Exposure-Response Analysis of Repotrectinib to Support the Dose Recommendation for Patients With ROS1-Positive NSCLC or NTRK-Positive Solid Tumors.

1
Introduction
Repotrectinib is a potent, small‐molecule inhibitor that targets tyrosine kinase encoded by c‐ros oncogene 1 (ROS1) and tropomyosin‐related kinase (TRK) gene families. It is specially designed to overcome on‐target resistance mutations in advanced solid tumors harboring ROS1 or neurotrophin receptor tyrosine kinase (NTRK1‐3) gene fusions [1, 2]. In TRIDENT‐1 (NCT03093116), a multicenter, single‐arm, open‐label trial, repotrectinib demonstrated robust efficacy and a manageable safety profile in patients with ROS1‐positive non‐small cell lung cancer (NSCLC) and NTRK‐positive solid tumors, including both tyrosine kinase inhibitors (TKI)‐naïve and TKI‐pretreated subjects [3, 4, 5]. Similarly, the CARE (NCT04094610) study, an open‐label trial evaluating safety and tolerability in pediatric (< 18 years) and young adult (≥ 18 and ≤ 25 years) patients, showed promising efficacy [6, 7]. Repotrectinib received FDA approval in November 2023 for the treatment of adults with advanced or metastatic ROS1‐positive NSCLC [8]. In June 2024, the FDA expanded its approval to include adult and pediatric patients ≥ 12 years with advanced or metastatic solid tumors harboring NTRK gene fusion [9]. Notably, it became the first treatment in the United States for patients previously treated with TKIs. Subsequently, in January 2025, EMA approved repotrectinib for treating adults and pediatric patients ≥ 12 years with advanced or metastatic ROS1‐positive NSCLC and NTRK‐positive solid tumors [10]. The approved dosing regimen, consistent across both FDA and EMA, is 160 mg orally once daily for 14 days, followed by 160 mg twice daily, with or without food (160 mg QD/BID).
Repotrectinib is primarily metabolized by cytochrome P450 3A4 (CYP3A4) and exhibits time‐dependent autoinduction in clearance [8]. A two‐compartment population pharmacokinetic (popPK) model with first‐order oral absorption was developed to characterize both time‐ and concentration‐dependent autoinduction [11]. The PopPK model suggested that body weight (BW), age, race, and food status affected exposure and warranted further evaluation in the context of exposure–response relationships for safety and efficacy. To evaluate the impact of the exposure difference across covariates or subgroups on efficacy and safety, and to confirm the recommended dose of 160 mg QD/BID, this paper describes the development of E–R models for repotrectinib. Efficacy models were developed to characterize the relationship between repotrectinib exposure and efficacy endpoints—objective response rate (ORR) and progression‐free survival (PFS)—in subjects with ROS1‐positive NSCLC or NTRK‐positive solid tumors, respectively. Additionally, safety models were developed to evaluate the relationship between repotrectinib exposure and the time to first occurrence of selected adverse events, including Grade 2 or higher (Gr2+) dizziness, Gr2+ anemia, Grade 3 or higher (Gr3+) treatment‐emergent adverse events (TEAEs), and Gr2+ neurologic adverse events (NEAEs: including dyspnea, dysgeusia, paresthesia, and ataxia), as well as dose reduction or interruption due to AEs (DRDIs) in the pooled patient population with advanced solid tumors harboring anaplastic lymphoma kinase (ALK), ROS1, or NTRK1‐3 fusions. This work emphasizes the importance of selecting an appropriate exposure measure when the early onset of safety events requires dose frequency changes, which can negatively impact the exposure–response relationships. The developed E–R models informed the benefit–risk assessment and supported confirmation of the optimal dosing regimen, dose modifications to maximize efficacy and manage safety, and flexibility in repotrectinib administration with or without regard to food. Understanding the exposure efficacy and safety relationships is critical for guiding treatment decisions and ensuring safe use of repotrectinib in different subpopulations.

2
Methods
2.1
Analysis Data
Due to the intended indication for repotrectinib that targets ROS1‐positive NSCLC and NTRK‐positive solid tumors, the E–R efficacy analyses were conducted separately by indication and endpoint. Specifically, ORR and PFS were each modeled independently in ROS1‐positive NSCLC (n = 127) and in NTRK‐positive solid tumors (n = 88), resulting in four distinct models. Data were not pooled across indications in the efficacy analyses to preserve indication‐specific characteristics (Table S1; data cutoff: March 2023 from TRIDENT‐1). The E–R safety analysis pooled data across tumor types and included an additional 287 subjects with ALK‐positive solid tumors, resulting in a total of 502 treated subjects with ALK, ROS1, or NTRK1–3 alterations from the TRIDENT‐1 study (Table S2). One model was developed for each of the five safety endpoints, resulting in five distinct safety models.
The probability of objective response (Pr[OR]), which reflects the probability of individual‐level response status (yes/no), underlies the observed ORR—the population‐level proportion of patients who achieve a confirmed complete or partial response. Pr(OR) was characterized using logistic regression. To select the most appropriate exposure measures to characterize the E–R relationship for (Pr[OR]), several exposure measures were assessed in the full ORR models (Table S3). These included simulated average concentrations on the first day (Cavg1), at steady state (Cavgss), and over the first 28 or 56 days using actual dosing records (Cavg28 and Cavg56). Cavg56 was selected as the final exposure metric, based on the strength of the E–R relationship, pharmacological rationale, and timing of tumor response, as well as alignment with the timing of dose reductions and QD‐to‐BID titration.
PFS and 5 safety endpoints were characterized using Cox proportional hazards (CPH) models. Time‐varying cumulative half‐daily average concentration (Cavgc) from Day 1 to the event or censor was used to capture the dynamic changes in drug exposure during treatment. Cavgc was simulated using Empirical Bayes estimates (EBEs) of pharmacokinetic parameters and the actual dosing history, accounting for dose interruptions and dose reductions; Figure S1 provides more detail. The popPK model and source NONMEM code are available in a previously published paper [11].

2.2
Analysis Methods
The relationships between Cavg56 and Pr(OR) were characterized by two separate logistic regression models—one for ROS1‐positive NSCLC and one for NTRK‐positive solid tumors—each representing the indication‐specific exposure–response relationship. The logit (log‐odds) of Pr(OR) is given by:where γ0 is a scalar parameter represent baseline log‐odds of response, γ is a vector of coefficients quantifying the effect of the predictor variables X, on the Pr(OR). The model parameters were estimated by maximum likelihood.
The relationship between Cavgc and time to first occurrence of PFS or safety events was characterized by semiparametric CPH models—two models for PFS (one for each indication) and five models for the safety (one for each endpoint). The hazard of event in the CPH model was expressed as:where λ0ti is the baseline hazard, and X is a vector of predictor variables, including time‐dependent exposure and other covariates (Table S4). The parameter vector β was estimated by maximum partial likelihood. An increased or decreased risk was determined based on HRs (hazard ratios), with values greater than 1 indicating increased risk and values less than 1 indicating decreased risk.
In both logistic regression and CPH full models, linear‐ and log‐transformed functional forms of exposure were assessed, along with pre‐specified covariate effects listed in Table S4. The potential interaction effect of exposure and the significant covariates were also evaluated. Model selection was guided by Bayesian information criteria (BIC) [12].

2.3
Model Evaluation
The model was evaluated using a visual predictive check (VPC). For ORR models, observed ORR was compared with the model‐predicted 5th and 95th percentiles, based on 1000 simulations of predicted Pr(OR) values for all subjects in the analysis dataset. For PFS or 5 safety endpoints, model‐predicted cumulative distributions of events were compared with the observed distribution determined by nonparametric Kaplan–Meier analysis. The 5th and 95th percentiles of the model‐predicted distributions of events were constructed based on 1000 simulations for each patient in the analysis dataset.

2.4
Model Application
2.4.1
Impact of Dosing Regimens and Food Status on Safety
To evaluate the impact of different dosing regimens and food status on efficacy and safety, the developed E–R full models were applied to predict and compare between the following scenarios:
Dosing regimen (under unknown food status; QD/BID: QD for 14 days then BID).
(1) 40 mg QD/BID; (2) 80 mg QD/BID; (3) 120 mg QD/BID; (4) 160 mg QD; (5) 160 mg QD/BID; (6) 200 mg QD/BID.
Food status (160 mg QD/BID).
(1) Unknown food status; (2) fasted food status (overnight fast); (3) fed status; (4) modified fasting status (no food or beverages 1 h before and 2 h after dosing).
Repotrectinib exposure (both Cavg56 and Cavgc) of the 502 adult patients from TRIDENT‐1 was simulated using individual EBEs from the popPK model [11]. All covariates in the E–R full model were sampled from the distribution of 502 adult patients.

2.4.2
Magnitude of the Exposure Effects on PFS and Safety (Typical Subjects)
To illustrate the impact of repotrectinib exposure on the probability of PFS and 5 safety endpoints, the time course of repotrectinib Cavgc was simulated using 502 adult subjects individual EBEs. In the ER model simulations, all other covariates were set to their reference values to ensure that any predicted differences are attributable to exposure. All analyses were performed on Intel Xeon‐based multi‐core Central Processing Unit servers running Ubuntu 18.04 on Amazon Web Services. Nonlinear mixed‐effects modeling software (NONMEM; Version 7.4.3; ICON, Hanover, MD, USA) was used for PopPK simulation. All exploratory data analyses, presentations of E–R analysis, VPCs, and model application simulations were performed using R (version 4.3.1). The E–R models were developed using the “rms” R package for logistic regression analyses of ORR and the “survival” R package for time‐to‐event analyses. Example R code for generating visual predictive checks (VPCs) and simulations was provided in Supplement S1.
The dataset used for the modeling and simulation work is available upon request through an Independent Review Committee (IRC) via the Bristol Myers Squibb data sharing portal: https://www.bms.com/researchers‐and‐partners/independent‐research/data‐sharing‐request‐process.html.

3
Results
3.1
OR Models Development
Different static exposure measures were evaluated in the logistic regression models (Table S3). For ROS1‐positive NSCLC, the odds ratio of OR (for every unit increase in log‐transformed exposure) increased in the following order: Cavg1, Cavgss, Cavg28, and Cavg56. For NTRK‐positive solid tumors, a negative exposure–response relationship was observed across all exposure metrics except Cavg56 (Table S3). The steeper exposure–response relationship observed for Cavg28 or Cavg56 in ROS1‐positive NSCLC was consistent with the fact that these metrics account for dose reduction and titration (i.e., switch from QD to BID on Day 14). Based on this result—along with the timing of tumor response and dose modifications, Cavg56 was selected as the most appropriate exposure measure for characterization of the E–R relationship for Pr(OR).
The models with log‐linear Cavg56 functional form had the lowest BIC value for both ROS1‐positive NSCLC and NTRK‐positive solid tumors; therefore, they were selected in the full models (Table S5). Full model parameter estimates and the odds ratio with 95% CIs of Pr(OR) across the predictor ranges are shown in Figure 1 and Table S6. The odds of Pr(OR) in TKI‐pretreated subjects were 85% lower than TKI‐naive subjects in ROS1‐positive NSCLC (Figure 1A); no significant interaction between TKI‐pretreatment and Cavg56 was observed, indicating the exposure–response relationship was the same for TKI naive and pretreated subjects and suggesting lower tumor response for TKI pretreated (versus TKI naive) subjects was not due to lower repotrectinib exposure (Table S5). While exposure (Cavg56) was not a statistically significant predictor on Pr(OR) for subjects with NTRK‐positive solid tumor (odds ratio 1.33 [95% CI 0.5266, 3.372] included the null value of 1), the odds of Pr(OR) were 9.57 times higher in subjects with NTRK3 mutation than those with NTRK1‐2 (Figure 1B) and no significant interaction was observed between genetic mutations and exposure (Table S5). Race, PS (baseline Eastern Cooperative Oncology Group performance status), age, baseline body weight, and baseline tumor size were not significant predictors of Pr(OR) for either the ROS1‐positive NSCLC or the NTRK‐positive solid tumor indications. VPCs of probability of achieving OR with respect to exposure quartiles are shown in Figure S2. The predictions are generally consistent with the observed data for ROS1‐positive NSCLC. For NTRK‐positive solid tumors, however, substantial variability of response rates was observed across different mutation types and quantiles of exposure, which may be attributed to the limited number of subjects in each category (Table S1: 83 NTRK‐positive solid tumors, with 36 responders and 47 non‐responders).

3.2
OR Models Application
Food status significantly affected repotrectinib exposure. On Day 1, subjects in the fed, modified fasted, and unknown food status groups had 2–3 times higher exposure than those in the fasted group. By Day 14 and at steady state, exposures across these groups remained comparable, though they had 1–2 times higher exposure than the fasted group [11]. Food status was not included as a covariate in the E–R models; instead, its potential effect on efficacy and safety was assessed indirectly through its impact on exposure. This approach was justified given that food status was not expected to influence efficacy or safety through biological mechanisms beyond its effect on drug exposure.
To evaluate the efficacy of different dosing regimens and the impact of food status on the efficacy, the two full models were applied to predict the probability of OR. For ROS1‐positive NSCLC, the 160 mg QD/BID regimen showed higher predicted Pr(OR) compared to the 160 mg QD, with 85.0% vs. 76.1% in TKI‐naive and 46.6% vs. 33.1% in TKI‐pretreated (Table S7, Figure 2A). The predicted Pr(OR) in TKI‐naive subjects was approximately 2‐fold higher than that in TKI‐pretreated subjects at 160 mg QD/BID. In subjects with NTRK‐positive solid tumors, Cavg56 was not a significant predictor of Pr(OR) and a flat E–R relationship was predicted at the evaluated dose (Table S7, Figure 2A), consistent with the observations in the VPC plots (Figure S2).
The median predicted probability of OR from low to high is in the order of fasted, unknown, modified fasted, and fed (at 160 mg QD/BID) (Table S8, Figure 2B). However, the predicted OR showed overlapping 90% PIs among the different food conditions with relatively small numeric differences.

3.3
PFS Models Development
For the time to event models, a time‐varying exposure measured was selected to account for the changing concentrations due to dose reductions to manage safety events and for the up titration from QD to BID on Day 15. The use of time‐varying exposures also avoids the immortal time bias (i.e., predicting past events with future exposures) that can result when using static measures of exposures [13]. The log‐linear function of Cavgc, which had the lowest BIC value, was selected as the full model (Table S9). Higher Cavgc was significantly associated with a lower risk of disease progression or death in subjects with ROS1‐positive NSCLC (Table S10). A similar trend was observed for NTRK‐positive solid tumors, although the effect was not statistically significant (HR 0.6073 [95% CI 0.2867, 1.286] included the null value 1) (Table S10). The use of time‐varying Cavgc in the E–R analysis posed challenges for visualizing the exposure effect on PFS, as exposure levels change over time rather than remaining constant like other covariates. To address this, we simulated the time course of predicted PFS corresponding to the time course of Cavgc, with all other covariates set to their reference values, using the full model (Figure S3).
The estimated effects of covariates that are constant over time in the full models are shown in Figure 3 and Table S10. In ROS1‐positive NSCLC, smaller baseline tumor size, female sex (compared to male), and TKI‐naive status (compared to TKI‐pretreated) were significantly associated with lower risk of disease progression or death. For subjects with NTRK‐positive solid tumors, the risk of disease progression or death was significantly lower for those with smaller baseline tumor size, TKI‐naive status (compared to TKI‐pretreated), and NTRK3 (compared to NTRK1‐2) (Figure 3, Table S10). Similar to ROS1‐positive NSCLC, no significant interactions were observed between Cavgc and the significant covariates, suggesting the differences in predicted PFS across the significant covariates were not due to exposure differences (Table S9).
The model‐predicted PFS was consistent with the observed K‐M across sub‐groups of both ROS1‐positive NSCLC and NTRK‐positive solid tumors, confirming adequate model performance; Figure S4.

3.4
PFS Models Application
The full models were used to predict the cumulative probability of PFS for different dosing regimens and food status listed in the methods section. The 160 mg QD/BID regimen demonstrated higher PFS than 160 mg QD for both ROS1‐positive NSCLC and NTRK‐positive solid tumors (Figure 4A and Table S11). For NTRK‐positive solid tumors, at 160 mg QD/BID, the highest 1‐year PFS probability was predicted in TKI‐naive NTRK3 subjects (0.788), followed by TKI‐naive NTRK1‐2 (0.440), TKI‐pretreated NTRK3 (0.437), and TKI‐pretreated NTRK1‐2 (0.0894). The model‐predicted probability of PFS across different food statuses at 160 mg QD/BID was similar, with minimal numeric difference across indications Figure 4B and Table S12.

3.5
Safety Models Development
The development of ER safety models for the 5 safety endpoints followed BIC selection criteria, with results detailed in Table S13. To illustrate the effect of time‐varying repotrectinib exposure on the risk of safety events, cumulative probabilities of safety events were simulated across Cavgc quantiles (at 160 mg QD/BID) using the full models (Figure S5). Subjects with higher repotrectinib exposure per unit increase had significantly increased risk of Gr2+ dizziness (HR: 1.002 [95% CI 1.001, 1.003]; Table S14) and DRDIs (HR: 1.002 [95% CI 1.001, 1.003]; Table S14). Exposure did not have a statistically significant effect on the remaining safety events, as the 95% CI for HR included the null value (Table S14). Additionally, no significant exposure–covariate interaction effects were identified, as indicated by the lack of BIC improvement during model development (Table S13).
The estimated effects of covariates that are constant over time in the full E–R safety models were summarized in Table S14.
Gr2+ dizziness: Risk increased with older age (HR: 2.59 for 95th percentile age, 75 years old vs. median age, 56 years old; Table S14, Figure 5A).
DRDIs: Risk increased with older age (HR: 1.64 for 95th percentile age, 75 years old vs. median age, 56 years old) and higher BW (HR: 1.71 for 95th percentile BW, 105 kg vs. median BW, 69 kg; Figure 5A). Additionally, NTRK‐positive subjects had a 63% higher risk of DRDI than ROS1‐positive subjects (Figure 5B).
Gr2+ anemia: baseline performance status fully active subjects (PS = 0) had 48% lower risk than the restricted subjects (PS = 1). Lower BW was associated with increased risk (HR: 1.74 for 5th percentile BW, 47.6 kg vs. median BW, 69 kg; Figure S6A).
Gr3+ TEAEs: Risk was 43% lower in the fully active subjects (PS = 0) compared to the restricted subjects (PS = 1), 33% lower in Asian compared to Caucasian, and 34% higher in males than females (Figure S6B).
Gr2+ NEAEs: Risk increased with age (HR: 1.41 for 95th percentile age, 75 years old vs. median age, 56 years old) and higher BW (HR: 1.84 for 95th percentile BW, 105 kg vs. median BW, 69 kg; Figure S6C).
VPCs showed the model predicted cumulative probabilities of safety events were generally in good agreement with observed data; indicating the reliable characterization of the probability of safety events (Figure S7).

3.6
Safety Models Application
The model‐predicted median cumulative probability of 5 safety endpoints across different dose regimens is shown in Figure 6A. The predicted probabilities of Gr2+ dizziness, Gr2+ NEAEs, and DRDI increased with higher doses, supporting the dose reduction strategy for managing AEs. However, the probabilities for 160 mg QD and 160 mg QD/BID were comparable, with overlapping 90% PIs. In contrast, Gr2+ anemia and Gr3+ TEAEs showed comparable probabilities across all dose levels. The effect of food status on AEs at the 160 mg QD/BID is shown in Figure 6B. The median probabilities of Gr2+ dizziness and DRDI were generally higher under fed, modified fasted, and unknown conditions compared to fasted conditions. In contrast, the probabilities of Gr3+ TEAEs, Gr2+ anemia, and Gr2+ NEAEs were comparable across different food statuses. Overall, the model‐predicted probabilities of all evaluated AEs were overlapping across food conditions, with only minor numeric differences.

4
Discussion
Selecting an appropriate exposure measure was crucial to account for dose reductions and titrations, and to minimize potential negative or shallow E–R relationships [14, 15]. Various exposure measures were evaluated in the OR models for ROS1‐positive NSCLC and NTRK‐positive solid tumors, including Cavg1, Cavg28, Cavg56, and Cavgss (Table S3). The exposure measure Cavg56 resulted in steeper E–R relationships for ROS1‐positive NSCLC and changed a negative E–R relationship into a positive one for NTRK‐positive solid tumors compared to the other measures. Cavg56 reflected the impact of dose reductions, which typically occurred within the first two months of treatment—as shown by the Kaplan–Meier plot of DRDIs (Figure S8), where the cumulative incidence of dose modifications or reductions due to AEs increased sharply during the first 2 months and plateaued thereafter, and also accounted for the scheduled dose titration on Day 15. Since the median time to respond to repotrectinib was approximately 2 months, Cavg56 was selected as the appropriate exposure measure for characterization of the E–R relationship for Pr(OR). In addition, this selection was supported by the pharmacological rationale of maintaining adequate exposure levels over the first 2 treatment cycles (56 days) before the first tumor assessment around Cycle 3 Day 1. The high correlation (r = 0.95) between Cavg56 and Cavg28 (Figure S9) further validated its selection as a reliable exposure measure for this analysis.
For PFS and the five safety endpoints, more than half of the subjects experienced DRDIs (Figure S8), highlighting the inadequacy of static summary exposure measures in capturing the time‐varying exposure of repotrectinib. To address this limitation, time‐varying exposure metrics were explored. Initially, time‐varying half‐daily average concentration (Cavg) was explored; however, this approach resulted in a negative E–R relationship for PFS and some of the safety endpoints (data not shown). This counterintuitive finding was likely due to early safety events that led to a DRDI reducing drug exposure, as over half of the subjects experienced the DRDI in the first 2 months of treatment (Figure S8). To mitigate the impact of dramatic exposure fluctuation associated with DRDI, a time‐varying cumulative half‐daily average concentration (Cavgc) was introduced that no longer resulted in a negative E–R relationship for PFS and safety events. Cavgc provided a more accurate representation of dynamic drug concentrations over the period of efficacy or safety assessment. This metric accounted for differences in exposure between QD and BID dosing frequency, enabling the assessment of different dosing regimens included in the analysis. Additionally, Cavgc avoided the immortal time bias that often is associated with static measures of exposures [13]. By incorporating the dynamic changes in dosing and drug concentrations, Cavgc provided a robust framework for assessing E–R relationships in both efficacy and safety time‐to‐event analyses.
The E–R models for efficacy and safety provided good descriptions of repotrectinib's effect in adult subjects with ROS1‐positive NSCLC or NTRK‐positive solid tumors. Higher repotrectinib exposure was associated with a higher probability of OR and longer PFS in ROS1‐positive NSCLC. While exposure was not significant for the probability of OR and PFS for NTRK‐positive solid tumors, a trend toward longer PFS with higher exposure was observed. The 160 mg QD/BID regimen demonstrated numerically improved efficacy over 160 mg QD for both ORR and PFS; although this difference was not statistically significant. Importantly, the switch from QD to BID dosing on Day 15 helps to mitigate the reduced exposure due to the autoinduction and the high incidence of DRDIs early in treatment. Although 200 mg QD/BID was included in the simulations to explore the upper limit of the E–R relationship, the 200 mg BID dose was associated with serious adverse events [16], and only two subjects received it in the clinical study, limiting confidence in model predictions for this dose (Table S2). Efficacy was comparable across food conditions.
The E–R safety models demonstrated that higher exposure increased the risk of AEs in adult subjects with advanced solid tumors harboring ALK, ROS1, or NTRK1‐3 rearrangements, such as Gr2+ dizziness, Gr2+ NEAEs, and DRDIs. While the risks of these AEs were higher at increased doses, the differences between 160 mg QD and 160 mg QD/BID were minimal (e.g., Gr2+ dizziness 15% vs. 13%), with overlapping 90% PIs. Safety outcomes were also consistent across food conditions.
Using clinical data from the TRIDENT‐1 trial, this comprehensive analysis confirmed a favorable benefit–risk profile for repotrectinib, supporting its use at the recommended dose of 160 mg QD/BID in ROS1‐positive NSCLC and NTRK‐positive solid tumors regardless of food intake. The identified significant covariate effects on efficacy and safety did not necessitate the need for dose adjustment.

Author Contributions
S.D. wrote the manuscript; S.D., Z.H., J.S., J.L., M.L., A.R., L.Z., A.K., and L.H. designed the research; S.D., Z.H., and L.H. performed the research; and S.D., Z.H., J.S., L.H., A.R., and A.K. analyzed the data.

Conflicts of Interest
All authors are or were employee and/or stock shareholders of Bristol Myers Squibb at the time of the work being conducted.

Supporting information

Data S1: psp470102‐sup‐0001‐Supinfo.zip.