo pathway signaling, the primary driver of MEK Inhibitors in Combination with YAP/TAZ dependency (Supplementary Fig. S2A and Sup- YAP1 Knockdown Enhance Response in plementary Table S2). Given Cluster 2 included many novel YAP1-Amplified Cancer Cell Lines genes not reported in previous gene sets, we next bench- As Hippo pathway inhibitors are under active development, marked our gene set compared with previously published identifying clinically actionable combinations becomes an gene sets (21, 31). We observed that our gene set performed important next step in augmenting therapeutic response. To better, independent of algorithm or training data (Sup- determine whether MAPK is a uniquely actionable pathway plementary Fig. S2B and S2C; Methods). Although we see that cross-talks with Hippo pathway dysregulation, we under- that the known genes (e.g., CTGF, CYR61, etc.) have high took a chemical genetic screening approach. We screened a importance/weight in the ML model (Supplementary Fig. drug library of 487 small-molecule compounds in Detroit S2D), many of the genes not found in previous gene sets (21, 562 cells stably transfected with an inducible YAP1 shRNA. 31) were among the greatest importance/weight in the ML Detroit 562 cells are very sensitive to YAP1 knockdown alone, model’s predictive power (Supplementary Fig. S2E and Sup- so we decided to knock down only YAP1 (Supplementary Fig. plementary Table S3), including CCDC42EP1, TNFRSF12A S3A) in our screen to reduce any small hairpin–related RNA (32), and PHLDB2I (33). toxicity. We assessed whether addition of each compound to Certain tissue lineages and histologic cell types were sig- the doxycycline-induced knockdown of YAP1 had a greater nificantly associated with YAP/TAZ dependency, including effect on cell viability than the noninduced shYAP1 arm. hematologic cell lines that are predicted to be not dependent We observed that MEK and ERK inhibitors were among the on YAP1/WWTR1 knockdown (P value < 10−47), whereas mes- highest-scoring hits showing the largest impact on viability othelioma histologic subtype was among the most predicted in combination with YAP1 knockdown (adjusted P < 0.1), to be YAP/TAZ dependent (Fig. 2A and B). We then validated whereas broad-spectrum cytotoxic chemotherapies did not our ML model by selecting 12 additional cell lines to con- modulate the effect of YAP1 knockdown, suggesting abrogating firm the predicted YAP/TAZ dependency: 6 cell lines which MAPK signaling further sensitizes cells to YAP1 knockdown were predicted to be YAP/TAZ dependent and 6 predicted (Fig. 3A; Supplementary Fig. S3B–S3D; and Supplementary independent from a variety of different lineages (Fig. 2C–E; Table S5). We sought to expand this observation to a larger Supplementary Fig. S2F–S2J). This provides a landscape of panel of YAP1-amplified cell lines treated with several MEK YAP/TAZ dependency across cancer cell line models and ena- inhibitors including cobimetinib, selumetinib, and PD-901 bles nomination of cell line models as well as prioritization (Fig. 3B–E; Supplementary Fig. S3E and S3F). Although this of cancer indications that would potentially benefit from a sensitization was observed in Hippo pathway–deregulated, Hippo pathway inhibitor. YAP1-amplified cell lines, this combination did not show To functionally annotate gene clusters, we performed a sys- further sensitization in squamous cell cancer lines that lack tematic gene signature correlation analysis of the gene clusters Hippo pathway alterations (Fig. 3C). The combination of MEK with other previously published gene signatures (Supplemen- inhibition and YAP1 knockdown promoted caspase-mediated tary Table S4). As expected, Cluster 2 scores were highly cell death that was measured by increase in caspase 3/7 activ- correlated with the previously published YAP gene signature ity, which was reversed upon treatment with a pan-caspase (CORDENONSI_YAP_CONSERVED_SIGNATURE, Pearson inhibitor (QVD; Fig. 3F). Clonogenic assays also confirmed ρ: 0.91; Supplementary Fig. S2K). Cluster 1 and 3 scores were the cooperation of YAP1 knockdown and MEK inhibitors in associated with several proliferation-associated gene signa- all three YAP1-amplified cell lines (Fig. 3G and H). This com- tures (HALLMARK_MYC_TARGETS_V2, Pearson ρ: 0.95) bination is unlikely due to general toxicity, as both SK-N-FI (a or epithelial-to-mesenchymal transition (HALLMARK_ Hippo-independent model) and MCF10-A (a nonmalignant EPITHELIAL_MESENCHYMAL_TRANSITION, Pearson ρ: 0.87; breast epithelial model) did not show further sensitization Supplementary Fig. S2L and S2M), respectively, both of (Supplementary Fig. S3G). Importantly, inducible YAP1 deple- which have been previously implicated in aberrant Hippo tion in vivo and cobimetinib combination exhibited significant pathway signaling (16, 34–36). Interestingly, Cluster 4 was tumor regression in Detroit 562 xenograft model (Fig. 3I; Sup- strongly associated with a KRAS dependency gene signature plementary Fig. S3H). Taken together, this small-molecule drug (SINGH_KRAS_DEPENDENCY_SIGNATURE, Pearson ρ: 0.87; library screen provided an orthogonal validation of the role of Fig. 2F), and although previous reports have suggested YAP1 the MAPK pathway in YAP-dependent cancers. overexpression as a bypass mechanism to KRAS activation (19), this result suggests that the MAPK pathway may play FOSL1/AP1 Is a Common Node in MAPK and a role in the context of Hippo signaling. We hypothesized Hippo Pathways the other gene clusters may also be associated with Hippo To further assess the impact of cobimetinib and YAP1 signaling although not directly downstream. Beyond Cluster knockdown, in Detroit 562 cells, we performed RNA-seq 2 scores as the strongest single predictor of YAP/TAZ depend- and ATAC-seq comparing YAP1 knockdown, or cobimetinib

782 | CANCER DISCOVERY March 2021 AACRJournals.org

Machine-Learning Approach Predicts Hippo Pathway Dependency RESEARCH ARTICLE

A OVCAR-8 HCC1954 0.9 SKG-II

DetroitSF268 562 0.8

0.7 ES-2

0.6

0.5

0.4 RVH-421

Predicted YAP/TAZ dependency YAP/TAZ Predicted 0.3 Hematologic cancers B Mesothelioma

D 1.5 SNU-423 C Observed YAP/TAZ dependency P < 0.001 0.0 0.2 0.4 0.6 0.8 1.0 1.0 SNU-423

NCl-H2373 . Rel RLU (%) PK-59 0.5 BICR 31 NCl-H2369 Cell viability HCC2450 0.0 siNTC siTox siYAP1 + WWTR1 Predicted dependent E 1.5 0.0 0.2 0.4 0.6 0.8 1.0 UMC-11

NB(TU)1-10 P < 0.001 1.0 SK-N-FI

G-361 . Rel RLU (%) UMC-11 0.5 HCSC-1 BEN Cell viability Predicted not dependent 0.0 siNTC siTox siYAP1 + WWTR1

FGLog2 fold change YAP1/WWTR1 KD sensitivity Cluster 4 0.20.4 0.60.8 1.00–.2–.4 –.6–.8 –1 1 OVCAR-8 0 Detroit 562

–1 SKG-II

_SIGNATURE HCC1954 –2 SF268

SINGH_KRAS_DEPENDENCY ES-2 –1.5 –1.0 –0.5 0.0 0.5 1.0 1.5 Cluster 4 RVH-421

Figure 2. ML approach defines Hippo pathway dependency and proposes parallel pathways for combination strategy. A, Landscape of predicted YAP/ TAZ dependency across cell lines based on parental RNA expression. B, Mesothelioma cell lines predicted to be significantly YAP/TAZ dependent, and hematologic cell lines predicted to be not dependent on YAP/TAZ. C, Validation of ML model prediction of YAP/TAZ dependency in 12 cell line models. D and E, Effect on cell viability of two cell lines from predicted dependent group and predicted not dependent group upon YAP1/WWTR1 knockdown. F, Cluster 4 score is significantly associated with a KRAS dependency signature. G, Average change in expression of Cluster 4 genes is an independent predictor of YAP/TAZ dependency.

March 2021 CANCER DISCOVERY | 783

RESEARCH ARTICLE Pham et al.

A YAP_plus_minus_DOX

0.2 0.1 0.0 –0.1 –0.2 MeanViabDiff 0 100 200 300 400 500

InhibCat ERK inhibitor MEK inhibitor PI3K inhibitor RAF inhibitor

B C Detroit 562 Detroit 562 COLO-680N COLO 680N 1.2 1.2 0.6 HEp-2 BICR10 T.T 0.9 0.9 0.4

0.6 0.6 0.2

Cell viability siNTC siNTC Cell viability 0.3 siYAP1 0.3 siYAP1 Mean differences (fraction of control) (fraction of control)

0.0 0.0 (fraction of cell viability) 0.0 0.0001 0.001 0.01 0.1 110 0.0001 0.001 0.01 0.1 110 0.0001 0.001 0.01 0.1 1 Cobimetinib (µmol/L) Cobimetinib (µmol/L) Cobimetinib (µmol/L)

DE F Selumetinib 10 1.2 Selumetinib/PD-901 0.5 PD-901 Detroit 562 Gemcitabine 8 COLO-680N 0.4 0.9 6 0.3 0.6 4 0.2 (fold increase) (fold Cell viability 2 0.3 siNTC Caspase 3/7 activity

Mean differences 0.1 (fraction of control) siYAP1

(fraction of cell viability) 0 0.0 0.0 0.001 0.01 0.1 11030 0.0001 0.001 0.01 0.1 1 Concentration (µmol/L) Concentration (µmol/L) siNTC/CobisiYAP/Cobi siNTC/CobisiYAP/Cobi siNTC/DMSOsiYAP/DMSO siNTC/DMSOsiYAP/DMSO siYAP/Cobi/QVD siYAP/Cobi/QVD

G I shYAP1/Sucrose DMSO Doxycycline Cobimetinib AZD6244 Dox/Cobi Dox/AZD 800 shYAP1/Doxycycline Cobimetinib 7.5 mg/kg shYAP1/Doxycycline + Cobimetinib 7.5 mg/kg 700 Detroit 562 )

shYAP1 3 600 500

Detroit 562 400 lume (mm P < 0.0001 shNTC 300 mor vo

Tu 200 P < 0.01 100 H siYAP1/ siYAP1/ siNTC/DM siNTC/Cobi DM Cobi 0 0 10 20 30 Time (days) COLO680N

HEp-2

784 | CANCER DISCOVERY March 2021 AACRJournals.org

Machine-Learning Approach Predicts Hippo Pathway Dependency RESEARCH ARTICLE treatment, and the combination of YAP1 knockdown and we observed a significant decrease in AP1/FOSL1 target gene cobimetinib treatment with the control treatment. Together, expression in the combination treatment (Supplementary the combination treatment significantly decreased expres- Fig. S4B) compared with YAP1 knockdown or cobimetinib sion of proliferation genes (Fig. 4A) compared with each treatment alone, suggesting the combination may further individual treatment, consistent with in vitro and in vivo contribute to loss of FOSL1 activity. As previous studies have observations (Fig. 3G–I; Supplementary Fig. S3C–S3F). Fur- suggested that TEADs (38) may have numerous dimeriza- thermore, Cluster 2 genes were downregulated in the YAP1 tion partners, we hypothesized that FOSL1 may interact with knockdown and the combination treatment but not in cobi- TEADs (33, 37). Consistent with previous reports, we confirm metinib treatment alone (Fig. 4A). Conversely, MAPK path- that TEAD directly interacts with FOSL1 (Supplementary Fig. way genes were downregulated upon MEK inhibition and S4C and S4D) through coimmunoprecipitation. Furthermore, combination treatment but not after YAP1 knockdown alone YAP1 is required for FOSL1–TEAD interaction, whereas YAP– (Fig. 4A). Although neither Hippo nor MAPK pathway genes TEAD interaction is independent of FOSL1 (Supplementary were further suppressed upon the combination treatment, we Fig. S4C). Consistent with these observations, FOSL1–TEAD hypothesized that YAP1 knockdown and cobimetinib jointly interaction is abolished upon cobimetinib treatment (Sup- affect a common node rather than further downregulating plementary Fig. S4D), and deletion of FOSL1, together with their individual pathways. We noted an overlapping 3,479 YAP1, mimics the synergistic effect observed with cobimetinib peaks exhibiting loss of chromatin accessibility in the combi- (Supplementary Fig. S4E and S4F). nation of YAP1 knockdown and cobimetinib treatment were also found in the single treatments alone (Fig. 4B). Motif MEK Inhibition and YAP1 Knockdown Decrease enrichment analysis revealed decreased chromatin accessibil- TEAD Protein Half-life ity at TEAD- and AP1-binding sites upon YAP1 depletion and Given that the combination of YAP1 depletion and cobi- MEK inhibition (Fig. 4C), respectively. Although enrichment metinib treatment affect TEAD interaction partners, we hypo was significant upon MEK inhibition, the combination treat- thesized that modulating these two interaction partners may ment exhibited greater significance of AP1 motif in peaks affect TEAD protein stability. Although neither YAP1 deple- with loss of chromatin accessibility (Fig. 4C). Previous stud- tion nor cobimetinib treatment alone changed TEAD protein ies have shown that TEADs and AP1 can coregulate gene levels (Fig. 5A and B), we observed a significant decrease in transcription through changes in enhancer and promoter TEAD protein half-life upon the combination of cobimetinib regions (33, 37). Taken together, these findings suggest that and YAP1 depletion following cycloheximide (CHX) chase concomitant YAP1 depletion and MEK inhibition serve to in YAP1-amplified cell lines (Fig. 5A and B; Supplementary further enhance the loss of AP1-binding sites through Hippo Fig. S5A), whereas TEAD transcript levels were unaltered and MAPK pathways, respectively. (Supplementary Fig. S5B). Decrease in TEAD protein half- As cobimetinib inhibitors affect MEK kinase activity, we life was reversed upon MG132 treatment 24 hours after CHX performed global phosphoproteomics analysis in Detroit 562 treatment (Fig. 5C and D; Supplementary Fig. S5C), suggesting to identify changes in phosphorylation sites across proteins that the decrease in TEAD half-life is mediated by proteasomal upon YAP1 depletion and/or MEK inhibition. This identi- degradation. These data imply that the combination of YAP1 fied significant changes in 18,800 phosphopeptides across depletion and cobimetinib treatment results in a decrease in 8,500 proteins. Consistent with the enriched AP1 motif in TEAD stability. Furthermore, we noted that several prolifera- peaks with loss of chromatin accessibility, we noted 2-fold tion genes such as MYC and FOSL2 (Fig. 5E and F) have nearby decrease in FOSL1 phosphorylation (Fig. 4D–F; Supplemen- TEAD- and AP1-binding sites with the potential to regulate tary Fig. S4A; Supplementary Tables S6 and S7). In addition, their expression. Together, our findings suggest that the

Figure 3. MEK inhibitors sensitize with YAP1 knockdown in YAP1-amplified cancer cell lines. A, A small-molecule library of 487 tool compounds was tested in Detroit 562 cells stably transfected with a doxycycline-inducible construct of YAP1 shRNA. The y axis represents differences in mean viability over a 9-point dose–response curve in either the presence or absence of doxycycline. Compounds are ranked from largest decrease in mean viability upon YAP1 knockdown to increase in viability upon YAP1 knockdown (antagonism). Several inhibitor classes are highlighted as indicated by colored circles. B, YAP1 knockdown and cobimetinib cooperatively inhibit cell growth in YAP1-amplified cancer cell lines. Two representativeYAP1 -amplified cell lines (Detroit 562 or COLO-680N) were transfected either with siNTC or siYAP1 followed by treatment with cobimetinib at various indicated concentrations for 48 hours after siRNA transfection. Cell viability was assessed at day 3 after treatment using Cell Titer-Glo [mean ± SD (n = 3) for each concentration point]. C, Cobimetinib selectively enhances cell growth inhibition effect of siYAP1 in YAP1-amplified cancer cells.YAP1- amplified cancer cells (red) and cancer cells without YAP1 alternation (blue) were cotreated with siYAP1 and cobimetinib. The mean differences between siNTC and siYAP1 at various concentrations of cobimetinib were assessed (n = 6). Note that the slopes of the mean differences in YAP1-amplified lines are positive, showing synergistic interaction of the combination. D, YAP1 knockdown and MEK inhibitors including selumetinib and PD-901 cooperatively inhibit cell growth in YAP1- amplified cancer cells. Detroit 562 cells were transfected with siNTC or siYAP1 in combination with MEK inhibitors at various concentrations. siRNA transfection was done 48 hours prior to drug treatment. Cell viability was assessed at day 3 after treatment using Cell Titer-Glo. Error bars, mean ± SD (n = 3). E, MEK inhibition selectively enhances cell growth inhibition effect of YAP1 knockdown in YAP1-amplified cancer cells. Detroit 562 cells were transfected with siYAP1 in the presence of MEK inhibitors or gemcitabine. Then, the mean differences between siNTC and siYAP1 at various concentrations of drugs were assessed (n = 6). F, siYAP1 and cobimetinib combination induces caspase activation. Detroit 562 cells were treated with either siYAP1 or cobimetinib (0.5 μmol/L) or in combination. Caspase 3/7 activity was assessed using Caspase-Glo 3/7. Error bars, mean ± SD (n = 3). G and H, Combina- tion of YAP1 knockdown and MEK inhibition significantly affects colony formation inYAP1 -amplified cancer cell lines. Detroit 562/shNTC and Detroit 562/shYAP1 were plated and subjected to MEK inhibitor treatment as previously described. Colony formation was then assessed with crystal violet stain (n = 3). The same was done for the other YAP1-amplified cell line (COLO680N and HEp-2;n = 3). I, Doxycycline-induced YAP1 knockdown inhibits Detroit 562 tumor growth. Detroit 562/shYAP1 cells were s.c. injected into the mice to allow tumor establishment. The total number of mice was then divided into group of n = 10 for sucrose, doxycycline to induce shYAP1, cobimetinib (7.5 mg/kg), and combination of doxycycline and cobimetinib. Tumor volume was measured and compared, and statistical analysis was done using unpaired two-sided Student t test.

March 2021 CANCER DISCOVERY | 785

RESEARCH ARTICLE Pham et al.

A Proliferation Cluster 2 MAPK activity B n = 2,117 0.0 0.4 0 n = 95 –0.4 0.2 –1 n = 3,479 0.0 –0.8

fold change fold –2 Cobimetinib 2 –0.2 –1.2 siYAP1 Log siYAP1 + Cobimetinib –0.4 –3 n = 5,595 Shared peaks –1.6

CobimetinibsiYAP1 siYAP1 + Cobimetinib

C 1e–288 1e–71 1e–253 D >30 Digest proteins/ Mix peptides/ Global proteome/ Samples label peptides fractionation PTM analysis 25 siNTC DMSO (2) ~ 8,800 proteins siNTC Cobi (2) P value) 20 ~ 19,000 phospho sites

10 siYAP DMSO (2) filtered by nuclear proteins siYAP Cobi (2) 15

10 E Proteome Phosphoproteome

5 3 3 Significance (–log 0 Not significant Not significant 2 2

1 FOSL 1 ERF 1 RAI1 TEAD-binding motif AP1-binding motif BCL9L KMT2D NCOR2 P1 0 KMT2D 0 SETX NCOR2 RAI1 BCL9L BCL3

Cobimetinib si YA ARNTL BCL3 –1 –1 KMT2D siYAP1 TSC22D4 FOSL 1 ratio (Cobi/DMSO)

2 SETX ERF –2 MKL2 –2 ARNTL siYAP1 + Cobimetinib TSC22D4

Log ERF F –3 –3 KMT2D BCL9L MKL2 FOSL1 protein quantitation FOSL1 s265 quantitation –3 –2 –1 0123 –3 –2 –1 0123 16 16 2.1× 1.5× Log2 ratio (Cobi/DMSO) Log2 ratio (Cobi/DMSO) 12 12 siNTC siNTC 8 8 4 4 0 0 Relative abundance Relative

siNTC + Cobi siNTC + Cobi YAP1 + DMSOsiYAP1 + Cobi YAP1 + DMSOsiYAP1 + Cobi siNTC + DMSO si siNTC + DMSO si

Figure 4. AP1/FOSL1 is a common node of MAPK and Hippo pathway activity. A, Log2 fold change of cobimetinib, YAP1 knockdown, or the combination in Cluster 2, MAPK, and proliferation gene expression. Combination does not further downregulate Hippo or MAPK pathway genes but enhances loss of proliferation. B, 3,479 peaks share loss of chromatin accessibility in the combination, cobimetinib, or YAP1 knockdown alone. C, HOMER motif analysis identified TEAD motif inYAP1 knockdown alone or in combination with MEK inhibition. AP1-binding site motif was more significant in the combination treatment as compared with YAP1 knockdown or cobimetinib alone. D, Schematic for global phosphoproteomic analyses for each condition is indicated. Replicates (n = 2) were used for each condition, and standard proteomic pipeline followed for analyses. PTM analysis, post-translational modification analysis by mass spectrometry. E and F, Combination of MEK inhibition and YAP1 knockdown decreases phosphorylation of FOSL1 TF. Cells were treated with cobimetinib for 1 hour before harvest. Relative abundance of total and phosphorylated FOSL1 in the indicated conditions and as determined by global proteomic analyses is shown.

convergence of YAP1 depletion and treatment with cobi- lineage-independent predictive Hippo pathway signature metinib is mediated through the cooperative interaction (Fig. 2) and nominated the MAPK pathway as a potential focus between AP1/FOSL1 and TEAD. for drug combinations that was orthogonally identified and confirmed through a small-molecule drug screen (Fig. 3). Further investigation revealed a novel mechanism in which Discussion both Hippo and MAPK pathways regulated TEAD function In this study, we developed an ML approach to understand through decreasing its stability with observed loss of chro- Hippo pathway activity (Fig. 1). This has identified a robust, matin accessibility at TEAD-binding motifs (Figs. 4 and 5).

786 | CANCER DISCOVERY March 2021 AACRJournals.org

Machine-Learning Approach Predicts Hippo Pathway Dependency RESEARCH ARTICLE

A C crCD81/DMSO crCD81/Cobi crYAP1/DMSO crYAP1/Cobi + + + + + + − − − − − − siNTC − − − − − − + + + + + + siYAP1

4 8 12 24 0 4 8 12 24 0 4 8 12 24 0 4 8 12 24 (CHX, hr) − − − + + + − − − + + + Cobi − + + − + + − + + − + + CHX Pan-TEAD (L.E.) − − + − − + − − + − − + MG132 Pan-TEAD Pan-TEAD (H.E.) YAP pERK pERK ERK β-actin YAP Detroit 562 β-actin

Detroit 562 D + + + + + + − − − − − − crCD81 − − − − − − + + + + + + crYAP1 − − − + + + − − − + + + Cobi − + + − + + − + + − + + CHX B − − + − − + − − + − − + MG132 Pan-TEAD (L.E.) Pan-TEAD (L.E.)

Pan-TEAD (H.E.) Pan-TEAD (H.E.)

pERK pERK

ERK ERK

YAP YAP

β-actin β-actin OVCAR-8 OVCAR-8 F E AP1TEAD TEADAP1

siNTC/DMSO siNTC/DMSO

siNTC/Cobi siNTC/Cobi

siYAP/DMSO siYAP/DMSO

siYAP/Cobi siYAP/Cobi

MYC FOSL2 G NF2/merlin

MST1/2 MAP4Ks Machine SAV1 learning TAOKs P P P P P YAP/TAZ P LATS1/2 P MOB1 14-3-3 P P YAP/TAZ YAP/TAZ P P

Predict sensitivity YAP/TAZ Cobi YAP/TAZ P FOSL1 MEK TEAD FOSL1 TEAD

High chromatin access Low chromatin access

Hippo transcriptional activity

Figure 5. Combination of MEK inhibition and YAP1 knockdown decreases TEAD protein half-life. A and B, Decrease in TEAD half-life by combination of cobimetinib and YAP1 knockdown. Indicated YAP1-amplified cells were treated withYAP1 guide RNA for 72 hours in the absence or presence of MEK inhibitors, cobimetinib (0.5 μmol/L) for 48 hours (n = 3). Then, CHX (100 μg/mL) was added at indicated times followed by immunoblotting with indicated antibodies. L.E., low exposure; H.E., high exposure. C and D, YAP1 knockdown/cobimetinib reduces TEAD half-life, which is rescued by MG132 treatment. YAP1-amplified cells were treated with siYAP1/YAP1 guide RNA or cobimetinib (0.5 μmol/L) or in combination for 48 hours. Note that 100 μg/mL of CHX or 10 μmol/L of MG132 was added 8 hours before cell harvest. Expression of indicated proteins was assessed by immunoblotting. β-Tubulin served as a loading control (n = 3). E and F, Examples of genomic region at the MYC and FOSL2 loci showing decreased accessibility of sites containing either the AP1 (blue arrow) or TEAD (red arrow) motifs with both siYAP1 and treatment with cobimetinib. G, Proposed model for integrated ML/experimental workflow.

March 2021 CANCER DISCOVERY | 787

RESEARCH ARTICLE Pham et al.

The Hippo pathway is emerging as an important area for was generated by transfecting in the shYAP1-pLKO lentiviral vector targeted drug discovery efforts, but greater understanding of and selecting for Puromycin-positive cells. Cell line authentication this pathway is warranted. As opposed to previous efforts that was conducted for short tandem repeat (STR) profiling using the have derived a curated list of Hippo pathway target genes, here Promega PowerPlex 16 System. This is performed once when receiv- we utilized an ML approach to systematically define four core ing a new cell line and compared with external STR profiles of cell lines (when available) to determine cell line ancestry. Cell line authen- target gene clusters that are altered as a direct result of loss of tication was routinely conducted by SNP-based genotyping using Hippo signaling. This lineage-independent approach identi- Fluidigm-multiplexed assays at the Genentech cell line core facility. fied many novel genes not reported in previous YAP/TAZ gene Antibodies used in this study are Pan-TEAD [13295, Cell Signal- sets. This has enabled accurate prediction of YAP/TAZ depend- ing Technology (CST)]; YAP (14074, CST); p44/42 MAPK (ERK1/2, ency in vitro and yielded a signature that can be used to define 4696, CST); Phospho-p44/42 MAPK (p-Erk1/2, 4370, CST); TAZ and prioritize cell line models and patient populations. We (70148, CST); MAX (sc-765 and sc-8011, Santa Cruz Biotechnology); validated our gene set to be robust across different genotypes α-tubulin (3873, CST); β-actin (3700, CST); cleaved PARP (9541, CST); and cancer types. To our knowledge, this is the first lineage- FRA1(5281, CST); Myc-tag (2278, CST); V5 (680602, Biolegend); independent, unbiased method that is predictive of Hippo anti-rabbit and anti-mouse horseradish peroxidase linked (7074 and pathway dependency and thus can serve as a valuable tool to 7076, CST), and IRDye anti-rabbit and anti-mouse (68070 and 32211, LI-COR). Cobimetinib was synthesized at Genentech. Selumetinib identify biomarkers of interest in a tumor-agnostic manner. (S1008), PD0325901 (S1036), and gemcitabine (S1714) were pur- We then uncovered the molecular mechanism underpin- chased from Selleckchem. CHX solution (C4859) and MG132 solu- ning the effects of combined inhibition of MAPK and Hippo tion (M7449) were purchased from Sigma-Aldrich. pathways through several orthogonal technologies. We performed ATAC-seq after YAP1 depletion and/or cobimetinib Cell Viability Assay and Colony Growth treatment; results here suggested that the combination con- Cells were seeded at 500 to 1,000/well on a 96-well plate for 24 verges on the loss of chromatin accessibility at AP1- and hours. They were then treated with indicated siRNAs (final concen- TEAD-binding sites. The combination of modulating both tration of 25 nmol/L) or indicated inhibitors (with indicated concen- TEAD interaction partners, YAP and FOSL1 (via MEK inhibi- tration). Cell growth was assessed using Cell Titer-Glo Luminescent tion), resulted in decreased TEAD protein stability (Fig. 5G). Cell Viability Assay (Promega). All cell viability data were collected Given previous studies have suggested differential regulation and calculated for at least six replicates per time point, per condition. of YAP and TAZ (39, 40), future studies will be necessary to IC50 for the inhibitors was determined by fitting the nonlinear regres- elucidate these mechanistic differences, if any, in the context sion curve generated by GraphPad Prism. Cells were seeded at 50,000 to 70,000/well on a 6-well plate for of MAPK pathway inhibition. 24 hours. They were then treated with indicated siRNAs (final con- Last, this approach serves as a framework for systematic centration of 25 nmol/L) or indicated inhibitors (with indicated characterization of signaling pathways. Here, we focused concentration) for 6 to 10 days. Colony formation was then accessed on the Hippo pathway in the context of combinations with with crystal violet stain. MAPK inhibitors. Earlier work has shown the cooperative activity of YAP/TAZ/TEAD and AP1 at many enhancers Caspase 3/7 Activation Assay and promoters (33, 37) as well as a role for Hippo pathway In a 96-well plate, cells were treated with indicated siRNAs, cobi- inhibitors to combat resistance to BRAFV600E inhibition (20). metinib, or Q-VD-OPH, a pan-caspase inhibitor (treatment can be As MEK inhibitors are currently in the clinic and Hippo alone or in the indicated various combinations). Caspase 3/7 activa- pathway inhibitors are in development, our studies suggest tion was measured using Caspase-Glo 3/7 assay reagent containing that cotargeting these pathways may achieve a deeper thera- proluminescent caspase 3/7 substrate, tetrapeptide sequence DEVD peutic response compared with single-agent treatment alone (Promega). The amount of luminescence is displayed as fold changes in Hippo pathway–dependent cancers. Our unbiased charac- of treatments to siNTC/DMSO-treated control. terization through a lineage-independent approach to study Chemical Genetics Screen Hippo pathway activity and dependency underscores the importance of understanding pathway cross-talk as a strategy A library comprising 485 compounds including targeted agents, to nominate potential treatment combinations (Fig. 5G). Our chemotherapeutics, and tool compounds was used to screen for inhibitors that exhibit enhanced efficacy in the context ofYAP1 findings here have translational impact not only on Hippo- knockdown. Compounds were obtained from in-house synthesis dependent cancers, but also on tumors with MAPK pathway or purchased from commercial vendors and managed as previously alterations in primary as well as resistance settings (20, 41). described (1). Detroit 562 cells harboring a doxycycline-inducible In summary, our study has made several significant contri- shRNA targeting YAP1 were treated with and without doxycycline butions to understanding the Hippo pathway, and in addi- for 72 hours prior to seeding into 384-well plates (BD Falcon) at a tion, we developed an approach to identify possible pathway density of 1,000 cells per well. Cells were maintained at 37°C with targets. This computational–experimental study defines a 5% CO2 and ± doxycycline throughout the course of the experiment. framework to establish a new paradigm to apply the ML Twenty-four hours after seeding, cells were treated with a nine-point toolbox to accelerate biology and drug development efforts. dose titration of each compound or DMSO control (0.1%) for 96 hours. Cell viability was then assessed using CellTiter-Glo (Promega).

IC50 (concentration yielding 50% reduction in viability) and mean via- Methods bility (roughly equivalent to the area under the dose–response curve; ref. 42) values were determined by fitting curves using Genedata Cell Lines, Antibodies, and Other Reagents Screener software (Genedata). Compounds exhibiting more activity Cell lines used in this study (Detroit 562, COLO-680N, HEp-2, and in the context of YAP1 knockdown were determined by calculating

OVCAR-8) were obtained from the ATCC. Detroit 562 with shYAP1 both the difference in mean viability and fold change in IC50 between

788 | CANCER DISCOVERY March 2021 AACRJournals.org

Machine-Learning Approach Predicts Hippo Pathway Dependency RESEARCH ARTICLE doxycycline-treated and nontreated arms. Screening results can be ACN into 96 fractions. The fractions were concatenated into either found in Supplementary Table S3. 24 or 12 samples for proteome and phosphoproteome samples, respectively. Fractions were concatenated by mixing different parts Animal Work of the gradient to produce samples that would be orthogonal to For tumor xenograft models, Detroit shYAP1 cells (4.25 × 106) downstream low pH reverse-phase LC-MS/MS. Samples were dried, were injected s.c. into right thoracic regions of 6- to 10-week-old Nu/ desalted by SPE, and dried again. Nu (nude-CRL) mice. When tumors reach a mean volume between 150 and 300 mm3, mice were then grouped into 10 per group for Quantitative Mass Spectrometry and Data Analysis further intervention treatments including doxycycline (1 μg/μL), NanoLC-MS/MS analysis was performed on an Orbitrap Fusion cobimetinib (7.5 mg/kg), and combination of doxycycline and cobi- Lumos mass spectrometer (Thermo Fisher Scientific) coupled to metinib. Tumor volume was collected and calculated after 3 days a Dionex Ultimate 3000 RSLC-nano (Thermo Fisher Scientific). and up to 28 days. A small group of mice was euthanized at 6 weeks Peptides were separated over a 100 μm × 250 mm PicoFrit column after treatment for immunoblotting. All animal experiments were (New Objective) packed with 1.7 μmol/L BEH-130 C18 (Waters) at approved by Genentech Animal Care and Use Committee. a flow rate of 450 nL/min for a total run time of 180 minutes. The gradient spanned from 2% Buffer B (0.1% FA/98% ACN/2% water) to Immunoprecipitation and Immunoblotting 30% B over 155 minutes and then to 50% B at 160 minutes. For mass 1 Cells were lysed in RIPA lysis buffer (89900, Thermo Fisher Scien- spectrometry analysis, peptides were surveyed within FTMS analyses 6 tific) containing protease inhibitor (Roche) and phosphatase inhibi- [120,000 resolution, AGC = 1 × 10 , and maximum injection time tor (Roche). Lysates were prepared by taking supernatants from (max IT) = 50 ms], and the top 10 peaks were selected for MS/MS, centrifugation at 12,000 × g for 15 minutes at 4°C. Equivalent ensuring that any given peak was selected only once in a 45-second amounts of proteins were loaded and separated by SDS-PAGE fol- window (ppm tolerance = 5 ppm). For dynamic exclusion, the “one lowed by transferring to membranes. precursor per charge state” was ON for proteome analysis and OFF 2 For endogenous coimmunoprecipitation experiments, 1 × 107 cells for phosphoproteome analysis. For MS analysis, precursors were were lysed using RIPA buffer (Thermo) and immunoprecipitation isolated using the quadrupole (0.5 Th window), fragmented using 4 with indicated antibody overnight at 4°C. After washing with RIPA CAD (NCE = 35, AGC = 2 × 10 , and max IT = 100 ms), and analyzed 2 buffer (Thermo), coimmunoprecipitated endogenous proteins were in the ion trap (scan speed = Turbo). Following MS analysis, the then detected by immunoblotting. top 8 (proteome) or 6 (phosphoproteome) ions were simultane- ously selected [synchronous precursor selection – SPS, AGC = 2.5 × 5 5 Global Proteome and Phosphoproteome 10 (proteome) or 3.0 × 10 (phosphoproteome), and max IT = 150 Sample Preparation (proteome) or 200 (phosphoproteome) ms] and fragmented by HCD (NCE = 55) before analysis in the Orbitrap (resolution = 50,000). Raw Cell pellets were lysed in 8 mol/L urea, 150 mmol/L NaCl, 50 data files are available via the MASSIVE data repository using the mmol/L HEPES (pH 7.2), complete-mini (EDTA free) protease inhibi- identifier MSV000085111. tor (Roche), PhosSTOP phosphatase inhibitor (Roche), 1 mmol/L All mass spectrometry data were searched using Mascot against a Na-orthovanadate, 2.5 mmol/L Na-pyrophosphate, and 1 mmol/L concatenated target-decoy human database (downloaded June 2016) beta-glycerol-phosphate by 15× passages through a 21 g needle. containing common contaminant sequences. For the database search, Protein concentrations were then estimated by BCA assay (Thermo a precursor mass tolerance of 50 ppm, fragment ion tolerance of 0.8 Fisher Pierce), and 2.5 mg of protein material was used for further Da, and up to 2 missed cleavages were used. For global proteome and sample preparation. Disulfide bonds were reduced by incubation phosphoproteome analysis, carbamidomethyl cysteine (+57.0214), and with 5 mmol/L DTT (45 minutes at 37°C), followed by alkylation TMT-labeled n-terminus and lysine (+229.1629) were applied as static of cysteine residues by 15 mmol/L IAA (30 minutes, RT Dark), and modifications, whereas methionine oxidation +( 15.9949) was set as a finally capped by the addition of 5 mmol/L DTT (15 minutes, RT dynamic modification. For phosphoproteome analysis, TMT-labeled Dark). Proteins were then precipitated by chloroform/methanol pre- tyrosine (+229.1629) and phosphorylation of serine, threonine, and cipitation and resuspended in digestion buffer (8 mol/L urea, 150 tyrosine (+79.9663) were also set as dynamic modifications. Peptide mmol/L NaCl, and 50 mmol/L HEPES, pH 7.2). Initial protein diges- spectral matches for each run were filtered using line discriminant tion was performed by the addition of 1:100 LysC followed by incuba- analysis to an FDR of 2% and subsequently as an aggregate to a protein tion at 37°C for 3 hours. Samples were then diluted to 1.5 mol/L urea level FDR of 2%. Localization of phosphorylation sites was performed with 50 mmol/L HEPES (pH 7.2) before the addition of 1:50 Trypsin using a modified version of the Ascore algorithm. TMT-MS3 quantifi- and incubation overnight at 37°C. Peptide mixtures were acidified cation was performed using Mojave, with only those peptide-spectrum and desalted via solid-phase extraction (SPE; SepPak–Waters). matches (PSM) possessing isolation specificities greater than or equal Before phosphopeptide enrichment, a 100 μg aliquot of pep- to 0.5 considered for the final dataset. Intensities of each PSM were tides was saved for global proteome analysis. Phosphopeptides were added to the peptide and then protein (proteome) or phosphoisoform enriched from the remaining material utilizing iron-IMAC–based (phosphoproteome) level. Expression is reported as relative abundance, magnetic beads as previously described (43). To enable multiplexed which is the measured intensity of any given channel divided by the quantitation, both the phosphopeptide-enriched and unenriched total intensity for that protein or phosphoisoform. peptide pools were resuspended in 200 mmol/L HEPES (pH 8.5) and mixed with tandem mass tags (TMT; Thermo Fisher Pierce) at a label to protein ratio of 2:1. After 1 hour of labeling, the reaction RNA Extraction, cDNA Synthesis, and Quantitative RT-PCR was quenched by the addition of 5% hydroxylamine and incubated at Tumors and cell lines were dissociated and lysed for RNA isola- room temperature for 15 minutes. Labeled peptides were then mixed, tion using RNAeasy Mini Kit (QIAGEN) with the on-column DNA acidified, and purified by SPE. elimination. cDNA was prepared by reverse transcription using the Labeled samples were separated by offline high pH reverse-phase iScript cDNA Synthesis Kit (Bio-Rad) as the manufacturer’s protocol. fractionation using an ammonium formate–based buffer system The quantitative RT-PCR was performed using QuantStudio 7 Flex delivered by a 1100 HPLC system (Agilent). Peptides were separated machine with TaqMan probes for YAP1 (Hs00902712_g1), WWTR1 over a 2.1 × 150 mm, 3.5 μm 300Extend-C18 Zorbax column (Agi- (Hs00210007_m1), TEAD1 (Hs00173359_m1), TEAD2 (Hs01055894_ lent) and separated over a 75-minute gradient from 5% ACN to 85% m1), TEAD3 (Hs00243231_m1), TEAD4 (Hs01125032_m1), GAPDH

March 2021 CANCER DISCOVERY | 789

RESEARCH ARTICLE Pham et al.

(Hs02786624_g1), CTCF (Hs00902016_m1), CYR61 (Hs00155479_ nity standard) generated on the basis of the start positions of the m1), CPA4 (Hs00275311_m1), THBS1 (Hs00962908_m1), PTRF mapped reads. Accessible peak locations were identified as described: (Hs00396859_m1), EPHA2 (Hs01072272_m1), EDN1(Hs00174961_ in brief, we called peaks on a group-level pooled sample containing m1), TSPAN3 (Hs00170681_m1), NDRG1 (Hs00608387_m1), and all pseudo-fragments observed in all samples within each group. EMP2 (Hs00171315_m1; Applied Biosystems). Relative expression of Peaks in the pooled sample that were independently identified in each gene to GAPDH of target genes was assessed for at least two to two or more of the constituent biological replicates were retained for three biological replicates. downstream analysis, using the union of all group-level reproduc- ible peaks (https://www.encodeproject.org/atac-seq/#standards). We RNA Library Preparation and Sequencing quantified the level of chromatin accessibility within each peak for Total RNA was extracted using the QIAGEN RNAeasy Mini Kit each replicate as the number of pseudo-fragments that overlapped (QIAGEN) with the on-column DNA elimination. The concentration the peak in question and normalized these estimates using the was identified using NanoDrop 8000 (Thermo Fisher Scientific). Trimmed Mean of M-values (TMM) method (50). We identified Quality control was done by determining RNA integrity with Bioana- differentially accessible peaks between groups in the framework of lyzer 2100 (Agilent Technologies). About 500 ng of RNA was used for a linear model implemented with the limma R package (51) and library synthesis using the TrueSeq RNA Sample Preparation Kit v2 incorporating precision weights calculated with the voom function (Illumina). Size of the libraries was confirmed using 2200 TapeSta- in the limma R package (49). We identified enriched TF motifs using tion and High Sensitivity D1K screen tape (Agilent Technologies), HOMER v4.7 (52). To evaluate the significance of the TF enrich- and the concentration was determined by a qPCR-based method ment, we defined peaks as significantly differentially accessible based using the Library Quantification Kit (KAPA). The libraries were mul- on a range of FDR-adjusted P value thresholds between 1 and 0.01 tiplexed and then sequenced on Illumina HiSeq2500 (Illumina) to and a |log2[fold change]| in accessibility ≥ 1 (estimated from the generate 30M of single-end 50-base pair reads. model coefficients). Given the strong enrichment of the top motifs The RNA-seq data have been deposited in the Gene Expression across a wide range of P-value cutoffs, we decided to consider peaks Omnibus (GEO) with the accession code GSE161019. Data or as different across groups for a |log2[fold change]| ≥ 1 and FDR- other materials are available from the corresponding authors upon adjusted P value ≤ 0.05 in subsequent analyses. request. The ATAC-seq data have been deposited in the GEO with the accession code GSE161019. Data or other materials are available from the corresponding authors upon request. RNA-seq Alignment and Feature Counting For RNA-seq data analysis, RNA-seq reads were first aligned to Hippo Pathway Cluster Analysis ribosomal RNA sequences to remove ribosomal reads. The remaining Statistically significant downregulated genes in three or more of the reads were aligned to the mouse reference genome (GRCm38) using seven cell lines were considered for common downregulated genes. GSNAP (44, 45) version “2013-10-10,” allowing a maximum of two Genes were subsequently filtered based on their expression, and genes mismatches per 75 base sequence (parameters: -M 2 -n 10 -B 2 -i 1 that were upregulated in any cell line upon YAP1/WWTR1 knockdown -N 1 -w 200000 -E 1 –pairmax-rna = 200000 –clip-overlap; refs. 44, were removed. Signed coexpression networks were built using the 46). Transcript annotation was based on the Ensembl genes database WGCNA package in R (minModuleSize = 10, reassignThreshold = 1e–6, (release 77). To quantify gene expression levels, the number of reads deepSplit = 2, mergeCutHeight = 0.15) using RNA expression of mapped to the exons of each RefSeq gene was calculated. pan-cancer cell lines (53).

RNA-seq Differential Gene Expression Permutation Analysis Differential gene expression was performed with DESeq2 (47). For each differential accessibility comparison (e.g., siYAP1/WWTR1 A prefilter was applied, so that only genes with at least a median vs. siNTC), ATAC-seq peaks were found within a window of 1,000 bp per kilobase per million mapped reads value of 10 in one condition before the start and 1,000 bp after the end of every gene in human were analyzed. P values for other genes were simply set to 1, and log GENCODE geneset, release 27 (genome assembly version GRCh38). fold changes were set to 0 for visualization purposes, but such genes The peaks were then defined as opening (log fold change> 1, adjusted were not included in the multiple testing correction. Q values were P value < 0.05), closing (log fold change < -1, adjusted P value < 0.05), obtained by correcting P values for multiple hypotheses using the or background (everything else), and only peaks that were annotated Benjamini–Hochberg procedure. Genes were considered if they had a as “protein-coding” were retained. The total number of closing peaks Q value of less than 0.05 and were protein coding. Counts were trans- within the 1,000 bp window of n genes in each Hippo pathway cluster formed to log2 counts per million, quantile normalized, and preci- was compared with the total number of closing peaks within 1,000 sion weighted with the “voom” function of the limma package (48). bp of n randomly sampled genes, repeated for 1 million permuta- tions. The P value was taken as the number of times the number of ATAC-seq and Data Analysis randomly sampled closing peaks was greater than that of the Hippo Reads were aligned to the human reference genome (NCBI Build cluster’s, divided by 1 million. 38) using GSNAP31 version “2013-10-10,” allowing a maximum of two mismatches per read sequence (parameters: -M 2 -n 10 -B 2 Prediction of YAP/TAZ Dependency -i 1–pairmax-dna = 1000–terminal-threshold = 1000–gmap-mode = YAP1/WWTR1 knockdown was performed on 42 cell lines, and none–clip-overlap). Reads aligning to locations in the human change in viability was assessed using Cell Titer Glo after 7 days. genome that contain substantial sequence homology to the MT Effect of viability was quantified as a ratio of fraction of remaining chromosome or to blacklisted regions identified by the ENCODE luminescence compared with nontargeting control (i.e., 1 would consortium were omitted from downstream analyses. Properly mean complete loss of luminescence and 0 as no change). To predict paired reads derived from nonduplicate sequencing fragments were this effect of viability, we constructed a random forest model using used to quantify chromatin accessibility according to the ENCODE the expression of Cluster 1 to 4 genes of 31 (70%) cell lines to predict pipeline standards with minor modifications as follows. Accessible the effect of viability. Predication accuracy was determined by mini- genomic locations were identified by calling peaks with Macs2 (49) mizing root mean square error from 5-fold cross-validation of the using insertion-centered pseudo-fragments (73 base pairs; commu- training data.

790 | CANCER DISCOVERY March 2021 AACRJournals.org

Machine-Learning Approach Predicts Hippo Pathway Dependency RESEARCH ARTICLE

Gene Set Benchmarking CRISPR RNA Transfection We identified two additional YAP/TAZ gene set: Cordenonsi and CRISPR transfection was performed in Detroit 562 and OVCAR-8 colleagues (21) or Wang and colleagues (TCGA; ref. 31). In addition, cell lines (obtained from the ATCC). Cells were seeded at 10,000/ we included our Cluster 0 gene set (i.e., genes that were significantly well on a 6-well plate for 24 hours. They were then transfected with downregulated by YAP1/WWTR1 knockdown but not assigned into the RNP complex that contains designed guide RNA siRNAs (final a gene cluster, likely not proximal to Hippo pathway activity) as a concentration of 1 μmol/L) using Lipofectamine CRISPR Max Cas9 negative control. First, we trained an ML model using either our gene transfection reagent for 72 hours before collection. Cas9 protein set (Cluster 2), Cordenonsi and colleagues’ gene set (21), or Wang and was purchased from Invitrogen (Truecut Cas9 protein v2, A364498). colleagues’ gene sets (TCGA; ref. 31) using the same training data and crRNA and transcrRNA were purchased from IDT. Multiple guide ML algorithm (parallel random forest, caret v6.0-85). The ML model RNAs for YAP1 were tested to get to the two optimized sequences of trained using Cluster 2 gene set outperformed the ML models trained gYAP1: ACATCGATCAGACAACAACA and gYAP2: CCACAGGGAG from Cordenonsi and colleagues (21) and Wang and colleagues (31) GCGTCATGGG. Transient knockout/knockdown was confirmed by in both root mean square error and R-squared metrics (Supplemen- Western blot with the following antibodies from CST: YAP (14074) tary Fig. S2B). This was not dependent on the ML algorithm used to and β-actin (3700). train the ML model—we repeated this using two other common ML algorithms: support vector machines and boosted generalized linear Authors’ Disclosures model. For both ML algorithms, our Cluster 2 gene set outperformed the other gene sets. Last, this difference in performance is not due to C.M. Rose, R. Piskol, V.C. Pham, M.T. Chang, and A. Dey report the initial training data used in model training. We created 500 ran- they are employees of Genentech and shareholders at Roche. S.E. dom iterations of training data (createDataPartition, caret v6.0-85), Martin reports personal fees from Genentech outside the submitted and our gene set significantly outperformed the other two gene sets work. Z. Modrusan reports personal fees from Roche stock outside the submitted work. C. Klijn reports he was an employee of Roche (Wang and colleagues, P value < 10−71; Cordenonsi and colleagues, P and shareholder at the time when the work was performed (currently value < 10−65; paired t-test). Importance of individual genes in a given no longer applicable). S. Malek reports other from Genentech/Roche ML model was calculated using varImp (caret v6.0-85). Genes within our gene were defined as novel if not previously reported in either during the conduct of the study and outside the submitted work, and Cordenonsi and colleagues (21) or Wang and colleagues (31) and is an employee and shareholder of Genentech/Roche. No disclosures known if present in one or both gene sets. were reported by the other authors.

Statistical Analysis Authors’ Contributions Besides the RNA-seq and ATAC-seq studies that have their own T.H. Pham: Formal analysis, validation, investigation, writing- statistical analysis, all of the statistical analysis for in vivo and in original draft, writing-review and editing. T.J. Hagenbeek: Valida- vitro studies was done using the Student t test (two-tailed, unpaired) tion, investigation, writing-original draft. H.-J. Lee: Conceptual- to compare treatment groups with control group. All of the in vitro ization, validation, investigation. J. Li: Software, formal analysis. experiments were performed at least 3 times. For the in vivo work, the C.M. Rose: Software, formal analysis, investigation. E. Lin: For- gender, age, and weight of animals were matched, and the sample size mal analysis, validation, investigation. M. Yu: Formal analysis, vali- is 3 to 10 mice per group. A P value of less than 0.05 was considered dation, investigation. S.E. Martin: Formal analysis, supervision, as statistically significant (*,P < 0.05). Significant values as well as writing-review and editing. R. Piskol: Formal analysis, supervision, number of replicates are noted for each experiment in the respective methodology. J.A. Lacap: Validation, investigation. D. Sampath: figure legends. Supervision, investigation. V.C. Pham: Validation, investigation. Z. Modrusan: Data curation, formal analysis, writing-review and siRNA Transfection and ATAC-seq editing. J.R. Lill: Supervision, visualization, writing-review and editing. C. Klijn: Formal analysis, supervision, methodology. S. Malek: siRNA transfection was performed in Detriot-562 and PATU-8902 Supervision, writing-review and editing. M.T. Chang: Conceptu- cell lines (obtained from the ATCC). Cells were seeded at 15,000/ alization, software, formal analysis, methodology, writing-original well on a 6-well plate or 1,000/well on a 96-well plate for 24 hours. draft, writing-review and editing. A. Dey: Conceptualization, super- They were then treated with the siRNAs (final concentration of vision, writing-original draft, writing-review and editing. 20 nmol/L) using Lipofectamine RNAi Max in serum-free RPMI media for 72 hours before collection. siRNAs were purchased from Received May 20, 2020; revised September 30, 2020; accepted Dharmacon, including siNTC (D-001810-10), siYAP1 (L-012200-00), November 13, 2020; published first November 18, 2020. siWWTR1 (L-016083-00), siTEAD1 (J-012603-05), siTEAD2 (J-012611- 09), siTEAD3 (L-012604-00), siTEAD4 (J-019570-09), and siFOSL1 (LQ-004341-00). Knockdown was confirmed by Western blot with the following antibodies from CST: Pan-TEAD (13295), YAP (14074), References TAZ (70148), FOSL1 (5281), and β-actin (3700). . 1 Chang MT, Bhattarai TS, Schram AM, Bielski CM, Donoghue MTA, ATAC-seq was performed as previously described (54, 55). A total Jonsson P, et al. Accelerating discovery of functional mutant alleles in of 1 105 cell pellets were washed with PBS, and cells were pelleted × cancer. Cancer Discov 2018;8:174. by centrifugation. Cell pellets were resuspended in 100 μL of ATAC- 2. Chang MT, Asthana S, Gao SP, Lee BH, Chapman JS, Kandoth C, resuspension buffer (10 mmol/L Tris HCl, pH 7.4, 10 mmol/L NaCl, et al. Identifying recurrent mutations in cancer reveals widespread and 3 mmol/L MgCl2), and nuclei were pelleted by centrifugation. lineage diversity and mutational specificity. Nat Biotechnol 2016; The nuclei were resuspended in 50 μL reaction buffer containing 34:155–63. Tn5 transposase (2.5 μL Tn5 transposase, 25 μL 2 × TD buffer, 3. Halder G, Johnson RL. Hippo signaling: growth control and beyond. 16.5 μL PBS, 0.5 μL 1% digitonin, 0.5 μL 10% Tween-20, and 5 μL Development 2011;138:9. H2O; Illumina). The reaction was carried out at 37°C for 30 minutes. 4. Justice RW, Zilian O, Woods DF, Noll M, Bryant PJ. The Drosophila Tagmented DNA was isolated by MinElute PCR Purification Kit tumor suppressor gene warts encodes a homolog of human myotonic (QIAGEN). Libraries were later generated and sequenced on Next- dystrophy kinase and is required for the control of cell shape and Seq500 (Illumina). proliferation. Genes Dev 1995;9:534–46.

March 2021 CANCER DISCOVERY | 791

RESEARCH ARTICLE Pham et al.

5. Tapon N, Harvey KF, Bell DW, Wahrer DCR, Schiripo TA, Haber DA, 28. Langfelder P, Horvath S. Fast R functions for robust correlations and et al. Salvador promotes both cell cycle exit and apoptosis in hierarchical clustering. J Stat Softw 2012;46:i11. drosophila and is mutated in human cancer cell lines. Cell 2002; 29. Langfelder P, Horvath S. WGCNA: an R package for weighted correla- 110:467–78. tion network analysis. BMC Bioinformatics 2008;9:559. 6. Xu T, Wang W, Zhang S, Stewart RA, Yu W. Identifying tumor sup- 30. Meyers RM, Bryan JG, McFarland JM, Weir BA, Sizemore AE, Xu H, pressors in genetic mosaics: the Drosophila lats gene encodes a puta- et al. Computational correction of copy number effect improves tive protein kinase. Development 1995;121:1053. specificity of CRISPR–Cas9 essentiality screens in cancer cells. Nat 7. Hansen CG, Moroishi T, Guan K-L. YAP and TAZ: a nexus for Hippo Genet 2017;49:1779–84. signaling and beyond. Trends Cell Biol 2015;25:499–513. 31. Wang Y, Xu X, Maglic D, Dill MT, Mojumdar K, Ng PK-S, et al. 8. Oka T, Mazack V, Sudol M. Mst2 and lats kinases regulate apop- Comprehensive molecular characterization of the hippo signaling totic function of yes kinase-associated protein (YAP). J Biol Chem pathway in cancer. Cell Rep 2018;25:1304–17.e5. 2008;283:27534–46. 32. St-Jean G, Tsoi M, Abedini A, Levasseur A, Rico C, Morin M, et al. 9. Cho E, Feng Y, Rauskolb C, Maitra S, Fehon R, Irvine KD. Delineation Lats1 and Lats2 are required for the maintenance of multipo- of a fat tumor suppressor pathway. Nat Genet 2006;38:1142–50. tency in the Müllerian duct mesenchyme. Development 2019;146: 10. Johnson R, Halder G. The two faces of Hippo: targeting the Hippo dev180430. pathway for regenerative medicine and cancer treatment. Nat Rev 33. Zanconato F, Forcato M, Battilana G, Azzolin L, Quaranta E, Bodega B, Drug Discov 2013;13:63. et al. Genome-wide association between YAP/TAZ/TEAD and AP-1 11. Zhao B, Ye X, Yu J, Li L, Li W, Li S, et al. TEAD mediates YAP- at enhancers drives oncogenic growth. Nat Cell Biol 2015;17:1218. dependent gene induction and growth control. Genes Dev 2008; 34. Camargo FD, Gokhale S, Johnnidis JB, Fu D, Bell GW, Jaenisch R, 22:1962–71. et al. YAP1 increases organ size and expands undifferentiated pro- 12. Zhang H, Liu C-Y, Zha Z-Y, Zhao B, Yao J, Zhao S, et al. TEAD genitor cells. Curr Biol 2007;17:2054–60. transcription factors mediate the function of TAZ in cell growth 35. Lei Q-Y, Zhang H, Zhao B, Zha Z-Y, Bai F, Pei X-H, et al. TAZ pro- and epithelial-mesenchymal transition. J Biol Chem 2009;284: motes cell proliferation and epithelial-mesenchymal transition and is 13355–62. inhibited by the hippo pathway. Mol Cell Biol 2008;28:2426. 13. Goulev Y, Fauny JD, Gonzalez-Marti B, Flagiello D, Silber J, Zider A. 36. Overholtzer M, Zhang J, Smolen GA, Muir B, Li W, Sgroi DC, et al. SCALLOPED interacts with YORKIE, the nuclear effector of the Transforming properties of YAP, a candidate oncogene on the hippo tumor-suppressor pathway in drosophila. Curr Biol 2008;18: chromosome 11q22 amplicon. Proc Natl Acad Sci U S A 2006;103: 435–41. 12405–10. 14. Steinhardt AA, Gayyed MF, Klein AP, Dong J, Maitra A, Pan D, et al. 37. Liu X, Li H, Rajurkar M, Li Q, Cotton Jennifer L, Ou J, et al. Tead Expression of Yes-associated protein in common solid tumors. Hum and AP1 coordinate transcription and motility. Cell Rep 2016;14: Pathol 2008;39:1582–9. 1169–80. 15. Zhao B, Wei X, Li W, Udan RS, Yang Q, Kim J, et al. Inactivation of 38. Lin KC, Park HW, Guan K-L. Regulation of the hippo pathway tran- YAP oncoprotein by the Hippo pathway is involved in cell contact scription factor TEAD. Trends Biochem Sci 2017;42:862–72. inhibition and tissue growth control. Genes Dev 2007;21:2747–61. 39. Hagenbeek TJ, Webster JD, Kljavin NM, Chang MT, Pham T, Lee H-J, 16. Dong J, Feldmann G, Huang J, Wu S, Zhang N, Comerford SA, et al. et al. The Hippo pathway effector TAZ induces TEAD-dependent liver Elucidation of a universal size-control mechanism in drosophila and inflammation and tumors. Sci Signal 2018;11:eaaj1757. mammals. Cell 2007;130:1120–33. 40. Kaan HYK, Chan SW, Tan SKJ, Guo F, Lim CJ, Hong W, et al. Crystal 17. Sanchez-Vega F, Mina M, Armenia J, Chatila WK, Luna A, La KC, structure of TAZ-TEAD complex reveals a distinct interaction mode et al. Oncogenic signaling pathways in the cancer genome atlas. Cell from that of YAP-TEAD complex. Sci Rep 2017;7:2035. 2018;173:321–37. 41. Hong X, Nguyen HT, Chen Q, Zhang R, Hagman Z, Voorhoeve PM, 18. Yu F-X, Guan K-L. The Hippo pathway: regulators and regulations. et al. Opposing activities of the Ras and Hippo pathways converge on Genes Dev 2013;27:355–71. regulation of YAP protein turnover. EMBO J 2014;33:2447–57. 19. Kapoor A, Yao W, Ying H, Hua S, Liewen A, Wang Q, et al. Yap1 activa- 42. Haverty PM, Lin E, Tan J, Yu Y, Lam B, Lianoglou S, et al. Reproduc- tion enables bypass of oncogenic kras addiction in pancreatic cancer. ible pharmacogenomic profiling of cancer cell line panels. Nature Cell 2014;158:185–97. 2016;533:333. 20. Lin L, Sabnis AJ, Chan E, Olivas V, Cade L, Pazarentzos E, et al. The 43. Rose CM, Venkateshwaran M, Volkening JD, Grimsrud PA, Maeda J, Hippo effector YAP promotes resistance to RAF- and MEK-targeted Bailey DJ, et al. Rapid phosphoproteomic and transcriptomic changes cancer therapies. Nat Genet 2015;47:250. in the rhizobia-legume symbiosis. Mol Cell Proteomics 2012;11:724. 21. Cordenonsi M, Zanconato F, Azzolin L, Forcato M, Rosato A, Frasson 44. Wu TD, Nacu S. Fast and SNP-tolerant detection of complex variants C, et al. The hippo transducer TAZ confers cancer stem cell-related and splicing in short reads. Bioinformatics 2010;26:873–81. traits on breast cancer cells. Cell 2011;147:759–72. 45. Wu TD, Reeder J, Lawrence M, Becker G, Brauer MJ. GMAP 22. Calses PC, Crawford JJ, Lill JR, Dey A. Hippo pathway in cancer: and GSNAP for genomic sequence alignment: enhancements to aberrant regulation and therapeutic opportunities. Trends Cancer speed, accuracy, and functionality. Methods Mol Biol 2016;1418: 2019;5:297–307. 283–334. 23. Zhou Z, Hu T, Xu Z, Lin Z, Zhang Z, Feng T, et al. Targeting Hippo 46. Wu TD, Reeder J, Lawrence M, Becker G, Brauer MJ.GMAP and pathway by specific interruption of YAP-TEAD interaction using GSNAP for genomic sequence alignment: enhancements to speed, cyclic YAP-like peptides. FASEB J 2014;29:724–32. accuracy, and functionality. In: Mathé E, Davis S, editors. Statistical 24. Wang W, Li N, Li X, Tran My K, Han X, Chen J. Tankyrase inhibi- genomics: methods and protocols. New York, NY: Springer New York; tors target YAP by stabilizing angiomotin family proteins. Cell Rep 2016. p. 283–334. 2015;13:524–32. 47. Love MI, Huber W, Anders S. Moderated estimation of fold change 25. Zhao Y, Khanal P, Savage P, She Y-M, Cyr TD, Yang X. YAP-induced and dispersion for RNA-seq data with DESeq2. Genome Biol 2014; resistance of cancer cells to antitubulin drugs is modulated by a 15:550. hippo-independent pathway. Cancer Res 2014;74:4493. 48. Law CW, Chen Y, Shi W, Smyth GK. voom: precision weights unlock 26. Harvey KF, Zhang X, Thomas DM. The Hippo pathway and human linear model analysis tools for RNA-seq read counts. Genome Biol 2014; cancer. Nat Rev Cancer 2013;13:246. 15:R29. 27. Hoadley KA, Yau C, Hinoue T, Wolf DM, Lazar AJ, Drill E, et al. Cell- 49. Zhang Y, Liu T, Meyer CA, Eeckhoute J, Johnson DS, Bernstein of-origin patterns dominate the molecular classification of 10,000 BE, et al. Model-based analysis of ChIP-Seq (MACS). Genome Biol tumors from 33 types of cancer. Cell 2018;173:291–304. 2008;9:R137.

792 | CANCER DISCOVERY March 2021 AACRJournals.org

Machine-Learning Approach Predicts Hippo Pathway Dependency RESEARCH ARTICLE

50. Robinson MD, Oshlack A. A scaling normalization method for dif- 53. Klijn C, Durinck S, Stawiski EW, Haverty PM, Jiang Z, Liu H, et al. ferential expression analysis of RNA-seq data. Genome Biol 2010; A comprehensive transcriptional portrait of human cancer cell lines. 11:R25. Nat Biotechnol 2014;33:306. 51. Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, et al. limma 54. Buenrostro JD, Wu B, Chang HY, Greenleaf WJ. ATAC-seq: a method powers differential expression analyses for RNA-sequencing and for assaying chromatin accessibility genome-wide. Curr Protoc Mol microarray studies. Nucleic Acids Res 2015;43:e47. Biol 2015;109:21.29.1–21.29.9. 52. Heinz S, Benner C, Spann N, Bertolino E, Lin YC, Laslo P, et al. Sim- 55. Corces MR, Trevino AE, Hamilton EG, Greenside PG, Sinnott-Arm- ple combinations of lineage-determining transcription factors prime strong NA, Vesuna S, et al. An improved ATAC-seq protocol reduces cis-regulatory elements required for macrophage and B cell identities. background and enables interrogation of frozen tissues. Nat Methods Mol Cell 2010;38:576–89. 2017;14:959–62.

March 2021 CANCER DISCOVERY | 793

Machine-Learning and Chemicogenomics Approach Defines and Predicts Cross-Talk of Hippo and MAPK Pathways

Trang H. Pham, Thijs J. Hagenbeek, Ho-June Lee, et al.

Cancer Discov 2021;11:778-793. Published OnlineFirst November 18, 2020.

Updated version Access the most recent version of this article at: doi:10.1158/2159-8290.CD-20-0706

Supplementary Access the most recent supplemental material at: Material http://cancerdiscovery.aacrjournals.org/content/suppl/2020/11/18/2159-8290.CD-20-0706.DC 1

Cited articles This article cites 54 articles, 16 of which you can access for free at: http://cancerdiscovery.aacrjournals.org/content/11/3/778.full#ref-list-1

Citing articles This article has been cited by 3 HighWire-hosted articles. Access the articles at: http://cancerdiscovery.aacrjournals.org/content/11/3/778.full#related-urls

E-mail alerts Sign up to receive free email-alerts related to this article or journal.

Reprints and To order reprints of this article or to subscribe to the journal, contact the AACR Publications Subscriptions Department at [email protected].

Permissions To request permission to re-use all or part of this article, use this link http://cancerdiscovery.aacrjournals.org/content/11/3/778. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) Rightslink site.