A Single-Step, High-Dose Selection Scheme Reveals Distinct Mechanisms of Acquired Resistance to Oncogenic Kinase Inhibition in Cancer Cells

Extensive research into the mechanisms that drive human oncogenesis has revealed numerous examples of changes in the levels or activity of key proteins that underlie the sustained proliferation and survival of cancer cells. Accordingly, drug development in oncology has focused heavily on the discovery of specific inhibitors of these so-called “drivers.” However, resistance to such targeted inhibitors continues to present a significant challenge to the effective long-term clinical management of cancer (1, 2). Even when primary resistance to a given inhibitor is not apparent, and the inhibitor yields an initial, often striking, clinical benefit, acquired resistance almost invariably develops.

One general approach taken by the field to address the problem of acquired drug resistance has been to identify the underlying mechanisms, potentially revealing targets for second-line (or greater) drug development. A variety of methods have been employed for identifying such resistance mechanisms, including genome-wide knockdown or knockout screens seeking sensitizing or buffering interactions, overexpression screens using cDNA libraries, characterization of clinical biopsy samples from patients who have developed resistance to primary treatment, and modeling of resistance in cell culture.

The most frequently employed approach encountered in the literature is the modeling of acquired drug resistance in cell culture. Generally, this involves exposing human tumor-derived cell lines in culture to a specific inhibitor that targets the protein(s) driving proliferation and/or survival of that cell line, isolating resistant subclones, and characterizing the means by which those subclones have gained resistance. Several variations on the approach have been employed, with the most frequent type being the so-called “ramp-up” method, in which cells are exposed to increasing concentrations of the inhibitor in a stepwise manner over an extended period of time.

This particular approach, however, has several potential drawbacks. Chief among these is the inherent bias imposed for selecting only the most robust resistance mechanism(s) from a population of cells, owing to the opportunity for competition among potential survivors during the extended selection process. Consistent with this concern, studies using this approach typically have identified a single mechanism of resistance for a given cell line. Importantly, although competition among cells likely plays a significant role in the evolution of resistance in real tumors, the nature of the competition allowed under the conditions of a ramp-up selection experiment does not well approximate the in vivo scenario. In a solid tumor, cells in distant regions will not directly compete with one another in the same manner that cells in a culture dish will when they are periodically trypsinized and remixed upon plating. In addition, the approach fails to approximate the typical clinical experience in that the cultured cells are gradually exposed to an increasing dose of inhibitor, often starting with relatively low drug concentrations, whereas patients are consistently administered a single therapeutically relevant dose throughout their treatment.

A less commonly employed alternative to the ramp-up method is the so-called “hard-hit” approach. This variation employs a single, therapeutically relevant concentration from the beginning of the selection process, continuing until the time of isolation of the resistant subclones. We suspect that this form of the general approach might uncover different resistance mechanisms than those identified using the ramp-up procedure, even for the same cell line. In addition, owing to the reduced opportunity for competition among potential survivors, we hypothesized that this variation would identify a wider variety of mechanisms in a given cancer cell line model.

Signaling through the MET tyrosine receptor kinase is frequently deregulated in cancer, both as a primary driver of oncogenesis and as a mechanism of resistance to primary therapies targeting other oncogenic drivers (3, 4). Accordingly, several pharmacologic inhibitors of MET signaling have been developed (3–6), and in turn there have been numerous studies aimed at identifying potential mechanisms of resistance to these inhibitors (7–20). These studies have revealed a wide variety of potential resistance mechanisms, including secondary mutations in the MET gene, activation of bypass pathways, and activation of downstream signaling components. Notably, although ramp-up selection, overexpression screens, knockdown screens, addition of activating ligands, and evaluation of a limited number of clinical samples have all been employed to investigate resistance to MET inhibition, we are unaware of any studies that employed a hard-hit selection approach.

Here we modeled mechanisms that confer resistance to inhibition of MET kinase activity using the hard-hit approach in a MET-driven gastric adenocarcinoma cell line. Using this approach, we identified three distinct resistance mechanisms in a single cell line model. Despite the intensive search for resistance mechanisms to MET inhibition prior to this work, including multiple studies using the cell line we employed, one of the specific mechanisms identified here has not been previously reported. These findings highlight the fact that multiple alterations can lead to resistance, even within a single clonally derived cell line model, demonstrating that employing a hard-hit selection approach can reveal resistance mechanisms that may be otherwise missed with more commonly used methodologies.

GTL-16 cells maintained by the Genentech cell bank are routinely genotyped to ensure their identity. Approximately 1 × 10⁷ GTL-16 cells were seeded into each of fifty 150-mm dishes in RPMI complete medium (per trial) and were allowed to reach approximately 70% to 80% confluency (approximately 20 × 10⁶ cells per dish). Crizotinib was then added to the medium to a final concentration of 1 μmol/L. Media (containing 1 μmol/L of crizotinib) was then exchanged every 3 to 4 days. After approximately 30 days, individual colonies of proliferating cells were isolated by ring cloning (six and five clones, respectively, from the two trials) and were transferred to new dishes for expansion and characterization. Resistant derivatives were verified to be of GTL-16 origin using short tandem repeat profiling. All resistant clones and parental GTL-16 cells in culture were subjected to routine Mycoplasma testing (monthly).

Screening and primary data analysis were carried out as previously described (21). Crizotinib was included at 1 μmol/L in the media used for the resistant clones but was omitted for the parental GTL-16 cells. Visualization was carried out using the heatmap.2 function of the gplots package within the R statistical computing environment, using the default clustering parameters.

For monoculture experiments, following the indicated treatment, cell viability was determined using a CellTiter-Glo (Promega) kit, following the manufacturer's protocol. For coculture experiments, parental GTL-16 cells labeled with a nuclear red fluorescent protein (RFP) marker were mixed with the indicated unlabeled cell line. Following the indicated treatment, relative cell viability for the labeled GTL-16 cells was determined by measuring total RFP signal per well. Data analysis was performed using GraphPad Prism. For all viability data, viability is presented as a percentage relative to the untreated control wells, or nontargeting siRNA control wells, of the same cell line. Figures show one representative trial, with samples in triplicate. Error bars displayed are ±1 SD.

Cells were cultured without drug for 7 to 10 days prior to the indicated treatment(s). Lysates were prepared on ice using 0.9% NP-40, 135 mmol/L NaCl, 45 mmol/L Tris-HCl pH 8.0, 4.5 mmol/L ETDA, and 10% glycerol [supplemented with HALT protease/phosphatase inhibitor cocktail (Pierce)]. Samples were vortexed and placed on ice for 10 minutes and then were frozen in liquid nitrogen. Samples were then quick-thawed and centrifuged for 15 minutes at 16,000 × g at 4°C to pellet insoluble material. Supernatants were transferred to new tubes, dithiothreitol was added to a final concentration of 100 mmol/L, and lithium dodecyl sulfate nonreducing loading buffer (Thermo Fisher Scientific) was added to 1×. Samples were heated to 95°C for 5 minutes prior to electrophoresis.

Antibodies recognizing total MET (#3127), phospho-MET_Y1234/1235 (#3077), total ERK1/2 (#9107), phospho-ERK1/2_T202/Y204 (#9101), total AKT (#2920), phospho-AKT_T308 (#13038), phospho-AKT_S473 (#4060), GAPDH (#2118), β-actin (#3700), phospho-CRAF_S338 (#9427), and phospho-CRAF_S289/296/301 (#9431) were from Cell Signaling Technology. Total CRAF (#sc-5284) antibody was from Santa Cruz Biotechnology. Total RAS antibody (ab52939) was from Abcam. FGF3 antibody (MAB1206) was from R&D Systems.

Active RAS pull-down experiments were performed according to the manufacturer's instructions using an Active Ras Pull-down and Detection Kit (Thermo Fisher Scientific).

SMARTpool siRNAs (Dharmacon) were used for knockdown experiments at a final concentration of 50 nmol/L using a reverse transfection protocol. Plates used for TaqMan analysis to assess knockdown efficiency were harvested 2 days following transfection. Plates used to assess the effect of knockdown on viability were read out 7 days following transfection using a CellTiter-Glo kit.

Total RNA was isolated using a RNeasy Kit (Qiagen), according to the manufacturer's instructions. Sequencing libraries were prepared as directed using an Illumina TruSeq Stranded mRNA kit and were run on a HiSeq 4000 Illumina sequencer using a 2 × 150 bp paired-end protocol. RNA-seq analysis was executed and visualized using an in-house, Web-based platform, employing the Salmon (22) and Sleuth (23) data analysis packages. These RNA-seq data are available through NCBI's Gene Expression Omnibus (24) and are accessible through the GEO series accession number GSE137290 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE137290).

RT-PCR was performed using a OneStep RT-PCR kit (Qiagen), according to the manufacturer's instructions (Primer A = 5′-GGGCCCCGGCTCTCGGTTAT-3′; Primer B = 5′-TGCTACTCTCCTGAACTCTCTCACTCA-3′). RT-PCR products were gel purified, and the region including nucleotide +1799 was sequenced using a primer with the sequence 5′-ATATCCTACTCTTCATGGGC-3′.

We chose to model mechanisms of frank resistance to MET inhibition in GTL-16 gastric adenocarcinoma cells, as this cell line has previously been shown to harbor an amplification of the gene encoding the MET receptor tyrosine kinase (25), and cells are consequently dependent upon MET-driven signaling for their survival and proliferation (26). Importantly, this cell line has been used previously by other groups for modeling resistance to MET inhibition (8, 9, 14, 17–19); however, none of those studies employed a hard-hit approach.

To identify mechanisms that can confer frank resistance to MET inhibition, we exposed GTL-16 cells to a continuous high dose (1 μmol/L) of the dual MET/ALK inhibitor, crizotinib (27), for a period of approximately 1 month, at which point, rare colonies (frequency of approximately 5 × 10⁻⁹) of proliferating cells were evident. As stated previously, we elected to utilize this “hard-hit” approach for isolating resistant clones, rather than the more commonly employed “ramp-up” method, as the hard-hit approach avoids the potential complication of competition that can occur between different resistant clones during the ramp-up approach. In the ramp-up scenario, if a given clone employs a resistance mechanism that confers less of a fitness advantage than those utilized by other clones, that clone may be lost during the selection process, as cells are simply out-competed by more fit clones. Indeed, this has been empirically demonstrated in at least one study, wherein an “early” mechanism driving resistance at lower drug concentrations was displaced by another as the dose of drug was increased (14). In contrast, the hard-hit approach should allow for the isolation of independent resistant clones, which can vary in their relative fitness and may correspond to distinct resistance mechanisms.

Isolated clones displayed marked resistance to crizotinib relative to the parental GTL-16 cell line, as well as to another inhibitor of MET kinase activity (Fig. 1A and B). For each of the isolated clones, MET remained dephosphorylated at tyrosine residues within the activation loop upon treatment with crizotinib, consistent with continued target engagement by the inhibitor (Fig. 1C). However, the MAPK pathway was reactivated in each of these clones, suggesting alterations either downstream of MET or in pathways parallel to the MET receptor that converge upon the MAPK pathway. The PI3K/AKT and mTORC2/AKT pathways, on the other hand, displayed little to no measurable reactivation in any of the resistant clones. Together, these results suggest that reactivation of the MAPK pathway is critical for driving resistance in each of the clones we isolated, and that this is achieved either through bypass of MET or activation of downstream components of the MAPK arm of the MET signaling network.

We reasoned that the crizotinib-resistant clones would have become dependent upon whichever bypass pathway had been engaged, or whichever downstream component of the signaling cascade had been reactivated, for their continued proliferation in the presence of MET inhibitors. Therefore, we screened each clone against a panel of approximately 350 small molecules, in combination with crizotinib, in an effort to identify new vulnerabilities not present in the parental cell line (Supplementary Table S1). The data from this screen revealed three distinct patterns of sensitivity. Most of the resistant clones (9 of 11) fell into a single group (Fig. 2A). This group uniquely displayed marked sensitivity to three compounds, all of which inhibit FGFR family kinase activity. Of the two remaining clones, one (designated GTL-16_D) was exquisitely sensitive to kinase inhibitors targeting ERK or MEK, as well as inhibitors specifically designed to target mutant forms of BRAF (Fig. 2A). For the final remaining clone (designated GTL-16_E), within the most obvious cluster of compounds displaying antiproliferative activity, three out of five compounds inhibited the kinase activity of the EGFR (Fig. 2A). These data are consistent with each resistant clone utilizing either an alternative receptor tyrosine kinase to engage MAPK signaling or directly activating the MAPK pathway downstream of MET to enable proliferation in the presence of crizotinib.

For each category of resistant clone, vulnerabilities identified in the primary screen were confirmed upon retesting of selected compounds in combination with crizotinib (Fig. 2B–D; Supplementary Fig. S1A–S1C). Notably, GTL-16_D proved to be almost as sensitive to inhibitors of ERK, MEK, or mutant BRAF used in isolation as it was to the combination of these compounds with crizotinib (Fig. 2C; Supplementary Fig. S1B). This single agent vulnerability is indicative of an “addiction switch” having occurred in these cells (discussed further below). On the other hand, the remaining crizotinib-resistant clones were no more sensitive to their respective group of inhibitors than the parental GTL-16 cells when crizotinib was omitted (Fig. 2B and D; Supplementary Fig. S1A and S1C). Consistent with the antiproliferative effects of the specific inhibitors identified, treatment with these inhibitors resulted in complete inhibition of MAPK signaling but only when combined with crizotinib in the case of the EGFR kinase inhibitor–sensitive and FGFR kinase inhibitor–sensitive clones (Fig. 2E–G).

To confirm that the observed antiproliferative effects of the inhibitors used in the screen were the result of on-target pharmacologic activity, we employed RNAi-mediated knockdown of the targets of those inhibitors. For the FGFR kinase inhibitor–sensitive group of clones, we found that they were dependent on FGFR2 expression but were insensitive to knockdown of FGFR1, FGFR3, or FGFR4 (Fig. 3A; Supplementary Fig. S2A). As an additional control, we confirmed the expected sensitivity to knockdown of FRS2 in these clones, which serves as a common downstream signal transducer for all FGFR family members. Knockdown of BRAF had a substantial impact on the viability of the parental GTL-16 cells; however, the effect on GTL-16_D was markedly greater, consistent with its increased dependence on BRAF-mediated signaling (Fig. 3B; Supplementary Fig. S2B). Knockdown of CRAF, on the other hand, had no significant effect on GTL-16_D or the parental GTL-16 cells. For GTL-16_E, we tested knockdown of EGFR, as well as the closely related HER2 and HER3 receptor tyrosine kinases. We observed an antiproliferative effect in GTL-16_E following knockdown of each of these three kinases, although most dramatically for EGFR (Fig. 3C; Supplementary Fig. S2C). These results are consistent with the fact that these kinases are known to form functional heterodimers with each other (28, 29).

To identify the proximate driver of resistance in the FGFR signaling–driven group of resistant clones, we performed RNA-seq for four clones from this group (GTL-16_A, GTL-16_B, GTL-16_C, and GTL-16_F), as well as the parental cell line (Fig. 4A; Supplementary Fig. S3A–S3C). We examined the list of significantly upregulated and downregulated genes in the resistant clones to identify those with any known role in FGFR-signaling. FGF3, which encodes a secreted ligand that activates the FGFR2-IIIb splice isoform of FGFR2 (30, 31), was substantially upregulated (effect size approximately 5.2–7.4 for the tested FGFR-signaling group clones) and, hence, stood out as a clear candidate for follow-up testing. Further examination of the raw data revealed that FGF3 is likely not detectably expressed in the parental GTL-16 cells, whereas transcription is robustly activated in the resistant clones. Consistent with FGF3 upregulation playing a key role in resistance to crizotinib in these clones, knockdown of FGF3 mRNA expression resulted in a substantial reduction in proliferation (Fig. 4B; Supplementary Fig. S4A).

To determine whether expression of FGF3 is sufficient to confer resistance to crizotinib in this cell line, we introduced an FGF3 overexpression lentiviral construct into GTL-16 cells (hereafter referred to as GTL-16_FGF3-O/E). Transduced cells displayed marked resistance to crizotinib, consistent with FGF3 expression being sufficient to drive resistance (Fig. 4C). Importantly, when cultured in the presence of crizotinib, the FGF3 overexpressing cells were sensitive to FGFR kinase inhibitors in a dose-dependent manner, demonstrating that the overexpression-driven resistance is dependent upon FGFR kinase activity (Supplementary Fig. S5A).

Given that FGF3 normally functions as a secreted ligand (30), we reasoned that conditioned medium from FGF3-expressing resistant clones applied to parental GTL-16 cells would confer a similar degree of resistance to crizotinib. However, we observed no such survival benefit (Fig. 4D). Similarly, when fluorescence-labeled, parental GTL-16 cells were cocultured with each of the resistant clones in a conventional two-dimensional (2D) format, no survival benefit was observed for the parental GTL-16 cells (Fig. 4E). We considered the possibility that the FGF3-driven resistant clones might produce only a small amount of ligand, which could restrict its ability to function over a very short distance, with the effective concentration of FGF3 dropping off rapidly with increasing distance from the producing cell. We reasoned that performing a coculture experiment in a three-dimensional (3D), spheroid format, enforcing direct contact between labeled parental GTL-16 cells and resistant clone cells, might lead to a measurable survival benefit for the parental cells. Again, no survival benefit for GTL-16 was observed (Fig. 4F).

Notably, recombinant human FGF3 added directly to the medium of parental GTL-16 cells could in fact mitigate sensitivity to crizotinib, confirming the ability of the ligand to confer resistance in a canonical manner (Fig. 4G). Taken together, these observations are most consistent with a mechanism in which the resistant cells are producing FGF3 at levels sufficient to act in an autocrine manner but insufficient to reach any neighboring cells at a concentration necessary to elicit a detectable biological effect. Consistent with this quantitative interpretation, media conditioned by GTL-16_FGF3-O/E cells, which produce more FGF3 than the resistant clones, confers a slight survival benefit in the presence of crizotinib when applied to parental GTL-16 cells (Fig. 4H; Supplementary Fig. S5B). Similarly, coculturing GTL-16_FGF3-O/E cells with labeled, parental GTL-16 cells in a 3D, spheroid format results in a slight survival benefit for the parental cells (Fig. 4I). Considering the conditioned media, coculturing, and recombinant protein treatment experiments together, the results indicate that FGF3 protein is sufficient to drive resistance to crizotinib in GTL-16 cells but only if its local concentration reaches a minimum threshold.

GTL-16_D cells displayed marked sensitivity to inhibitors that target mutant forms of BRAF specifically, most notably the V600E variant. Therefore, we performed RT-PCR and sequencing of BRAF transcripts from GTL-16_D to determine whether there were any mutations in the coding sequence. Indeed, we identified a T to A transversion at nucleotide +1799 of the CDS, resulting in a V600E substitution in at least one allele of BRAF in GTL-16_D (Fig. 5A).

To determine whether or not expression of BRAF^V600E is sufficient to confer resistance, we transduced GTL-16 cells at low multiplicity of infection with lentiviral constructs tuned to express near physiologic levels of either the V600E variant of BRAF or the wild-type allele, or with an empty vector control, isolated positively transduced cells by flow cytometry, and then immediately assessed their sensitivity to crizotinib (Supplementary Fig. S6A and S6B; ref. 32). Cells expressing the V600E variant displayed substantial resistance to crizotinib (Fig. 5B).

We note, however, that the degree of resistance conferred under these conditions by the BRAF^V600E transgene was dramatically less than that observed for GTL-16_D (Fig. 1A). We suspect that this difference is driven by a selection bias in favor of transduced cells expressing crippled versions of the transgene, resulting in a mixed population of cells wherein only a portion represents the state of interest. We attempted to generate clonal isolates expressing BRAF^V600E from parental drug-sensitive GTL-16 cells following transduction with the lentiviral constructs used in the short-term experiment described previously; however, every surviving clone was found to harbor alterations within the transgene (Supplementary Table S2). For the equivalent wild-type BRAF construct, on the other hand, no changes affecting the amino acid sequence were detected. Moreover, transduction of GTL-16 cells with lentiviruses bearing the V600E version resulted in far fewer surviving cells relative to the wild-type version, suggesting that expression of BRAF^V600E is toxic to GTL-16 cells (Supplementary Fig. S6C). It is presently unclear whether GTL-16_D harbors any additional mutation(s) that enable long-term tolerance of the expressed V600E allele.

As mentioned previously, GTL-16_D cells displayed a switch in oncogenic addiction from MET-driven to BRAF^V600E-driven proliferation, as evidenced by their sensitivity to BRAF, MEK, or ERK inhibitors used as single agents (Fig. 2C). Hyperactivated ERK signaling driven by BRAF^V600E is known to result in feedback inhibition of RAS and CRAF (33–38), which we suspected might underlie the observed addiction switch. We therefore assessed the (in)activation status of CRAF in these cells by immunoblotting and found that CRAF is largely in an inhibited state (Fig. 5C). GTL-16_D cells also have substantially less RAS-GTP than parental GTL-16 cells, consistent with the feedback inhibition mechanism (Fig. 5D). These results are consistent with the observed addiction switch in GTL-16_D being driven by a break in the signaling cascade from activated MET to the MAPK pathway, owing to feedback inhibition from hyperactivated ERK onto intermediate elements of that cascade, including RAS and CRAF.

The remaining type of resistance mechanism identified in the initial screen involved signaling by EGFR. RNA-seq analysis of this clone revealed strong upregulation of the gene encoding the HER family receptor ligand HBEGF (Fig. 6A; effect size approximately 2.6), which functions directly as an EGFR and HER4 agonist but does not act directly on HER2 or HER3 (39). Knockdown of HBEGF in GTL-16_E cells using siRNAs resulted in a substantial decrease in proliferation in the presence of crizotinib, consistent with a major role for HBEGF in driving the observed resistance (Fig. 6B; Supplementary Fig. S7).

We performed coculture experiments with GTL-16_E and parental GTL-16 cells, just as we did for the FGF3-expressing resistant clones. GTL-16_E cells failed to confer any survival benefit upon the cocultured parental GTL-16 cells in either 2D or spheroid coculture formats (Fig. 6C and D). However, just as for FGF3, addition of recombinant, soluble HBEGF to the culture medium protected GTL-16 cells from the antiproliferative effects of crizotinib (Fig. 6E). These results together suggest that the presence of HBEGF at a sufficient local concentration can confer resistance to crizotinib in GTL-16 cells, and that the GTL-16_E cells produce only enough HBEGF such that this concentration threshold is crossed for the producing cell and not its nearby neighbors.

Aside from GTL-16_D cells, which harbor a point mutation in BRAF, we could not identify genetic alterations that readily explain the drug-resistant phenotype that we observed in the other clones. We wondered whether the FGF3 upregulation or HBEGF upregulation mechanisms identified were genetic in nature, or epigenetic. We reasoned that genetic alterations driving resistance would be durable over time, whereas epigenetic changes would be more labile and, therefore, reversible. To address this, we cultured a subset of the clones (GTL-16_A through GTL-16_F) without drug for varying lengths of time and then reassessed their sensitivity to crizotinib. GTL-16_E quickly relinquished resistance, displaying near parental levels of sensitivity after only 3 weeks without drug treatment (Supplementary Fig. S8). This rapid loss of resistance is inconsistent with a resistance mechanism driven by a stable mutation. Rather, it suggests that a reversible, epigenetic mechanism underlies the observed resistance, the nature of which remains unknown. Among the remaining clones, all displayed long-term stability of the resistance phenotype, with the exception of GTL-16_C, which showed some loss of resistance over the course of the drug holiday experiment, though not to the extent observed for GTL-16_E. Thus, these results reveal that both genetic and nongenetic mechanisms can drive resistance to MET inhibition in GTL-16 cells, and both classes can be uncovered with a hard-hit selection approach.

As stated previously, our choice to employ a hard-hit approach for selection of resistant clones was partially driven by a concern that potential resistance mechanisms may be missed in a ramp-up scenario if particular clones were to out-compete others. In an effort to model such competitive growth, we tagged three of our resistant clones representing distinct mechanisms of resistance (GTL-16_A, GTL-16_D, and GTL-16_E) with different fluorescent protein parkers (mCherry-NLS, mVenus-NLS, and mCerulean-NLS, respectively) and then grew these tagged clones together, with or without the addition of 1 μmol/L of crizotinib, and monitored the composition of the population over time using flow cytometry (Fig. 7). Over 3 weeks under these competitive growth conditions, we observed a steady decline in the proportion of the total population represented by GTL-16_D (BRAF^V600E) in the absence of selective pressure. In the presence of 1 μmol/L of crizotinib, however, GTL-16_D appeared to have the greatest fitness, whereas the fraction of the population represented by GTL-16_E (HBEGF driven) declined over time. These results demonstrate that particular clones can indeed be lost over time in the presence of other, more fit clones, and that the relative fitness of the clones is affected by the growth conditions (i.e., drug concentration). It therefore seems reasonable to suggest that over the long period of time taken for typical ramp-up selection, and the changing selective pressures during that selection, some clones may be lost, despite their capacity to proliferate at the highest drug concentration reached.

Modeling-acquired resistance to various anticancer drugs through the selection of spontaneously resistant cells in culture has proved to be a powerful approach for the field at large, identifying many potential drug resistance mechanisms. The vast majority of these studies have employed a ramp-up selection protocol, with only a small fraction utilizing the “hard-hit” approach, despite the latter more closely approximating the clinical experience. Indeed, a sampling of the relevant, recent literature shows that approximately 80% of relevant published studies employed the ramp-up method (Supplementary Table S3). Here, we employed the hard-hit variation in an effort to identify mechanisms that can confer resistance to the MET kinase inhibitor crizotinib. We hypothesized that employing this variation on the typical cell line modeling approach might uncover resistance mechanisms distinct from those previously identified and might also allow us to identify a larger number of resistance mechanisms in parallel. Indeed, we identified three distinct mechanisms of resistance, and despite the fact that we employed a cancer cell line model extensively used in prior resistance modeling studies, two of these mechanisms, upregulation of FGF3 or HBEGF, had not previously been associated with resistance in this cell line, although activation of various members of the FGFR and HER families of kinases by other means has been previously associated with resistance to MET inhibition in a variety of cell lines (12, 13, 15, 16, 18–20). Significantly, upregulation of FGF3 has not been previously reported for any model of MET inhibitor resistance. These results highlight the fact that the specific nature of the selection approach applied can influence the type of resistance mechanism(s) identified.

Both of these mechanisms of resistance, upregulation of FGF3 or HBEGF, involve secreted ligands that act on alternative receptor tyrosine kinases expressed by GTL-16 cells, enabling bypass of MET signaling blockade. This general class of resistance mechanism, upregulation of a receptor-activating ligand, has been frequently observed in other resistance modeling studies, including those investigating resistance to MET inhibition. However, neither of the two cases described here conforms to the general expectations for this class of mechanism. In both cases, conditioned medium taken from resistant clones and applied to parental cells fails to mitigate drug sensitivity, as does coculturing parental cells with resistant clones. That is, in both cases, these ligands function in a purely autocrine manner in the resistant clones we isolated and fail to act in a paracrine or juxtacrine fashion, as one would expect. It is unclear to what extent these results are truly peculiar, as the type of conditioned medium and coculture experiments performed here are often not carried out. Rather, administration of recombinant protein to address sufficiency is more typical. However, we suspect that FGF3 and HBEGF may indeed represent special cases owing to their particular known processing and secretion characteristics. Secretion of murine FGF3 has been shown to be very inefficient, with a large fraction of the protein accumulating in the Golgi (40). The portion that is successfully secreted is then bound up within the extracellular matrix. For HBEGF, conversion from the pro-, membrane-bound form to the free, soluble form requires the action of ADAM-family metalloproteases (41, 42). Although three of the ADAMs associated with HBEGF processing (ADAM9, ADAM10, and ADAM17) are transcriptionally expressed in GTL-16_E cells, their level of proteolytic activity in these cells was not explored. So, for both of the ligands identified as drivers of resistance in our experiments, their specific biology may underlie the observed autocrine-only signaling.

The third resistance mechanism that we observed, activation of BRAF through a V to E substitution at position 600, has been previously associated with resistance to MET inhibition in GTL-16 cells. In that prior study, the activated allele was expressed transgenically, and the overexpressing cells were shown to be resistant to MET inhibition (8). However, we note that in the previous work, the authors selected for the BRAF^V600E overexpressing cells following transduction by treating with a MET inhibitor, complicating interpretation of the subsequent assay assessing resistance. We attempted a similar experiment to address the sufficiency of the BRAF^V600E variant to confer resistance to MET inhibition wherein we introduced lentiviral constructs to transgenically express BRAF^V600E. However, we were surprised to find that expression of BRAF^V600E, even at near physiologic levels, was too toxic for cells to tolerate long enough to isolate stable lines. We suspect that the resistant clone in which we detect the V600E variant, GTL-16_D, bears one or more unknown additional mutations that enable it to tolerate BRAF^V600E.

Notably, although a common driver of oncogenesis in melanoma, colon cancer, and thyroid tumors, acquisition of a BRAF^V600E variant as a mechanism of resistance to primary treatment has been reported only in three instances clinically to our knowledge (43–45). Moreover, we were unable to find any reports of spontaneous mutations resulting in a V to E substitution in BRAF in cell culture modeling studies. We hypothesize that the potency of the V600E variant underlies the toxicity in GTL-16 cells observed in our study and speculate that this toxicity may play a role in restricting this mutation to the limited number of tissue types in which it has been observed clinically.

In addition to resistance mechanisms that are likely genetic in nature, as evidenced by their long-term stability in the absence of selective pressure, we also identified a putatively epigenetic mechanism of resistance, namely, upregulation of HBEGF. Elevated expression of HBEGF was recently identified as a resistance mechanism to MET inhibition in EBC-1 cells (12), although it was not determined in that study if resistance in those cells was genetic or epigenetic in nature. Generally, this distinction is not explored in resistance modeling studies, so it is unclear to what extent these two classes contribute to drug resistance.

To conclude, our results suggest that modeling drug resistance in cell culture using a “hard-hit” methodology may uncover potential resistance mechanisms that might otherwise be missed using the standard ramp-up approach. Moreover, owing to the single step nature of the selection process, a hard-hit approach limits the potential for competition between clones that may employ different resistance mechanisms, which we surmise played an important role in allowing us to identify multiple resistance mechanisms in parallel within a single cell line population. In future cell culture–based resistance modeling exercises, it will be important to consider the specific nature of the selection scheme and its influence on the array of potential resistance mechanisms identified.

No potential conflicts of interest were disclosed.

Conception and design: K.J. Finn, J. Settleman

Development of methodology: K.J. Finn

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K.J. Finn, S.E. Martin

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K.J. Finn

Writing, review, and/or revision of the manuscript: K.J. Finn, S.E. Martin, J. Settleman

Study supervision: J. Settleman

We thank the gCSI team at Genentech for carrying out the small-molecule synthetic-lethality screen. We thank the genomics core at Calico Labs for processing and running the RNA-seq samples. We thank the platform team at Calico Labs for performing the analysis of the RNA-seq data. We are grateful to members of the Settleman lab for helpful discussion.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.