Outlier Kinase Expression by RNA Sequencing as Targets for Precision Therapy

The dependence of cancers on a primary driver, most often a kinase (1, 2), forms the guiding principle of targeted therapy that has had some notable clinical successes, such as imatinib for BCR-ABL–positive chronic myeloid leukemia, trastuzumab and lapatinib for ERBB2-positive breast cancers, gefitinib for lung cancers with kinase domain mutations in EGFR (3, 4), and, more recently, crizotinib for lung cancers with ALK gene fusions (5). Thus, protein kinases are the mainstay of a majority of the current targeted therapeutic strategies for cancers, and inhibitors of several oncogenic kinases such as AKT, BRAF, CDKs, KIT, RET, SRC, MAPKs, MET, PIK3CA, PLKs, AURKs, S6Ks, and VEGFR are in various stages of clinical use, trials, or development (4, 6). While activating somatic mutations are associated with a few of these genes, overexpression of kinases (resulting from genomic amplification or other underlying somatic aberrations) is often a strong indicator of aberrant activity that may impart dependence of cancer cells.

Pancreatic cancer is the fourth leading cause of cancer-related deaths in the United States, with the worst prognosis (5-year survival <3%) of all major malignancies (7), due to diagnosis of the disease at an advanced, unresectable stage and poor responsiveness to chemo-/radiotherapy (8, 9). The overarching oncogenic driver of pancreatic cancer is mutant KRAS, which has eluded therapeutic interventions (10, 11), spurring the search for alternative targets (11). The identification of distinct kinases in independent screens for synthetic lethal interactors of KRAS (12–14) led us to systematically explore the expression profiles of all 468 human kinases (the kinome) to identify and test “personalized kinase targets” in a panel of pancreatic cancer cell lines.

Next-generation sequencing of transcriptomes offers significant advantages over microarrays in terms of throughput, elimination of probe bias, and simultaneous monitoring of diverse components of transcriptome biology (15), including gene expression (15–18), alternative splicing (19, 20), chimeric/read-through transcripts (21, 22), and noncoding transcripts (23, 24). Furthermore, transcriptome sequencing affords a direct and quantitative readout of transcript abundance, facilitating sample-wise gene expression analyses using a digital metric of normalized fragment reads, which are not possible using microarrays. Here, we set out to use transcriptome data from a compendium of 482 cancer and benign samples from 25 different tissue types to carry out gene expression profiling of the complete complement of kinases in the human genome, the kinome, to identify “individual sample-specific outlier kinases” inspired by the concept of cancer outlier profile analysis (COPA; refs. 25, 26). Importantly, while COPA analysis was used to identify subsets of “samples displaying outlier expression of candidate genes,” here, we interrogated subsets of “outlier genes in individual samples,” focusing on kinases that display the highest levels of absolute expression among all the kinases in a sample and the highest levels of differential expression compared with the median level of expression of the respective gene(s) across the compendium. As proof-of-concept, we observed outlier expression of the therapeutic target ERBB2 specifically in all the breast cancer cell lines analyzed that are known to be ERBB2 positive. Thus, we hypothesized that specific outlier kinases in other samples may also impart “dependence” owing to clonal selection for extremely high expression and may thereby represent personalized therapeutic targets.

Here, we analyzed kinome expression profiles of breast and pancreatic cancer samples to identify sample-specific outlier kinases. Next, focusing on cell lines displaying outlier expression of kinases with available therapeutics or pharmacologic inhibitors, we tested their dependence on specific outlier kinases compared with nonspecific targets using short hairpin RNA (shRNA) or siRNA and/or small-molecule inhibitors to assess their effects on cell proliferation. Using this approach, we identified several cell line–specific dependencies as well as kinase targets showing enhanced effects with ERBB2 inhibition in breast and KRAS knockdown in pancreatic cancer cells.

Taking advantage of the direct and unbiased readout of gene expression in terms of defined RNA sequencing (RNA-Seq) reads, we carried out a systematic analysis of the human kinome expression in cancer. RNA-Seq–based, normalized read-counts of all 468 kinases available in our transcriptome compendium, composed of 482 samples from 25 different tissue types, revealed distinct kinases expressed at very high levels—both in absolute terms and in the context of their typical range of expression levels—in virtually all the samples examined (Supplementary Table S1).

Querying individual breast cancer samples (43 cell lines and 67 tissues) for kinases that display the highest levels of absolute expression [>20 reads per kb transcript per million total reads in the given sequencing run (RPKM)] among all the kinases in an individual sample and the highest levels of differential expression compared with the median level of expression of the respective gene across the compendium (>5-fold), we identified outlier kinases across the cohort of breast cancer samples (Fig. 1A and Supplementary Table S2). In addition, each of the outliers was assessed for significant Mahalanobis distance from the center of the scatter plot distribution (χ² test, P < 0.05) to prioritize sample-specific kinase outliers. For example, in the breast cancer cell line BT-474, ERBB2 is the predominant outlier kinase (Fig. 1A, inset). Remarkably, with this approach, all breast cancer cell lines known to be ERBB2-positive were scored as displaying an outlier expression of ERBB2. Interestingly, many ERBB2-positive cell lines also displayed outlier expression of additional kinase genes like CDK12 (Fig. 1A, inset), FGFR4, and/or RET, among others (Supplementary Table S2). Similar to the well-known case of ERBB2, we hypothesized that, in general, outlier kinases specific to individual cancer samples could represent additional therapeutic avenues and were thus explored further.

Likewise, kinome expression data from 22 pancreatic cancer cell lines and 13 pancreatic tissue samples also revealed a set of outlier kinases specifically overexpressed in pancreatic cancers (Fig. 1B and Supplementary Table S3), with the outlier kinase profile of a representative pancreatic cancer cell line AsPC-1 depicted in the inset (Fig. 1B). Assessment of outlier kinases in pancreatic and breast cancer cohorts revealed distinct outlier kinase profiles between the 2 diseases. For example, common outlier kinases in breast cancer included ERBB2, FGFR4, and RET, whereas kinases displaying outlier expression across multiple pancreatic cancer samples included EPHA2, MET, PLK2, MST1R, and AKT2. Interestingly, AXL and EGFR showed outlier expression in both pancreatic and breast cancer samples.

Before proceeding to test outlier kinase–specific dependencies in individual cell lines, we validated the gene expression readout provided by the RNA-Seq data. First, comparing the gene expression profiles of a prostate cancer cell line DU145 across 4 independent RNA-Seq runs, we observed a robust correlation (R² > 0.96) between the technical replicates (Supplementary Fig. S1A). Next, we analyzed the variance across RNA-Seq data from a breast cancer cell line, MCF-7, treated with estrogen (0, 3, and 6 hours) as biologic quasi-replicates. Interestingly, we observed an overall high correlation (R² > 0.91) here also, albeit less than the technical replicates (Supplementary Fig. S1B). Next, we validated the expression profiles of kinase genes derived from RNA-Seq by quantitative reverse-transcription PCR (qRT-PCR) and Western blot analyses. As an example, a strong correlation (R² > 0.88) was observed between the levels of MET expression by RNA-Seq and qRT-PCR, over a range of expression values across a panel of samples (Fig. 2A). In addition, individual samples showing outlier expression of MET by RNA-Seq showed distinctly higher expression by qRT-PCR, compared with nonoutlier samples (Fig. 2B). Similarly, we conducted qRT-PCR validation of RNA-Seq data from multiple samples for 8 additional kinases, again showing strong, statistically significant correlations with overall gene expression levels (Supplementary Fig. S2) as well as outlier calls (Supplementary Fig. S3). Furthermore, extending the correlation of outlier expression to protein levels, cell lines with outlier expression of MET were found to display higher levels of total as well as phosphorylated MET, compared with cells without outlier expression of MET (Fig. 2C). Finally, to assess the feasibility of identifying outlier kinases in cancer tissue samples in the backdrop of underlying benign stromal, vascular, and immune cells, we observed a strong correlation between the RNA-Seq data and outlier calls between a primary tumor-derived xenograft tissue, DS-08-947, and its derivative cell line (Supplementary Fig. S4A and Supplementary Table S4). Similar correlation was observed between BxPC-3 and PANC-1 cell lines and xenograft tissues derived from them (Supplementary Fig. S4B).

Among the ERBB2-positive breast cancer cell lines analyzed by RNA-Seq, ZR-75-30 exhibited singular outlier kinase expression of ERBB2, whose knockdown resulted in a strong growth inhibition (Fig. 3). However, knockdown of RPS6KB1, another oncogenic kinase on chromosome 17 located near the ERBB2 amplicon and overexpressed in 40% to 50% of breast cancers, did not affect the proliferation rate of ZR-75-30 cells, which do not show outlier expression of RPS6KB1 (Fig. 3). Many other ERBB2-positive cell lines, however, displayed outlier expression of additional kinases, frequently including FGFR4, such as MDA-MB-361 and MDA-MB-453 (Fig. 3), as well as MDA-MB-330, HCC202, and HCC1419 (Supplementary Table S2). To assess the dependence on the outlier expression of FGFR4 in the backdrop of ERBB2 overexpression, multiple shRNA-encoding lentiviral constructs were used to knock down FGFR4 in MDA-MB-361 and MDA-MB-453 cells exhibiting outlier expression of both ERBB2 and FGFR4, as well as in CAMA-1, with outlier expression of FGFR4 but not ERBB2. Target knockdown for all siRNA and shRNA experiments were assessed by qRT-PCR and/or Western blot analysis (Supplementary Fig. S5A–S5H). Remarkably, knockdown of FGFR4 resulted in decreased cell proliferation in all 3 cell lines with FGFR4 outlier expression (Fig. 3), whereas treatment of these cells with ERBB2-targeting trastuzumab had no effect on the proliferation of CAMA-1 and MDA-MB-361 cells. In contrast, MDA-MB-453 cells showed diminished cell proliferation rates independently upon FGFR4 knockdown as well as trastuzumab treatment and showed an additive effect upon combined treatment.

To further examine the dependence of a subset of ERBB2-positive cells on FGFR4, we generated trastuzumab-resistant sublines of MDA-MB-453 and BT-474, an ERBB2-positive breast cancer cell line that does not exhibit FGFR4 outlier expression (Fig. 4A). Consistent with the experiments involving trastuzumab and shRNA-mediated knockdown of FGFR4 (Fig. 3), MDA-MB-453 cells were found to be independently responsive to both trastuzumab and PD173074, a small-molecule inhibitor of FGFR, whereas a combined treatment with both of these reagents provided the strongest effect on cell proliferation (Fig. 4B, left). Interestingly, MDA-MB-453 cells, grown to be resistant to trastuzumab, continued to display responsiveness to PD173074 (Fig. 4B, right), suggesting that FGFR4 represents an independent therapeutic target in a subset of ERBB2-positive breast cancer cells. Similar results were obtained with another FGFR inhibitor, dovitinib, which significantly decreased cell proliferation in both the MDA-MB-453 parental and trastuzumab-resistant subline (Fig. 4C, left) but did not affect the BT-474 parental or trastuzumab-resistant subline, neither of which displays FGFR4 outlier expression (Fig. 4C, right). Next, we carried out dose–response experiments using specific pharmacologic inhibitors against outlier kinases (Supplementary Fig. S6A–S6C). Cell lines exhibiting outlier expression of FGFRs displayed a dose-dependent response to PD173074 and dovitinib, with significantly lower IC₅₀ values, as compared with cell lines without outlier expression (Supplementary Fig. S6A and S6B). Taken together, these results suggest that a subset of ERBB2-positive breast cancers that display outlier expression of FGFR4 may specifically respond to combined treatment with ERBB2 and FGFR inhibitors more effectively than to ERBB2-directed therapy alone.

We next extended our kinome outlier analysis to pancreatic cancer, a tumor type critically lacking in rational therapeutic options, particularly in the realm of actionable kinases. Kinome expression profiles of individual pancreatic cancer cell lines were used to identify sample-specific outlier kinases (Fig. 5, left). The pancreatic cancer cell lines were then tested for effects on cell proliferation following siRNA-based knockdown of sample-specific outlier and nonoutlier kinases. Knockdown of the sample-specific outlier kinases—for example, EGFR in L3.3, PLK2 in MIA-PaCa-2, MET in BxPC-3, and AKT2 in PANC-1 cells—inhibited the proliferation of respective cells (Fig. 5, middle). A similar growth inhibition was observed following knockdown of MET in HPAC and AXL in Panc-08.13 and PL45 cells (Supplementary Fig. S7). Conversely, knockdown of the nonoutlier kinases AXL in L3.3, MET in MIA-Paca-2, PLK2 in BxPC-3, and PANC-1 cells did not significantly affect cell growth (Fig. 5, right). Also, L3.3 cells remained unaffected by knockdown of the nonoutlier PLK2 (Supplementary Fig. S7). These observations strongly support the notion that outlier kinases represent specific therapeutic targets in individual cancer samples.

Notably, knockdown of the outlier kinase PLK2 in MIA-PaCa-2 cells did not have as profound an effect on cell proliferation as outlier kinase targeting in many other samples (Fig. 5, middle). We hypothesized that this could be due to a pervasive influence of oncogenic KRAS activity in these cells. To test this idea next, we analyzed the effect of KRAS knockdown in pancreatic cancer cell lines with PLK outlier expression.

A panel of pancreatic cancer cell lines with and without PLK outlier expression was stably transduced with 2 independent inducible shRNAs against KRAS and assessed for sensitivity to KRAS knockdown and/or the PLK inhibitor BI-6727 (Fig. 6). Following induction by doxycycline, the cells expressing KRAS shRNAs were distinguished by red fluorescence, resulting from the red fluorescence protein (RFP) tag coexpressed with the shRNA (Fig. 6, middle). KRAS knockdown efficiency of approximately 50% or more was obtained in all the cells tested (Supplementary Fig. S5H). Of the cell lines tested, knockdown of KRAS significantly inhibited the proliferation of L3.3, MIA-PaCa-2, and Panc-03.27, which all harbor oncogenic mutations in KRAS and were therefore designated as KRAS dependent (Fig. 6A). BxPC-3 cells, which have wild-type KRAS, as well as HPAC and PANC-1 cells, which have mutant KRAS, were not affected by KRAS knockdown and were therefore categorized as KRAS independent (Fig. 6B). Incidentally, all 3 PLK outlier cell lines tested here—L3.3, MIA-PaCa-2, and Panc-03.27—were found to be in the KRAS-dependent category based on their reduced proliferation upon KRAS knockdown (Fig. 6A). Furthermore, treatment with the PLK inhibitor BI-6727 significantly inhibited proliferation in cell lines with PLK outlier expression (Fig. 6A, right) but had no effect in cell lines without PLK outlier expression (Fig. 6B, right). The decrease in cell proliferation following BI-6727 treatment was associated with increased apoptosis, as measured by the flow cytometry of Annexin V/propidium iodide–stained cells (Supplementary Fig. S8A). Finally, treatment with BI-6727 in combination with knockdown of KRAS enhanced the inhibition of cell proliferation in the KRAS-dependent, PLK outlier cells (Fig. 6A, right) but had no effect in the KRAS-independent cells without PLK outlier expression (Fig. 6B, right). Investigating the likely reason for the lack of sensitivity to KRAS knockdown in a subset of pancreatic cancer cells harboring oncogenic KRAS, we observed that following KRAS knockdown, the levels of phospho-ERK, one of the major downstream effector proteins in the RAS signaling pathway, were reduced in the KRAS-dependent cell lines L3.3 and MIA-PaCa-2, but not in the KRAS-independent cell line PANC-1 (Supplementary Fig. S8B), suggesting that ERK activity in PANC-1 cells may be sustained by other convergent pathways. Notably, the KRAS-independent cell lines BxPC-3 and PANC-1 did respond to inhibition of their respective outlier kinases, both in vitro (Fig. 5, middle) and in vivo, as described below.

To test the effect of inhibiting sample-specific outlier kinases in vivo, we treated orthotopic tumor xenografts of 2 KRAS-independent pancreatic cancer cell lines, BxPC-3 and PANC-1, established in nonobese diabetic/severe combined immunodeficient (NOD/SCID) mice, with the MET inhibitor XL184. BxPC-3 cells and, to a lesser but significant degree, PANC-1 cells, were found to have MET outlier expression by RNA-Seq, which was validated by qRT-PCR and Western blot analyses (Fig. 2). Notably, both of these cell lines also displayed a dose-dependent response to XL184 in vitro, with significantly lower IC₅₀ values compared with the L3.3 cell line that does not have outlier expression of MET (Supplementary Fig. S6C). Consistent with our hypothesis of dependence on outlier kinases, growth of both BxPC-3 and PANC-1 xenografts was also significantly inhibited by treatment with XL184, as measured by tumor volume and weight (Fig. 7A–C). Of note, no significant difference was found in the body weight of XL184-treated and untreated mice, suggesting that the effective dose of the inhibitor caused no measurable toxicity in vivo (Fig. 7D).

The specificity of response to the MET inhibitor XL184 was analyzed by Western blot analysis, which showed a sharp decrease in phospho-MET levels in BxPC-3 and to a relatively lesser extent in PANC-1 cells following treatment with XL184 (Fig. 7E). Considering that AKT2 represents the predominant outlier kinase in PANC-1 cells (Mahalanobis distance 217.6, P ∼ 0; Supplementary Table S3), lending significant dependence on AKT2 (Fig. 5), we queried whether the profound inhibitory effect of XL184 on PANC-1 xenografts was also mediated through nonspecific targeting of AKT. Western blot analysis of PANC-1 xenograft tumor lysates revealed a markedly decreased level phospho-AKT following XL184 treatment (Fig. 7F). This supports the notion that XL184 suppresses PANC-1 proliferation through inhibition of both AKT and MET signaling. Thus, PANC-1 represents an example of a cancer sample showing dependency on multiple actionable outliers that may respond to a combinatorial therapeutic option or appropriate pan-kinase inhibitors.

The advent of high-throughput sequencing enables a comprehensive characterization of the genomic and transcriptomic landscape of individual cancer samples, inexorably leading to the challenge of defining and prioritizing clinically relevant findings to translate into improved diagnostic and therapeutic options (27, 28). Clinical sequencing of cancers aims to identify actionable genomic aberrations and match patients with available therapies. Protein kinases, being central to biologic and disease processes, including cancer, and being therapeutically targetable, constitute a large proportion of available and potential targets; thus, any novel disease-specific kinase aberrations are of great clinical interest. This study proposes and tests the hypothesis that specific kinases showing outlier expression in individual cancer samples impart “dependence” on the cells, which may be targeted in combination with existing treatment modalities. Importantly, a case is made for considering the entire profile of kinome aberrations to prioritize potentially effective targets.

The “sample-centric” analysis of kinome expression revealed unique profiles of outlier kinases that were tested for dependency. The receptor tyrosine kinase ERBB2 overexpressed in 20% to 30% of breast cancers confers a more aggressive phenotype, increased metastasis, and worse patient prognosis (29, 30). In our outlier kinase analysis, several well-known “ERBB2-positive” breast cancer cell lines, including MDA-MB-361 and MDA-MB-453, were found to display outlier expression of ERBB2, as expected, but frequently also an outlier expression of the therapeutic target FGFR4. Notably, a survey of microarray-based gene expression data in Oncomine (31, 32) also displayed a subset of ERBB2-positive primary breast cancer samples with outlier expression of FGFR4 (data not shown), emphasizing the clinical relevance of our observations. Targeting outlier FGFR4 in ERBB2-positive breast cancer samples was found to confer independent as well as additive inhibitory effects upon their combined knockdown (Fig. 3), highlighting the potential of combining 2 or more outlier kinase targets in treating cancer, even in cases with a predominant driver such as ERBB2. Interestingly, we also observed that the ERBB2-positive MDA-MB-453 cells grown resistant to trastuzumab treatment continued to remain dependent on FGFR4 and responded to FGFR inhibitors (Fig. 4). In clinical trials with ERBB2-positive metastatic breast cancer, 50% to 74% patients have been reported as not responsive to trastuzumab monotherapy or in combination with chemotherapy (33, 34). Our results suggest that the ERBB2-positive breast cancers may be partly dependent on additional drivers, such as FGFR4, RET, EGFR, and MET, which may sustain these cancers following therapeutic abrogation of ERBB2 activity. Another important corollary to our observations is that combinatorial targeting of ERBB2 and additional outlier kinases at the outset may be much more effective than approaching a single target at a time, a concept that warrants further study. Furthermore, each cancer sample needs to be investigated individually to rationally determine patient-specific unique target combinations.

Next, we extended the approach of nominating sample-specific outlier kinases to pancreatic cancer, which is characterized by a bleak prognosis due to presentation at an advanced stage and resistance to traditional chemotherapy and radiation in the setting of its pancreatic cancer sanctuary, encompassing tumor stroma, extracellular matrix, tumor-infiltrating immune cells, and cancer stem cells. Given the paucity of effective targets in pancreatic cancer, the strong response of pancreatic cancer cell lines to knockdown or inhibition of a priori designated outlier kinases is a promising lead. Our results also underscore the importance of matching sample-specific actionable targets with the appropriate therapeutics. For example, targeting MET was found to be more effective in pancreatic cancer cell lines with MET outlier expression than in nonoutlier samples. Notably, many of our experimental results are consistent with several anecdotal studies using kinase inhibitors against EGFR, MET, and AKT2 (35–39).

We also examined the effect of targeting sample-specific outlier kinases in conjunction with the oncogenic KRAS mutation that is present in virtually all cases of pancreatic cancer. Consistent with previous reports (40–42), we observed that only a subset of KRAS-mutant cells display KRAS dependency. Using tetracycline (tet)–shKRAS stable cell lines, we determined L3.3, MIA-PaCa-2, and Panc-03.27 cells to be KRAS dependent, whereas BxPC-3 cells (the only pancreatic cancer cell line in our panel with wild-type KRAS) as well as PANC-1 and HPAC were KRAS independent. Interestingly, comparing our results with the published literature, we noted a general lack of consensus in the “KRAS dependence” status of pancreatic cancer cell lines (10, 14, 40–45). For example, whereas 2 prior studies using siRNA-mediated knockdown of KRAS in the KRAS-mutant cell line MIA-PaCa-2 designated it as KRAS dependent, based on reduced cellular proliferation, invasion, and colony formation assays (10, 44), more recently, Collisson and colleagues (40) observed no significant effect on proliferation in MIA-PaCa-2 cells transduced with shKRAS lentivirus. Similarly, PANC-1 was identified as KRAS dependent in 4 different studies by both siRNA- and shRNA-mediated knockdowns, as assessed by cellular proliferation, colony formation, invasion, and xenograft tumor growth (10, 14, 43, 44), whereas 3 studies found PANC-1 to be KRAS independent by shRNA-mediated knockdown and farnesyl transferase inhibitor treatment using similar in vitro assays (40–42). Conversely, the KRAS wild-type cell line BxPC-3 has been consistently reported to be KRAS independent (14, 44), similar to our findings. Interestingly, HPAC was described as KRAS dependent by Collisson and colleagues (40) but was found to be KRAS independent in our assays. No published references were found for L3.3 and Panc-03.27, which we report as KRAS dependent.

Several KRAS synthetic lethal screens and DNA microarray analyses have been used to describe genes and gene signatures associated with KRAS dependence (12–14, 40, 41, 46) and include kinase genes such as PLK1, MST1R, and SYK (12, 40, 41). Interestingly, we observed outlier expression of PLK to be restricted to KRAS-dependent cells, and these cells showed higher sensitivity to the pan-PLK inhibitor BI-6727 both alone and in combination with KRAS knockdown, as compared with KRAS-independent cells. Previously, Luo and colleagues identified PLK1 as a RAS synthetic lethal interactor in a lung and a colorectal cancer cell line, although they did not test any pancreatic cancer cell lines (12). Our results additionally show that cells respond to the pan-PLK inhibitor BI-6727 only if they have outlier expression of either PLK1 or PLK2 (Fig. 6A and B). This finding highlights the importance of using therapeutic targets in a sample-specific manner.

Overall, our study provides a generalizable metric to define and prioritize personalized target spectra specific to individual tumors. The recent report of a remarkably successful treatment of a patient with acute lymphoblastic leukemia with sunitinib targeting “wildly active” expression of FLT3 kinase identified by RNA-Seq when whole-genome sequencing failed to identify any actionable aberrations (47), provides an anecdotal yet powerful illustration of the potential application of the systematic identification of outlier kinases proposed in our study.

Transcriptome sequencing data from 482 cancer and benign samples from 25 different tissue types previously generated on Illumina GA and GAII platforms were mapped using Bowtie (48) against University of California Santa Cruz (Santa Cruz, CA) Genome Browser genes in the hg18 human genome assembly (49). Unique best-match hit sequences normalized for the number of RPKM (16) were used to generate a gene expression data matrix for the entire compendium (24). The expression data for the complete list of kinase genes (50) were used to identify “outlier kinases” in individual samples based on their absolute expression within the sample and differential expression (defined as absolute expression divided by median expression level of that gene across the compendium). GraphPad Prism software was used to generate kinome expression profiles for each sample, plotting absolute expression versus differential expression for all kinases.

Statistical significance of outlier expression was quantified using a Mahalanobis distance metric [D² = (x − μ)′Σ⁻¹(x − μ); Σ = covariance matrix, D = Mahalanobis distance of the point x to the mean μ; refs. 51, 52), to measure the “distance” of each kinase's absolute and differential expression from the center of the scatter plot distribution. P values were calculated assuming a χ² distribution, with 2 degrees of freedom. Kinases with absolute expression of more than 20 RPKM, differential expression of more than 5-fold, and P < 0.05 were nominated as having “outlier expression.” R language (53) was used to conduct statistical analysis.

All human breast and pancreatic cancer and benign epithelial cell lines were purchased from the American Type Culture Collection (ATCC), except the benign immortalized pancreatic epithelial cell line HPDE and the xenograft cell lines derived from primary pancreatic adenocarcinoma tissues, which were provided by D.M. Simeone (University of Michigan, Ann Arbor, MI). The pancreatic adenocarcinoma cell line L3.3 was obtained from the University of Texas MD Anderson Characterized Cell Line Core (Houston, TX). All cell lines were grown in recommended culture media and maintained at 37°C in 5% CO₂. To ensure cellular identities, a panel of cell lines was genotyped at the University of Michigan Sequencing Core using Profiler Plus (Applied Biosystems) and compared with the short tandem repeat (STR) profiles of respective cell lines available in the STR Profile Database (ATCC).

ON-TARGETplus siRNA against AKT2, AXL, EGFR, MET, and PLK2, and nontargeting control (siNTC) from Dharmacon (Supplementary Table S5A) were used at 100 nmol/L. Cells were transfected in 6-well plates at a density of 50,000 cells per well using Oligofectamine (Invitrogen), according to the manufacturer's protocol. Transfection was repeated 24 hours later; the cells were grown for an additional 48 hours and replated at a density of 5,000 cells per well in 24-well plates. Cells were counted over a period of 1 to 6 days using a Beckman Coulter cell counter. Transient transductions with shRNA against ERBB2, RPS6KB1, and FGFR4, or nontargeting control (shNTC), were carried out in 6-well plates in the presence of 8 μg/mL hexadimethrine bromide (Polybrene; Sigma). For trastuzumab (Herceptin; Roche) experiments, cells were grown for 3 days in 24-well plates with and without trastuzumab (100 μg/mL), in combination with the FGFR inhibitor PD173074 (TOCRIS Bioscience) at 1 μmol/L or TKI-258 (dovitinib; Selleck Chemicals) at 0.1 μmol/L. Trastuzumab-resistant cell lines were generated from MDA-MB-453 and BT-474 by maintaining the cells in the continuous presence of 100 μg/mL trastuzumab over 1 month. Cell proliferation assays were carried out over a period of 1 to 7 days, using a Beckman Coulter cell counter, and growth curves were plotted using GraphPad Prism software. Statistical comparisons were conducted using one-way ANOVA.

Doxycycline-inducible shRNAmir-TRIPZ lentiviral constructs targeting KRAS or nontargeting control (Open Biosystems) tagged with RFP were used to transduce a panel of pancreatic cell lines in the presence of 8 μg/mL Polybrene (Supplementary Table S5A). Forty-eight hours after transduction, cells were selected in medium containing 1 μg/mL puromycin (Invitrogen) for 4 days. The shRNA expression was induced by growing cells in medium containing 1 μg/mL doxycycline (Sigma) for 72 hours. The enrichment of stable cells and efficiency of shRNA induction were assessed by measuring the percentage of cells displaying red fluorescence by flow cytometry (FACSAria Cell Sorter; BD Biosciences). Experiments with stable cell lines were carried out in the presence of 1 μg/mL doxycycline, refreshed daily. Experiments with the PLK inhibitor BI-6727 (volasertib; Selleck Chemicals) were carried out with cells plated in 96-well culture plates at a density of 3,000 to 4,000 cells per well and treated with 10 nmol/L BI-6727 or dimethyl sulfoxide (DMSO). This concentration was selected on the basis of IC₅₀ values calculated from prior proliferation assays using 1 to 500 nmol/L BI-6727 (data not shown). At 0, 1, 3, and 5 days following drug treatment, viable cells were quantified using WST-1 reagent (Roche) and absorbance was measured at 440 nm, per the manufacturer's protocol. Growth curves were plotted using GraphPad Prism software. Statistical comparisons were conducted using one-way ANOVA.

Cell or tissue lysates were separated on 4% to 12% SDS polyacrylamide gels (Novex) and blotted on polyvinylidene difluoride membranes (Amersham) by semi-dry transfer. Antibodies to FGFR4 (Santa Cruz), phospho-AKT, total AKT, phospho-ERK, total ERK, phospho-MET, and total MET (Cell Signaling Technology) were used at 1:1,000 dilutions for standard immunoblotting and detection by enhanced chemiluminescence (ECL Prime), per the manufacturer's protocol. For phospho-MET blots, cells treated with 10 μmol/L XL184 for 12 hours were stimulated with 100 ng/mL human recombinant hepatocyte growth factor (Invitrogen) for 1 hour before harvesting in radioimmunoprecipitation assay RIPA buffer.

RNA was isolated from cell lysates by the RNeasy Micro Kit (Qiagen), and cDNA was synthesized from 1 μg RNA using SuperScript III (Invitrogen) and Random Primers (Invitrogen), per the manufacturer's protocol. qRT-PCR was carried out on the StepOne Real-Time PCR system (Applied Biosystems) using gene-specific primers designed with Primer-BLAST (Supplementary Table S5B and S5C) and synthesized by IDT Technologies. Validation of RNA-Seq results was carried out using TaqMan Universal PCR Master Mix II with uracil-N-glycosylase (Applied Biosystems) and Universal ProbeLibrary System probes (Roche), following the manufacturer's protocol. Validation of siRNA- and shRNA-mediated knockdown was carried out using Fast SYBR Green Master Mix (Invitrogen), per the manufacturer's protocol. qRT-PCR data were analyzed using the relative quantification method and plotted as average fold-change compared with the control. Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) was used as an internal reference. For qRT-PCR validation studies, GraphPad Prism software was used to conduct linear regression and calculate R² correlation coefficients.

Experiments with the FGFR inhibitors PD173074 and dovitinib and the MET inhibitor XL184 were carried out with cells seeded at a density of 3,000 to 4,000 cells per well, plated in 96-well culture plates, and treated with concentrations from 100 to 0.1 μmol/L. WST-1 assay (Roche) was conducted after 72 hours, and readings were recorded at 440 nm. GraphPad Prism software was used to generate nonlinear regression curves and calculate IC₅₀ values.

The apoptosis assay was carried out using ApoScreen Annexin V Apoptosis Kit (Southern Biotech), per the manufacturer's protocol. Briefly, cells treated for 48 hours with DMSO or increasing concentrations of BI-6727 were washed with cold PBS, suspended in cold 1× binding buffer, stained with Annexin V and propidium iodide, and subjected to flow cytometry by FACSAria Cell Sorter (BD Biosciences). Results were analyzed and plotted using Summit 6.0 Software (Beckman Coulter).

Six-week-old male NOD/SCID mice (Taconic) were housed under pathogen-free conditions approved by the American Association for Accreditation of Laboratory Animal Care in accordance with current regulations and standards of the U.S. Department of Agriculture and Department of Health and Human Services. Animal experiments were approved by the University of Michigan Animal Care and Use Committee and carried out in accordance with established guidelines. Mice anesthetized with an intraperitoneal injection of xylazine (9 mg/kg) and ketamine (100 mg/kg body weight) were implanted with 1 × 10⁶ BxPC-3 or PANC-1 cells suspended in 50 μL 1:1 mixture of Media 199 and Matrigel (BD Biosciences) injected subcutaneously into their flanks using a 30-gauge needle. When tumor size reached 0.4 mm, mice were randomized into control and treatment groups (n = 8 per group). The MET inhibitor XL184 (Exelixis Chemicals) was orally administered at 30 mg/kg body weight twice per week for 3 weeks. Tumor growth was monitored weekly. Tumor caliper measurements were converted into tumor volumes using the formula ½[length × (width)²] mm³ and plotted using GraphPad Prism software. At 3 weeks of treatment, mice were weighed and euthanized and the tumors harvested. Statistical comparisons were conducted using one-way ANOVA.

No potential conflicts of interest were disclosed.

Conception and design: D.R. Robinson, X. Cao, D.M. Simeone, A.M. Chinnaiyan, C. Kumar-Sinha

Development of methodology: I. Wei, S. Shankar, L.W. Ma, D.R. Robinson, Y.-M. Wu, D.M. Simeone, C. Kumar-Sinha

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): V. Kothari, I. Wei, S. Shankar, S. Kalyana-Sundaram, L. Wang, L.W. Ma, D.R. Robinson, X. Cao, D.M. Simeone, A.M. Chinnaiyan, C. Kumar-Sinha

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): V. Kothari, I. Wei, S. Shankar, S. Kalyana-Sundaram, L. Wang, L.W. Ma, P. Vats, C.S. Grasso, D.M. Simeone, C. Kumar-Sinha

Writing, review, and/or revision of the manuscript: V. Kothari, I. Wei, S. Kalyana-Sundaram, D.M. Simeone, A.M. Chinnaiyan, C. Kumar-Sinha

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): I. Wei, S. Shankar, S. Kalyana-Sundaram, L.W. Ma, Y.-M. Wu, D.M. Simeone, C. Kumar-Sinha

Study supervision: D.M. Simeone, A.M. Chinnaiyan, C. Kumar-Sinha

The authors thank Terrence Barrette, Michael Quist, Robert Lonigro, and Sheeba Powar for bioinformatics help; Mark Hynes for help with animal work; and Irfan A. Asangani and Filip Bednar for useful discussions. Trastuzumab (Herceptin; Roche) was kindly provided by Dr. Max Wicha (University of Michigan Cancer Center).

This work was supported in part by NIH 5-R21-CA-155992-02 (to C. Kumar-Sinha), NIH 2T32CA009672-21 (to I. Wei), NIH R01CA131045-01 and NIH P50CA130810-1A (to D.M. Simeone), and the Department of Defense Era of Hope grant W81XWH-08-0110 (to A.M. Chinnaiyan). D.M. Simeone is also supported by the Rich Rogel Fund for Pancreatic Cancer Research; A.M. Chinnaiyan is supported by the Doris Duke Charitable Foundation Clinical Scientist Award and is an American Cancer Society Research Professor and A. Alfred Taubman Scholar; and C. Kumar-Sinha is supported by University of Michigan Gastrointestinal (GI) Specialized Programs of Research Excellence (SPORE) Career Development Award and is a recipient of Lustgarten Foundation Award.