Crizotinib-Resistant ROS1 Mutations Reveal a Predictive Kinase Inhibitor Sensitivity Model for ROS1- and ALK-Rearranged Lung Cancers

Alterations in tyrosine kinases or in their downstream effectors drive the activation of intracellular pathways culminating in neoplastic transformation (1, 2). As cancer cells can strictly rely on these specific alterations in order to proliferate, survive, and invade, the ability to block their function became of crucial therapeutic interest. This kind of biological dependency, termed oncogene addiction, has been documented in a broad spectrum of solid and hematologic malignancies. In non–small cell lung cancer (NSCLC), the first evidence of a targetable activating mutation was EGFR-mutated tumors (3). Later, ALK and ROS1 gene rearrangements were documented, and fusions involving these genes lead to abnormal expression of constitutively activated kinases, thus resulting in oncogene addiction (4). Patients with ALK- and ROS1-rearranged diseases are treated with shared tyrosine kinase inhibitors (TKI). First- (crizotinib), second- (ceritinib, brigatinib), and third- [lorlatinib (PF-0646392)] generation compounds are either approved or currently under evaluation in clinical trials. As for the other TKIs, patients ultimately develop resistance to the therapy. Parallel therapeutic approaches against the two receptors can be explained by their common phylogenic origin. ALK and ROS1 are indeed evolutionary conserved and share 77% of amino acid identity in their ATP-binding sites (5, 6), specifically where TKIs exert their inhibitory activity.

If 11 different mutations in the ALK kinase domain explain a relevant part of acquired resistance to TKIs in ALK-rearranged tumors, only two mutations (G2032R, D2033N) have so far been reported as mechanism of crizotinib resistance in ROS1-rearranged NSCLC (7, 8).

Here, we describe the case of a patient suffering from an NSCLC harboring the ezrin (EZR)–ROS1 fusion gene with acquired crizotinib resistance. Sequencing of the tumor revealed the presence of novel S1986Y and S1986F mutations. Given the sequence homology, mentioned above, codon 1986 of ROS1 is the equivalent of codon 1156 of ALK (6). The C1156Y ALK mutation has already been reported in crizotinib-resistant patients (9, 10). This ALK mutant, also conferring ceritinib resistance in preclinical models (11), remained however sensitive to lorlatinib in both in vitro models (12) and in the clinical setting (10). We therefore hypothesized that the ROS1 mutations we detected would exhibit a superposable pattern of response to the different ALK/ROS1 TKIs. Experiments on engineered Ba/F3 cells confirmed our conjecture, and the patient was indeed successfully addressed to lorlatinib treatment. We therefore predicted overlapping profiles of sensitivity of ROS1 and ALK corresponding mutations to the different TKIs.

Comparative genomic hybridization (CGH) analysis (Agilent technology) was performed following standard procedures (13). ROS1 and ALK rearrangements were evaluated by FISH, as previously described (5, 14), and threshold for positivity attribution was defined according to current recommendations. The fusion partner of the ROS1-rearranged gene was identified by RNAseq analysis (Illumina technology), and ROS1 resistance mutations were identified by whole-exome sequencing (WES; Illumina technology).

The patient provided written informed consent to participate in the molecular screening studies MOSCATO (NCT01566019) and MATCH-R (NCT02517892) of serial biopsy specimens. Both trials are ongoing at Gustave Roussy Cancer Center. Lorlatinib is administered in the phase I/II clinical trial (NCT01970865), enrolling patients suffering from ALK- or ROS1-rearranged NSCLC.

Ba/F3 cells were purchased from DSMZ cell bank less than 6 months ago and confirmed as mouse origin with IEF of AST, MDH, NP, PEP B, and with species PCR. Cells were confirmed Mycoplasma negative using microbiological culture, RNA hybridization, and PCR assays. Ba/F3 cells were transduced with lentiviral constructs in order to express the EZR–ROS1 fusion protein harboring different ROS1 resistance mutations. Cell-survival assays were performed as previously described (11).

Native, EZR-ROS1^WT, EZR-ROS1^G2032R, EZR-ROS1^S1986F, and EZR-ROS1^S1986Y Ba/F3-expressing cells were treated with the indicated concentrations of inhibitors for 3 hours. Immunoblotting was performed as previously described (11). Phospho-ERK (T202/Y204), ERK, S6, phospho-S6 (S240/244), phospho-AKT (S473), AKT, phospho-ROS1 (Y2274), and ROS1 antibodies were obtained from Cell Signaling Technology.

ALK and ROS1 proteins share 77% of their active site amino acid sequence. Sequence alignment for ALK and ROS1 was described by Ou and colleagues (6) and obtained by Log-Expression (MUSCLE) software.

In-silico studies were performed by Life Chemicals. The initial ROS1 structure 4UXL was downloaded from the protein data bank. Mutated proteins and reconstructed gaps were generated with the Swiss-model server. Gromacs package 4.5.5 and amber99sb force field were used to perform fast simulations of the wild-type and mutant ROS1 structures. We used TIP3P water model to get the protein solved in the octahedron with a 10 Å radius. AmberTools 14 antechamber module was used to generate topology of ligands in AM1-bcc approach. A total of 25,000 steps of steepest descent energy minimization were carried before the equilibration step at constant temperature and pressure. The long-range electrostatic interactions were handled by particle-mesh Ewald algorithm, whereas constrains bond lengths by LINCS algorithm. A heat-annealing code was used for conformational mobility of the protein with periodic temperature changes for each 100 picoseconds (ps) from 100 to 320 Kelvin (K), as a modification of default parameters. The simulation time was set to 5,000 ps with integration time step of 2 femtoseconds. A free molecular dynamics (MD) simulation without any intervention in the system was then performed in a 10 ns time-lapse. Root mean square deviation (RMSD), root mean square fluctuation (RMSF), and protein–ligand interaction energies were analyzed starting from free MD with the respective Gromacs tools. Average structures were generated from last 5 ns of free MD with g-cluster tool utilizing a RMSD cut-off of 0.1 nmol/L and were then visualized with PyMol software.

In August 2010, a 63-year-old male never-smoker was diagnosed with stage IV lung adenocarcinoma with diffuse lymph node and pleural involvement. No EGFR, KRAS, PI3K, BRAF, and HER2 mutations were found (Sanger sequencing) in the diagnostic specimen, and a search for an ALK translocation proved negative. First-line treatment, including cisplatin, gemcitabine, and the PARP inhibitor, iniparib (NCT01086254), started in September 2010. Therapy was administered up to the sixth cycle with good clinical and radiological responses (–44% according to RECIST 1.1 as the best response). When pulmonary disease progression was documented in April 2012, second-line treatment with pemetrexed was initiated leading to a prolonged partial response.

A new CT scan–guided biopsy of the left pleura was performed in January 2013. Histology confirmed the presence of a TTF1-positive lung adenocarcinoma, whereas CGH analysis led to the detection of HOXA7 gene amplification and deletion of the 3′ region of ROS1. FISH showed ROS1 rearrangement in 40% of tumor cells (Fig. 1A), and EZR was identified as the ROS1 fusion partner by RNAseq. Targeted next-generation sequencing on a panel of 50 cancer genes did not reveal any mutation. ROS1 rearrangement was confirmed by FISH in circulating tumor cells (15).

After a total of 17 pemetrexed cycles from April 2012 to September 2013, manifest bilateral lung progression was detected, and third-line therapy with off-label crizotinib, 250 mg twice daily, was initiated. Previous coughing symptoms disappeared a few days after treatment initiation, and after 40 days of treatment, CT scans depicted a profound thoracic response, with clear shrinkage of parenchymal lesions dimensions. Crizotinib exerted a clinical activity maintained over time, with the best radiological response achieved in March 2014, after 6 months of therapy (–75% according to RECIST 1.1).

Disease control and excellent clinical tolerability persisted over 22 months of TKI therapy, and disease progression was first detected in July 2015; crizotinib administration was maintained beyond tumor progression with still a major clinical benefit. In order to explore the molecular events explaining crizotinib secondary resistance, a new CT scan–guided biopsy was performed on a left lung cancer nodule. WES analysis uncovered the presence of molecular alterations in exon 37 of the ROS1 gene, corresponding to the ROS1 tyrosine kinase domain. Mutations consisted of serine-to-tyrosine and to-phenylalanine substitutions in codon 1986 (S1986Y/F) of ROS1 (Fig. 1B). The specimen analyzed contained 70% of cancer cells, and tyrosine and phenylalanine substitutions were found respectively in 18% and 17% of WES reads. No other relevant molecular alterations were identified (Supplementary Table S1).

The patient was therefore enrolled in the phase I/II clinical trial evaluating lorlatinib activity (NCT01970865), given the strong in vitro evidence of the activity of this compound against the analogous C1156Y ALK mutation.(12) As depicted in Fig. 1B, although not strictly conserved, native 1986 ROS1 and 1156 ALK codons respectively harbor serine (S) and cysteine (C) amino acid residues, whose structures are markedly similar. A single missense mutation generates the substitution of serine to tyrosine (Y)/phenylalanine (F) for ROS1 and from cysteine to tyrosine for ALK. Tyrosine and phenylalanine have comparable structures with a benzoic group, indicating that they act in a superposable manner modifying ROS1 S1986Y/F and ALK C1156Y conformations. Lorlatinib administration, 100 mg daily, started at the beginning of December 2015 with a suboptimal tolerance profile, characterized by grade I hypercholesterolemia and peripheral neuropathy. The first evaluation CT scan performed 40 days after treatment initiation showed an impressive disease response, confirmed at the second radiologic evaluation at the third month of lorlatinib treatment (–89% according to RECIST 1.1; Fig. 1C). Disease remains controlled without any signs of progression after 6 months of therapy (June 2016).

In order to investigate the effect of ROS1 mutations on ROS1 sensitivity to TKIs, we generated Ba/F3 cells expressing EZR-ROS1^WT, EZR-ROS1^G2032R, EZR-ROS1^S1986F, and EZR-ROS1^S1986Y. The cells were treated with increasing concentrations of crizotinib, ceritinib, and lorlatinib (Fig. 2A). The results of viability assays confirmed the superiority of lorlatinib (≅7–10-fold lower IC₅₀) against ROS1^WT-expressing cells compared with crizotinib and ceritinib. The latters showed similar IC₅₀ values in inhibiting ROS1^WT, with contrast to the pronounced biological superiority of ceritinib compared with crizotinib against the native ALK form (11). As suggested by structure analogies with ALK, S1986F/Y substitutions in the ROS1 kinase domain led to crizotinib and ceritinib resistance. However, lorlatinib maintained its strong inhibitory activity against S1986F and S1986Y mutations (Fig. 2A). The first and most reported ROS1 G2032R mutation (7, 16, 17) engendered resistance to the three compounds (Fig. 2A).

Immunoblot analyses of ROS1 and downstream pathway activation corroborated the cell survival assays. Lower doses of lorlatinib, compared with crizotinib and ceritinib, were required to decrease ROS1 phosphorylation, whereas only a high concentration of lorlatinib appeared to have a mild effect on ROS1^G2032R phosphorylation (Fig. 2B). Crizotinib and ceritinib had only a mild effect on ROS1 phosphorylation in cells harboring S1986F/Y mutations, which led to persistent S6 phosphorylation and cell survival. Only lorlatinib was able to completely switch off ROS1, AKT, and ERK signaling leading to inhibition of S6 phosphorylation in the presence of S1986Y or S1986F mutations (Fig. 2B).

Taken together, these in vitro results revealed crizotinib and ceritinib resistance conferred by ROS1 S1986F/Y mutations, against which lorlatinib maintained its inhibitory potency.

We performed molecular studies to acquire information about the conformational changes induced by the two uncovered mutations on ROS1 structure. In the absence of ligands, we observed that mutant structures underwent some motions with comparison to the wild-type ROS1 protein (Fig. 3). Importantly, we noticed that these mutations do not directly occur within the enzyme active site, but do actually impact on the alphaC helix of the kinase domain, causing its movement with a following change in the tip of the close glycine-rich region, involving the residues from 1950 to 1960 (Fig. 3). As suggested by their similar structures, S1986Y and S1986F substitutions generated perfectly comparable conformations in ROS1 kinase domain.

We also performed conformational studies involving protein–inhibitor complexes (Supplementary Fig. S1). With the intrinsic limits of the complexity concerning the mutation site, as already reported for the corresponding ALK C1156Y substitution (11), S1986Y/F mutations prevent crizotinib access to the kinase domain, whereas lorlatinib can fix and maintain its ability to induce a conformational shift in both wild-type and mutated ROS1 forms. In the case of crizotinib binding, the positioning of the glycine-rich loop tip of the alphaC helix is inside-out, whereas it does present oppositely in the case of lorlatinib interaction, potentially explaining the differential capacity in inhibiting ROS1 mutants (Supplementary Fig. S1).

All mutations loci reported so far in crizotinib-resistant ROS1-rearranged NSCLC [G2032R (7), D2033N (8), and the present S1986Y/F, chronologically] have already been reported in the clinic at the corresponding ALK kinase domain codons (G1202R, D1203N, and C1156Y, respectively; Fig. 4). In addition to a perfect match in involved amino acids, functional consequences, in terms of structure and TKI sensitivity, overlapped between ROS1 S1986Y/F and ALK C1156Y, as depicted above (Figs. 1B and 2). Therefore, ROS1 and ALK kinase domains do not only share phylogenic origin and structure homology, but also mutational hotspots and TKI sensitivity. We can therefore predict that any of the ALK-resistant mutations could be identified in ROS1-rearranged tumors and confer specific TKI resistance. Those mutations would be the 1981Tins, L1982F, S1986Y/F (reported here), M2001T, F2004C/V, L2026M, G2032R (7), D2033N (8), and G2101A (Fig. 4).

With regard to ALK analogy, the 1981Tins, L1982F, S1986Y/F, F2004C/V, and D2033N ROS1 mutations (homologous to the 1151Tins, L1152R, C1156Y, F1174C/V, and D1203N ALK mutations, respectively) would be ceritinib resistant and lorlatinib sensitive (Fig. 4; refs. 11, 12).The M2001T, the gatekeeper L2026M, and the G2101A ROS1 mutations (corresponding to I1171T, L1196M, and G1269A ALK mutations, respectively) would still be prone to ceritinib inhibition (Fig. 4), due to smaller dimensions allowing access to the ATP-binding pocket (11). Remarkably, recent evidence confirmed our predictions concerning ceritinib lack of activity against D2033N and L1982F mutations (8, 16), as we functionally validated here for S1986Y/F mutants and ceritinib efficacy against the L2026M gatekeeper mutation (18).

Of note, not all described ALK secondary mutations can be acquired in ROS1 kinase domain. An arginine (R) residue cannot substitute the leucine (L) of L1982 ROS1 codon as a consequence of a single nucleotide base mutation as observed for ALK L1152R mutation. However, Katayama and colleagues have reported a L1982F mutation conferring crizotinib and ceritinib resistance in a mutagenesis screening (16). ROS1 cannot either acquire the homologous of L1198F (resistant to ceritinib and lorlatinib, crizotinib re-sensitizing; ref. 10) and S1206Y ALK mutation, as a the result of a single-nucleotide modification. Similarly, M2001 ROS1 codon can give rise to M2001T but not to M2001S or M2001N as reported for ALK I1171 codon (Supplementary Table S2; refs. 19, 20).

Growing attention is currently being devoted to the identification of resistance mechanisms to targeted therapies, to better guide treatment strategies and impact favorably on patient outcomes. We report two novel mutations in the ROS1 kinase domain capable of conferring crizotinib resistance but allowing lorlatinib efficacy. Using Ba/F3 cells expressing native or mutated EZR-ROS1, we functionally demonstrated that S1986F/Y substitutions confer crizotinib and ceritinib resistance. Lorlatinib, more potent against WT ROS1 compared with crizotinib and ceritinib, maintained strong growth inhibition of ROS1^S1986F/Y cells (Fig. 2A). Immunoblot analyses confirmed that lorlatinib was the only ALK/ROS1 TKI tested capable of completely switching off ROS1 and downstream signaling phosphorylation in engineered Ba/F3 cells (Fig. 2B). Early clinical data supported lorlatinib strong efficacy by reporting remarkable tumor response rates, including partial and complete intracranial remissions (21).

We moreover confirmed that G2032R, the first documented and most reported ROS1 secondary mutation, confers resistance to crizotinib and ceritinib. Lorlatinib activity against this mutation (IC₅₀ ≅ 500 nmol/L) suggested that the third-generation compound would not be able to reverse resistance in the clinic. As other in vitro data (18) suggested a potential lorlatinib efficacy on this mutation, clinical activity remains to be sorted out. Recent evidence suggests potential activity of board-spectrum TKIs such as foretinib and cabozantinib against G2032R ROS1 mutation (16–18, 22). Similarly, cabozantinib has recently been shown to overcome D2033N, the other ROS1-resistant mutation, both in in vitro models and in a patient (8). The therapeutic index of these mutitargeted kinase inhibitors against the G2032R ROS1 mutation has not yet been clinically investigated.

Crystallographic features and docking studies for ROS1 kinase domain have already been described (7, 22). Regarding the current S1986F/Y mutations, Shaw and colleagues recently described structural alteration of the analogous C1156Y ALK mutation (10). Generating co-crystal structures of the native or mutated kinase domains coupled with crizotinib or lorlatinib, the authors observed that substitution at codon C1156 would probably not directly affect drug binding, due to its relatively remote distance from the inhibitor binding site. The structural studies we performed revealed a similar situation concerning S1986Y/F ROS1 mutations, supposed to trigger a conformational shift in the alphaC helix and the glycine-rich region close to the active site of the enzyme (Fig. 3). The latter event appears to impact on crizotinib binding while allowing the access and the inhibitory activity of lorlatinib into the kinase domain (Supplementary Fig. S1).

In their report, Shaw and colleagues reported two mutations within the same ALK allele, C1156Y/L1198F (10). In our case, the two mutations at the exact same position can either be on two distinct translocated ROS1 alleles within the same cell or in two individual cells. In addition to being technically challenging to investigate, it did not seem crucial here to examine whether the two mutations were in the same cells. Indeed, the two mutants displayed identical sensitivity to the different TKIs tested, and cell sensitivity would be the same with one or two mutants of the ROS1 protein.

Along with the functional characterization of the novel S1986F/Y ROS1 mutations, the present study focuses on the transposability of ALK-validated therapeutic approaches to ROS1 secondary mutations. Based on ALK and ROS1 similarity in kinase domain sequences, common pharmacologic susceptibility, and homologous hotspots loci involved in molecular events, we developed a model predicting sensitivity of 9 ROS1 mutants to different generations of ALK/ROS1 TKIs (Fig. 4). Six of these putative ROS1 secondary mutations have been studied for some TKI in preclinical models, and the data the authors obtained matched our prediction based on ALK knowledge (Fig. 4; refs. 7, 8, 16–18). We cannot however exclude that other ROS1 mutations could occur, and functional studies will be required to directly assess TKI sensitivity. If 11 different mutations conferring resistance to TKIs have already been reported in ALK-rearranged NSCLC thus far (9–11, 19, 23, 24), three secondary mutations have been documented in ROS1-rearranged tumors [G2032R (7), D2033N (8), and the present S1986Y/F]. This discrepancy is ostensibly due to the number of patients with ALK-rearranged disease, about 4 to 5 times higher compared with ROS1-rearranged NSCLC (21), and the fact that therapeutic targeting against ALK began earlier.

Besides ROS1 kinase domains mutations, crizotinib resistance has been reported to emerge due to EGFR (25), RAS (26), or KIT signaling activation (27). Remarkably, the three events have also been reported in ALK-rearranged tumors and suggest the potential of dual target inhibition (19, 23). Pharmacokinetic issues can account for a significant part of resistance to crizotinib and other inhibitors in ALK-rearranged NSCLC, and the equivalent scenario can be depicted for tumors harboring ROS1 fusion gene. The significant number of patients developing brain metastases during crizotinib treatment (28) reflects the scarce penetration of the compound across the blood–brain barrier (29). The novel ALK/ROS1 inhibitors have been designed to obtain a better performance over brain metastases and are indeed showing remarkable results, with a special mention accorded to lorlatinib (12, 21). Recently, ALK-rearranged cells have been proven to adapt to both crizotinib and ceritinib treatment by the overexpression of the p-glycoprotein, whereas lorlatinib and the second-generation ALK inhibitor alectinib are not suitable substrates for this efflux pump (30).

Anyhow, as further exhaustion of novel generation compounds does eventually lead to disease progression, treatment combinations with downstream signaling cascade inhibitors, cytotoxic chemotherapies, or antiangiogenic antibodies represent potential strategies to approach the issue of acquired resistance (31).

As previously suggested (18), our experiments indicate that crizotinib and ceritinib display similar potency on WT ROS1. This is in contrast with the superior efficacy of ceritinib compared with crizotinib on WT ALK both in preclinical and clinical studies (11, 28, 32). Therefore, the ability of ceritinib to overcome crizotinib resistance in ROS1-rearranged tumors without secondary mutations, as it is the case for ALK-positive patients, would need to be directly addressed in the clinic.

Besides defining the pharmacologic susceptibility of ROS1 S1986F/Y mutations, the present report highlights the unavoidable contribution of molecular analyses of specimens obtained at disease progression under TKI therapy. The availability of tumor material thus far allows us to address treatment decisions and to undertake functional validations. The current study enabled us to confirm that the ROS1 S1986Y/F mutations confer resistance to both crizotinib and ceritinib and allowed more efficient patient treatment by avoiding a ceritinib therapy, which would very likely not have afforded a clinical benefit. It is actually challenging to drive specific indications for the treatment of ROS1-rearranged NSCLC patients, given the relative small incidence (1%–2% of all NSCLC), potentially designating ROS1 tumors as a “rare molecularly subtype of common cancers” (33). Given the absolute incidence of NSCLC, the number of patients harboring ROS1 rearrangement worldwide is still relevant. Validated data concerning ALK inhibition can be advantageously transposed to guide treatment decisions in ROS1-rearranged tumors. The current widespread access to deep molecular analyses would ostensibly reveal, for every patient, the specific ROS1 mutation responsible for crizotinib resistance, when it is actually due to kinase domain modifications. We therefore provide a predictive model which finds in the clinical setting its best field of both validation and exploitation. However, performing repeated biopsies in patients is not always feasible. A noninvasive method to detect molecular alterations conferring ALK and ROS1 resistance to TKIs, such as that finely developed for EGFR NSCLC with cell-free circulating tumor DNA (34), would be of inestimable interest.

D. Planchard is a consultant/advisory board member for Pfizer. B. Besse reports receiving commercial research grants from Novartis and Pfizer. J. Remon is a consultant/advisory board member for osepharma. J-C. Soria is a consultant/advisory board member for Pfizer. No potential conflicts of interest were disclosed by the other authors.

Conception and design: F. Facchinetti, Y. Loriot, L. Lacroix, D. Planchard, B. Besse, F. André, J.-C. Soria, L. Friboulet

Development of methodology: F. Facchinetti, Y. Loriot, M.-S. Kuo, L. Lacroix, D. Planchard, J.-C. Soria, L. Friboulet

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): F. Facchinetti, Y. Loriot, M.-S. Kuo, L. Mahjoubi, L. Lacroix, D. Planchard, B. Besse, F. Farace, N. Auger, J.-Y. Scoazec, J.-C. Soria

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): F. Facchinetti, Y. Loriot, M.-S. Kuo, L. Lacroix, D. Planchard, B. Besse, N. Auger, J.-C. Soria, L. Friboulet

Writing, review, and/or revision of the manuscript: F. Facchinetti, Y. Loriot, D. Planchard, B. Besse, N. Auger, J. Remon, J.-Y. Scoazec, F. André, J.-C. Soria, L. Friboulet

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): F. Facchinetti, M.-S. Kuo, F. André, J.-C. Soria

Study supervision: F. André, J.-C. Soria, L. Friboulet

Other (patient CTC testing): F. Farace

We thank Roman Chabanon for graphical assistance, Ken Olaussen for helpful discussions, and Lorna Saint-Ange for article editing. In-silico studies were performed by Life Chemicals' computational chemistry team. Especially, Alexey Rayevsky generated the predicted ROS1 kinase domain structure with S1986 mutation and ligand binding stability analysis for crizotinib and lorlatinib inhibitors.

This work is supported by Institut National du Cancer (INCa-DGOS-INSERM 6043) and Lombard Odier Foundation. F. Facchinetti is financed by the University of Modena and Reggio Emilia (Italy), through its association with the University Hospital of Parma (Italy).

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.