Poziotinib for EGFR exon 20-insertion NSCLC: Clinical efficacy of the phase 2 ZENITH trial and differential impact of EGFR exon 20 insertion location on sensitivity

Abstract

EGFRex20 insertions (EGFRex20ins) can be classified as near- and far-loop based on the insertion location, however, the impact of location on responses to various EGFR tyrosine kinase inhibitors (TKIs) is poorly understood. In vitro studies show that afatinib, poziotinib, and zipalertinib more potently inhibited near-loop than far-loop insertions, whereas mobocertinib has similar IC₅₀ in both groups. Molecular dynamics simulations reveal that near-loop insertions have multiple conformational states and lower transitional energy than far-loop insertions. ZENITH20 trial cohort 1 (NCT03318939) evaluates poziotinib in EGFRex20 NSCLC patients (n = 115) and demonstrates an objective response rate of 14.8% (95% Confidence Interval [CI], 8.9 to 22.6, primary endpoint of the trial). Although the study’s primary efficacy endpoint was not met in the overall cohort, the exploratory analysis indicates poziotinib has superior benefit in EGFRex20 near- versus far-insertions showing greater mean tumor size reduction (−25.9% vs. −9.8%, p = 0.0014) and progression-free survival (PFS, 11.1 vs. 3.5 months, p = 0.016). In comparison, in the previously published EXCLAIM trial (NCT02716116), mobocertinib demonstrates similar activities across both groups in tumor size reduction (−38.5% vs. −34.1%, p = 0.59) and PFS (12.0 vs. 13.0 months, p = 0.99). Therefore, EGFRex20ins location differentially impacts the sensitivity of TKIs.

The location of EGFR exon 20 loop insertions (EGFRex20ins) has been shown to alter sensitivity to lung cancer therapy. Here, the authors report the results of the ZENITH20 clinical trial investigating poziotinib (EGFR TKI) in lung cancer patients and, combining with a similar trial, investigate how structural differences due to location of EGGFRex20ins alters sensitivity to EGFR TKI.

Introduction

Mutations in the epidermal growth factor receptor (EGFR) gene occur in approximately 15–40% of patients with non-small cell lung cancer (NSCLC), and historically, patients have been separated into two groups based on their mutation subtype: patients with classical EGFR mutations (exon 19 deletions and L858R point mutation) and patients with atypical EGFR mutations. Patients with classical EGFR mutations comprise approximately 70% of EGFR mutant NSCLC¹ and have a collection of standard care treatment options, including tyrosine kinase inhibitors (TKIs) such as osimertinib, which has high clinical efficacy^2,3. Recently, we have shown that patients with atypical EGFR mutations constitute approximately 30% of the EGFR mutant NSCLC population, have heterogeneous drug sensitivity, and can be divided into four distinct subgroups based on their drug sensitivity and structural location of the mutation¹, including classical-like EGFR mutations, T790M-like mutations, P-loop and αC-helix Compressing (PACC) mutations, and exon 20 loop insertions.

EGFR exon 20 loop insertions (EGFRex20ins) are a distinct subgroup of EGFR mutations with unique structural underpinnings, which lack sensitivity to standard EGFR TKIs for classical EGFR mutations⁴. Therefore, several EGFRex20ins targeting agents have been developed, including amivantamab with FDA accelerated approval for treating EGFRex20ins NSCLC. Amivantamab, an EGFR/MET bispecific antibody, achieved an ORR of 40% and median progression-free survival (mPFS) of 8.3 months in a Phase I study (NCT02609776)^5–7 in treating of EGFRex20ins mutant NSCLC with prior therapy. It also demonstrated an ORR of 73% and PFS of 11.4 months when used in combination with chemotherapy for treatment naïve EGFRex20ins mutation NSCLC patients (04538664)⁸. Mobocertinib is a small molecule EGFR TKI, and showed confirmed ORR of 28% and mPFS of 7.3 months in the platinum-pretreated population included in the EXCLAIM trial (NCT02716116)^9–11. However, because of the lack of PFS benefit of mobocertinib compared to chemotherapy, mobocertinib was withdrawn from the US market¹². Encouragingly, additional TKIs are being evaluated in clinical studies for EGFRex20ins NSCLC, including zipalertinib (CLN-081/TAS6417, NCT04036682)¹³, poziotinib (NCT03318939 and NCT03066206)^4,14,15, sunvozertinib (DZD9008, NCT03974022) and furmonertinib (NCT05607550)¹⁶. Recent clinical study reports of these inhibitors demonstrated ORRs of 15–40% for patients receiving poziotinib (depending on trial and dosing regimen)^14,15,17,18, 31% for patients receiving zipalertinib¹⁹; 40% in patients receiving sunvozertinib^20,21, 36–55% receiving furmonertinib, in previously treated patient populations and even higher responses in treatment-naïve populations¹⁶. Notably, zipalertinib, sunvozertinib, and furmonertinib have obtained FDA Breakthrough Designation (BTD) for EGFRex20ins NSCLC.

Through structure/function analysis of atypical EGFR mutations, including exon 20 loop insertions, we have previously shown that exon 20 loop insertions can be separated into two distinct groups: near-loop (A767-P772) and far-loop (H773-R776) insertions¹. Poziotinib demonstrated an improved response rate in EGFRex20 near-loop insertions in an investigator-initiated trial²². Therefore, we hypothesized that the structure/function difference between the near-loop and far-loop insertions could lead to differential responses to various EGFR TKIs.

Here, we show that EGFR TKI drug sensitivity is highly dependent on the location of the EGFRex20ins. The majority of TKIs are more active in near-loop insertions compared to far-loop insertions, including poziotinib, afatinib, and zipalertinib. In contrast, mobocertinib has slightly greater activity in far-loop insertions in preclinical studies. Long-time scale molecular dynamics (MD) simulations reveal that EGFR with a near-loop insertion has enhanced flexibility compared to the EGFR wild type and far-loop insertions, which in turn influence the thermodynamics and kinetics of TKI binding and sensitivities. Clinical data from the EXCLAIM trial indicate that mobocertinib demonstrates comparable efficacy for both near- and far-loop insertions, whereas the ZENITH20 trial data supportes that poziotinib has more pronounced activity in the near-loop insertions. Taken together, this work shows that EGFRex20ins are a heterogeneous population of mutations, and that insertion location alters TKI binding, emphasizing the importance of personalizing TKIs to patient’s tumor mutations even within a subset of EGFRex20ins.

Results

Association between EGFRex20ins location and differential TKI sensitivity in vitro

Previous studies have shown that EGFRex20ins mutations occurring in the αC-helix or the αC-β4 loop have significant differences in the overall structure and drug sensitivity²³. The helical exon 20 insertions, such as A763_Y764insFQEA, are classical-like mutations¹ with sensitivity to first-, second-, and third-generation EGFR TKIs in preclinical models^4,24 and patients^23,25–29. The exon 20 loop insertions can be grouped as near-loop (A767-P772) and far-loop (H773-R776) insertions based on location. To test our hypothesis that the structure/function difference between the near-loop and far-loop insertions could lead to differential responses to various EGFR TKIs, we generated a panel of Ba/F3 cell lines expressing twenty-two different near- and far-loop exon 20 insertion mutations and eight exon 20 point mutations and screened these cell lines against a panel of EGFR TKIs. Since helical exon 20 insertions are classical-like mutations, these mutations were excluded from the analysis. We observed that mobocertinib demonstrated a trend of greater sensitivity with far-loop insertions in vitro (R = − 0.43, p = 0.023, Fig. 1a, g). In contrast, other EGFR TKIs, afatinib, zipalertinib, osimertinib, tarlox-TKI, and poziotinib, all demonstrated higher sensitivity in near-loop insertions than far-loop insertions (R ranges from 0.45–0.79, p < 0.05, Fig. 1b–f and Fig. 1g). These data suggest that the specific location of the EGFRex20ins may impact drug sensitivity, and that near- and far-loop insertions may have different binding kinetics and thermodynamics of ATP and TKIs.

Fig. 1. EGFR TKI activity correlates with exon 20 insertion mutation location.

a–f Dot plots of average IC₅₀ values for Ba/F3 cell lines expressing various EGFR exon 20 insertion mutations (N = 22 cell lines) treated with the indicated inhibitors compared to the amino acid residue of the EGFR exon 20 insertion mutation. The red dashed line indicates the average IC₅₀ value for the TKI against Ba/F3 cells expressing WT EGFR + 10 ng/ml EGF. Dots are representative of the average IC₅₀ value determined in biological triplicate. a–f Data was fit to a linear regression model, and a two-tailed Pearson correlation was determined. Each specific mutation was denoted as a colored dot, and a key was provided to the right of the figure. a mobocertinib, (b) afatinib, (c) zipalertinib (CLN-081), (d) osimertinib, (e) Tarlox-TKI, (f) poziotinib. g Boxplots were shown and t test was used to compare the IC50 of each drug in near- (n = 17) vs far-loop (n = 5) mutations.

Conformational differences between wildtype EGFR, exon 20 near- and far- insertions

To elucidate the impact of insertion location on EGFR structure and dynamics, we next leveraged molecular dynamics simulations to quantify the differences in the conformational free energy landscapes and conformational flexibility of the EGFR kinase domain harboring near- and far-loop mutations. Because interactions between drug and protein are highly dependent on the size, shape, and accessibility of the binding pocket, any structural changes resulting from mutations, whether within or distal to the binding pocket, can influence the thermodynamics and kinetics of drug binding and thus affect the pharmacological activity and duration of pharmacological action^30,31.

First, we assessed the conformational free energy landscapes of wildtype (WT) EGFR, near-loop insertion (S768dupSVD and A767dupASV), and far-loop insertion (H773insNPH and H773insAH), to understand the mutation-induced structural changes and the energy barriers between the active and inactive forms of the protein. The bias exchange meta-dynamics (BE-MetaD) approach was used, which depicted the different accessible conformational states (active versus inactive), the relative stability of these states, and the transition pathways. Our free energy landscape analysis revealed multiple conformational states for WT EGFR, near- and far-loop insertions, some were shared, and others were unique (Fig. 2 and Supplementary Results). For the WT and both the mutant EGFR proteins, conformations in state I corresponded to a previously described “Src-like” inactive state³² with the phenylalanine (F) in DFG motif packed towards the αC-helix (“DFG-in” conformation) and the αC-helix is displaced away from the catalytic site (“αC-out” conformation). In this form, the receptor is deemed as not actively transducing downstream oncogenic signals. State II corresponded to the active kinase conformation, which represents the oncogenic form, also observed in all three proteins. In addition, both near- and far-mutant proteins featured a state III that constituted the transition path (dashed lines in Fig. 2) from the inactive to the active state. This intermediate conformation was not very prominent in the WT protein (Fig. 2a). The near-loop insertion also displayed an additional intermediate state (state IV) along the transition pathway that was not observed in either the WT or far-loop mutant (Fig. 2b). Further details on the structural differences between the conformational states are provided in Supplementary Results and Supplementary Fig. S1. Note that the Src-like inactive conformation is stabilized relative to the active state for the WT and all the mutants (indicated by the deeper free energy minima of state I compared to state II) as expected for the monomeric form of the EGFR proteins^32,33. Next, we assessed the energy barriers between each state for the WT and mutants. In the near-loop mutant, the conformational states were separated by lower free energy differences when compared to the far-loop and the WT (Supplementary Fig. S2 and Supplementary Results). This can be inferred from the energy difference between the states, as determined from the minimum free energy path (dashed lines in Fig. 2). Therefore, in the near-loop insertion (S768dupSVD and A767dupASV), the critical structural elements undergo frequent transitions between different conformations (Fig. 2b). In comparison, the far-loop insertion (H773insNPH and H773insAH) featured well-defined active and inactivate states in addition to a relatively stable intermediate state (State III, Fig. 2c), exhibiting higher free energy differences and thus, less frequent transitions between them. Far-loop mutants, especially H773insNPH, displayed the most stable active state compared to the WT and near-loop insertion mutant.

Fig. 2. The free energy (FE) landscape of WT and EGFRex20ins mutations.

a WT EGFR, (b) near-loop insertion mutants (S768dupSVD and A767dupASV), and (c) far-loop insertion mutants (H773insNPH and H773insAH). The prominent FE basins are marked as states (I–IV), and the dashed black lines denote the most probable transition path between the inactive and the active state. States I and II predominantly correspond to a Src-like inactive state and the active state, respectively, while states III and IV are the intermediate states.

The observed conformational difference was then further quantified with long-timescale classical MD simulations (~3 µs simulations without TKI bound to the protein), initiated from the active conformation of each EGFR protein. These analyses were performed only for representative near- and far-loop mutants (S768dupSVD and H773insNPH, respectively). Structural analysis revealed a partial loss of helicity for the αC-helix in the near-loop insertion compared to a more stable αC-helix in the case of the far-loop insertion. In addition, the conserved E745-K762 salt bridge is more stable in the far-loop than the near-loop insertion (Supplementary Fig. S3 and Supplementary Fig. S4). Furthermore, the aspartic acid residue (D858) of the DFG motif forms hydrogen-bond interactions with the rest of the protein less frequently (~9% of the time in the observed trajectory) in the near-loop mutant when compared to the far-loop (~35% of the time), consistent with a higher flexibility of the near-loop insertion (Supplementary Fig. S5). Taken together, the above observations suggest that the near-loop insertion mutant was relatively more flexible and conformationally diverse compared to the far-loop insertion, and this difference in the structural heterogeneity might be one of the underlying factors for drug binding ability.

Differential TKI binding affinity to EGFR exon 20 near- and far- insertions

To understand the effect of these mutations on the binding affinity of TKIs’, classical MD simulations³⁴ were performed on modeled EGFRex20ins S768dupSVD (near-loop) and H773insNPH (far-loop) docked to zipalertinib and mobocertinib in a non-covalent state, which presents the first step of covalent inhibition. Simulations were performed on a total of four non-covalent complexes (S768dupSVD and H773insNPH bound to zipalertinib and mobocertinib, Fig. 3a–d). All four modeled protein-drug complexes remained stable during the entire simulation based on time-dependent backbone RMSD plots. Binding energy calculations revealed that mobocertinib had a better binding affinity for H773insNPH than S768dupSVD, whereas zipalertinib demonstrated a better binding affinity for S768dupSVD than H773insNPH (Fig. 3e, f). The calculated binding energy of poziotinib demonstrated a similar pattern to zipalertinib²², favoring near-loop insertion mutant (Fig. 3f). Tables 1, 2 Next, we performed covalent docking of the TKI followed by several short MD simulations of the bound complex to predict the interactions between the EGFR mutant proteins and the drug. Zipalertinib exhibited a relatively stronger interaction propensity in forming a hydrogen bond with the near-loop insertion (Supplementary Fig. S6). While poziotinib formed frequent interactions with both the near- (D858 - part of the DFG motif and K745- salt-bridge forming residue in the αC-helix) and far-loop (Y727 residue), it interacted slightly more favorably with the near-loop (Supplementary Fig. S6). For mobocertinib, the far-loop insertion interacted slightly more favorably with the drug, forming hydrogen-bond interactions in the P-loop region of the protein (Supplementary Fig. S6). These results suggested that zipalertinib had the propensity to form a stronger interaction with the near-loop mutant, and thus, enhanced binding affinity. Also, poziotinib formed slightly more favorable interactions with the near-loop mutant. In comparison, mobocertinib displayed similar hydrogen and hydrophobic bonded interactions with both near- and far-loop mutants, with slightly favorable binding affinity to far-loop insertion. Combining all these data, we note that the drug-protein binding affinities from simulations were completely aligned with the in vitro observations wherein the near-loop insertions were more sensitive to zipalertinib and poziotinib than the far-loop insertions, yet mobocertinib exhibited notable activity in both near- and far-loop mutants.

Fig. 3. Exon 20 insertion location effects drug-receptor interactions.

a, b Ribbon diagram representation of CLN-081 (purple) bound to (a) S768dupSVD and (b) H773insNPH. The black dashed line indicates the distance between the reactive acrylamide group (purple arrow) and reactive cysteine (dark blue arrow). c, d Ribbon diagram representation of mobocertinib (purple) bound to (c) S768dupSVD and (d) H773insNPH. The black dashed line indicates the distance between the reactive acrylamide group (purple arrow) and reactive cysteine (dark blue arrow).

Table 1.

Free binding energy calculations in kcal/mol for mobocertinib and CLN-081 for the indicated mutations using MM/GBSA and MM/PBSA

Binding Energy	S768dupSVD(near)		H773insNPH (far)		Δ (far-near)
kcal/mol	MM/GBSA	MM/PBSA	MM/GBSA	MM/PBSA	MM/GBSA	MM/PBSA
Mobocertinib	− 45	− 38	− 49	− 41	− 4	-3
Zipalertinib	− 35	− 32	− 31	− 28	4	4

Table 2.

Free binding energies calculations in kcal/mol for mobocertinib, CLN-081, and poziotinib for the indicated mutations using funnel metadynamics

Binding Free Energy	S768dupSVD(near)	H773insNPH (far)
kcal/mol	Funnel Metadynamics	Funnel Metadynamics
Mobocertinib	− 7.1 ± 0.2	− 11.4 ± 0.2
Zipalertinib	− 5.9 ± 0.2	− 4.5 ± 0.1
Poziotinib	− 12.3 ± 0.2	− 4.4 ± 0.3

Clinical response with poziotinib in near- vs. far- EGFRex20ins NSCLC

Finally, we sought clinical evidence to validate our hypothesis that near- vs. far-loop insertion location of the EGFRex20ins may impact TKI drug effectiveness. First, we reported clinical response data from the ZENITH20 study (NCT03318939): a single-arm, open-label international multi-center, multi-cohort study to evaluate the efficacy of poziotinib in EGFR and HER2 exon 20 insertion mutation NSCLC^35,36. Cohort 1 enrolled 115 patients with EGFRex20ins NSCLC who had prior treatments (Supplementary Table 1). Median age was 61 (range 33–83) years, 67% were female, and 69% were never-smokers. 98% of patients had adenocarcinoma histopathology, 2% had squamous cell carcinomas. 10% had stable CNS metastasis. All patients had received one to six prior lines of treatment. ZENITH20 cohort 1 efficacy and safety data were presented as an oral abstract at the AACR annual meeting 2020³⁵, but not previously published in journal articles. In the primary analysis population, the primary endpoint objective response rate (ORR) was 14.8% (95%CI, 8.9 to 22.6), with 17 of 115 treated patients achieving partial response (PR) and none achieving complete response. Other key clinical efficacy endpoints were evaluated as secondary endpoints, including: the disease control rate (DCR) was 68.7% (95%CI, 59.4 to 77, Fig. 4a, and Supplementary Table 2). Among the 17 responders, the median duration of response (DoR) was 7.4 months (95% CI, 3.7 to 9.7, Fig. 4b), and median progression-free survival (PFS) for all treated patients was 4.2 months (95% CI, 3.7 to 6.6). The clinical efficacy were similar regardless prior lines of therapy: ORR was 16.2% (n = 37) in those who had ≥ 3 prior lines of therapy, 13.8% (n = 29) in those who had two lines, and 14.3% (n = 49) in those with one prior systemic therapy (Supplementary Table 2). Treatment-emergent AEs occurred in all patients. Treatment-related AEs (TRAEs) were reported in 114 (99%) patients, with 70 (61%) having grade 3 and two patients (2%) having grade 4 TRAEs (Supplementary Table 3). The most common TRAEs were diarrhea (79%), rash (60%), and stomatitis (52%), which was consistent with the prior safety profile reported for poziotinib. Dose interruption was frequent (88% in 97 patients) with the median day of first dose interruption being day 16. Dose reduction rate was 68% and permanent discontinuation rate to treatment-related AE was 10% (12 patients).

Fig. 4. Tumor response in ZENITH20 trial cohort 1.

a Best percentage change from baseline in target lesion size (RECISTv1.1 by Independent Review Committee) in patients with advanced EGFR exon 20 insertions NSCLC (N = 98 for having target lesion and having at least one tumor response assessment scans). N denotes near-loop insertion, F denotes far-loop insertion, there were cases without annotation (b). Swimmer plot for individual responders showing days on treatment and strength of treatment. Gaps indicate dose interruption. Color lines indicate different dose levels.

Based on the preclinical evidence and our previous report from a single-center clinical trial²², we hypothesized that the near-loop EGFR exon 20 insertions could be associated with better responses to poziotinib, and performed an exploratory analysis. In this cohort, 98 patients had at least one tumor measurement after poziotinib treatment, among those, 82 patients had exon 20 insertions location information per local next-generation sequencing (NGS) testing, with 63 (77%) patients whose tumor mutation were in the near-loop amino acids A767-P772 and 19 (23%) in the far-loop H773-C775 amino acids. The mean of the best tumor reduction was significantly different between near- vs. far-loop groups, with 25.9% tumor reduction for near-loop vs. 9.8% for far-loop (p = 0.0014, Fig. 5a, b and Supplementary Data 1). The median progression-free survival (PFS) was also significantly different in patients with near- vs far-loop insertions at 5.5 months (near) vs. 3.5 months (far) (Log Rank p = 0.015), with a PFS hazard ratio of 0.42 (near- vs. far- 95% Confidence Interval [CI] 0.21-0.86, Fig. 5c). The response rates for patients with near vs. far-loop mutations were overall similar at 17.5% (near) vs. 10.5% (far, p = 0.7136). The DCR demonstrated a similar trend for patients with near-loop versus far-loop mutations at 87.3% vs. 68.4% (p = 0.1166). Furthermore, patient’s tumor reduction per RECIST measurement positively correlated with mutation location (R = 0.33, p = 0.0026, Supplementary Fig. 9a). Therefore, the clinical outcomes, including tumor size reduction and PFS, were significantly better in the near-loop group than the far-loop in EGFRex20ins NSCLC patients treated with poziotinib.

Fig. 5. Differential clinical responses with poziotinib and mobocertinibin EGFR exon 20 near- vs. far- insertion lung cancers.

Poziotinib in (a–c) and mobocertinib (d–f). a Waterfall plot of the best response by tumor size reduction by RECIST at the individual patient level in the total population with known EGFRex20 insertion location. b Bar graph of best response by tumor size divided by mutation location. Chi-square test was used to determine the p-value. Far-loop n = 19, near-loop n = 63. c Kaplan-Meier plot of progression-free survival of patients with near loop and far loop mutants. The long-Rank Mantel-Cox approach was used to determine the p-value. d Waterfall plot of the best response by tumor size reduction by RECIST at the individual patient level in the total population with centrally confirmed EGFRex20 insertion location. e Bar graph of the best response by tumor size divided by mutation location. Chi-square test was used to determine the p-value. Far-loop n = 19, near-loop n = 57. f Kaplan-Meier plot of progression-free survival of patients with near-loop and far-loop mutants. The long-rank Mantel-Cox approach was used to determine the p-value. All tests were performed with two-sided hypotheses.

Clinical responses with mobocertinib in near- vs. far-EGFRex20ins NSCLC

Next, we evaluated clinical response data from the previously reported EXCLAIM trial (NCT02716116), a single-arm, open-label international multi-center, multi-cohort study evaluating mobocertinib in NSCLC patients whose tumors harboring EGFRex20ins. We performed similar analyses to investigate the clinical responses association with EGFRex20ins location. In the platinum-pretreated patients (PPP) cohort of 114 patients, the ORR was 28% (95% CI: 20–37), the DCR was 78% (95% CI:69–85), and the median PFS was 7.3 months¹¹. The results from the trial led to the FDA’s initial accelerated approval of mobocertinib in patients with metastatic NSCLC with EGFRex20ins, whose disease has progressed on or after platinum-based chemotherapy. In this cohort, 76 patients had central confirmation of the exon 20 insertion location, with 57 (75%) in the near-loop and 19 (25%) in the far-loop. There was no significant difference in the best tumor reduction between tumors with near- vs. far-loop mutations (38.5% vs. 34.1%, respectively, p = 0.59) and no difference in the DCR between the two groups, 91.1% (near-loop) vs. 94.4% (far-loop, p = 1.00) (Fig. 5d, e and Supplementary Data 2). The median PFS was also not significantly different at 12.0 (near) vs. 13.0 (far) months respectively (Log Rank p = 0.99), with a PFS hazard ratio of 1.0 (far- vs. near-, 95% CI 0.44-2.28, Fig. 5f). In the EXCLAIM trial, there was no correlation between RECIST responses to insertion location (R = 0.061, p = 0.57, Supplementary Fig. S9b). Taken together, the results from the two large clinical trials validated our preclinical observations that EGFRex20ins location can differentially impact TKI sensitivity. Poziotinib was more active in near-loop EGFR exon 20ins with better tumor size reduction and improved PFS, whereas mobocertinib’s clinical activity was demonstrated in both near- and far-groups across EGFRex20ins locations.

Discussion

We and others have reported the heterogeneity of EGFRex20ins sensitivity to TKIs both in clinical and preclinical studies^{4,9,10,34,37–41}. Our data support that EGFRex20ins can be divided structurally into three distinct groups: helical, near-loop, and far-loop mutations, and that these three groups display distinct patterns of TKI sensitivity in vitro. Furthermore, we show that insertion location can have a differential impact on TKI sensitivity in a drug-specific manner. Using MD simulations, we investigated these differences and determined the structural features that provide a potential explanation for the differential drug sensitivity between near- and far-loop EGFRex20ins mutations. The differential impact of insertion location was validated using two large clinical cohorts of patients treated with poziotinib and mobocertinib, respectively. This work established that EGFRex20 near-loop and far-loop insertion mutations have distinct structures and conformations, therefore, TKI development strategies may differ for each group.

In our Ba/F3 cell line system expressing over twenty different exon 20 loop insertions, we showed that the majority of EGFR TKIs have better sensitivity to near-loop insertions. In the ZENITH20 trial cohort 1, the clinical efficacy was limited in the total population, not meeting the previously defined efficacy endpoint, however, the near-loop insertions responded better to poziotinib both at tumor size reduction and at PFS (11.1 months [near] vs. 3.5 months [far]). Interestingly, mobocertinib is the only TKI we evaluated displaying superior in vitro efficacy in far-loop insertions than near, consistent with previously shown efficacy across both near- and far-loop EGFRex20ins mutations⁹. Our analysis from the EXCLAIM trial cohort demonstrated a similar clinical benefit of mobocertinib in both near- and far-loop EGFRex20ins lung cancers, suggesting that mobocertinib is an EGFR TKI preserving efficacy regardless of the insertion locations. The discrepancy between in vitro data showing higher efficacy in far-loop insertions and clinical data could be related to many different factors that also impact clinical outcomes in a moderate size study cohort, including tumor intrinsic factors, such as co-mutation and tumor lineage status, as well as clinical factors, such as patients’ performance status, organ function, and comorbidities, in addition to tumors size, clinical stages, and locations.

For the rest of the tested TKIs, although the drug sensitivity varies in near- vs far-loop insertions, it is important to note that if the drug has a sufficiently large therapeutic window (dictated by the IC₅₀ ratio for mutant EGFR vs. wildtype EGFR), the drug can still be highly effective for both near- and far-loop insertions. In the case of poziotinib, the therapeutic window is narrow due to side effects, so that dose administered in patients was limited, and sufficient inhibition in far-loop insertions was not achieved. Newer TKIs, such as zipalertinib, appear to have a larger therapeutic window, although our in vitro and molecular simulation results indicate higher efficacy in near-loop insertions, it remains to be seen whether the drug can render a similar magnitude of benefit in both near- and far-loop insertions as clinical data becomes available. One of the novel TKIs, sunvozertinib, has demonstrated good clinical efficacy across near- vs. far-loop EGFRex20ins mutations in NSCLC⁴².

While our data applies to EGFR TKIs, these findings might not be applicable to the antibody-based targeting approach. Amivantamab is a bispecific EGFR and MET antibody approved for EGFRex20 mutation lung cancer. The mechanism of action of amivantamab is mediated by extracellular binding of the bispecific antibody and not through small molecule competitive inhibition at the EGFR kinase domain ATP pocket; therefore, the near- and far-loop distinction is likely not be applicable to the antibody-based drugs.

The differential sensitivity of TKIs is driven by the insertion location’s impact on the EGFR kinase domain’s conformational change and the resulting change in the population of active and inactive states. Our observations from the free energy analysis and MD simulations indicated that the near-loop insertion had multiple transition states between the inactive and active states with low free energy barriers showing conformational heterogeneity; in contrast, the far-loop insertion featured fewer conformational states and a high free energy barrier for transition between them, exhibiting rigidity. This difference in conformational flexibility might be one of the underlying factors for the experimental observations with multiple EGFR TKIs having increased efficacy for near-loop insertions. Our simulation results on drug-protein interactions with multiple inhibitors further supported this hypothesis. Generally, drug binding is facilitated by hydrogen bonding and hydrophobic interactions. For zipalertinib and poziotinib, the binding affinities were more favorable in the near-loop insertions than in the far-loop. Though, mobocertinib showed stronger binding affinity towards the far-loop mutant, it also exhibits noticeable affinity towards the near-loop mutant (Fig. 3e, f). Also, mobocertinib had similar hydrogen bonding propensities with both the mutants (Supplementary Fig. 7). Interestingly, zipalertinib and poziotinib had strong hydrogen bonding at different amino acid sites, indicating that each TKI fits the receptor configuration differently, although having similar results in the near- vs. far-loop insertion’s differential drug-protein interaction. The detailed binding modes, which could be obtained through experimental methods such as X-ray crystallography or by TKI binding and kinetics calculations, could further confirm the above observations in the future. It is important to highlight that in our EGFR simulations, we focus solely on the kinase domain and do not consider co-receptor binding or receptor dimerization, which can be important in vivo. By modeling drug binding to the kinase domain in EGFR’s active state, we capture the effects of receptor binding crucial for EGFR kinase activity. Using this approach, our models exhibit strong correlation with in vitro experimental data. While considering entire receptor models for drug binding is ideal, it is computationally impractical at the resolution of our all-atom models and warrants future research.

Other factors could potentially impact drug sensitivity, especially in the clinical setting. At the EGFR mutation level, other than insertion location, the size and specific amino acid sequence of the insertion could also alter the receptor structure and impact drug-protein interaction. Furthermore, other factors extrinsic to EGFR mutations, such as co-occurring genomic alterations (e.g., TP53, MET amplification) as well as lineage status (e.g., mesenchymal, small cell neuroendocrine transformation), impact drug sensitivity and need to be considered independently. Beyond tumor genomics, clinical features (e.g., age, performance status, sites of metastases) influence clinical outcomes as well. Comparing our in vitro vs. clinical trial data, poziotinib preclinical cell line data and clinical tumor size reduction data demonstrated excellent consistency; whereas for mobocertinib, the preclinical data showed favorable efficacy toward far-loop insertions, R = − 0.48, p = 0.02, but the clinical data indicated no difference in near- vs. far-loop insertions (R = 0.061, p = 0.57). This discrepancy could be related to the above-discussed factors. Nonetheless, the totality of the data supports that grouping EGFRex20ins by location (helical and near- vs far-loop insertions) reduces the heterogeneity within groups compared with treating all exon 20 insertions as a single entity and can readily be adapted to clinical use and drug development. Our study was limited to the currently available inhibitors for preclinical evaluation and clinical trial data availability. Newer and higher potency inhibitors can potentially overcome the difference by having a large therapeutic index. Furthermore, due to the technical limitations with current simulation, more detailed descriptions on the drug-protein interaction is needed to enhance our understanding.

In conclusion, this work established that EGFRex20 near-loop and far-loop insertion mutations are distinct in their structures and responses to different TKIs. Our finding of EGFRex20 far-loop insertions are conformationally rigid compared with near-loop insertions, and that most TKIs inhibit far-loop insertions with less potency, provides an opportunity to more precisely tailor the use of TKIs for EGFRex20ins lung cancer patients.

Methods

Ethics statement

The clinical trial data were obtained from the sponsors, Spectrum Pharmaceuticals and Takeda Pharmaceutical Company. The clinical trials were conducted under the supervision of the sponsors after IRB approval. The primary outcomes from the Takeda trial (EXCLAIM trial) have been reported previously PMID34647988. All clinical trial conducts were compliant with the declaration of Helsinki. This research work complies with all relevant ethical regulations, including approval from the Institutional Review Board (IRB) for clinical trials and other protocol guidelines for non-human research per MD Anderson Cancer Center regulations.

Ba/F3 cell generation and IC₅₀ approximation

As previously described^4,34 Ba/F3 cells were from Dr. Gordon Mills (MD Anderson Cancer Center), and cultured in RPMI (Sigma) with 10% FBS, 1% penicillin/streptomycin, and 10 ng/ml recombinant mIL-3 (R&D Biosystems). Stable Ba/F3 cell lines were generated, and drug screening was performed as previously described using Cell Titer Glo^1,4,34. Raw luminescence values were normalized to DMSO control-treated cells, and values were plotted in GraphPad Prism. Non-linear regression models were used to fit the normalized data, and IC_5o values were determined by GraphPad Prism by interpolation of concentrations at 50% inhibition. Drug screens were performed in technical triplicate on each plate and three independent replicates. Correlations between drug sensitivity and mutation location were completed by plotting the calculated average IC₅₀ values against the amino acid residue number using an x-y scatter plot. Using GraphPad Prism, the data was fit to a linear regression, and Pearson’s correlation was used to determine the R and p-values indicated on the plots.

Tyrosine kinase inhibitors

Tyrosine kinase inhibitors were purchased from Selleck Chemicals, with the exception of Tarlox-TKI which was purchased from MedChem Express. Inhibitors were reconstituted in DMSO to 10 mM and stored at − 80 °C as single use aliquots.

Protein modeling and molecular dynamics (MD) simulations

Classical MD and enhanced sampling calculations of monomer WT and mutant proteins without ligands

Unbiased classical molecular dynamics (MD) and enhanced sampling simulations without the bound ligand were performed to elucidate the structural differences in the kinase domains of EGFR wild type (WT), near-loop (S768dupSVD and A767dupASV), and far-loop (H773insNPH and H773insAH) mutations. Each of these systems were simulated from their respective inactive and active conformations to map their underlying free energy (FE) landscape and thereby identify the structural consequence of the mutations.

Initial structure preparation

Inactive and active conformations of wild-type (WT) EGFR domains were obtained from PDB IDs 3W32 and 2GS6, respectively^43,44. In both the PDBs, the crystallographic structure consists of EGFR proteins co-crystallized with either an adenosine triphosphate (ATP) analog or a ligand in the binding pocket. These ligands were removed, and the resulting WT EGFR domain sequence ranged from Q701-L1017. Any missing residues were modeled using the kinematic loop closure algorithm of Rosetta as needed^45,46. For near-loop and far-loop mutants, as their crystallographic structures are not available, the inactive and active conformations of WT EGFR were used as templates for constructing their respective structures using Rosetta. In both the mutants, the three amino acids insertions – SVD and NPH – occurs near the αC helix thereby moving the subsequent residues towards to C-terminus.

Simulation setup

Following the initial structure generation, the proteins (inactive and active structures of WT, near-loop, and far-loop mutants) were solvated with TIP3P water molecules for at least 1 nm on all their sides, and the systems were charge neutralized by adding counter-ions. Periodic boundary conditions were considered in all directions to mimic an infinitely large solvated system. The parameters necessary to compute the inter- and intra-molecular interactions were assigned using the AMBER99SB-ILDN force-field⁴⁷. A timestep of 2 fs was used, and the bonds involving hydrogen atoms were constrained using the LINCS algorithm⁴⁸. All the simulations were performed at a temperature of 310 K and at a pressure of 1 atm. The temperature and pressure were maintained constant using Nose-Hoover thermostat (with time constant $\tau =0.2$ ps) and Parinello-Rahman barostat (with time constant $\tau =2$ ps), respectively⁴⁹. The electrostatic and Lennard – Jones (LJ) interactions were explicitly calculated up to a cutoff of 1 nm. Beyond the cutoff, the electrostatic interactions were handled by the Particle Mesh Ewald (PME) algorithm^50,51 with a Fourier spacing of 0.14 nm. LJ interactions, on the other hand, were shifted by a constant value such that it is zero at and beyond the cutoff (i.e., 1 nm). Prior to initializing any production simulations, all the systems were energy minimized, followed by short MD simulations in the NVT ensemble for 100 ps and in the NPT ensemble for 200 ps. In each of these steps, the protein backbone atoms were harmonically restrained to their current positions with a force constant of 1000 kJ mol^-1nm^-2 while allowing the remaining atoms to relax. Unbiased MD and enhanced sampling simulations pertaining to WT, near-loop, and far-loop mutants were performed using Gromacs 2019.4⁵² and Gromacs 2019.4 + Plumed-2.5.4^53,54, respectively, with the above set of parameters.

Bias exchange metadynamics (BE-MetaD)

To rigorously map the free energy (FE) landscape of WT, near-loop, and far-loop mutants, BE-MetaD⁵⁵ simulations were performed. In the current work, the following two collective variables (CVs) – (1) pseudo dihedral angle formed by the alpha carbons of DFG and FGL residues and (2) difference in salt bridge distance formed between K860/E762 and K745/E762 (for WT), and K863/E762 and K745/E762 (for near-loop and far-loop mutants) – were used to map the FE landscape modulating the inactive – active transitions in the WT and mutant systems. For WT and mutant systems, BE-MetaD simulations were performed for ~ 13–15 $\mu s$. (See Supplementary Note 2 for more details on BE-MetaD simulations). Finally, BE-MetaD simulations were reweighted to obtain the free energy landscapes. The transition pathways alone the free energy landscapes were obtained using the Minimum Energy Path Surface Analysis (MEPSA) program⁵⁶.

Protein unbiased simulations

In addition to the BE-MetaD, unbiased classical MD simulations initialized from the active state of each system (WT, near-loop, and far-loop) were also performed to identify structural differences between them. These simulations were run for ~3$\mu s$ in the NPT ensemble. During these runs, intermediate snapshots were saved every 1 ns and were used for the analyses.

MD simulations of non-covalent protein-ligand complex

Classical MD simulations were performed to estimate the binding affinity of zipalertinib and mobocertinib binding to EGFR near and far insertion mutants. Structural models of EGFR near and far insertion mutants bound to zipalertinib and mobocertinib in their pre-reactive state (non-covalent complex) were considered for MD simulations. In the absence of experimental structure of EGFR S768dupSVD and EGFR H773insNPH mutants, the Prime homology modeling tool available in the Schrodinger package was used for building homology models of the EGFR mutants. X-ray structure of EGFR complexed to a pyrazolopyrimidine inhibitor (PDB entry 5J9Y), which is structurally similar to zipalertinib, was considered as a template for modeling the structure of EGFR near and far insertion mutants bound to CLN-081. The flexible loop regions around the binding pocket were subsequently refined using the Prime loop refinement protocol. A docked model of zipalertinib covalently bound to EGFR mutants was obtained using the covalent docking program CovDock⁵⁷. The covalent link was removed, and the pre-reactive complex was used for subsequent MD simulations.

PDB entries 6JWL and 3UG2 were considered for modeling the structure of EGFR S768dupSVD and EGFR H773insNPH insertion mutant bound to mobocertinib. The crystal structure of AZD9291 (osimertinib) bound to mutant EGFR (PDB entry 6JWL) was considered as a template for modeling the structure of EGFR near and far insertion mutants bound to mobocertinib. AZD9291 (osimertinib) is structurally similar to mobocertinib; hence, structural modifications were carried out on the X-ray structure to generate a non-covalent mobocertinib complex. The “Copy ligand” option was used during homology modeling to transfer the mobocertinib coordinates onto the EGFR near and far insertion models. Following this, 500 ns of production MD runs were carried out using the AMBER simulation package (Version 18). The simulation setup and procedure employed were based on our previous publication¹⁹. The binding free energies of zipalertinib and mobocertinib towards the mutants were calculated using two end-point-based methods, namely MM/PBSA and MM/GBSA, as described earlier¹⁹.

Absolute Binding free energy calculations of non-covalent protein-ligand complex using funnel meta-dynamics

In addition to end-point-based binding free energy methods, we also calculated the absolute binding free energies of the non-covalent protein-ligand complexes using the more accurate funnel metadynamics simulations (FMetaD)⁵⁸. The distance between the center of mass of the binding pocket residues and the center of mass of the ligand was used as the one-dimensional collective variable for the FMetaD calculations. The ligand was restrained to sample the defined funnel region close to the protein binding pocket and the cylindrical solvent region using strong biasing forces^58,59. Further details on the calculations are provided in Supplementary Note 2.

MD simulations mimicking covalent complex

To understand the detailed interactions between the TKI and the protein residues accounting for the covalent linkages, we construct an approximate model of EGFR-bound TKIs and perform several short MD simulations (10 independent simulations each of 50 ns length) to identify the key interactions. Covalent linkages between the protein and the ligand were mimicked using harmonic restraints between 2the reacting atoms (the details of model are provided in Supplementary Note 2). Following the MD simulations, protein-ligand interaction fingerprints were calculated using ProLIF⁶⁰ to classify each interaction as hydrogen-bonded (both donor and acceptor), hydrophobic, etc.

Clinical trial procedure, endpoints, and analysis for ZENITH20 (NCT03318939)

Patients aged 18 years or older were eligible if they were previously treated for locally advanced or metastatic NSCLC with documented EGFR exon 20 insertion mutations. The enrollment period was from October 2017 and March 2021. Patients must have measurable NSCLC disease (per RECIST Guidelines, v1.1). Patients with known brain metastases were eligible if the patient’s condition was stable. Patients received 16 mg poziotinib orally once daily. The dose could be reduced in 2 mg increments, if necessary, in the presence of toxicity. Response evaluation was conducted by a blinded independent committee review using RECIST v1.1 at baseline, after 4 and 8 weeks of treatment, and every 8 weeks thereafter for up to 24 months. Adverse events (AEs) were monitored throughout the study and for 35 days after poziotinib discontinuation and, along with laboratory abnormalities, were graded by investigators according to the National Cancer Institute Common Terminology Criteria for Adverse Events version 4.03. The primary endpoint was objective response rate (ORR). Secondary outcome measures were disease control rate (DCR), duration of response (DoR), progression-free survival. (PFS), and safety and tolerability. The study was conducted in accordance with Good Clinical Practice guidelines and the Declaration of Helsinki and in discussion with the US Food and Drug Administration. All patients provided written informed consent. The study was conducted in accordance with the criteria set by the Declaration of Helsinki. The primary efficacy and safety results from this cohort of the ZENITH20 study was reported in this paper.

Sample size justification

Each cohort will be a single-arm, and the primary test of hypotheses will be based on a single proportion. Sample size calculation is based on single-arm hypothesis testing to reject non-desired ORR of 17% vs. clinically meaningful ORR of 30%. A sample size of 87 patients in each cohort will provide 85% power, for a two-sided test with a significant level of 5% to reject a non-desired ORR of 17% if the observed ORR is 30% or higher. A sample size of 87 patients in each cohort will also provide a 95% CI that contains 30% and rules out 17% as the lower bound is above 17% (Details please see the redacted protocol (Supplementary Note 3, session 8.1, the population definition in session 8.4, and statistical methods in 8.6).

Efficacy analysis

The primary efficacy variable, ORR, will be analyzed descriptively along with the 95% CI for each cohort. The determination of ORR will be based upon assessment by an Independent Radiologic Review Committee. The other outcomes, including PFS, will be analyzed using descriptive statistics. Biomarker analysis is exploratory and descriptive. The analyses performed comply with ICMJE guidelines.

Retrospective analysis of patient outcomes by mutation location

Treatment outcomes, including responses as measured by RECIST 1.1⁶¹ and progression-free survival (PFS), which was measured as time from the beginning of TKI treatment until radiologic progression or death, were collected from patients with EGFR exon 20 mutations enrolled in ZENITH20 NCT03318939 (cohorts 1) and EXCLAIM NCT02716116 clinical trials. Patients were stratified by the location of the mutation within exon 20 into two groups: near-loop (patients with tumors harboring mutations between A767-P772) and far-loop (patients with tumors harboring mutations between H773-R776), as defined by a previous publication¹.

Statistics and reproducibility

For exploratory analyses from both ZENITH and EXCLAIM trials, near- and far-loop mutation best responses, response rates and PFS were determined per clinical trials. Median PFS was determined using the Kaplan-Meier method. Hazard ratios and p-values were determined using GraphPad Prism software and the Mantel-Cox Log Rank method. Statistical software SAS 9.4 (SAS, Cary, NC) and S-Plus 8.0 (TIBCO Software Inc., Palo Alto, CA) were used for all the analyses.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization: X.L., J.P.R., and J.V.H.; Experiments: J.P.R., J.H., and R.S.K.V.; Clinical Analysis: X.L., Y.Y.E., J.P.R., and S.P.; Statistical Analysis: X.L., L.F., and J.P.R.; Software: R.S.K.V. and J.P.; Visualization: X.L., J.P.R., R.S.K.V., and J.P.; Writing – Original Draft: X.L., J.P.R., R.S.K.V., J.B.C., and J.V.H.; Writing – Review and Editing: X.L., J.P.R., R.S.K.V., J.B.C., and J.V.H.; Project Administration: X.L., J.P.R., M.N., J.C., and J.V.H.; Funding Acquisition: X.L., J.P.R., and J.V.H.; Supervision: X.L., J.P.R., J.B.C., and J.V.H.

Peer review

Peer review information

Nature Communications thanks Balazs Halmos, Hiroyuki Yasuda and the other, anonymous, reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The data sets supporting the results reported in this article are available: the redacted ZENITH20 clinical trial protocol can be found in Supplementary Note 3. The deidentified response data in ZENITH and EXCLAIM trials are available in the Supplementary Data 1 and 2. The data are provided in compliance with applicable laws, data protection, and requirements for consent and anonymization. All remaining data can be found in the Article, Supplementary and Source Data files. Source data are provided in this paper.

Competing interests

Xiuning Le, Consulting/advisory fees from Eli Lilly, EMD Serono (Merck KGaA), AstraZeneca, Spectrum Pharmaceutics, Novartis, Regeneron, Boehringer Ingelheim, Hengrui Therapeutics, Bayer, Teligene, Taiho, Daiichi Sankyo, Janssen, Blueprint Medicines, Sensei Biotherapeutics, SystImmune, ArriVent, Abion, BlossomHill, and AbbVie. Research Funding to Institution from Eli Lilly, EMD Serono, ArriVent, Dizal, Teligene, Regeneron, Janssen, ThermoFisher, Takeda, and Boehringer Ingelheim. Travel Support from EMD Serono, Janssen, and Spectrum Pharmaceutics. Jacqulyne P. Robichaux, Employment and stock/shares with AstraZeneca; inventor on patents held by UT MD Anderson Cancer Center licensed to Spectrum Pharmaceuticals for treatment of EGFR/HER2 exon 20 mutant cancers and inventor on patent held by UT MD Anderson Cancer Center regarding EGFR mutation subtypes and methods of treatment. Monique Nilsson, Patent and license fees from Spectrum. R.S.K.Vijayan, No competing interest. Ashwin Ravichandran, No competing interest. Rizwan Qureshi, No competing interest. Jia Wu, Siemens Healthcare. Yasir Y. Elamin, reports honoraria: CME Matters, Clinical Care Options CME, Research to Practice CME, Medscape CME, Biomedical Learning Institute CME, MLI Peerview CME, Prime Oncology CME, Projects in Knowledge CME, Rockpointe CME, MJH Life Sciences CME, Medical Educator Consortium, HMP education consulting or advisory role: AstraZeneca, Genentech (Roche), Exelixis, Takeda Pharmaceuticals, Eli Lilly, Amgen, Iovance Biotherapeutics, Blueprint Pharmaceuticals, Regeneron Pharmaceuticals, Natera, Sanofi, D2G Oncology, Surface Oncology, Turning Point Therapeutics, Mirati Therapeutics, Gilead Sciences, Abbive, Summit Therapeutics, Novartis, Novocure, Janssen Oncology, Anheart Therapeutics Reserach Funding: Genentech/Roche, Merck, Novartis, Boehringer Ingelheim, Exelixis, Nektart Therapeutics, Takeda Pharmaceuticals, Adaptimmune, GlaxoSmithKline, Janssen Pharmaceuticals, Abbvie Pharmaceuticals and Novocure. M.V.N. reports research funding to institution: Mirati, Novartis, Checkmate (ended), Alaunos, AstraZeneca, Pfizer, Genentech, Navire; consultant or advisory board: Mirati, Merck/MSD, Novartis and Genentech. Lingzhi Hong, No competing interest. Jun Pei, No competing interest. Jun He, No competing interest. Sonia Patel. Employment from BioNTech. Hibiki Udagawa, research grants from Takeda and Boehringer Ingelheim outside the current study. Sriramvignesh Mani, No competing interest. Chang Woon Jang, No competing interest. Jeffrey M. Clarke, Bristol-Myers Squibb, Genentech, Spectrum, Adaptimmune, AbbVie, Moderna, AstraZeneca, Grid Therapeutics, Abel Zeta, Amgen, Pfizer. Speaking: Merck, AstraZeneca, Amgen. Travel: AstraZeneca, BMS, Amgen. Advisory: AstraZeneca, Merck, Pfizer, Spectrum, Genentech, Novartis, Turning Point, G1 Therapeutics, Vivacitas, Omega, Amgen, Corbus, Sanofi, Coherus, CDR Life, AbbVie, Blackdiamond. Nishan Tchekmedyian, Stock and Other Ownership Interests: Portola Pharmaceuticals, Halozyme (I), Infinity Pharmaceuticals (I), Global therapeutics (I), Biomarin (I), Exelixis (I), Trillium Therapeutics (I)Consulting or Advisory Role: Seattle Genetics, IntrinsiQ, Foundation Medicine, Amgen. Research Funding: Bristol Myers Squibb (Inst), Spectrum Pharmaceuticals (Inst), Turning Point Therapeutics (Inst), Cullinan Oncology (Inst), Rain Therapeutics (Inst), Takeda (Inst), Amgen (Inst). Travel, Accommodations, Expenses: IntrinsiQ. Jonathan W. Goldman, Consulting or Advisory Role: AstraZeneca, Bristol Myers Squibb, Lilly, Amgen, Pfizer. Research Funding: Lilly (Inst), Genentech/Roche (Inst), Bristol-Myers Squibb (Inst), AstraZeneca/MedImmune (Inst), Array BioPharma (Inst), AbbVie, Corvus Pharmaceuticals (Inst), Spectrum Pharmaceuticals (Inst), Advaxis (Inst), Pfizer (Inst). Travel, Accommodations, Expenses: AstraZeneca. Mark Socinski, Honoraria: Genentech, Bristol Myers Squibb, Celgene, AstraZeneca, Guardant Health, Bayer, Merck, Roche/Genentech, Lilly, Genentech, AstraZeneca, MedImmune, Lilly, Janssen, Novartis. Speakers’ Bureau: Genentech, Bristol Myers Squibb, AstraZeneca, Boehringer Ingelheim, Bayer, Merck, Amgen, Blueprint Medicines, G1 Therapeutics, Guardant Health, Lilly, Regeneron/Sanofi, Jazz Pharmaceuticals, Janssen Oncology. Research Funding: Genentech (Inst), Spectrum Pharmaceuticals (Inst), AstraZeneca/MedImmune (Inst). Gajanan Bhat, Employment from Spectrum Therapeutics. Sharon Leu, Employment from Spectrum Therapeutics. Veronica Bunn, Employment from Takeda. Zhenqiang Su, Employment from Takeda. Sylvie Vincent, Employment from Takeda. John W. Lawson, No competing interest. Jason B. Cross, No competing interests. John V. Heymach, Advisory Committees: Genentech, Mirati Therapeutics, Eli Lilly, Janssen, Boehringer Ingelheim, Regeneron, Takeda, BerGenBio, Jazz, Curio Science, Novartis, AstraZeneca, BioAlta, Sanofi, Spectrum, GlaxoSmithKline, EMD Serono, BluePrint Medicine and Chugai; research support: AstraZeneca, Boehringer Ingelheim, Spectrum, Mirati, Bristol Myers Squibb and Takeda; licensing or royalties: Spectrum.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-025-61817-8.