Spectrum of Mechanisms of Resistance to Crizotinib and Lorlatinib in ROS1 Fusion–Positive Lung Cancer

ROS proto-oncogene 1, receptor tyrosine kinase (ROS1) gene fusions define a unique molecular subset of non–small cell lung cancer (NSCLC) highly sensitive to treatment with ROS1 tyrosine kinase inhibitors (TKI; refs. 1, 2). Crizotinib, a multitargeted anaplastic lymphoma kinase (ALK)/ROS1/MET proto-oncogene, receptor tyrosine kinase (MET) inhibitor, was the first TKI to receive FDA and European Medicines Agency approval for the treatment of advanced ROS1 fusion–positive (ROS1⁺) lung cancer (3, 4). In a phase I expansion cohort of patients with ROS1⁺ lung cancer, crizotinib resulted in a significant objective response rate (ORR; 72%) and progression-free survival (PFS; median, 19.2 months; refs. 3, 4), and additional single-arm studies of crizotinib have supported its efficacy (5–7). More recently, a ROS1/TRK/ALK inhibitor, entrectinib, has also attained approval on the basis of efficacy demonstrated in an integrated analysis of three phase I–II trials (ORR, 77%; median PFS, 19 months; ref. 8).

Despite these advances and the significant initial benefit seen in most patients with ROS1⁺ lung cancer treated with ROS1 inhibitors, TKI resistance remains a major hurdle limiting duration of benefit to these therapies. Given the significant clinical and biological similarities between ROS1 and ALK, efforts to combat ROS1 inhibitor resistance have been commonly drawn from the experiences in ALK fusion–positive NSCLC. Lorlatinib, a highly brain-penetrant, next-generation ALK and ROS1 inhibitor with FDA approval in ALK fusion–positive lung cancer, has recently demonstrated efficacy in ROS1⁺ disease, with an ORR of 62% among treatment-naïve and 35% among crizotinib-pretreated patients (duration of response, 25.3 and 13.8 months and central nervous system ORR, 64% and 50%, respectively) in a phase I–II study (9). The National Comprehensive Cancer Network guidelines list lorlatinib as a subsequent option after disease progression on crizotinib or entrectinib in ROS1⁺ lung cancer (10). Nonetheless, resistance inevitably develops on lorlatinib as well.

Thus far, efforts to understand mechanisms of resistance to ROS1 inhibitors have been centered on ROS1-dependent resistance to the oldest TKI, crizotinib. A number of crizotinib-resistant ROS1 resistance mutations have been identified in patient samples, including the most commonly occurring ROS1 G2032R (11–13), in addition to ROS1 D2033N (14), L2026M (15), L1951R (16), and S1986F/Y (12, 16). However, the frequency of ROS1 mutations reported in these small case series has ranged widely from 8% to more than 50% (12, 13, 15). Furthermore, outside of an isolated case report of ROS1 G2032K mutation conferring resistance to lorlatinib (17), mechanisms of resistance to lorlatinib in ROS1⁺ lung cancer have not yet been elucidated.

Here, we present the largest multi-institutional series to date of repeat biopsies from patients with ROS1⁺ NSCLC progressing on crizotinib or lorlatinib, analyzed by gene sequencing to characterize the frequency and spectrum of resistance mechanisms.

We identified patients with advanced or metastatic NSCLC harboring ROS1 fusion detected by local molecular profiling [e.g., FISH, DNA-based next-generation sequencing (NGS), or targeted RNA sequencing]. Patients must have received ROS1 inhibitor(s) with posttreatment tumor or plasma analyzed by genotyping. Patients were identified at three institutions: Massachusetts General Hospital (MGH; Boston, MA; n = 37), Memorial Sloan Kettering Cancer Center (MSKCC; New York, NY; n = 11), and University of California Irvine (UCI; Irvine, CA; n = 7). All studies were performed under institutional review board–approved protocols at respective institutions.

Medical records were retrospectively reviewed and data were extracted on clinical, pathologic, and molecular features. Data were updated as of September 2020. Time to progression (TTP) was measured from the time of therapy initiation to clinical/radiographic disease progression. Patients without documented progression were censored at last follow-up. Duration of therapy was measured from the time of therapy initiation to therapy discontinuation. Patients continuing on therapy were censored at last follow-up.

All patients underwent tumor or plasma biopsy after treatment with crizotinib and/or lorlatinib and genotyping after providing informed consent. Of the 56 tissue biopsies, 54 were analyzed using: the MGH SNaPshot DNA-based genotyping panel and a separate RNA-based NGS Assay (Solid Fusion Assay; n = 27; ref. 18), FoundationOne (n = 11; Foundation Medicine, Inc.), MSK IMPACT (n = 9; ref. 19), OncoPanel (n = 4; ref. 20), Ion Ampliseq Comprehensive Cancer Panel (n = 1; Thermo Fisher Scientific), Moffitt STAR Solid Tumor Assay (n = 1; Moffitt), and University of Vermont Medical Center Solid Tumor Gene Panel (n = 1; University of Vermont, Burlington, VT). Two tumor biopsies were analyzed by whole-exome sequencing as described previously (21). Sixteen of the 27 samples analyzed by MGH SNaPshot additionally underwent analysis by Sanger sequencing of the ROS1 kinase domain (Supplementary Materials and Methods). A total of 23 liquid biopsies were genotyped using the Guardant360 Cell-free DNA (cfDNA; n = 17; Guardant Health, Inc.), FoundationACT (n = 3; Foundation Medicine, Inc.), or Resolution Bioscience (n = 3; Resolution Bioscience, Inc.) assays. MET FISH was performed on two posttreatment tumor samples as described previously (ref. 22; Supplementary Materials and Methods).

Ba/F3 cells expressing nonmutant CD74-ROS1 or various mutant CD74-ROS1 were generated as in Supplementary Materials and Methods (reagents are also included therein) and as described previously (23), and were plated (2,000–10,000) in triplicate into 96-well plates. Two days after drug treatment, the cell viability was measured by using CellTiter-Glo (Promega) and a SpectraMax M5 Multi-Mode Microplate Reader (Molecular Devices, LLC). The luminescence values were normalized to untreated wells. GraphPad Prism (GraphPad Software) was used to display data and determine IC₅₀ values by the four-parameter logistic regression model. Western blotting was performed per Supplementary Materials and Methods.

Co-crystal structures of lorlatinib in ROS1 (PDB 4UXL) and a close-in analogue of cabozantinib, foretinib, in MET (PDB 3LQ8) were used as starting points for modeling. Details are provided in Supplementary Materials and Methods.

TTP and duration of therapy medians were estimated using the Kaplan–Meier method (Stata version 14.2). Fisher exact test was used to compare the frequency of ROS1 mutations between lorlatinib-resistant and crizotinib-resistant specimens based on a two-sided hypothesis.

Between 2014 and 2020, a total of 55 patients with advanced ROS1⁺ NSCLC who were treated with crizotinib and/or lorlatinib underwent posttreatment biopsies with sequencing analyses. Baseline characteristics are shown in Table 1. The median age at diagnosis was 50 years (range, 22–81). All patients had adenocarcinoma and the majority (96%) were never or light smokers. The ROS1 fusion partner was known for most patients (85%), CD74 being the most common fusion partner detected in 44% of cases.

All patients in the cohort received crizotinib. The median TTP on crizotinib was 10.1 months [95% confidence interval (CI), 6.9–12.4 months], and the median duration of therapy was 13 months (95% CI, 10–18.8 months). Fourteen of 55 patients did not undergo a post-crizotinib biopsy, but were included as they had a subsequent post-lorlatinib biopsy. Among 41 patients who did have post-crizotinib biopsies, the median TTP on crizotinib was 10.5 months (95% CI, 6.9–18.8 months), with median duration of therapy of 14.7 months (95% CI, 10.6–24.3 months). Twenty-five patients received lorlatinib and had post-lorlatinib biopsies. All 25 patients had received prior crizotinib; 3 had additionally received prior entrectinib (Supplementary Fig. S1). The median TTP on lorlatinib was 8.5 months (95% CI, 5.1–13.8 months), with median duration of therapy of 13.7 months (95% CI, 8.4–18.7 months).

From this cohort, 11 (20%) underwent both post-crizotinib and post-lorlatinib biopsies, 30 (55%) had at least one post-crizotinib biopsy (thus, a total of 41 patients had post-crizotinib biopsies), and 14 (25%) had at least one post-lorlatinib biopsy (thus, a total of 25 patients had post-lorlatinib biopsies; Fig. 1; Supplementary Fig. S1). A subset of patients had paired tissue and plasma analyses at the time of disease progression, or underwent serial repeat biopsies on the same TKI (Fig. 1; Supplementary Fig. S1). In total, 47 post-crizotinib and 32 post-lorlatinib biopsies were analyzed.

We first examined the distribution of ROS1-dependent resistance in patients progressing on crizotinib. The 47 post-crizotinib biopsies included five sets of paired tissue and plasma collected at the same timepoint, thus representing 42 timepoint distinct biopsies. The gene alterations detected in the crizotinib-resistant biopsies are summarized in Supplementary Fig. S2 and Supplementary Table S1.

ROS1 resistance mutations were identified in 16 cases (38%). Consistent with the prior literature (12, 13), the most common ROS1 resistance mutation detected in approximately one-third of cases was the solvent front mutation, ROS1 G2032R. Other ROS1 resistance mutations included ROS1 D2033N (2%) and ROS1 S1986F (2%; Fig. 2A), both of which have been described in ROS1⁺ NSCLC (12, 14, 16). Of note, one case (MGH0003.A) was found to harbor ROS1 G2032R in the post-crizotinib plasma [variant allele fraction (VAF), 0.4%], but not in the paired crizotinib-resistant liver specimen. This patient subsequently continued on crizotinib with the addition of chemotherapy and had initial disease response. Nine months later, the patient again experienced disease progression, and repeat paired biopsies of the plasma and growing pleural effusion (MGH0003.B) identified ROS1 G2032R in both specimens (VAF, 5.7% in the plasma; described in ref. 13). Therefore, the discordance in the detection of ROS1 G2032R in the initial crizotinib-resistant liver versus plasma samples (MGH0003.A) likely represented either tumor heterogeneity or early emergence of ROS1 G2032R not captured by tissue genotyping.

Next, we analyzed 32 lorlatinib-resistant tissue and plasma samples which included four sets of paired tissue/plasma, representing a total of 28 distinct timepoint lorlatinib-resistant cases. ROS1 resistance mutations were detected in 13 post-lorlatinib cases (46%; Fig. 2; Supplementary Fig. S2). While the frequency of ROS1 kinase domain mutations was numerically higher in lorlatinib-resistant as compared with crizotinib-resistant specimens (46% vs. 38%, respectively), this difference was not statistically significant (P = 0.621).

The most common ROS1 mutation identified in nine post-lorlatinib samples (32%) was again ROS1 G2032R. Four of these nine cases with ROS1 G2032R had matched post-crizotinib/pre-lorlatinib specimens analyzed (Supplementary Tables S1 and S2). In three of these four cases, ROS1 G2032R was preexisting in the post-crizotinib samples, and TTP on lorlatinib for these patients ranged from 3.2 to 6.1 months. In the remaining case, MGH0024, the post-crizotinib/pre-lorlatinib liver specimen (MGH0024.A) was found to harbor a ROS1 S1986F mutation, while whole-exome sequencing of the subsequent lorlatinib-resistant liver specimen (TTP on lorlatinib, 5.1 months) revealed ROS1 G2032R and L2086F mutations in addition to the previously seen ROS1 S1986F. We were unable to determine whether the mutations were present in cis or in trans due to the distance between the nucleotides. Of note, sequencing of the post-lorlatinib plasma (MGH0024.B) from this patient also detected ROS1 G2032R (VAF, 12.3%), which had been absent in the post-crizotinib/pre-lorlatinib plasma (MGH0024.A). This liquid NGS platform did not assess the exons covering ROS1 S1986 or L2086 mutations. Therefore, in the case of MGH0024, ROS1 G2032R was most likely acquired on lorlatinib.

Outside of ROS1 G2032R, we identified a ROS1 L2086F mutation, alone or together with other ROS1 mutation(s), in three lorlatinib-resistant biopsies (11%): the aforementioned MGH0024, MGH0026, and MGH0035 (Supplementary Fig. S2). ROS1^L2086F is analogous to ALK^L1256F, which confers resistance to crizotinib and lorlatinib in ALK fusion–positive lung cancer in preclinical models (24, 25). In addition, a ROS1 double mutation, S1986F and L2000V in cis, was observed in the lorlatinib-resistant left-sided pleural fluid sample of MGH0032.B (Fig. 2A). This patient's crizotinib-resistant/pre-lorlatinib right pleural fluid detected no ROS1 mutations. In total, three post-lorlatinib biopsies (11%) were found to harbor 2 ROS1 resistance mutations (Fig. 2A), of which two included G2032R. All three cases had received prior crizotinib.

Overall, 11 of the 28 lorlatinib-resistant cases had matched post-crizotinib/pre-lorlatinib biopsies analyzed by NGS (Supplementary Table S2). Six of these 11 cases (55%) had no ROS1 mutation detected in either post-crizotinib or post-lorlatinib specimen. Three cases (27%) acquired new ROS1 mutations on lorlatinib. Specifically, in two cases (18%), no ROS1 mutations were detected post-crizotinib, whereas either ROS1 G2032R (MGH0039.B) or S1986F/L2000V (MGH0032.B) mutations were identified post-lorlatinib; and in one case (9%), ROS1 S1986F was detected pre-lorlatinib, with the identification of two additional ROS1 mutations, G2032R and L2086F, post-lorlatinib (MGH0024.B). The remaining three cases had ROS1 G2032R in both post-crizotinib and post-lorlatinib biopsies (MGH0036, MGH0039, and UCI0002).

To assess the functional role and potential clinical implications of the spectrum of ROS1 mutations detected in our cohort, we generated Ba/F3 cell lines expressing nonmutant CD74-ROS1 or mutant CD74-ROS1 harboring G2032R, L2000V, L2086F, S1986F/L2000V, S1986F/L2086F, S1986F/G2032R, G2032R/L2086F, or S1986F/G2032R/L2086F. Cells were treated with the FDA-approved ROS1 inhibitors, crizotinib and entrectinib, or other TKIs with activity against ROS1 (lorlatinib, repotrectinib, cabozantinib, ceritinib, brigatinib, and taletrectinib), as well as alectinib (an ALK TKI without any ROS1 activity), and cell viability was measured after 2 days to determine the potency of these TKIs against the various mutant ROS1 kinases.

We and others have reported previously ROS1 G2032R as a solvent front mutation conferring resistance to crizotinib (11, 12, 26). In the Ba/F3 models, ROS1^G2032R was indeed refractory to crizotinib, entrectinib, ceritinib, and brigatinib. Lorlatinib was not potent against this mutation, with an IC₅₀ value of 196.6 nmol/L (vs. 0.7 nmol/L with ROS1^wild-type; Fig. 3A and B). The next-generation ROS1/TRK inhibitor, repotrectinib, and a type II TKI, cabozantinib, maintained potency against ROS1^G2032R (IC₅₀ value, 23.1 and 17.5 nmol/L, respectively), while taletrectinib had moderate potency against ROS1^G2032R (IC₅₀ value, 53.3 vs. 2.6 nmol/L with ROS1^wild-type).

ROS1^L2086F conferred resistance to lorlatinib in this Ba/F3 model (Fig. 3A and B; Supplementary Fig. S3), in concordance with the clinical detection of ROS1 L2086F in the lorlatinib-resistant biopsies. ROS1^L2086F was notably refractory to the majority of other currently available ROS1 inhibitors. Brigatinib had a relatively lower IC₅₀ value for ROS1^L2086F, although remained 17-fold less potent against this mutant compared with the nonmutant kinase (IC₅₀ value, 159.3 nmol/L with ROS1^L2086F vs. 9.4 nmol/L with ROS1^wild-type). The L2086F-harboring compound mutant kinases, ROS1^{S1986F/L2086F}, ROS1^{G2032R/L2086F}, and ROS1^{S1986F/G2032R/L2086F}, were also highly refractory to lorlatinib, crizotinib, and entrectinib, as well as ceritinib, taletrectinib, and repotrectinib. Only cabozantinib, a type II multitargeted TKI, maintained strong potency across these ROS1 mutations. Consistent with the observed effects on cell viability, the phosphorylation of ROS1^L2086F was sustained in the presence of most ROS1 inhibitors even at the concentrations of 300 nmol/L (Fig. 3C). In contrast, cabozantinib suppressed ROS1^L2086F (and all L2086F-harboring compound mutant ROS1 kinases) at 30 nmol/L (Supplementary Fig. S4).

As ROS1 L2000V mutation has not been described previously, we also assessed the functional consequence of the ROS1 S1986F/L2000V mutation detected in a lorlatinib-resistant case. Both ROS1^L2000V and ROS1^{S1986F/L2000V} resulted in relatively decreased sensitivity to crizotinib (IC₅₀ value, 37.1 or 159.4 nmol/L, respectively, vs. 5.4 nmol/L with ROS1^wild-type; Fig. 3A; Supplementary Figs. S3 and S5). On the other hand, neither mutant kinase caused resistance to lorlatinib in the Ba/F3 model (IC₅₀ value, 2.5 or 2.4 nmol/L, respectively, vs. 0.7 nmol/L with ROS1^wild-type), highlighting the importance of functionally validating on-target genetic alterations.

Next, we performed structural modeling of the refractory ROS1 L2086F mutation. ROS1^L2086F was predicted to cause severe steric clash against the binding of lorlatinib, in line with the Ba/F3 drug sensitivity results. In particular, the fluorophenyl group of lorlatinib that fills the pocket formed by G2101, N2084, and R2083 clashes with ROS1^L2086F (Fig. 4A), with the rigid, cyclic nature of lorlatinib not permitting independent bond rotation for better accommodation of the larger phenyl side chain at L2086 (Supplementary Fig. S6A). On the other hand, the binding of cabozantinib could be less affected, as modeled using a close analogue, foretinib (27), and MET in DFG-out state (there are no reported structural data for cabozantinib bound to ROS1, or for ROS1 DFG-out conformation; Supplementary Fig. S6B and S6C).

MGH0026 was one of the lorlatinib-resistant cases in this cohort found to have an acquired ROS1 L2086F mutation. This patient with advanced ROS1⁺ NSCLC had received prior chemotherapy, immunotherapy, followed by crizotinib with disease progression in the brain (Fig. 4B). Approximately 17 months after subsequent therapy with lorlatinib, biopsy of a progressing lorlatinib-resistant lymph node revealed the CD74-ROS1 fusion and the ROS1 L2086F mutation (other alterations shown in Supplementary Fig. S2). On the basis of the above preclinical data, this patient was then treated with cabozantinib. The patient had disease stabilization on cabozantinib, with duration of therapy lasting nearly 11 months (Fig. 4C).

ROS1 resistance mutations were not identified in approximately 50%–60% of this cohort, suggesting a role for ROS1-independent or bypass mechanism (Fig. 2B). Therefore, we investigated potential ROS1-independent mechanisms of resistance in these specimens using the NGS results.

MET amplification is a known mechanism of bypass signaling across multiple subsets of NSCLC, including EGFR-, ALK-, and RET-driven NSCLC (22, 28–32). MET amplification was identified in two (2.9%) biopsies of our ROS1 cohort. MGH0034 with advanced CD74-ROS1⁺ NSCLC had no evidence of MET amplification in the treatment-naïve pleural fluid (MET FISH negative with MET:CEP7 ratio of 1.0), and received crizotinib for 4.7 months, followed by lorlatinib. The lorlatinib-resistant brain specimen demonstrated the known CD74-ROS1 fusion and acquired MET gain confirmed by FISH (MET:CEP7 ratio of 6.3; Supplementary Fig. S2). In addition, acquired MET amplification was detected in the crizotinib-resistant adrenal biopsy of patient MGH0014 by NGS and MET FISH (MET:CEP7 ratio >25:1; Supplementary Fig. S2). In this patient, paired crizotinib-resistant plasma analysis revealed MET gain, as well as an MET L1195V mutation (VAF, 3.6% and VAF of ROS1-SLC34A2, 0.6%; Fig. 2B), which is known to cause resistance to MET inhibitors, including crizotinib (33). While we identified additional MET alterations in three (4.3%) biopsies, these consisted of MET V1271M (UCI0003) and MET R469Q (MGH0037.B) variants of unclear significance, and a WAC intron 3 to MET intron 17 fusion of unclear significance in a lorlatinib-resistant case with ROS1 G2032R (MGH0022; Supplementary Fig. S2; Supplementary Table S1).

Although activating EGFR mutations, E709K and L858R, were detected in the crizotinib-resistant lung nodule of UCI0004.A (sample insufficient for ROS1 fusion assessment), this case was from a patient who initially presented with multifocal bilateral lung nodules. The biopsied oligoprogressing lung nodule was treated with stereotactic body radiotherapy, and the patient continued on crizotinib for >4 years thereafter before switching to lorlatinib. Plasma biopsy at the time of progression on lorlatinib detected neither ROS1 fusion nor EGFR mutations, but rather identified a KRAS Q61H mutation, raising the possibility of multifocal lung cancer. Three additional cases had EGFR variants of unlikely functional consequence (MGH0039.A with EGFR R521K polymorphism; and UCI0003 and UCI0007 with EGFR S1120R and R377S mutations, respectively, both outside the protein kinase domain).

Alterations in genes affecting the RAS-MAPK signaling node (such as NF1, MAP2K1, MAP3K1, and KRAS) may mediate resistance to ROS1 inhibitors (34). As an example, KRAS G12D mutation has been reported in a patient with crizotinib-resistant ROS1⁺ lung adenocarcinoma (35). Overall, we observed KRAS mutations in six time distinct biopsies (8.6%), two of which were canonical alterations KRAS G12C and Q61H (MSKCC51 with acquired KRAS G12C which was absent in the treatment-naïve tissue and UCI0004.B with KRAS Q61H, respectively), and four of which involved variants of unknown significance (VUS; V8I in MGH0003.A and MGH0003.B, I24N in MGH0018, and V114A in MGH0035; Supplementary Fig. S2). In addition, KRAS amplification concomitant with CD74-ROS1 fusion was detected in the lorlatinib-resistant lung and plasma biopsies of MGH0028, whose treatment-naïve biopsy did not harbor KRAS amplification. One lorlatinib-resistant case (MSKCC11; published previously; ref. 34) had the MAP2K1 E41_L54del in-frame deletion mutation, a recently described oncogenic driver mutation. Another lorlatinib-resistant plasma (MGH0038.B) had an acquired pathogenic NRAS G60E mutation, which affects the nucleotide binding domain and results in activation of downstream signaling (Fig. 2B). Finally, NF1 alterations were identified in five TKI-resistant cases (7%; all crizotinib resistant), including two cases with loss-of-function mutations [NF1 Q756Ter (MGH0004) and NF1 R69fs*7 (UCI0003); Fig. 2B], one case with possibly loss-of-function mutation [NF1 c.7458–1G>C splice acceptor variant (MGH0014, also with MET gain and MET L1195V)], and two cases with VUS [NF1 E1516D (MGH0018) and P1867 L (UCI0007)]. It is not known whether the NF1 alterations may have existed pre-crizotinib in these cases.

We did not identify canonical pathogenic mutations in other drivers, such as ALK, BRAF, or ERBB2 (those marked in Supplementary Fig. S2 for these genes were VUS). Oncogenic fusions, involving ALK, ROS1, NTRK1–3, RET, BRAF, or NRG1 genes, were not detected.

Since the initial report of oncogenic ROS1 fusions in lung cancer in 2007 (36), two TKIs with ROS1 activity, crizotinib and entrectinib, have received approval across multiple countries for the treatment of advanced ROS1⁺ NSCLC and are considered standard of care (3, 4, 8). In addition, lorlatinib, a highly brain-penetrant, next-generation ALK/ROS1 TKI, has demonstrated efficacy in patients with advanced ROS1⁺ lung cancer, including in a subset of patients who had received prior crizotinib (9). However, resistance to these TKIs is inevitable, causing disease relapse in most patients. While ROS1 mutations causing resistance to crizotinib have been reported, the studies to date have been limited to isolated case reports or small series. Questions remain regarding the scope of mechanisms of resistance to other ROS1 inhibitors, such as lorlatinib in ROS1⁺ disease, and ultimately regarding how to consider the optimal sequencing of various ROS1 inhibitors currently in clinical development.

In this multi-institutional study, we evaluated a cohort of patients with ROS1⁺ NSCLC treated with crizotinib and/or lorlatinib who had at least one posttreatment biopsy. We identified ROS1 kinase domain mutations in 38% of crizotinib-resistant biopsies and in 46% of lorlatinib-resistant biopsies. In both instances, ROS1 G2032R represented the predominant mutation, detected in approximately one-third of cases. This solvent front mutation has been described previously in the setting of crizotinib resistance, where it has been shown to cause steric clash with crizotinib binding (11). In our series, ROS1 G2032R was identified in 32% of post-lorlatinib samples, and lorlatinib was unable to suppress ROS1 G2032R in patients known to have this mutation pre-lorlatinib. Furthermore, we observed development of this mutation in a patient on lorlatinib therapy. Thus, in contrast to ALK fusion–positive lung cancer, in which lorlatinib potently overcomes solvent front resistance mutations (e.g., analogous ALK G1202R), the experience with lorlatinib and ROS1 G2032R in ROS1⁺ lung cancer appears distinct. This is further validated by our preclinical models, which demonstrate the decreased potency of lorlatinib, as well as crizotinib, entrectinib, ceritinib, and brigatinib against ROS1 G2032R. In a phase I/II study of lorlatinib, 6 patients were known to have baseline ROS1 G2032R from cfDNA or tumor analysis; of these, 5 patients had stable disease lasting between 2.9 and 9.6 months, 1 patient had primary progressive disease, and no objective responses were observed on lorlatinib (9), again consistent with modest potency of this TKI against ROS1^G2032R.

Outside of ROS1 G2032R, ROS1 L2086F was detected as a recurring resistance mutation in the post-lorlatinib biopsies. ROS1 L2086F is analogous to the ALK L1256F mutation, which causes steric clash with the fluorobenzene group of lorlatinib thereby conferring resistance to lorlatinib in ALK fusion–positive lung cancer (24, 25), and this mutation has been described in one case resistant to taletrectinib (37). Our structural modeling predicts that ROS1^L2086F causes steric interference with the binding of lorlatinib in addition to other type I ROS1 inhibitors, such as crizotinib and entrectinib, but may accommodate a type II inhibitor, cabozantinib. Indeed, in Ba/F3 models, ROS1^L2086F single mutant kinase and ROS1^L2086F-based compound mutants were highly refractory to the type I ROS1 inhibitors, whereas cabozantinib maintained potency against the mutant ROS1 kinases. Furthermore, a patient with prior progression on crizotinib and lorlatinib was able to receive cabozantinib with disease control lasting nearly 11 months. These findings are reminiscent of observations in TRK fusion–positive tumors, wherein TRK xDFG mutations conferred resistance to type I TRK inhibitors through steric hindrance and represented a shared liability, but sensitized tumors to type II TRK inhibitors (38).

Our findings may help inform the approach of sequential TKI therapy using alternative ROS1 inhibitors in ROS1⁺ lung cancer, similar to the previously modeled paradigm in EGFR-mutant and ALK fusion–positive disease. For patients with disease progression on prior crizotinib or entrectinib and known ROS1 G2032R, lorlatinib (as with ceritinib or brigatinib) is unlikely to provide significant durable benefit, and alternative inhibitors, such as repotrectinib or taletrectinib, may be explored on a clinical trial. Repotrectinib is a next-generation ROS1/TRK inhibitor that has demonstrated preliminary efficacy in ROS1 TKI-naïve and TKI-pretreated NSCLC; responses have been observed in patients with baseline ROS1^G2032R (39, 40), although the extent and durability of these responses await further determination. In the setting of a known ROS1 L2086F mutation, either alone or together with another mutation (e.g., G2032R or S1986F), other investigational inhibitors, including lorlatinib, taletrectinib, and repotrectinib, are predicted to be ineffective. Instead, a role for switching to a type II ROS1 inhibitor, cabozantinib, may be studied in this setting and potentially against xDFG ROS1 mutations (G1201A and G1201C; ref. 38). Cabozantinib as a multikinase inhibitor can notably be associated with significant and at times intolerable toxicities that necessitate dose interruptions and reductions or even treatment discontinuations (41). Therefore, there remains a need to develop novel, selective, and tolerable ROS1 TKIs including rationally designed type II inhibitors, which retain potency against the spectrum of known ROS1 resistance mutations, such as the recurring G2032R and L2086F.

In approximately two-thirds and one-half of cases with resistance to crizotinib and lorlatinib, respectively, a ROS1 kinase domain mutation was not identified, suggestive of ROS1-independent resistance mechanisms. We investigated potential ROS1-independent mechanisms mediating resistance in this series using genetic sequencing. RAS-MAPK pathway activation has been implicated previously in ROS1 inhibitor resistance (34, 35, 42). In this cohort, pathogenic KRAS G12C mutation, KRAS amplification, NRAS G60E mutation, and MAP2K1 E41_L54del in-frame deletion mutation were identified in one lorlatinib-resistant case each, and NF1 alterations were identified in five crizotinib-resistant cases, providing further support for the potential role of RAS-MAPK pathway dysregulation in these tumors. We additionally identified high-level MET amplification in one lorlatinib-resistant case without concurrent ROS1 kinase domain mutations. Interestingly, acquired high-level MET amplification was also detected in one case resistant to crizotinib, an ALK/ROS1/MET inhibitor. In this crizotinib-resistant case, the MET amplification cooccurred with an MET L1195V mutation (detected in plasma), which is known to alter MET topology and cause decreased sensitivity to MET inhibitors, including crizotinib (33, 43). It is, therefore, plausible that this patient's tumor was able to overcome the constraint on MET and ROS1 imposed by crizotinib via selective expansion of a more highly MET-amplified subclone additionally harboring the MET resistance mutation. Thus, MET pathway, known to mediate resistance in multiple other lung cancers, including EGFR-mutant, ALK fusion–positive, and RET fusion–positive NSCLC (28–32), may represent another recurring bypass mechanism in ROS1⁺ lung cancer after treatment with ROS1 inhibitors. Indeed, the true frequency of MET amplification in ROS1 TKI-resistant cases may be underestimated in this series as all patients herein (including those who had post-lorlatinib biopsies) received prior crizotinib. Broadly, these cases with RAS-MAPK or MET alterations, including a KRAS G12C mutation which is now druggable with covalent inhibitor of KRAS^G12C (44, 45), support further investigation of potential combination strategies after resistance to ROS1 TKI(s) develops. It is additionally worth noting that cabozantinib as a multikinase inhibitor could be considered not only for certain ROS1 kinase domain mutations, but also for acquired MET dependency in ROS1⁺ tumors (46).

Our study had several important limitations. It was a retrospective study with a relatively small sample size, although remains the largest multi-institutional dataset analyzed to date. Second, selection bias cannot be excluded particularly in the lorlatinib-resistant cohort, as patients known to have ROS1 resistance mutations in the tumor may have been preferentially directed toward lorlatinib as subsequent therapy; this may have resulted in a higher frequency of ROS1 mutations in post-lorlatinib biopsies. Third, matched sequencing data from pre-lorlatinib biopsies were generally not available, and, therefore, we could not always determine whether gene alterations were preexisting versus acquired on lorlatinib. We have indicated which cases had prior biopsies available and the corresponding sequencing results to inform data interpretation. In certain cases, discrepancies were identified between matched tumor and plasma analyses, likely owing to multiple factors, including tumor heterogeneity. In addition, not all NGS platforms consistently assessed for certain gene alterations, including the ROS1 L2086F mutation (likely because its incidence and consequence were not appreciated previously, and because L2086 resides in a separate exon from the previously known G2032). Recognizing this, in at least a subset of the cases analyzed by NGS platform where L2086 was not covered, we performed Sanger sequencing of the ROS1 kinase domain. It remains possible that ROS1 L2086F may mediate resistance in higher proportion of cases than is appreciable from this study. Furthermore, this limitation highlights the importance of elucidating functional and clinical implications of diverse gene alterations, including and extending beyond ROS1, to ensure that clinical NGS panels incorporate the assessment of these alterations and do not inadvertently miss a finding that could guide treatment recommendations. Finally, approximately 40% of the lorlatinib-resistant cohort had unknown mechanisms of resistance based on gene sequencing alone. Future studies will need to further investigate mechanisms of signaling dysregulation at the epigenetic, RNA, protein, and phospho-protein levels.

In summary, we demonstrated that ROS1 kinase domain mutations are identified in more than one-third of crizotinib-resistant and nearly one-half of lorlatinib-resistant cases, and remain an important hurdle to overcome in the treatment of ROS1⁺ lung cancer. Our clinical and preclinical findings suggest lorlatinib is not sufficiently potent against the known crizotinib-resistant mutation ROS1 G2032R, and additionally lacks potency against a previously less well-known mutation ROS1 L2086F, which was refractory to all ROS1 inhibitors tested in this study, except for the type II multikinase inhibitor, cabozantinib. This work highlights the potential utility of rebiopsy and sequential TKI therapy in ROS1⁺ lung cancer. Moving forward, it will be important to continue the ongoing efforts to develop novel, selective ROS1 inhibitors with potency against the known spectrum of ROS1 resistance mutations. In addition, our work adds insights into potential ROS1-independent resistance mechanisms, such as MET and RAS-MAPK alterations, which may support the exploration of combination strategies if further validated.

J.J. Lin reports personal fees from Genentech, C4 Therapeutics, Nuvalent, Blueprint Medicines, and Pfizer; personal fees and other from Turning Point Therapeutics and Elevation Oncology; and other from Hengrui Therapeutics, Novartis, Neon Therapeutics, Relay Therapeutics, Bayer, and Roche outside the submitted work. S. Yoda reports grants from Lung Cancer Research Foundation during the conduct of the study and personal fees from Pfizer Japan outside the submitted work. V.W. Zhu reports personal fees from AstraZeneca, Blueprint, Roche Foundation Medicine, Roche/Genentech, Takeda, TP Therapeutics, and Xcovery outside the submitted work. T.W. Johnson reports personal fees from Pfizer, Inc. outside the submitted work. I. Dagogo-Jack reports grants and personal fees from Pfizer; grants from Array, Novartis, and Genentech; and personal fees from AstraZeneca, Novocure, Syros, Catalyst Pharmaceuticals, Xcovery, and BostonGene outside the submitted work. S.R. Digumarthy reports other from Merck, Pfizer, Bristol Myers Squibb, Novartis, Roche, Polaris, Cascadian, Zai Laboratories, Clinical Bayer, Gradalis, AbbVie, Siemens Healthineers, and Jenssen and grants from Lunit INC outside the submitted work. L.V. Sequist reports grants and personal fees from AstraZeneca and Genentech and grants from Novartis, Boehringer Ingelheim, LOXO, Blueprint Medicines, and Merrimack Pharmaceuticals outside the submitted work. B.Y. Yeap reports grants from NIH during the conduct of the study. A.T. Shaw reports grants and personal fees from Pfizer during the conduct of the study, as well as other from Novartis and personal fees from Genentech/Roche, TP Therapeutics, Blueprint, KSQ Therapeutics, Chugai, Bayer, and Syros outside the submitted work. S.-H.I. Ou reports personal fees from Pfizer, Roche/Genentech, Takeda/ARIAD, AstraZeneca, Daiichi Sankyo, and Janssen/JNJ and other from Turning Point Therapeutics and Elevation Oncology outside the submitted work. A.N. Hata reports grants from Pfizer, Amgen, Roche/Genentech, Blueprint Medicines, Eli Lilly, and Relay Therapeutics outside the submitted work. A. Drilon reports personal fees from Pfizer during the conduct of the study; personal fees from Ignyta/Genentech/Roche, Loxo/Bayer/Lilly, Takeda/Ariad/Millenium, TP Therapeutics, AstraZeneca, Blueprint Medicines, Helsinn, BeiGene, BerGenBio, Hengrui Therapeutics, Exelixis, Tyra Biosciences, Verastem, MORE Health, AbbVie, 14ner/Elevation Oncology, Remedica Ltd., ArcherDX, Monopteros, Novartis, EMD Serono, Melendi, Liberum, and Repare RX outside the submitted work; associated research to institution from Pfizer, Exelixis, GlaxoSmithKline, Teva, Taiho, and PharmaMar; royalties from Wolters Kluwer; other from Merck, Puma, Merus, and Boehringer Ingelheim; and CME honoraria from Medscape, OncLive, PeerVoice, Physicians Education Resources, Targeted Oncology, Research to Practice, Axis, Peerview Institute, Paradigm Medical Communications, WebMD, and MJH Life Sciences. J.F. Gainor reports personal fees and other from Bristol-Myers Squibb, Genentech/Roche, Blueprint, Jounce, and Merck; other from Tesaro, Moderna, and Alexo; and personal fees from Takeda, Loxo/Lilly, Oncorus, Regeneron, Gilead, AstraZeneca, Pfizer, Novartis, Agios, Amgen, and Array outside the submitted work, and has an immediate family member who is an employee with equity in Ironwood Pharmaceuticals. No disclosures were reported by the other authors.

J.J. Lin: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. N.J. Choudhury: Resources, data curation, investigation, writing–review and editing. S. Yoda: Resources, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. V.W. Zhu: Resources, data curation, investigation, writing–review and editing. T.W. Johnson: Formal analysis, investigation, visualization, writing–original draft, writing–review and editing. R. Sakhtemani: Formal analysis, investigation, writing–review and editing. I. Dagogo-Jack: Resources, data curation, investigation, writing–review and editing. S.R. Digumarthy: Investigation, visualization, writing–review and editing. C. Lee: Data curation, investigation, writing–review and editing. A. Do: Data curation, writing–review and editing. J. Peterson: Data curation, writing–review and editing. K. Prutisto-Chang: Data curation, investigation, writing–review and editing. W. Malik: Data curation, investigation, writing–review and editing. H.G. Hubbeling: Data curation, writing–review and editing. A. Langenbucher: Formal analysis, investigation, writing–review and editing. A.J. Schoenfeld: Data curation, investigation, writing–review and editing. C.J. Falcon: Data curation, writing–review and editing. J.S. Temel: Resources, writing–review and editing. L.V. Sequist: Resources, writing–review and editing. B.Y. Yeap: Formal analysis, investigation, methodology, writing–review and editing. J.K. Lennerz: Resources, writing–review and editing. A.T. Shaw: Conceptualization, resources, writing–review and editing. M.S. Lawrence: Software, formal analysis, supervision, investigation, writing–review and editing. S.-H.I. Ou: Resources, supervision, writing–review and editing. A.N. Hata: Resources, supervision, project administration, writing–review and editing. A. Drilon: Resources, supervision, writing–review and editing. J.F. Gainor: Conceptualization, resources, supervision, funding acquisition, project administration, writing–review and editing.

This work was supported by the Targeting a Cure for Lung Cancer Research Fund and by Be A Piece of the Solution at Massachusetts General Hospital. J.J. Lin and B.Y. Yeap were partially supported by funding from R01-CA164273. S. Yoda was partially supported by Lung Cancer Research Foundation Annual Grant Program award.

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.