We aimed to identify gain-of-function mediators of resistance to ALK inhibition by systematic perturbation of gene expression. The Center for Cancer Systems Biology (CCSB)-Broad lentiviral expression library is a publicly-available large-scale open reading frame (ORF) library consisting of 15,885 ORFs representing 12,800 human genes (Yang et al., 2011). To identify transcripts sufficient to drive resistance to ALK inhibition, we individually introduced each ORF into an ALK-dependent lung adenocarcinoma cell line (H3122) with marked sensitivity to ALK inhibitors. ORF-expressing cells were assayed for sensitivity both to crizotinib and to the second-generation ALK inhibitor, TAE684 (Figure 1A). A mutant form (L1152R) of EML4-ALK known to confer resistance to both crizotinib and second-generation ALK inhibitors was used as a positive control (Figure 1B) (Sasaki et al., 2011). See Experimental Procedures for a full description of the experimental design.

Statistical analysis of correlation among screening replicates and estimation of lentiviral infection efficiency were performed as quality-control measures (Figure S1). 95% of assayed ORFs met quality-control thresholds and were evaluated for effects on sensitivity to ALK inhibition (see Experimental Procedures). Most ORF-expressing cells remained sensitive to crizotinib and TAE684 (Figure 1C). For H3122 cells expressing each individual ORF in the library, the differential proliferation in drug was expressed as the ratio of cell viability in the presence of drug divided by cell viability in the absence of drug. Differential proliferation for each ORF in crizotinib and TAE684 was then expressed as a z-score (% viability z-score) to normalize across the entire dataset acquired during the screen (see Experimental Procedures). ORFs conferring a differential proliferation in the presence of crizotinib or TAE684 ≥3 standard deviations (SD) above the mean (% viability z-score score ≥3) were selected as candidate mediators of resistance to ALK inhibition. A total of 75 ORFs representing 67 genes met this criterion, with 43 ORFs yielding a % viability z-score ≥3 for both drugs (Figure 1D, 2A).

We next performed a validation study in which the 75 candidate ORFs were subjected to 7-point growth inhibition assays in H3122 with crizotinib and TAE684. The area under the curve (AUC) was determined for each ORF. ORFs associated with an AUC ≥2.5 SD above the mean for either drug (determined in parallel from a cohort of control ORFs) were considered validated resistance drivers in H3122. Overall, 61 ORFs representing 54 genes met this criterion (81% of candidate resistance ORFs identified from the initial screening study; Figure 2A).

Validated resistance genes reflect diverse protein classes (Figure 2B). Prominent among these are transcription factors, serine-threonine and tyrosine kinases, growth factors, and signal transduction adapter proteins. Resistance profiles of the 54 validated resistance genes against crizotinib and TAE684 in H3122 are shown in Figure 2C. Of note, we confirmed that validated genes also drive resistance to the second-generation ALK inhibitor ceritinib with results comparable to TAE684 (Figure S2).

A subset of the identified ALK resistance genes overlap with resistance drivers to MAP kinase pathway inhibition identified in BRAF-mutant melanoma (Johannessen et al., 2010; Johannessen et al., 2013). Among these are kinases (such as AXL, MAP3K8, and RAF1), adapter proteins (including CRKL and GAB1), and transcription factors (such as ETV1, FOS, and YAP1). This finding is not surprising given the central role of MEK/ERK signaling in both contexts. However there are noteworthy differences between the datasets. For example, growth factors are more prevalent in the present study and include five EGF-related peptides (Figure S2). This is consistent with prior reports suggesting a role for EGF ligands and EGFR activation in resistance to ALK inhibition (Katayama et al., 2012; Sasaki et al., 2011; Tanimoto et al., 2014; Tanizaki et al., 2012; Voena et al., 2013; Yamada et al., 2012). More generally, these results corroborate reports implicating paracrine and autocrine mechanisms of resistance to targeted therapies (Harbinski et al., 2012; Straussman et al., 2012; Wilson et al., 2012).

We compared resistance profiles for crizotinib and TAE684 and identified three ORFs (AXL, MST1R, and HGF) that validated strongly as drivers of resistance to TAE684 (% viability z-score >5.5) but not crizotinib (% viability z-score <1; Figure 2D). This discordance likely reflects the pharmacologic activity of crizotinib against AXL, MST1R, and MET (the kinase activated by HGF) (Peters and Adjei, 2012). Similar findings for HGF-mediated resistance to ALK inhibition have been reported (Tanimoto et al., 2014; Yamada et al., 2012).

Since oncogenic ALK dysregulation activates the MEK/ERK and PI3K pathways (which promote cell proliferation and survival), we next determined whether ectopically expressed resistance ORFs might confer resistance by reactivation of ERK and/or AKT in H3122 in the presence of TAE684. Approximately half of the resistance ORFs (including a subset of growth factors, kinases, and adapter proteins) produced sustained ERK and/or AKT activation despite ALK inhibition, underscoring the importance of these pathways in mediating ALK-driven tumorigenesis (Figure 3). The remaining resistance ORFs had minimal effect on these pathways, raising the possibility of resistance modules that bypass these effectors.

To identify resistance mechanisms more broadly operant in ALK-rearranged lung cancer, we assayed the 54 validated resistance drivers in MGH006, a second ALK-rearranged NSCLC cell line (Sequist et al., 2010). Of these, 23 conferred resistance to ALK inhibition in MGH006, as defined by a differential proliferation in drug ≥2.5 SD above the mean (determined from control ORFs) with a single dose of crizotinib or TAE684 (Experimental Procedures; Figure 4). MGH006 is reported to have a secondary dependence on EGFR activation and demonstrates reduced sensitivity to ALK inhibition compared to H3122 (Figure S3) (Katayama et al., 2012). This observation, coupled with the finding that ectopic ORF expression was generally lower in MGH006 compared to H3122 (Figure S3), suggests the resistance effects observed in MGH006 could be an underestimate. Alternatively, our findings might indicate that certain molecular contexts enable some resistance mechanisms over others.

Several ALK resistance drivers identified in H3122 (including AXL, CRKL, and HGF) have previously been implicated in resistance to EGFR inhibitors in EGFR-mutant lung cancer (Cheung et al., 2011; Engelman et al., 2007; Zhang et al., 2012). In addition, both the MEK/ERK and PI3K pathways are up-regulated by ALK or EGFR activation in lung cancer. We hypothesized that a subset of identified mediators of resistance to ALK inhibition in H3122 might also drive resistance to EGFR inhibition in EGFR-mutant lung cancer. To test this, we expressed the ALK resistance ORFs in three EGFR-mutant NSCLC lines (HCC827, HCC4006, and PC-9) and performed inhibition studies with erlotinib (Figure 4, Figure S3). Most ORFs mediating ALK resistance in H3122 also conferred resistance to erlotinib in EGFR-mutant lines, with 38 of 54 (70%) validating in at least one such line (Experimental Procedures). As the activating effects of EGF ligands are presumably blunted by EGFR inhibition, no EGF ligands identified from the ALK inhibitor screen induced resistance to erlotinib. Together, these results raise the possibility of shared mechanisms of resistance to ALK and EGFR inhibition in ALK-rearranged and EGFR-mutant lung cancers, respectively.

We evaluated expression of each resistance ORF in all 5 cell lines assayed in the initial screening and subsequent validation studies (Figure S3). The majority of assayed ORFs are expressed in all cell systems, although a subset of ORFs demonstrate differential expression across cell lines.

The ORF that most strongly induced resistance to ALK inhibition in both H3122 and MGH006 was neuregulin-1 (NRG1), the ligand that activates HER3. Upon ligand binding, HER3 forms heterodimers with HER2 to activate downstream signaling pathways. To confirm the ORF screening and validation findings for NRG1, we treated ALK-dependent cell lines with recombinant NRG1 peptide and assayed resistance to TAE684 (Figure 5A, B). Recombinant NRG1 peptide drives resistance to TAE684 in ALK-dependent NSCLC cell lines (H3122, H2228, and MGH006). NRG1 had no impact on sensitivity to TAE684 in KELLY cells (an ALK-dependent neuroblastoma cell line), suggesting a context-dependent role for NRG1 in driving resistance to ALK inhibition. The effect of NRG1 in ALK-dependent NSCLC cell lines was abrogated with the combination of TAE684 and lapatinib, which inhibits the activity of HER2 (the dimerization partner of HER3; Figure 5C, D). Together, these findings suggest a role for HER2/HER3 activation in resistance to ALK inhibition. Of note, NRG1 has recently been identified as a putative driver alteration in the somatic gene fusion CD74-NRG1 in a subset of invasive mucinous lung adenocarcinoma (Fernandez-Cuesta et al., 2014).

We also examined less-characterized ALK resistance ORFs and associated pathways. Among these is P2Y2 (encoded by P2RY2), a member of the P2Y subfamily of P2 purinergic receptors. P2Y receptors are membrane-bound G-protein coupled receptors (GPCRs) activated in response to extracellular nucleotides including ATP. They are widely-expressed in a variety of tissues including lung epithelia (Schafer et al., 2003). A role has been postulated for P2Y receptors in mediating the activating effects of ATP in the tumor microenvironment on tumor growth (Di Virgilio, 2012). However, a role for these proteins in drug resistance has not previously been demonstrated. There are multiple P2Y family members which are further subdivided based on coupling to either Gq, Gs, or Gi (Baroja-Mazo et al., 2013). ORFs encoding ten P2Y family members are included in the CCSB-Broad lentiviral expression library. Strikingly, two additional P2Y family members (P2Y1 and P2Y6) scored highly in the initial screening study (% viability z-score 2–3) but did not meet our threshold for initial follow-up (z-score ≥3; Figure S4).

Given this observation suggesting three P2Y family members capable of impacting sensitivity to ALK inhibition in H3122, we examined these GPCRs in further detail. H3122 cells expressing these P2Y receptors demonstrated a modest but reproducible decrease in sensitivity to crizotinib in short-term assays (Figure S4). P2Y-mediated resistance was more apparent in colony-formation assays, where we observed increased survival of cells expressing P2Y receptors in crizotinib or ceritinib compared to the Lac Z control (most notably for P2Y1 and P2Y2; Figure 6A–B).

We sought to explore the mechanism by which P2Y receptors mediate resistance to ALK inhibition. Notably, P2Y1, P2Y2, and P2Y6 each couple to Gq to activate phospholipase C (Baroja-Mazo et al., 2013). Prior studies suggest that P2Y receptors activate phospholipase C and protein kinase C (PKC) (Falkenburger et al., 2013). More specifically, P2Y2 has been shown to activate PKCδ in intestinal cells (Amin et al., 2013). Indeed, we found that H3122 cells overexpressing P2Y1 or P2Y2 (although not P2Y6) had increased levels of activated PKCδ (Figure 6C). To determine whether PKC activation alone is sufficient to drive resistance to ALK inhibition, we treated H3122 cells with ceritinib in the presence of the PKC agonist phorbol 12-myristate 13-acetate (PMA). PMA increased the GI₅₀ of ceritinib by 10-fold (Figure 6D). PKC activation also conferred resistance to ceritinib in MGH006 (Figure 6D). PMA had no impact on basal cell proliferation in the absence of ceritinib (Figure 6E). Using activated PKCδ as a marker for PKC activity, we also observed increased PKCδ activation in populations of H3122 cells cultured to resistance with ceritinib (in addition to increased EGFR activation as previously reported; Figure 6F; Figure S4; Experimental Procedures) (Katayama et al., 2012; Sasaki et al., 2011).

Sotrastaurin (AEB071) is a pan-PKC inhibitor in clinical trials (Buder et al., 2013; Chen et al., 2014). The resistance induced by PMA in H3122 and MGH006 was reversed by sotrastaurin, consistent with a PKC-dependent phenotype (Figure 6D, G). Furthermore, H3122 cells ectopically expressing P2Y receptors were sensitive to the combination of crizotinib and sotrastaurin (Figure 6H). The effect appeared specific for P2Y receptors, as this combination did not impact resistance conferred by RAF1 (another resistance driver identified from the study; Figure 6H).

As PKCδ is not included in the CCSB-Broad ORF library, we tested whether overexpression of PKCδ (encoded by PRKCD) is sufficient to confer resistance to ALK inhibition in H3122. We observed that overexpression of PKCδ alone is not sufficient to induce resistance to TAE684 in our assay, suggesting that other PKC isoforms activated by PMA and inhibited by sotrastaurin may be driving the observed resistance (Figure S4). Of note, the initial screening study also identified another PKC isoform, PKCε (encoded by PRKCE), as sufficient to drive resistance to ALK inhibition in H3122 (Figure 2C).

We next asked whether identified resistance drivers and associated pathways might play a role in clinical resistance to ALK inhibition. We used RNAseq to generate gene expression profiles from a cohort of ALK-rearranged lung tumors with acquired resistance to crizotinib and from treatment-naïve ALK-rearranged controls (Table 1). Specimen DFCI76 harbors a secondary EML4-ALK mutation associated with crizotinib resistance (Sasaki et al., 2011); all other tumors lacked secondary ALK mutations (and activating EGFR or KRAS mutations), suggesting alternative resistance mechanisms. We used single-sample Gene Set Enrichment Analysis (ssGSEA) to query gene expression signatures for enrichment in resistant tumors compared to controls (Barbie et al., 2009; Konieczkowski et al., 2014; Subramanian et al., 2005). A total of 286 oncogenic gene signatures were evaluated, including 189 from the Molecular Signatures database (Liberzon et al., 2011). These include signatures associated with activation of a subset of resistance drivers/pathways identified from our screen, along with a P2Y signature we derived from overexpression of P2Y receptors in H3122 (Experimental Procedures; Figure 7A).

Among the gene signatures enriched in resistant tumors compared to controls (with FDR ≤0.2) were 5 signatures associated with resistance drivers/pathways identified from our functional study. Two of these were independent EGFR signatures (Creighton et al., 2006; Gatza et al., 2010; Gatza et al., 2014), consistent with prior reports of an association between EGFR activation and clinical resistance to ALK inhibition (Figure 7B). We also observed enrichment of HER2 and RAF1 signatures in resistant tumors, consistent with our findings of NRG1 (which activates HER2/HER3) and RAF1 as strong drivers of resistance to ALK inhibition (Figure 4, 5) (Creighton et al., 2006; Gatza et al., 2010; Gatza et al., 2014). Finally, the P2Y signature we identified in H3122 cells overexpressing P2Y receptors was also enriched in crizotinib-resistant tumors (Figure 7B, C; Figure S5; Table S1, S2). These findings suggest activation of transcriptional states associated with EGFR, HER2, P2Y, and RAF1 in crizotinib-resistant tumors and likely reflect underlying biological relationships between these pathways. Examination of expanded patient cohorts will be necessary to confirm these findings. Of note, many genes and pathways identified from our functional studies as resistance drivers are not represented among gene signatures used for ssGSEA analysis, limiting a computational assessment of those genes/pathways in resistant tumors compared to controls.

Given these results, we asked if transcripts encoding EGFR, HER2, RAF1, or P2Y receptors were increased in resistant tumors. Among these, expression of both P2RY6 and P2RY2 was significantly increased in crizotinib-resistant tumors compared to treatment-naïve controls (p=0.021 and 0.033, respectively; Figure 7D, E). This observation provides further support for activation of a P2Y gene signature in resistant tumors.

Expression signatures associated with PKC activation were not represented in the reference collection of gene signatures used for the ssGSEA analyses described above. Therefore, we queried tumor expression profiles against the compendium of expression signatures associated with chemical perturbations in the Connectivity Map (CMap) that included PKC agonists (Experimental Procedures; Table S3). Among the ALK-rearranged tumors were a relapsed post-ceritinib (DFCI133) and a pre-ceritinib (DFCI107) specimen obtained from a single patient who previously experienced disease progression on crizotinib (Table 1). Among >3200 compounds profiled in CMap, the top 3 chemical signatures most similar to DFCI133 (compared to DFCI107) were PKC agonists (Table S4). Although limited to a single patient, this preliminary observation is suggestive of PKC activation in a tumor with acquired resistance to ceritinib, consistent with our functional data. However, we did not identify a connection between gene signatures induced by PKC agonists and the entire cohort of unmatched resistant ALK-rearranged tumors compared to treatment-naïve controls.

H3122 cells were obtained from the National Cancer Institute. Generation of ceritinib-resistant H3122 cells is described in the Supplemental Experimental Procedures. H2228, HCC827, and HCC4006 were purchased from the American Type Culture Collection (ATCC). PC-9 cells were obtained from Public Health England, and KELLY neuroblastoma cells were obtained from Sigma-Aldrich. Cells were maintained in RPMI-1640 (Cellgro) with 10% fetal bovine serum (Gemini Bioproducts) and penicillin (100 units/mL) / streptomycin (100 µg/mL; Cellgro). MGH006 cells have been previously reported and were maintained in DMEM (Cellgro) with 10% fetal bovine serum, penicillin, and streptomycin (Sequist et al., 2010). All cell lines were tested to confirm the absence of mycoplasma contamination. Crizotinib, TAE684, ceritinib, erlotinib, lapatinib, and sotrastaurin were purchased from Selleck Chemicals. Blasticidin was obtained from Life Technologies. Phorbol 12-myristate 13-acetate (PMA) was purchased from Santa Cruz Biotechnology.

The Center for Cancer Systems Biology (CCSB)-Broad Lentiviral Expression library has been previously described (Yang et al., 2011). Additional information is included in the Supplemental Experimental Procedures. H3122 cells were seeded in white clear-bottom 384-well microtiter plates at 1500 cells per well using a Multidrop Combi (Thermo-Scientific). The next day, cells were spin-infected with the expression library (or L1152R EML4-ALK as a positive control) in the presence of 4 ug/mL polybrene at 2250 rpm for 30 minutes at 37°C. Cells were infected in 7 replicates. Virus was removed after 24 hours and replaced with standard growth media (containing 12 ug/mL blasticidin for one replicate). After an additional 48 hours, the remaining 6 replicates of ORF-expressing cells were treated with 0.01% DMSO (drug vehicle), 1 µM crizotinib, or 30 nM TAE684 (2 replicates each). Lentiviral infection, media change, and drug addition were all performed with robotics. Following 5 days of drug exposure, cell viability was determined using the Cell Titer-Glo luminescent assay (Promega) according to the manufacturer’s instructions. Luminescence was measured with an EnVision Multilabel Reader (Perkin-Elmer). Screening of the entire CCSB-Broad Lentiviral Expression library (which is arrayed in forty-seven 384-well plates) was performed in 7 batches. Identification and validation of candidate resistance ORFs (and associated statistical analyses) are described in the Supplemental Experimental Procedures.

Cells were lysed in 50 mM HEPES (pH 7.5), 1% Triton X-100, 10% glycerol, 150 mM NaCl, 2 mM EDTA, and 10 mM NaF with protease inhibitors (Roche) and phosphatase inhibitors (Calbiochem). Lysates were fractionated by SDS-polyacrylamide gel electrophoresis and transferred to nitrocellulose membranes using the iBlot system (Life Technologies). Two-color immunoblotting was performed using LI-COR reagents (Odyssey Blocking Buffer and IRDye 800CW and IRDye 680RD secondary antibodies) according to the manufacturer’s instructions (LI-COR Biosciences). Fluorescence detection was performed using an Odyssey CLx Infrared Imaging System, and quantitation was performed using Image Studio software (LI-COR). Antibodies against ALK (#3791), AKT (#2920), phospho-AKT (Ser 473; #4060), EGFR (#2239), phospho-EGFR (Tyr 1068; #3777), PKCδ (#9616), and phospho-PKCδ (Tyr 311; #2055) were obtained from Cell Signaling. ERK2 antibody (sc-154) was purchased from Santa Cruz Biotechnology. Antibodies against phospho-ERK1/ERK2 (Thr 185/Tyr 187; #M8159) and vinculin (#V9131) were obtained from Sigma. V5 antibody (#R960-25) was obtained from Life Technologies. In-cell Western blotting is described in the Supplemental Experimental Procedures.

H3122 cells were seeded in 12-well plates at a density of 10–20,000 cells per well. The following day, cells were spin-infected with lentivirus harboring the indicated ORFs in the presence of 4 µg/mL polybrene. Virus was removed after 24 hours and replaced with standard growth media. After an additional 24–48 hours, cells were treated with drug or DMSO as indicated. Cells were exposed to drug or DMSO for 14–21 days, with media change and fresh drug addition every 3 days. Cells were fixed with 4% formaldehyde and stained with 0.5% crystal violet. Cells were photographed using a Leica microscope and imaging software. Quantification of crystal violet uptake for each sample was determined by de-staining cells with 10% acetic acid and measurement of absorbance at 595 nm using a SpectraMax 190 instrument (Molecular Devices).

Patients with ALK-rearranged NSCLC were identified from the Lowe Center for Thoracic Oncology at Dana-Farber Cancer Institute. Tumor samples were obtained under an institutional IRB-approved protocol, and all participating patients provided written informed consent. Total RNA was extracted from fresh frozen tumor specimens using TRIzol (Life Technologies) according to the manufacturer’s instructions. The presence of the EML4-ALK fusion transcript was confirmed by RT-PCR for each tumor.

We also prepared total RNA from H3122 cells expressing P2Y receptors for RNAseq. Parental H3122 cells or H3122 expressing ectopic Lac Z, P2Y1, or P2Y2 were exposed to 100 nM TAE684 for 12 hours. (Experimental samples were H3122 cells expressing P2Y1 or P2Y2; control samples were parental H3122 and H3122 cells expressing Lac Z.) Total RNA was extracted using the RNeasy kit (Qiagen) with on-column DNase treatment.

cDNA libraries were generated for tumor and cell line samples using the TruSeq Sample Preparation kit (Illumina) with polyA selection according to the manufacturer’s protocol (Revision A 2010). Libraries were pooled using indexed adapters for multiplexing. Pooled libraries were normalized to 2 nM and denatured using 0.2 N NaOH prior to sequencing. Flowcell cluster amplification and sequencing were performed according to the manufacturer’s protocols using either the HiSeq 2000 or HiSeq 2500 to a depth of 45–50 million reads per sample (paired-end with 76 base pair reads).

Data was analyzed using the Broad Picard Pipeline which includes de-multiplexing and data aggregation. Reads were mapped to the reference human genome (hg19) using TopHat 1.4.1. Expression levels for each gene were denoted as RPKM values (reads per kilobase of transcript per million mapped reads) determined using Cufflinks 2.0.2. Genes that were not expressed in any patient-derived tumor samples (RPKM=0) were removed from the tumor dataset; genes not expressed in the cell line samples were removed from the cell line dataset. Differential transcript expression was determined with the signal-to-noise test statistic with standard deviation at least 20% of the class mean. A description of the methods used to analyze RNAseq gene expression data is provided in the Supplemental Experimental Procedures.