Zidesamtinib Selective Targeting of Diverse ROS1 Drug-Resistant Mutations

Abstract

Zidesamtinib (NVL-520) is a ROS1-selective macrocyclic tyrosine kinase inhibitor designed with the aim to address clinical challenges for patients with non–small cell lung or other cancers that are ROS1 fusion–positive. These challenges include emergent ROS1 resistance mutations and brain metastases that can lead to disease progression and central nervous system adverse events attributed to off-target tropomyosin-related kinase inhibition that can be treatment-limiting. We evaluated zidesamtinib in accelerated mutagenesis screens and a brain tumor model, comparing it with other approved or investigational ROS1 inhibitors. At clinically relevant concentrations, zidesamtinib robustly inhibited >1,500 pooled ROS1 mutants with virtually no resistance emerging (≤1%), outperforming comparators crizotinib, entrectinib, and repotrectinib. Zidesamtinib also induced more durable responses than repotrectinib and taletrectinib in an aggressive intracranial ROS1 G2032R xenograft model. A 2.2 Å cocrystal structure with ROS1 G2032R, the most frequently identified ROS1 resistance mutation, reveals that zidesamtinib uniquely accommodates the mutated residue while potentially clashing with tropomyosin-related kinases, consistent with its selective ROS1-targeting design and supported by computational modeling. Taken together, these data support zidesamtinib’s potential as a novel best-in-class ROS1 inhibitor.

Introduction

ROS1 receptor tyrosine kinase fusions are clinically validated oncogenic drivers (1). Present in 1% to 3% of advanced non–small cell lung cancer (NSCLC) and in other solid tumors, ROS1 fusions activate downstream signaling and promote cancer cell transformation (2–9). Tyrosine kinase inhibitors (TKI) crizotinib, entrectinib, and repotrectinib are approved by the U.S. FDA for the treatment of patients with ROS1 fusion–positive NSCLC (10–12), and lorlatinib and taletrectinib are undergoing clinical evaluation for this indication (13–15). However, none of these treatment options addresses the combined medical need that includes broad activity against ROS1 resistance mutations, brain penetrance, and avoidance of tropomyosin-related kinase (TRK) inhibition. Acquired ROS1 kinase domain mutations account for 30% to 50% of all patients progressing on crizotinib, entrectinib, and lorlatinib, with the ROS1 G2032R “solvent-front” substitution occurring most frequently (16, 17). Brain metastases are identified in 20% to 40% of patients at diagnosis and represent the sole site of progression in 50% of crizotinib-treated patients (18, 19). Finally, concurrent inhibition of TRKA/B/C due to structural similarities to ROS1, particularly TRKB in the central nervous system (CNS), has been associated with neurologic toxicities (20, 21) observed with entrectinib (22), repotrectinib (23), lorlatinib (24), and taletrectinib (25). These toxicities have the potential to impact durability of therapeutic response and have been reported to be dose-limiting for entrectinib (22), repotrectinib (23), and lorlatinib (26).

Zidesamtinib (NVL-520) is a novel, selective, and brain-penetrant macrocyclic ROS1 TKI. Preclinically, zidesamtinib inhibits cancer cells harboring diverse ROS1 fusions and resistant mutations, including in a brain tumor model, with selectivity over TRKB (27). Computational modeling suggests that zidesamtinib achieves selectivity by accommodating the mutant Arg2032 in ROS1 while clashing with Tyr591 in TRKA (27). In the ongoing ARROS-1 phase 1/2 clinical trial (NCT05118789), zidesamtinib has demonstrated preliminary efficacy and tolerability in heavily pretreated patients, including those who had received prior lorlatinib or repotrectinib, had brain metastases, or harbored ROS1 G2032R tumors (28).

The current work seeks to deepen the preclinical understanding of zidesamtinib as a potential best-in-class ROS1 therapy by addressing three gaps in the current literature. First, collective experience in the NSCLC TKI space suggests that a TKI’s ability to delay resistance, both peripherally and intracranially, is a key driver of response durability (29), and this has not been fully evaluated for zidesamtinib preclinically. Second, previous studies did not consider activity at clinically relevant concentrations of all ROS1 TKIs, which have only recently been reported for repotrectinib (12) and zidesamtinib (28). Third, prior structural studies have been hindered by a lack of ROS1 G2032R crystal structures. Our findings in this work show that, at clinically relevant concentrations, zidesamtinib suppressed on-target resistance and inhibited ROS1 G2032R brain tumors more effectively than other ROS1 TKIs. We also present the first crystal structure of ROS1 G2032R in complex with zidesamtinib, providing a model for computationally predicting small-molecule binding. Together, these studies support zidesamtinib’s potential as a best-in-class ROS1 TKI and offer a framework for understanding ROS1 TKI activity against ROS1 G2032R.

Materials and Methods

Compounds

Compounds were purchased from commercial sources: crizotinib (Combi-Blocks, #QE-4059; MedChemExpress, #HY-50878; SelleckChem), entrectinib (Ark Pharm Inc., #AK547001; Advanced ChemBlocks, #K13380; SelleckChem), lorlatinib (Ark Pharm Inc., #AK175603; eNovation Chemicals, #Y0975931; SelleckChem), taletrectinib (MedChemExpress, #HY-131003), repotrectinib (Ark Pharm Inc., #AK689211; MedChemExpress, #HY-103022; SelleckChem, #S8583; DC Chemicals, #DC10463), and N-ethyl-N-nitrosourea (ENU; Sigma, #D2650). Zidesamtinib (NVL-520) was identified from a structure–activity relationship effort to discover kinase inhibitors potent against wild-type (WT) and mutant ROS1 and selective against the TRK family of receptor tyrosine kinases. Synthesis of zidesamtinib can be found in U.S. Patent No. 11542278, Example 5 (30).

Cell culture

All cells were maintained at 37°C with 5% CO₂ and used within 6 months after thawing. Ba/F3 CD74-ROS1 (generated in August 2019), Ba/F3 CD74-ROS1 G2032R (generated in August 2019), and Ba/F3 CD74-ROS1 G2032R luciferase cells (generated in May 2020) were generated as previously described from Ba/F3 cells (biological sex not specified; obtained in October 2018; RRID: CVCL_0161) provided by the RIKEN BRC through the National Bio-Resource Project of the MEXT, Japan, validated for target gene expression by Sanger sequencing and Western blot, maintained in RPMI 1640 + 10% FBS, and used at passage number 12 or less (27). IL3 (R&D Systems, 403-ML-100) was used at 10 ng/mL. MDCK–MDR1 cells were obtained from the Netherlands Cancer Institute, Amsterdam (Borst strain; female; obtained in January 2015, September 2019, and March 2021; RRID: CVCL_S586) or the NIH, United States (NIH strain; biological sex not specified; obtained in July 2019), validated for transporter expression by qPCR and Western blot, maintained in DMEM + 10% FBS, and used at passage number 28 or less (within 5 passages after thawing). All cells were inspected for Mycoplasma contamination using ABM PCR Mycoplasma Detection Kit (cat. #G238) or Lonza MycoAlert Detection Kit (LT07–710 or LT07–318) at a frequency of at least once every 3 months.

Estimating clinical exposure

Average unbound plasma exposure (C_avg,u) was calculated by dividing the AUC at steady state (AUC_ss) with the corresponding dosing interval (τ), adjusted with fraction unbound (f_u) in human plasma and molecular weight to obtain a nmol/L unit. AUC_ss and τ values were obtained from scientific publications for crizotinib (31) and taletrectinib (25, 32) and from the FDA prescribing information for entrectinib (11), lorlatinib (33), and repotrectinib (12). Values of f_u in human plasma were obtained from the FDA prescribing information for crizotinib (10), lorlatinib (33), and repotrectinib (12) and from internal testing for entrectinib and taletrectinib via rapid equilibrium dialysis. The C_avg,u of zidesamtinib was determined internally as of a May 17, 2023 data cut.

Estimating in vitro potency

Ba/F3 CD74-ROS1 and Ba/F3 CD74-ROS1 G2032R cells were seeded at 2 × 10³ cells per well in 96-well plates and treated with a titrated range of TKIs in full media containing 10% FBS, in duplicate or triplicate. For Ba/F3 CD74-ROS1, TKIs included crizotinib (40–800 nmol/L), entrectinib (20–400 nmol/L), taletrectinib (10–200 nmol/L), repotrectinib (5–100 nmol/L), and zidesamtinib (1–20 nmol/L). For Ba/F3 CD74-ROS1 G2032R, TKIs included taletrectinib (50–1,000 nmol/L), repotrectinib (25–500 nmol/L), and zidesamtinib (2.5–100 nmol/L). Full TKI concentrations were converted to unbound TKI concentrations as indicated on graphs by adjusting for f_u in 10% FBS, determined internally via rapid equilibrium dialysis. Media and TKIs were refreshed approximately once a week, and cell numbers were determined by CellTiter-Glo (Promega) approximately twice a week.

Mutagenesis screen

ENU mutagenesis

We treated 3 × 10⁸ Ba/F3 CD74-ROS1 (in ∼300 mL media) and 1.5 × 10⁸ Ba/F3 CD74-ROS1 G2032R cells (in ∼150 mL media) with ENU (100 μg/mL) in complete medium for 16 hours. Cells were then washed with DPBS, resuspended in complete medium, and left to recover for 24 hours. After a media refresh, 3 × 10⁶ cells from each screen were saved for next-generation sequencing (NGS) analysis. To prepare for compound treatment, remaining cells were divided into 12 and 6 aliquots for CD74-ROS1 and CD74-ROS1 G2032R, respectively, for a total of 3 × 10⁷ cells in 60 mL per aliquot. The 12 aliquots of Ba/F3 CD74-ROS1 were treated with crizotinib (80, 200, and 400 nmol/L), entrectinib (40, 100, and 200 nmol/L), repotrectinib (10, 25, and 50 nmol/L), and zidesamtinib (2, 5, and 10 nmol/L). The six aliquots of Ba/F3 CD74-ROS1 G2032R were treated with repotrectinib (120, 250, and 500 nmol/L) and zidesamtinib (10, 25, and 50 nmol/L). Each aliquot was distributed into 5 × 96-well plates (total 480 wells per aliquot) at 5 × 10⁴ cells in 100 μL per well. Because there were 18 aliquots, this translated to 90 plates in total. Cells were incubated for 28 days. The medium was refreshed approximately once a week. Wells exhibiting outgrowth were counted on day 28.

Sanger sequencing

In up to 50 resistant wells from each of the 18 experimental arms, 5 × 10⁴ cells were collected for Sanger sequencing. For experimental arms with more than 50 resistant wells, only 50 were arbitrarily selected as representative. Genomic DNA was obtained using QuickExtract DNA Extraction Solution (Epicentre #QE09050) at 65°C for 15 minutes followed by 98°C for 2 minutes. The CD74-ROS1 locus was amplified by PCR using the primer set 5′-CCC​TTG​AAC​CTC​CTC​GTT​CGA​CC-3′ and 5′-GCG​AAG​GGG​CCA​CCA​AAG​AA-3′ using Ex Taq DNA Polymerase (Takara, #RR001B) with 35 rounds of a 3-minute elongation window at 72°C. PCR products were subjected to Sanger sequencing using 4 primers spanning the entire CD74-ROS1 locus to identify any putative resistant kinase domain mutations: 5′-GCA​CCA​TTG​GCT​CCT​GTT​TG-3′, 5′-TCA​TCA​GAT​GTG​CCT​CCT​TC-3′, 5′-GGA​ATG​AGC​TGG​ATA​AGG​CTG-3′, and 5′-CTA​AAG​CGC​ATG​CTC​CAG​ACT-3′. Only kinase domain mutations, defined as ROS1 residues 1945 to 2222 based on the Uniprot.org annotation, were presented.

TKI profiling against resistant clones

Clones harboring distinct ROS1 kinase domain mutations, identified from Sanger sequencing, were expanded. For mutations that were present in more than one clone, only one clone was arbitrarily selected as representative. Cells were seeded into 384-well plates, and test compounds were added in a 3-fold dilution series in complete culture medium. After a 72-hour incubation, cell viability was measured using CellTiter-Glo (Promega). Untreated wells served as negative controls (no inhibition of proliferation), whereas wells treated with high concentrations of the nonspecific kinase inhibitor staurosporine served as positive controls (full inhibition of proliferation). The IC₅₀ value was calculated from percent inhibition and log (inhibitor concentration) using four-parameter logistic regression.

NGS

DNA was extracted from starting cell pools using DNeasy Blood & Tissue Kit according to the manufacturer’s instructions (Qiagen, #69504). The CD74-ROS1 locus was amplified by PCR using the primer set 5′-CCC​TTG​AAC​CTC​CTC​GTT​CGA​CC-3′ and 5′-GCG​AAG​GGG​CCA​CCA​AAG​AA-3′ using Ex Taq DNA Polymerase (Takara, #RR001B) with 35 rounds of a 3-minute elongation window at 72°C. Concentrations were confirmed by NanoDrop to be ∼500 ng/μL. PCR products were fragmented by Covaris S220 ultrasonication to lengths of 200 to 300 base pairs. Fragments were treated with End Prep Enzyme Mix for end repair, 5′-phosphorylation, and dA-tailing, followed by T-A ligation and purification by DNA Clean Beads (Vazyme, #ND627 and #N411). A second PCR reaction was carried out with P5 and P7 primers followed by purification by beads. The qualified libraries were subjected to 150 base paired-end sequencing on the Illumina NovaSeq System.

NGS analysis

Reads were trimmed using Cutadapt v1.9.1 (34) and aligned using Sentieon Genomics software (www.sentieon.com). Variants were detected with VarScan v2.4.2 (github.com/dkoboldt/varscan) using the parental amplicon sequence as reference. Mutant nucleotide frequencies were calculated using the BCFtools (samtools.github.io/bcftools/) and mutant nucleotide sequences were translated to amino acid changes using the R package Biostrings (bioconductor.org/packages/Biostrings; RRID: SCR_016949).

Intracranial xenograft study

Efficacy

All the procedures related to animal handling, care, and treatment in this study were performed according to guidelines approved by the Institutional Animal Care and Use Committee of Pharmaron following the guidance of the Association for Assessment and Accreditation of Laboratory Animal Care International. The intracranial efficacy study was performed similarly to a previously reported protocol (27). Briefly, 1 × 10⁵ Ba/F3 CD74-ROS1 G2032R luciferase cells were stereotactically implanted into the right forebrains of 6- to 8-week-old female Balb/c nude mice. After 5 days, mice were randomized based on mean bioluminescence signal into three groups of n = 6 mice each and received vehicle (twice daily), taletrectinib (100 mg/kg once daily), repotrectinib (15 mg/kg twice daily or 75 mg/kg twice daily), zidesamtinib (3 mg/kg twice daily), or repotrectinib (15 mg/kg twice daily) for the first 7 days followed by zidesamtinib (3 mg/kg twice daily). Taletrectinib was formulated as suspension in 0.5% methyl cellulose (viscosity: 400 cP) in water. Repotrectinib was formulated as suspension in 0.5% carboxymethyl cellulose and 1% Tween-80 in water. Zidesamtinib was formulated as solution in 20% 2-hydroxypropyl-β-cyclodextrin (HP-β-CD) in water, the same formulation used for vehicle treatment. Bioluminescence and body weight were measured at regular intervals until the end of the study (on treatment day 25) or until animals met the criteria for euthanasia. Pairwise statistical analyses were performed using TumGrowth open-access web tool (35).

Pharmacokinetics

At the time of euthanasia, plasma and brain samples were collected from each mouse in the efficacy study, with n = 3 collected at 1 hour and n = 3 collected at τ/2 after dosing (i.e., τ/2 = 6 hours for twice daily and 12 hours for once daily). The three mice were assigned at random for each collection time point. Drug concentrations were measured by LC/MS.

Preclinical brain penetrance

MDCK–MDR1 permeability assay

Permeability studies were conducted with human MDR1-transfected MDCK (MDCK–MDR1–NIH) cells from the NIH (NIH, Bethesda, MD). MDCK–MDR1–NIH cells were diluted to 1.56 × 10⁶ cells/mL with culture medium, and 50 μL of cell suspension was dispensed into each filter well of a 96-well HTS transwell plate. The plate was incubated for 4 to 8 days in a cell culture incubator at 37°C, 5% CO₂, and 95% relative humidity before use. The cell culture medium was replaced every other day during the incubation period. Permeability experiments were conducted using well-established protocols for MDCK permeability assays (36). Experiments were performed by adding 1 μmol/L test compound in Hank’s balanced salt solution added to donor wells and measuring appearance in receiver wells after incubation for 2 hours at 37°C. All incubations were performed in duplicate. Lucifer yellow was used as a marker to confirm the integrity of the cell monolayers after 2 hours of incubation. P_app values and the corresponding efflux ratios (ER) were calculated using the method described (36).

K_p,uu measurement

All the procedures related to animal handling, care, and the treatment in these studies were performed according to guidelines approved by the Institutional Animal Care and Use Committee of Pharmaron following the guidance of the Association for Assessment and Accreditation of Laboratory Animal Care International. Zidesamtinib (suspension), repotrectinib (suspension), and taletrectinib (solution) were formulated at 1 mg/mL in 20% HP-β-CD in deionized water. Lorlatinib was formulated as a 1 mg/mL solution in two equivalents of HCl +20% HP-β-CD in deionized water. Compounds were administered orally to male Wistar Han rats (n = 3 each). After 1 hour, brain samples and plasma were collected, and brain samples were homogenized in PBS. Brain and plasma samples were precipitated by acetonitrile and centrifugation (4,700 rpm, 15 minutes). Drug concentrations in the supernatants were quantified by LC/MS-MS. Unbound fractions were determined using rapid equilibrium dialysis. K_p,uu was calculated as the ratio between unbound drug concentration in the brain and unbound drug concentration in the plasma.

Crystallography and biochemistry

Expression and purification

ROS1 G2032R kinase domain (residues 1934–2232) was cloned into a pFastBac vector with an N-terminal His-tag. Recombinant bacmids were used to generate viruses. Sf21 cells were infected with viruses containing ROS1 G2032R and YopH (tyrosine phosphatase) at 0.13% and 0.04%, respectively, for 66 hours in the presence of 0.5 μmol/L zidesamtinib. ROS1 G2032R was purified as previously described (37). Pellets were lysed in Buffer A containing 50 mmol/L Tris, pH 8.0, 150 mmol/L NaCl, 10% glycerol, and complete protease inhibitor mixture tablet (Roche Applied Science). Clarified lysate was immobilized onto 5 mL of Ni-NTA resin and then washed and eluted with Buffer A supplemented with 20 mmol/L and 300 mmol/L imidazole, respectively. The His-tag was cleaved via tobacco etch virus protease digestion, and samples were loaded onto 5 mL of the HisTrap column for subtractive immobilized metal affinity chromatography in the presence of Buffer A with 20 mmol/L imidazole. Pooled samples were then concentrated and loaded onto a HiLoad Superdex 75 pg column and eluted in storage buffer containing 25 mmol/L HEPES, pH 7.4, 350 mmol/L NaCl, 5% glycerol, and 2 mmol/L tris(2-carboxyethyl)phosphine. Samples were then concentrated to 11 mg/mL, flash-frozen, and stored at −80°C.

Crystallization and structural determination

ROS1 G2032R in the presence of zidesamtinib was crystallized via sitting-drop vapor diffusion. Initially, ROS1 G2032R was preincubated with 2.0 mmol/L zidesamtinib for 1 hour. Drops were formed by combining 0.02 μL of pre-formed crystals, 0.2 μL of the protein solution, and 0.08 μL of the reservoir solution (100 mmol/L citric acid and trisodium citrate, pH 4.83, 600 mmol/L potassium thiocyanate, and 22% v/v PEG 3350). Well-diffracting crystals formed after 39 days when incubated at 12°C. Crystals were briefly swiped through a 25% ethylene glycol solution before mounting. Crystallographic diffraction data were collected at the European Synchrotron Radiation Facility (beamline ID23.1). Indexing, analysis, and scaling were performed using XDS (RRID: SCR_015652), Pointless (RRID: SCR_014218), and STARANISO (staraniso.globalphasing.org/cgi-bin/staraniso.cgi; RRID: SCR_018362), respectively (38–40). The structure was then solved via molecular replacement using WT ROS1 (PDB:4UXL) as a search model. Refinement and model building were conducted using REFMAC5 (41) with restraints for zidesamtinib determined by the program GRADE. Model statistics are defined by REFMAC5.

Michaelis–Menten kinetics

Kinetics was studied using the PhosphoSens platform (42). In a typical reaction, 0.3 nmol/L of WT ROS1 (CarnaBio, Cat #08-163) or 1.5 nmol/L of ROS1 G2032R (CarnaBio, #08-585) were mixed with reaction buffer containing 54 mmol/L HEPES, 1.2 mmol/L dithiothreitol, 0.012% Brij-35, 1% glycerol, 0.2 mg/mL BSA, 0.5 mmol/L triethylene glycol diamine tetraacetic acid, 10 mmol/L MgCl₂, and 13 μmol/L C-Sox sensor (AssayQuant, #AQT0101) with increasing concentrations of ATP. Samples were read continuously for 240 minutes at 30°C using a Biotek Synergy Neo 2 microplate reader at excitation and emission wavelengths of 360 and 485 nm, respectively. Linear regions of fluorescence over time were used to determine Michaelis–Menten kinetics.

Computational modeling

Structure visualization and overlay

Structures were visualized and overlayed using Molecular Operating Environment (MOE; chemcomp.com; RRID: SCR_023932) or the PyMOL Molecular Graphics System (pymol.org; RRID: SCR_000305). Protein–ligand interactions and clashes were visualized using the Contacts function in MOE. Structures used for comparison included PDB codes 3ZBF (37), 4UXL (43), 7Z5W (44), 7Z5X (44), and 7VKO (45). B-factor and crystal packing analyses were conducted in PyMOL. To model in ATP, ROS1 G2032R was aligned with ERK2 bound to ATP (PDB:4GT3) in MOE. The optimal state of the superimposed structure of ROS1 G2032R with ATP was determined in MOE.

Molecular dynamics simulations

System preparation was carried out using MOE, which included running the standard protocols for system setup and protonation. The NAMD setup within MOE was used to create a periodic box, adding a 10 Å margin of TIP3P water and utilizing the Amber10 EHT force field. Simulations were equilibrated for 50 nanoseconds, with results evaluated from three independent MD runs of 20 nanoseconds each. Initial input coordinates were sourced from PDB structures of 4UXL (ROS1), and the G → R mutation was introduced using the MOE protein builder. The initial orientation of the arginine side-chain was determined by using the MOE side-chain rotamer generation tool and selecting the lowest energy rotamer in the presence of the ligand. Model orientations of the other ligands were obtained by direct superposition of the published crystal structures of lorlatinib in ROS1 (4UXL; ref. 43), crizotinib in ROS1 (3ZBF; ref. 37), entrectinib in ALK (5FTO; ref. 46), or repotrectinib in TRKA (7VKN; ref. 45) onto the 4UXL structure.

Free-energy perturbation

System preparation

The initial modeling orientation of zidesamtinib was prepared using the MOE docking protocol into the 4UXL structure of ROS1 (43). Model orientations of the other ligands were obtained by direct superposition of the published crystal structures of lorlatinib in ROS1 (4UXL; ref. 43), crizotinib in ROS1 (3ZBF; ref. 37), entrectinib in ALK (5FTO; ref. 46), or repotrectinib in TRKA (7VKN; ref. 45) onto either the 4UXL structure or the ROS1 G2032R structure. As no X-ray structure of taletrectinib was available, this ligand was manually built in MOE based on the overlay of repotrectinib. We performed mutational free-energy calculations in each direction using 4UXL as the starting protein for the G → R direction and our own ROS1 G2032R cocrystal with zidesamtinib as the starting protein for the R → G direction. Superposition of entrectinib onto the ROS1 G2032R structure resulted in a direct overlay of the of entrectinib piperazine ring with R2032, making this simulation untenable. Using these superpositions, input structures were created using the default protein preparation wizard in Maestro 13.3 and served as the starting points for the subsequent free-energy perturbation (FEP) calculations. This preparation included adding hydrogens, filling missing side-chains, and assigning proper ionization states for the amino acids and the ligands at physiologic pH. Prior to running the FEP simulations, missing ligand torsion parameters for all compounds were computed with the Force Field Builder that utilized the M06-2X/cc-pVTZ9(-f) functional. Crizotinib, entrectinib, and taletrectinib were prepared in their cationic form with a +1 charge.

FEP calculations

The FEP methodology has previously been described (47), including for protein side-chain modifications (48) as opposed to the more commonly applied ligand-based changes. FEP in the 2022–3 release with the OPLS4 force field (49) was used to predict the relative binding free energies for the ROS1–ligand complexes. A detailed description of the FEP methodology (47, 50–53) and its use to predict ligand selectivity (54, 55) have been reported. The Desmond MD calculations and FEP processing of results were accomplished on the Schrodinger-hosted AWS GPU server. During these calculations, the atoms of the arginine side-chain were grown via the windowed perturbation scheme either to or from glycine. Default values of 5 nanoseconds per lambda window and 24 lambda windows were used. Each system was solvated in an orthogonal box of explicit solvent [SPC water (56)] with box size determined to ensure that the solute atoms were no less than 5 Å (complex leg) or 10 Å (solvent leg) from an edge of the box. Counterions were included to neutralize the charge of the system, and additional Na⁺ and Cl^– ions were added to achieve 0.15 mol/L excess NaCl.

Data availability

The data generated in this study are available upon request from the corresponding author. NGS data generated in this study are publicly available in the Sequence Read Archive database with accession number PRJNA1240840. The atomic coordinates and structure factors of the ROS1 G2032R crystal structure have been deposited in the Protein Data Bank, www.pdb.org (PDB ID code 9QEK).

Results

Clinical exposure

To enable analysis of preclinical data in the context of clinically relevant concentrations, we first calculated the C_avg,u for each ROS1 TKI based on reported human pharmacokinetics studies (Supplementary Table S1; refs. 10–12, 31–33, 57). The C_avg,u is the time-averaged AUC_ss plasma drug concentration adjusted for protein binding, representing a balanced view between maximal and minimal drug exposures and accounting for differences in media compositions across biological contexts. The C_avg,u is regarded as a suitable surrogate for the clinically relevant concentration of reversible molecularly targeted therapies (58, 59).

We then determined the efficacious unbound concentration (C_eff,u) of each TKI, defined as the minimal concentration required to prevent Ba/F3 cell growth over 21 days (Fig. 1A; Supplementary Fig. S1; Supplementary Table S2). Zidesamtinib potently suppressed Ba/F3 CD74-ROS1 with C_eff,u = 1.6 nmol/L, which was 4× to 41× more potent than the other TKIs. Zidesamtinib also potently suppressed Ba/F3 CD74-ROS1 G2032R with C_eff,u = 7.2 nmol/L, which was 10× to 22× more potent than taletrectinib and repotrectinib, which have demonstrated clinical activity in patients whose tumors harbored ROS1 G2032R (14, 23). Zidesamtinib’s favorable potency against both cell lines is consistent with previously reported data (Supplementary Fig. S2; ref. 27).

Figure 1.

Clinical exposure vs. in vitro potency of ROS1 TKIs. A, Titration of zidesamtinib against Ba/F3 CD74-ROS1 (left) and CD74-ROS1 G2032R (right) cells over 21 days. Viability was measured by CellTiter-Glo luminescence and plotted as the mean ± SEM (n = 3). Representative examples shown (N = 2 repeats each). C_eff,u is indicated in green. B, Clinical exposure (measured as C_avg,u) and in vitro potency (measured as C_eff,u) against Ba/F3 CD74-ROS1 (left) or CD74-ROS1 G2032R (right) for each ROS1 TKI. The C_avg,u ÷ C_eff,u window represents drug exposure relative to potency, with higher numbers indicating more favorable target inhibition.

Comparing average human plasma exposure (C_avg,u) with in vitro potency (C_eff,u) can provide an estimate for the degree of target inhibition in patients, as higher exposure relative to potency suggests more complete target inhibition. Zidesamtinib’s C_avg,u exceeded its C_eff,u by 104× for ROS1 and 24× for ROS1 G2032R (Fig. 1B; Table 1), suggesting a potential for deep and durable inhibition of target oncoproteins in patients. The other ROS1 TKIs had narrower C_avg,u ÷ C_eff,u windows (range = 0.3–7×), suggesting potentially less optimal inhibition compared with zidesamtinib.

Table 1.

Comparing clinical exposure (measured as C_avg,u) vs. in vitro potency (measured as C_eff,u) against Ba/F3 CD74-ROS1 (top) or CD74-ROS1 G2032R (bottom) for each ROS1 TKI. The C_avg,u ÷ C_eff,u window represents drug exposure relative to potency, with higher numbers indicating more favorable target inhibition.

Cell line	ROS1 TKI				Cavg,u ÷ Ceff,u window
		Cavg,total (Average total plasma exposure; nmol/L)	Cavg,u (Average unbound plasma exposure; nmol/L)	Ceff,u (Efficacious unbound concentration; nmol/L)
Ba/F3 CD74-ROS1	Crizotinib	719	65	66	1×
	Entrectinib	2,000	5	6.4	0.8×
	Taletrectinib	772	48	31	1.5×
	Repotrectinib	846	39	5.6 (N = 2)	7×
	Zidesamtinib	Not reported	166	1.6 (N = 2)	104×
Ba/F3 CD74-ROS1 G2032R	Taletrectinib	772	48	155	0.3×
	Repotrectinib	846	39	70 (N = 2)	0.6×
	Zidesamtinib	Not reported	166	7.2 (N = 2)	24×

Mutagenesis screens

Predicting resistance in a TKI-naïve setting

ENU mutagenesis screens have been used to predict clinically relevant on-target resistance to TKIs, including ROS1 TKIs (60–65). Using Ba/F3 CD74-ROS1 as a model for TKI-naïve disease, we introduced random genome-wide mutations with ENU, distributed cells into 480 wells, selected for resistance after 28-day TKI treatment, and sequenced the ROS1 kinase domain (defined as residues 1945–2222) to identify mutations (Fig. 2A). This process was replicated across multiple TKIs (crizotinib, entrectinib, repotrectinib, and zidesamtinib) at 3 concentrations each. Because these unbound concentrations were selected to exceed C_eff,u required to prevent long-term growth, we expected cell numbers in all wells to decrease in the absence of TKI-resistant mutations. Instead, we observed outgrowth in many wells, indicating presence of TKI-resistant subpopulations (Fig. 2B and C). We examined each TKI with respect to the number of resistant wells and kinase domain mutations identified at various concentrations.

Figure 2.

Zidesamtinib suppressed ROS1 mutations in ENU mutagenesis screens. A, Workflow. ENU-exposed cells were collected for NGS to determine mutations in the starting cell pool. Remaining cells were distributed in 480 wells (5 × 96-well plates) and treated with a TKI at a single concentration for 28 days. Resistant wells were counted, and some wells were sequenced for kinase domain mutations. B and C, Results from a TKI-naïve screen. Graph showing the number of resistant wells out of 480 (B) and the distribution of kinase domain mutations among sequenced wells (C) for each TKI at each concentration. Blue triangles indicate clinical exposures and are positioned relative to screen concentrations (not to scale). D, Unbound IC₅₀ value heatmap for ROS1 TKIs against clones harboring indicated mutations isolated from the TKI-naïve screen. These clones were generated by ENU mutagenesis and may also harbor uncharacterized mutations outside ROS1. Parental indicates Ba/F3 CD74-ROS1 cells. E, Same as B and C but for a post-TKI screen. For B–E, full numerical data and total concentrations can be found in Supplementary Tables S1, S3, and S4.

Crizotinib

Resistant outgrowth was observed in 375/480 (78%) wells treated with 66 nmol/L unbound crizotinib (Fig. 2B; Supplementary Table S3), indicating a high likelihood of resistance developing at the clinical exposure (C_avg,u = 65 nmol/L). Most of the 50 sequenced wells harbored ROS1 mutations spread across residues L1951, G1957, V1979, L1982, S1986, E1990, F2004, L2026, G2032, and F2075 (Fig. 2C); of these, S1986F, L2026M, and G2032R have been implicated in clinical resistance to crizotinib (16, 66). A small number of resistant wells did not harbor ROS1 kinase domain mutations, suggesting off-target bypass mechanisms (Fig. 2C). At higher crizotinib concentrations, the total number of resistant wells decreased, bypass resistance disappeared, and the mutational landscape became less diverse, with only ROS1 L1982W and G2032R remaining at 332 nmol/L unbound crizotinib treatment, suggesting that they induced larger potency losses compared with the other mutations (Fig. 2C).

Entrectinib

All 480/480 (100%) wells exhibited resistant outgrowth following 6.4 nmol/L unbound entrectinib treatment, indicating ineffective suppression of resistance at its clinical exposure (C_avg,u = 5 nmol/L). Diverse ROS1 mutations were identified across residues L1982, S1986, L2000, F2004, H2074, F2075, D2113, M2128, and H2165. As with crizotinib, resistant well numbers, mutational diversity, and bypass incidence all decreased at higher entrectinib concentrations. Only ROS1 F2004C/V and G2032R remained at 32 nmol/L unbound entrectinib treatment, suggesting strong resistance consistent with identification of these mutations after relapse on entrectinib (Fig. 2B and C; Supplementary Table S3; ref. 17).

Repotrectinib

Only 1/480 (0.2%) wells exhibited resistant outgrowth following 28 nmol/L unbound repotrectinib treatment. Whereas this suggests effective prevention of resistance at the clinical exposure (C_avg,u = 39 nmol/L, based on 160 mg twice daily), such prevention may not be robust, as a decrease in drug exposure by only a few fold could result in ROS1 resistance emerging (Fig. 2B). ROS1 G2032R was recurrently identified after 5.6, 14, and 28 nmol/L unbound repotrectinib treatment, suggesting its potential as a resistance mechanism (Fig. 2C; Supplementary Table S3). Recent data showed acquisition of ROS1 G2032R in TKI-pretreated patients after disease progression on repotrectinib (23).

Zidesamtinib

Only 1/480 (0.2%) wells exhibited resistant outgrowth following 8.2 nmol/L unbound zidesamtinib treatment, suggesting effective prevention of resistance at the clinical exposure (C_avg,u = 166 nmol/L; Fig. 2B). Compared with repotrectinib, zidesamtinib’s prevention of resistance is expected to be more robust. Its clinical exposure exceeded the unbound concentrations used in this screen (1.6–8.2 nmol/L) by a factor of 20 to 100×, making the results less susceptible to pharmacokinetic variability (Fig. 2B and C; Supplementary Table S3).

TKI profiling in resistant clones

We isolated 28 resistant clones harboring distinct ROS1 kinase domain mutations spanning 19 residues from the mutagenesis screen and profiled 6 ROS1 TKIs. Most of these mutations reside in the N-terminal lobe (Supplementary Fig. S3), and mutations at 8 of the 19 residues have been identified clinically: L1951 (67), G1957 (68), L1982 (68), S1986 (69), F2004 (17), L2026 (67), G2032 (16), and G2101 (70). Importantly, these 28 clones were generated by ENU mutagenesis and may also harbor other mutations outside ROS1 that were not characterized in this study.

Zidesamtinib potently inhibited all clones with unbound IC₅₀ values (IC_50,u) ≤ 2 nmol/L (Fig. 2D; Supplementary Table S4). The clinical exposure of zidesamtinib (C_avg,u = 166 nmol/L) exceeded this IC_50,u value threshold by >80×. All these mutant clones, including ROS1 G2032R, are expected to be inhibited by ≥ 90% at the clinical exposure, suggesting that on-target resistance pathways are not predicted to emerge for zidesamtinib based on this mutagenesis screen.

The other TKIs showed limited activity against one or more ROS1-mutant clones, suggesting a higher likelihood that on-target resistance can emerge in the clinic. The clinical exposures of crizotinib and entrectinib are lower than their IC_50,u values for ROS1 F2004X (X = C, I, L, or V), L2026M, G2032R, and G2101A clones (Supplementary Table S4). These mutations would be inhibited by ≤50% at the clinical concentrations and hence may emerge as resistance. Lorlatinib, taletrectinib, and repotrectinib showed broad activity against all ROS1-mutant clones with the notable exception of ROS1 G2032R (Supplementary Table S4). The average clinical exposure of repotrectinib exceeded its IC_50,u value for ROS1 G2032R by less than 3-fold (C_avg,u = 39 vs. IC_50,u = 14 nmol/L), and the margin was further reduced for lorlatinib (C_avg,u = 197 vs. IC_50,u = 201 nmol/L) and taletrectinib (C_avg,u = 48 vs. IC_50,u = 74 nmol/L).

Predicting resistance in a post-TKI setting

We also performed mutagenesis screens in Ba/F3 CD74-ROS1 G2032R, modeling disease with acquired ROS1 G2032R after relapse on a prior TKI. ENU-exposed cells were treated with repotrectinib or zidesamtinib. Even though resistant wells emerged, none of the 208 sequenced clones across both TKIs harbored any ROS1 kinase domain mutations beyond the preexisting G2032R (Fig. 2E). This suggests that compound mutations (≥2 mutations residing on the same allele) may not be a major mechanism of resistance to ROS1 TKIs in later-line treatment. Although not the focus of this screen, off-target resistance was much more frequent for repotrectinib (158/480 wells = 33% at 70 nmol/L) than zidesamtinib (6/480 wells = 1% at 41 nmol/L) at clinically relevant concentrations (Fig. 2E).

Bioinformatics analysis

To ensure that ENU produced diverse ROS1 mutations in our resistance screens, we collected ENU-exposed cells before TKI selection and analyzed them by targeted NGS of the CD74-ROS1 amplicon at >3,000× coverage. ENU induced all six types of base-pair changes, with overrepresentation of A:T → G:C, G:C → T:A, and A:T → C:G (range = 19%–29%) relative to G:C → A:T, A:T → T:A, and G:C → C:G (range = 6%–12%; Supplementary Fig. S4A). The mutation rate was 1.5 per 1,000 bases (Supplementary Fig. S4B), equivalent to an average of 1.2 mutations in the ROS1 kinase domain (834 bases long). Mutations were roughly evenly distributed across the gene and were similar across both screens (Supplementary Fig. S4C).

ENU introduced 5 to 7 amino acid changes per residue (Supplementary Fig. S4D), translating to >1,500 distinct ROS1 kinase domain mutational variants across 278 residues in the starting cell pool. Of these, only 28 variants at 19 residues were identified after TKI treatment (Fig. 2C), indicating that a small fraction of all mutations conferred TKI resistance. Similarly, at hotspot residues S1986, F2004, and G2032, there were 22 variants in the starting cell pool, but only 7 remained after TKI treatment (Supplementary Fig. S4D; Fig. 2C). Many of the surviving mutations matched case reports of clinical resistance to ROS1 TKIs (16, 17, 71, 72), suggesting clinical translatability of the screens. Notably, ROS1 L2086F was not identified in any resistant wells even though it was present in the starting pool and has been reported to confer resistance to many ROS1 TKIs (Supplementary Fig. S4D; refs. 25, 27, 68, 73–75). ROS1 compound mutations (G2032R + other mutation in cis) were also not identified in any resistant wells despite presence in the starting pool and potential resistance to repotrectinib (Supplementary Fig. S4D; Fig. 2E; ref. 76). This suggests that other factors beyond resistance may play a role in enriching or depleting clones, representing a potential limitation of pooled screening assays. Overall, our screens profiled an extensive spectrum of ROS1 kinase domain mutations.

Intracranial activity in a ROS1 G2032R brain tumor model

Among all ROS1 mutations, ROS1 G2032R plays a predominant role in TKI resistance. It was the only mutation found recurrently across all TKIs in our ENU screens, conferring resistance even at high TKI concentrations (Fig. 2C). Whereas taletrectinib, repotrectinib, and zidesamtinib have demonstrated clinical responses in patients whose tumors harbored the ROS1 G2032R mutation (14, 23, 28), our preclinical data suggest that zidesamtinib has the potential to induce the deepest and most durable inhibition at clinically relevant concentrations (Fig. 1B). We sought to further compare the three TKIs using an intracranial Ba/F3 CD74-ROS1 G2032R luciferase model. Achieving robust activity in this model is difficult because: (i) the model harbors the drug-resistant G2032R mutation, (ii) it grows and metastasizes rapidly, (iii) it is partly protected by the blood–brain barrier which may impede drug distribution into the brain, and (iv) this cell line displays reduced TKI sensitivity similar to the patient-derived cell line MGH9018-1, which also harbors CD74-ROS1 G2032R (27). Together, these factors counter the effectiveness of a TKI, making this model uniquely difficult to treat.

Intracranial response

Intracranially implanted Ba/F3 CD74-ROS1 G2032R luciferase cells were extremely aggressive, exhibiting a >10,000% (100-fold) increase in average brain bioluminescence imaging (BLI) signal over 12 days and metastasizing across the body (Fig. 3A and B; Supplementary Table S5). All mice in the control group rapidly lost weight and succumbed between treatment days 15 to 18 (19–22 days after implantation; Fig. 3C and D).

Figure 3.

Zidesamtinib induced durable intracranial responses in a Ba/F3 CD74-ROS1 G2032R luciferase model. A, Bioluminescence images. B, Quantification of brain bioluminescence signals plotted as the mean ± SEM (n = 6). Horizontal gray line denotes initial tumor burden. C, Weight changes plotted as the mean ± SEM (n = 6). Horizontal gray line denotes initial weight. D, Survival analysis. BID, twice daily; QD, once daily.

Taletrectinib (100 mg/kg once daily) and repotrectinib [15 and 75 mg/kg twice daily, matching doses used in other preclinical studies (15, 77, 78)] showed initial suppression of brain tumors, but their efficacy was short-lived. By day 24, brain BLI signals had rebounded to 300% to 3,000% of the initial value, indicating intracranial progression (Fig. 3A and B; Supplementary Table S5). Mice treated with taletrectinib and repotrectinib (15 mg/kg twice daily) exhibited an 8% mean body weight loss, and three of six mice treated with taletrectinib succumbed between treatment days 23 to 25 (Fig. 3C and D).

Zidesamtinib at 3 mg/kg twice daily suppressed brain BLI signal to ≤4% of the initial value through day 24, indicating deep and durable regression. Bioluminescence images indicated that all mice were virtually free of tumors both intracranially and peripherally. All mice survived until the end of the experiment with a 5% gain in mean body weight, suggesting that the treatment was well-tolerated (Fig. 3A–D; Supplementary Table S5). Zidesamtinib had significantly better intracranial efficacy compared with taletrectinib and repotrectinib in this model (P < 0.0001 on day 24; Fig. 3A and B; Supplementary Table S5).

Continuous treatment with repotrectinib (15 mg/kg twice daily) did not confer durable efficacy. However, switching from repotrectinib (15 mg/kg twice daily) to zidesamtinib (3 mg/kg twice daily) on day 8 significantly improved response (P < 0.0001), suppressing the mean BLI signal on day 24 to ≤15% of the initial value. The treatment-switched group showed comparable BLI signal, body weight, and survival to the continuous zidesamtinib treatment group, highlighting the potential benefit of zidesamtinib even after a period of pretreatment with repotrectinib (Fig. 3A–D; Supplementary Table S5).

Pharmacokinetics analysis

Plasma

We determined steady-state drug exposures in the plasma of mice at the endpoint of the efficacy study. Zidesamtinib at 3 mg/kg twice daily and repotrectinib at 75 mg/kg twice daily were clinically relevant dosages, as they provided exposures within 3× of the clinical exposures (C_avg,u) at both 1 and 6 hours after administration. Repotrectinib at 15 mg/kg twice daily was a subclinical dosage, providing exposure close to C_avg,u at 1 hour but appreciably below C_avg,u at 6 hours. Taletrectinib at 100 mg/kg once daily exceeded clinical dosage, delivering exposures ≥5× above C_avg,u at both 1 and 12 hours, even though this dosage had been used for efficacy studies in other reports (Supplementary Fig. S5A; Supplementary Table S6; refs. 15, 77). Despite this super-clinical dosage, response to taletrectinib was not maintained over the dosing period (Fig. 3A and B).

Brain

We also determined steady-state drug exposures in the brain at the endpoint of the efficacy study. Zidesamtinib (3 mg/kg twice daily) unbound brain concentrations were above ROS1 G2032R IC_50,u value at both 1 and 6 hours, suggesting sustained oncoprotein inhibition, and were substantially below unbound TRKB phosphorylation (pTRKB) IC₅₀ value at both time points, suggesting the potential to spare TRKB. By contrast, repotrectinib (15 and 75 mg/kg twice daily) brain concentrations were below ROS1 G2032R IC₅₀ value but above pTRKB IC₅₀ value at both 1 and 6 hours, indicating poor oncoprotein inhibition concurrent with the potential for TRKB inhibition. Taletrectinib (100 mg/kg once daily) brain concentrations reached ROS1 G2032R IC₅₀ value at one of the two time points, suggesting partial oncoprotein inhibition (Supplementary Fig. S5B; Supplementary Table S6). These pharmacokinetic profiles are consistent with zidesamtinib yielding more durable intracranial responses than repotrectinib and taletrectinib in this mouse model.

Preclinical brain penetrance

We speculated that zidesamtinib’s sustained intracranial efficacy was a result of two factors: stronger ROS1 G2032R potency (27) and higher brain penetrance than taletrectinib or repotrectinib. To better understand brain penetrance, we profiled the three ROS1 TKIs in MDCK–MDR1 permeability and rat CNS pharmacokinetic assays. MDCK–MDR1 cells express PgP, an important blood–brain barrier efflux transporter, and lower ERs in MDCK–MDR1 permeability assays have been correlated with higher brain penetrance (36). Zidesamtinib showed low ERs that were more favorable than taletrectinib or repotrectinib and were comparable to lorlatinib, a TKI widely recognized for its brain penetrance (Supplementary Fig. S6A; refs. 26, 29). We also administered Wistar Han rats with a single 10 mg/kg TKI dose and measured the unbound brain-to-plasma partitioning coefficient (K_p,uu) at 1 hour. Zidesamtinib exhibited K_p,uu = 0.16, which was more favorable than taletrectinib or repotrectinib and comparable to lorlatinib (Supplementary Fig. S6B). These preclinical experiments demonstrated zidesamtinib’s potential for increased brain penetrance compared with repotrectinib and taletrectinib, which may translate into improved potential for intracranial activity against ROS1 and ROS1 resistance mutants for patients.

ROS1 G2032R structural analysis

Crystal structure

Given zidestamtinib’s increased activity for ROS1 G2032R compared with other TKIs, we next sought to understand its structural mechanism of inhibition. We solved a crystal structure of the ROS1 G2032R kinase domain in complex with zidesamtinib to 2.2 Å resolution (PDB: 9QEK; Fig. 4A and B; Supplementary Fig. S7A and S7B; Supplementary Table S7). To our knowledge, this represents the first structure of ROS1 G2032R or any mutant ROS1 reported.

Figure 4.

Structural analysis illuminates binding of zidesamtinib with ROS1 G2032R. A, Cocrystal structure of zidesamtinib and ROS1 G2032R. B, Each region of zidesamtinib complements the shape of the ROS1 G2032R ATP-binding pocket. C and D, Hinge (C) and other (D) intermolecular interactions.

ROS1 G2032R is structurally similar to WT ROS1 (PDB:3ZBF, 4UXL, 7Z5W, and 7Z5X; refs. 37, 43, 44), with average α-carbon root-mean-square deviation of 0.6 Å (Supplementary Fig. S8A). There are slight differences in the positioning of the αC helix, but these are also observed among WT ROS1 structures, suggesting that the helix may inherently assume several crystallographic states (Supplementary Fig. S8A). The ATP-binding pocket remained largely unchanged in both polarity and size, and docking indicated that the G2032R mutation is unlikely to impede ATP binding (Supplementary Fig. S8B and S8C). Recombinant ROS1 displayed slightly higher catalytic efficiency than ROS1 G2032R (k_cat/K_M = 26 vs. 7.5 L/μmol/s) but similar Michaelis constants for ATP (K_M = 25 vs. 14 μmol/L), suggesting that the ROS1 G2032R mutant maintains kinase activity and ATP binding (Supplementary Fig. S8D).

The aminopyridine forms two canonical hydrogen bonds with residues E2027 and M2029 in the hinge region (Fig. 4C), and the fluorobenzene inserts into a shallow cavity at the floor of the pocket, positioning the ortho-hydrogen atom for interaction with the carbonyl of R2083 (Fig. 4D). CH–π interactions were observed between the R2032 side-chain and the π-system of the N-ethylpyrazole and between the terminal methyl groups of L2086 and the π-systems of the aminopyridine and fluorobenzene moieties (Fig. 4D). The P-loop of ROS1 G2032R is shifted ∼1.5 Å upward compared with WT ROS1 bound to lorlatinib and displays a lower B-factor, suggesting that the N-methyltriazole group of zidesamtinib may reduce the P-loop’s degrees of freedom and lock it in this crystallographic state (Supplementary Fig. S9A and S9B).

We observed that the R2032 side-chain extended into the solvent-front, which may be a result of an intermolecular electrostatic interaction between the positively charged R2032 and the negative dipole of a neighboring α-helix near residue Y2227 (Supplementary Fig. S10A). The α and β carbon atoms of R2032 adopt a gauche⁺ conformation (χ1 angle = 306°), allowing it to engage in multiple CH–π interactions with the N-ethylpyrazole ring of zidesamtinib (Supplementary Fig. S10B; ref. 79). An overlay of ROS1-bound lorlatinib (PDB:4UXL) and crizotinib (PDB:3ZBF) crystal structures onto our ROS1 G2032R protein structure indicated potential steric clashes between these TKIs and R2032 (Supplementary Fig. S10C; refs. 37, 43). In further support of these observations, MD simulations revealed that R2032 adopts the energetically more favorable gauche⁺ conformation with zidesamtinib, forming favorable hydrogen bond interactions in some frames (Supplementary Fig. S10D). By contrast, R2032 adopts the energetically less favorable gauche^– conformation with lorlatinib, presumably to avoid clashes in the solvent-front region, which may contribute to its potency loss against ROS1 G2032R (Supplementary Fig. S10D).

Zidesamtinib was designed to spare inhibition of TRK family kinases by sterically clashing with the key differentiating residue Y591 on TRKA and Y619 on TRKB/TRKC, analogous to the smaller residue L2028 on ROS1. Overlay of the zidesamtinib-bound ROS1 G2032R structure onto the repotrectinib-bound TRKA structure (PDB:7VKO; ref. 45) indicated that zidesamtinib’s N-ethylpyrazole clashed with TRKA Y591 but accommodated ROS1 L2028, consistent with its selective inhibition (Supplementary Fig. S11; ref. 27). Repotrectinib showed no clash with either ROS1 or TRKA, consistent with its dual TRK/ROS1 inhibitory design and with reported TRK-related adverse events (27, 78).

FEP calculations

ROS1 WT and G2032R crystal structures not only improved our understanding of zidesamtinib binding but also enabled the computational prediction of ROS1 ligand-binding affinity. We applied the FEP methodology (47, 55) to study the impact of the G2032R modification on TKI binding. Beginning with the WT ROS1 crystal structure, FEP examined the energetic changes both upon gradual growth of glycine 2032 into arginine and upon TKI binding (Supplementary Fig. S12A and S12B). We also performed FEP in the reverse direction, beginning with our ROS1 G2032R crystal structure and gradually shrinking arginine 2032 into glycine. The output of these FEP calculations is a predicted change in binding free energy (∆∆G) between WT ROS1 and ROS1 G2032R, with smaller ∆∆G values corresponding to a smaller influence of the mutant arginine on binding affinity.

In a set of six approved or investigational ROS1 TKIs, FEP predicted affinity changes (∆∆G_FEP) that correlated well with reported experimental data (∆∆G_biochem; Fig. 5 and Table 2). Zidesamtinib and repotrectinib clustered toward the bottom-left corner of the correlation plot (∆∆G ≤ 1.9 kcal/mol), consistent with tolerance for the G2032R mutation imparted by minimal steric bulk in the solvent-front region. By contrast, crizotinib, entrectinib, and lorlatinib clustered toward the top-right corner (∆∆G ≥ 2.9 kcal/mol), consistent with G2032R resistance imparted by steric clashes in the solvent-front region.

Figure 5.

FEP calculations predict potency loss due to G-to-R mutation. Experimental (∆∆G_biochem) and predicted (∆∆G_FEP) changes in binding free energy for each ROS1 TKI shown as the mean ± SEM. ∆∆G_FEP was obtained through simulations involving growing glycine to arginine (G → R) or truncating arginine to glycine (R → G), or the average of the two directions. A–F represent crizotinib, entrectinib, lorlatinib, taletrectinib, repotrectinib, and zidesamtinib, respectively. R² denotes correlation coefficient, and RMSD denotes the root-mean-squared deviation between predicted and experimental data.

Table 2.

FEP calculations predict potency loss due to G-to-R mutation. Experimental (∆∆G_biochem) and predicted (∆∆G_FEP) changes in binding free energy for each ROS1 TKI shown as the mean ± SEM. ∆∆G_FEP was obtained through simulations involving growing glycine to arginine (G → R) or truncating arginine to glycine (R → G), or the average of the two directions. “Not available” denotes calculations not completed due to extensive steric clashing between entrectinib and arginine in the R → G direction.

Ligand	∆∆Gbiochem (kcal/mol)	∆∆GFEP (kcal/mol)
		G → R direction	R → G direction	Average
Crizotinib	3.7 ± 0.5	4.6 ± 0.4	2.9 ± 0.4	3.8 ± 0.3
Entrectinib	3.4 ± 0.5	4.2 ± 0.4	Not available	4.2 ± 0.4
Lorlatinib	4.7 ± 0.7	3.6 ± 0.4	3.4 ± 0.4	3.5 ± 0.3
Taletrectinib	2.4 ± 0.6	2.5 ± 0.4	2.5 ± 0.4	2.5 ± 0.3
Repotrectinib	1.9 ± 0.4	0.2 ± 0.4	1.1 ± 0.4	0.7 ± 0.3
Zidesamtinib	1.4 ± 0.5	1.5 ± 0.4	1.4 ± 0.4	1.5 ± 0.3

Availability of the ROS1 G2032R structure improved prediction accuracy compared with relying on the existing WT ROS1 structure alone. Whereas FEP predictions were often similar in both G → R and R → G directions, large differences were observed for repotrectinib and crizotinib, indicating that the two directions were reciprocal but not necessarily identical (Fig. 5; Table 2). In this set, the R → G prediction displayed better correlation and concordance with experimental data than the G → R prediction (R² = 0.85 vs. 0.62 and root-mean-square deviation = 0.77 vs. 0.96). These data highlight FEP as a reliable computational approach for predicting changes in ligand binding affinity between WT ROS1 and ROS1 G2032R.

Discussion

An emerging paradigm in NSCLC treatment is the preference for later-generation over earlier-generation TKIs as first-line therapies. In 2018, the National Comprehensive Cancer Network designated osimertinib (a third-generation EGFR TKI) and alectinib (a second-generation ALK TKI) as the preferred first-line treatment for EGFR mutations and ALK rearrangement, respectively, displacing first-generation TKIs gefitinib, erlotinib, and crizotinib (80). Osimertinib and alectinib share key advancements, including improved potency against the oncogenic target, activity for certain on-target resistance mutations, brain penetration, and safety profiles, which translated into longer median progression-free survival and more manageable adverse events in randomized controlled trials (81, 82). These same features are likely to underlie best-in-class TKIs for other oncogenic drivers.

Such a paradigm shift has yet to occur with ROS1-driven NSCLC. Since its approval in 2016, crizotinib has remained an National Comprehensive Cancer Network–preferred first-line treatment alongside entrectinib (approved in 2019) and repotrectinib (approved in 2023), consistent with real-world practice (80, 83, 84). Each of these options presents unique challenges, with crizotinib having limited brain penetration (18, 19); crizotinib and entrectinib having limited activity against resistance mutations, especially ROS1 G2032R (16, 17); and entrectinib and repotrectinib, which also inhibit TRK, inducing CNS-related adverse events such as dizziness, ataxia, and cognitive impairment (11, 12). Zidesamtinib was designed to address these limitations, offering broad activity against resistance mutations and brain penetration while sparing TRK inhibition (27).

Whether newer-generation ROS1 TKIs will replace crizotinib as the preferred first-line therapy remains to be seen. However, preclinical experiments can enable data-driven predictions about the potential of each ROS1 TKI. Desirable attributes such as delay of resistance and CNS activity can be respectively assessed using preclinical accelerated mutagenesis screens and intracranial xenograft models. A key to meaningful preclinical evaluation is the use of clinically relevant concentrations, reflecting the physiologic drug levels experienced by tumor cells in patients. With the availability of recommended clinical dosages and human pharmacokinetic data for several ROS1 TKIs, preclinical evaluation of the ability of ROS1 TKIs to prevent resistance and control brain metastases at clinically relevant concentrations is possible.

Our pooled mutagenesis screen evaluated >1,500 distinct mutations distributed across the ROS1 kinase domain. At clinically relevant concentrations, 28 resistance mutants spanning 19 residues emerged with crizotinib and entrectinib but not repotrectinib or zidesamtinib in the TKI-naïve setting, aligning with reported clinical data (16, 17, 23). However, repotrectinib’s ability to prevent resistance was not robust in our study, as resistance could emerge at concentrations only slightly lower than its clinical exposure. Factors that could contribute to decreased repotrectinib exposure include dose reductions (reported in 35% of patients; ref. 12), treatment below recommended dosing (17% of patients reported to not escalate to 160 mg twice daily; ref. 85), and lower blood–brain barrier penetration (Supplementary Fig. S6B). In contrast, zidesamtinib demonstrated more robust prevention of resistance, with a wider margin between clinical exposure and effective concentration (104× for zidesamtinib compared with 7× for repotrectinib), ensuring adequate coverage to suppress mutations. Most mutations cluster in the N-terminal lobe. Among these, ROS1 G2032R was the most prevalent and most resistant across TKIs, whereas ROS1 F2004X uniquely conferred resistance to high-dose entrectinib, consistent with both preclinical and clinical findings (16, 17, 23, 63). Both mutations remain highly sensitive to zidesamtinib.

ROS1 G2032R compound mutants (i.e., ≥2 cis-allelic mutations) and ROS1 L2086F did not emerge in our screens despite being present in the starting cell pool and expected to confer resistance to ROS1 TKIs. There are two potential explanations. First, this observation signals a limitation of pooled screens involving competition among thousands of clones. Some clones may disappear for reasons unrelated to TKI sensitivity, such as reduced fitness. Second, this observation suggests that compound mutations may be less likely than bypass resistance in a post-TKI setting. To date, no ROS1 compound mutations have been confirmed in patients, in contrast to ALK compound mutations, which have been identified both preclinically and clinically (86, 87).

In the ROS1 G2032R intracranial study, whereas taletrectinib and repotrectinib induced an initial response, only zidesamtinib provided a sustained response at clinically relevant doses. This outcome aligns with zidesamtinib’s higher preclinical brain exposure relative to its potency threshold. Furthermore, zidesamtinib showed a low ER in the MDCK–MDR1 permeability assay and high brain-to-plasma partitioning in rats, with both parameters comparable to those of lorlatinib, a highly brain-penetrant TKI (26). By combining high brain penetration with broad activity against resistance mutations, zidesamtinib has the potential to delay progression and address existing resistant disease, both peripherally and intracranially.

This study reports the first crystal structure of ROS1 G2032R, validating zidesamtinib’s molecular design and providing structural insights into how the G2032R mutation affects TKI binding. As designed, zidesamtinib’s N-ethylpyrazole serves the dual function of accommodating the G2032R substitution while clashing with Y591/Y619 of TRKs (27). The structure also revealed interactions with the binding pocket floor and the P-loop. FEP calculations based on this structure accurately predicted affinity changes caused by the G2032R mutation across approved and investigational ROS1 TKIs with diverse chemotypes, highlighting the potential of this in silico method for interrogating novel binders targeting solvent-front mutant kinases.

In summary, this work connects the chemical design of zidesamtinib with its advantageous preclinical biological activity compared with other TKIs, highlighted by preclinical coverage of ROS1 resistance mutations, brain penetration, and selective inhibition of ROS1 over TRKs. This favorable preclinical activity suggests the potential for delayed tumor progression with zidesamtinib, both peripherally and intracranially. The prospect of sustained clinical responses is being assessed in an ongoing clinical trial (NCT05118789).

Supplementary Material

Supplementary Table S1

Supplementary Table S1: Determining Cavg,u (clinical average unbound plasma exposure)

Supplementary Table S2

Supplementary Table S2: Determining Ceff,u (efficacious unbound concentration)

Supplementary Table S3

Supplementary Table S3: Number of resistant wells and distribution of ROS1 kinase-domain mutations identified in ENU mutagenesis screens

Supplementary Table S4

Supplementary Table S4: Potency against resistant clones from the ENU mutagenesis screen

Supplementary Table S5

Supplementary Table S5: Quantification of brain BLI from intracranial Ba/F3 CD74-ROS1 G2032R luciferase study

Supplementary Table S6

Supplementary Table S6: Pharmacokinetics measurements from intracranial Ba/F3 CD74-ROS1 G2032R luciferase study

Supplementary Table S7

Supplementary Table S7: Crystallographic and refinement statistics

Supplementary Figure S1

Supplementary Figure S1: Determining efficacious unbound concentration (Ceff,u)

Supplementary Figure S2

Supplementary Figure S2: Dose-response curves for 6 TKIs

Supplementary Figure S3

Supplementary Figure S3: Residues with mutations identified in mutagenesis screens

Supplementary Figure S4

Supplementary Figure S4: Bioinformatic analyses of starting cell pools from ENU mutagenesis screens

Supplementary Figure S5

Supplementary Figure S5: Pharmacokinetics measurements from the intracranial study

Supplementary Figure S6

Supplementary Figure S6: Preclinical measurements of brain penetration potential

Supplementary Figure S7

Supplementary Figure S7: ROS1 G2032R protein purification

Supplementary Figure S8

Supplementary Figure S8: Similarities between wild-type ROS1 and ROS1 G2032R

Supplementary Figure S9

Supplementary Figure S9: P-loop differences between wild-type ROS1 and ROS1 G2032R

Supplementary Figure S10

Supplementary Figure S10: Structural impact of ROS1 G2032R mutation on TKI binding

Supplementary Figure S11

Supplementary Figure S11: Structural basis for TRK selectivity

Supplementary Figure S12

Supplementary Figure S12: Free-energy perturbation (FEP) workflow

Authors’ Disclosures

A. Tangpeerachaikul reports personal fees from Nuvalent, Inc. during the conduct of the study; in addition, A. Tangpeerachaikul has a patent for US 11542278 issued, a patent for US 20240398768 pending, and a patent for PCT/US2025/022255 pending. S. Mente reports personal fees from Nuvalent, Inc. during the conduct of the study; in addition, S. Mente has a patent for US 11542278 issued and a patent for PCT/US2025/022255 pending. J. Magrino reports personal fees from Nuvalent, Inc. during the conduct of the study; in addition, J. Magrino has a patent for PCT/US2025/022255 pending. F. Gu reports personal fees from Nuvalent, Inc. during the conduct of the study. J.C. Horan reports personal fees from Nuvalent, Inc. during the conduct of the study; in addition, J.C. Horan has a patent for US 11542278 issued and a patent for US 12043626 issued. H.E. Pelish reports personal fees from Nuvalent, Inc. during the conduct of the study; in addition, H.E. Pelish has a patent for US 11542278 issued, a patent for US 20240398768 pending, and a patent for PCT/US2025/022255 pending.

Authors’ Contributions

A. Tangpeerachaikul: Conceptualization, resources, data curation, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. S. Mente: Data curation, software, investigation, visualization, methodology, writing–original draft, writing–review and editing. J. Magrino: Conceptualization, resources, data curation, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. F. Gu: Resources, data curation, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. J.C. Horan: Conceptualization, project administration, writing–review and editing. H.E. Pelish: Conceptualization, supervision, investigation, methodology, writing–original draft, project administration, writing–review and editing.