FDA-AACR Strategies for Optimizing Dosages for Oncology Drug Products: Selecting Optimized Dosages for Registrational Trials

Abstract

The maximum tolerated dose has historically been the recommended phase two dose, and this dosage has typically been evaluated in registrational clinical trials for oncology drugs. With the emergence of targeted therapies, this approach may lead to the investigation of unnecessarily high dosages that elicit additional toxicity without added benefit. The utilization of innovative trial designs and model-informed approaches during clinical development can potentially lead to more informed dosage selection. Exposure-response analyses, clinical utility index, and other model-informed approaches have been successfully applied to understand preliminary activity and safety data for various classes of modern oncology drugs, providing insight to support the proposed dosage(s) for the registrational trial. Seamless trial designs have also played an important role in dosage selection by leveraging pre-planned flexibilities and statistical procedures to increase efficiency during the conduct of trials. Critically, both approaches can be fit for purpose, allowing for adaptation and the usage of the totality of relevant clinical and nonclinical data. Despite this, the evaluation of maximum tolerated dose remains prevalent in registrational trials. This article, the third in a series of three describing best-practice approaches to dosage optimization in oncology drug development, highlights successful applications of and relevant considerations for innovative trial designs and model-based approaches to aid the selection of better optimized dosages for evaluation in registrational clinical trials.

Introduction

Registrational trials for oncology drugs typically evaluate a single dosing regimen to establish safety and efficacy. This single dosage is historically the maximum tolerated dose (MTD), determined by the observance of dose-limiting toxicities in a dose-finding trial wherein a dose range is evaluated in a small number of patients for a short duration. This approach adequately identified appropriate dosages for cytotoxic chemotherapy agents where the therapeutic window is narrow but may unnecessarily select higher dosages for modern oncology drugs where lower dosages can exhibit similar activity with less toxicity. Given the shift from traditional chemotherapy agents toward targeted therapies over the past 20 years, it is time to reexamine this dosage-selection paradigm, especially given patient investment in the reevaluation of dosing in oncology drug development (1,2). Encouraging an innovative environment where oncology drug development programs identify optimized dosages that maximize efficacy and safety before beginning registrational trials will be critical for future development programs and is central to U. S. Food and Drug Administration (FDA) initiatives focused on dosage optimization such as Project Optimus, Model-Informed Drug Development Paired Meeting Program (an FDA program that affords sponsors who are selected for participation the opportunity to meet with FDA to discuss model-informed approaches to support the development of a drug product), and Fit-for-Purpose Initiative (an FDA initiative that provides a pathway for regulatory acceptance of dynamic tools for use in drug development) (3–5).

In February 2024, the FDA Office of Clinical Pharmacology and American Association for Cancer Research (AACR) co-sponsored a public workshop titled “Optimizing Dosages for Oncology Drug Products: Quantitative Approaches to Select Dosages for Clinical Trials” to discuss challenges and opportunities in the dosage optimization of oncology drugs (6). In this article, the third and final in a series of three discussing approaches to dosage optimization at various stages of clinical development, we present considerations and learnings regarding model-based approaches and innovative trial designs that support the selection and investigation of an optimized dosing regimen in registrational trials. For deeper discussions regarding dosing selection for the dose-escalation portion of first-in-human clinical trials or for selecting dosages for further exploration based on nonclinical and early clinical data such as alternative safety and activity endpoints, please see the first and second articles in this series respectively (7,8).

Considering the Totality of Efficacy and Safety Data to Aid Registrational Trial Designs Through Modeling

The MTD approach briefly described above focuses only on initial safety data to select the dosing regimen for evaluation in subsequent clinical trials. A holistic approach that utilizes all available nonclinical and clinical data can lead to more optimized dosages that maximize benefit/risk profile by integrating both safety and efficacy. This alternative approach requires study designs capable of comprehensively and systematically collecting and evaluating all applicable data iteratively as it emerges during drug development – from the first-in-human (FIH) trial through the registrational trial, and potentially including orthogonal data from drugs in the same class. Relevant data types are summarized in Table 1 and are discussed throughout this article alongside considerations for their inclusion in decision-making.

Table 1.

Non-exhaustive list of data for potential inclusion in future model-based approaches to facilitate dosing decisions for registrational clinical trials in oncology. Many of these data types can be leveraged across both the drug being evaluated and previous drugs in the same class.

Key Data Area	Data Subtype	Example Assays/Datapoints
Nonclinical Data	Pharmacokinetics	Plasma drug concentration Tumor partitioning
	Pharmacodynamics	Target expression Target engagement/occupancy
	Efficacy in model systems	Tumor growth inhibition - in vitro or in vivo Validated biomarker response - in vivo
Clinical Pharmacology	Pharmacokinetics	Maximum concentration Time to maximum concentration Trough concentration Elimination half-life Area under the curve
	Pharmacodynamics	Effect on pharmacodynamic biomarker Clinical receptor occupancy
Clinical Safety	Landmark dosage modifications	Incidence of interruption Incidence of dose reduction Incidence of discontinuation Incidence of dropout Dosage intensity
	Longitudinal dosing parameters	Time to first dosage modification Duration of modification/interruption
	Adverse events	Grade 3+ severe AEs Low grade AEs Time to toxicity Duration of toxicity Number of risk reduction interventions Duration of risk reduction interventions
	Patient reported outcomes	Symptomatic AEs Impact of AEs Physical function Reason for dosage modification/interruption
Clinical Efficacy	Preliminary efficacy	Overall response rate Effect on surrogate endpoint biomarker Preliminary registrational endpoint data from earlier trials
	Patient reported outcomes	Disease symptoms Quality of life

Model-informed approaches are an instrumental mechanism for systematically evaluating and integrating the appropriate data to select an optimized dosing regimen(s) and inform trial design for registrational trials. Approaches including quantitative exposure-based, biological, and statistical models derived from broad nonclinical and clinical data can reduce uncertainty and thereby enable more informed decision-making. Specifically, model-informed approaches can be used to predict drug concentrations and responses at doses and regimens not studied, identify differences within a population by testing the effects of covariates, characterize dose- and exposure-response relationships, and facilitate a thorough understanding of therapeutic index. The following examples demonstrate how model-informed approaches can synthesize the totality of all relevant data, including a thorough understanding of a drug, the disease, and the response of the body to gain better understanding of the relationship between drug exposure and preliminary activity or adverse reactions (ARs) to inform dosage selection for the registration trial (Table 2).

Table 2.

List of model-based approaches used to support drug development and their potential uses.

Model-Based Approach	Goals/Use Case
Population pharmacokinetics modeling	Aims to describe the pharmacokinetics and interindividual variability for a given population, as well as the sources of variability
	Can be used to select dosing regimens likely to achieve target exposure
	Can be used to evaluate multiple analytes or parent drug and its metabolites
	Can be used to transition from weight-based dosing regimen to fixed dosing regimen
	Can be used to identify specific populations with clinically meaningful differences in pharmacokinetics and support an alternative recommended dosage
Exposure-response modeling	Aims to determine the clinical significance of observed differences in drug exposure
	Can incorporate nonclinical and emerging clinical data to support selection of dosing regimens to be evaluated in a trial(s)
	Can predict the probability of adverse reactions as a function of drug exposure
	Can be coupled with tumor growth models to understand the response as a function of drug exposure
	Can simulate the potential benefit risk for possible dosing regimens
	Can account for confounding factors, including prior therapy
	Includes population pharmacokinetic-pharmacodynamic modeling and QSP modeling
Population pharmacokinetic-pharmacodynamic modeling	Aims to correlate or link changes in exposure to changes in a clinical endpoint, such as measures of safety or efficacy
	Can integrate nonclinical and emerging clinical and data to support selection of dosing regimen to be evaluated in a trial(s)
	Can predict the probability of adverse reactions as a function of drug exposure and time-course
	Can be coupled with tumor growth models to understand the response as a function of drug exposure
	Can simulate the potential benefit risk for possible dosing regimens
	Can simulate dosing regimens to minimize adverse reactions or recommend a starting dose for a combination therapy
	Considerations include which adverse reactions, activity measures, covariates and exposure metrics to include in the model
	Can account for confounding factors, such as concomitant therapies
Quantitative systems pharmacology	Aims to incorporate biological mechanisms and evaluate complex interactions to understand and predict therapeutic and adverse effects of drug with limited or no clinical data
	Can be used to develop a dosing strategy regimen to reduce the risk of an adverse reaction of special interest
	May consider clinical data from another drug within the same class
Other advanced analytical techniques (e.g., model-based meta-analysis (MBMA), disease progression modeling (DPM), artificial intelligence and machine learning (AI/ML))	Can analyze large datasets of patient and tumor information and allow for more personalized treatment approaches

A common and simple safety-based model-informed approach to identify an optimized dosing regimen is the logistic regression analysis of key landmark safety data from early trials across dosages (9). Important considerations for this method include which AR to model, which exposure metrics best correlate with the chosen AR, and the identification and incorporation of clinical covariates. Given incidence rates of individual ARs are often not sufficient to have robust exposure-response assessments, the regression analysis typically focuses on the combined absence or presence of total severe ARs. Ultimately, the dosing regimens to be further evaluated are generally selected by balancing the modeled probability of an adverse reactions in the patient population with likelihood of therapeutic response.

Alternatively, activity-centric approaches may be more appropriate when there is no clear dose-response relationship. For example, modeling and simulation were used to select the dosing regimen for further clinical development of the HER2-targeting monoclonal antibody pertuzumab, since the MTD was not reached and no clear dose-safety relationships were observed during its dose-finding trial (10–12). Specifically, models based on clinical pharmacokinetics (PK) and an efficacious target exposure derived from nonclinical data facilitated the transition from a body weight-based dosing regimen used in the FIH trial to a fixed dosing regimen used in subsequent trials. Initially, clinical and nonclinical data from the previous generation drug trastuzumab were leveraged to determine the appropriate efficacious exposure target that may translate to clinical efficacy for trastuzumab. Then, population PK modeling and simulations using data from the dose-ranging trial of pertuzumab were used to suggest that an 840 mg loading dose followed by a 420 mg fixed dosage every three weeks would maintain trough exposures above the target exposure level in more than 90% of patients in all cycles, a comparable level to the clinically tested weight-adjusted dose schemes (13,14). These model-based approaches were key to identifying and simplifying an optimized dosing regimen for pertuzumab for evaluation in later trials, including the registration trial that supported its initial approval (15).

More complex mechanistic model-based approaches may be required to select dosing regimens for drug classes with complicated mechanisms of action such as Bispecific T-cell Engagers (BiTEs). For example, the development of mosunetuzumab, a CD3/CD20 BiTE, required the consideration of multiple variables specific to its drug class and target disease (16,17). Initial efforts to identify the starting dose for the FIH trial were centered on reduction of risk for cytokine release syndrome (CRS), a common AR for BiTEs, with subsequent efforts focused on identifying an effective and safe dosing regimen. Using in vitro and in vivo nonclinical data, plus model calibration using clinical data from a similar BiTE, a quantitative systems pharmacology (QSP) model was developed to determine a dosing strategy that could mitigate CRS risk (18). Indeed, emerging clinical data from patients in the FIH trial revealed a trend towards decreased AR with a model-informed step-up dosing regimen. However, no clear exposure-response relationship was observed with efficacy. Given receptor occupancy may better correspond to activity (19,20), a mechanism-based receptor-occupancy driven exposure-response model was subsequently developed using emerging clinical data to complement the QSP model and support the dosing selections for the ensuing trials. An important consideration in the development of this model was prior treatment with rituximab, a monoclonal antibody that binds one of mosunetuzumab’s targets and which many trial participants had previously received, requiring the model to be specifically designed to account for this confounding exposure to a prior therapy that competitively binds the same target to reveal the actual underlying mosunetuzumab exposure-response relationship (21). Together, these two models accelerated the development of mosunetuzumab by identifying a potentially safe and effective dosing regimen to be evaluated in the registration trial.

Given that patients are being exposed to drugs for longer and longer time periods as outcomes have improved, approaches that better recognize and react to time-dependent toxicities have been developed. Traditionally, logistic regression exposure-response models for safety have been used to predict the probability of the emergence of clinically important ARs as a function of drug exposure, and when coupled with tumor growth models for efficacy, simulations have been used to evaluate the benefit and risk for clinically useful dosing regimens. However, such approaches are limited in that they do not fully capture the onset/offset of the AR as well as other potentially confounding factors. More recently, the development of longitudinal pharmacokinetic-pharmacodynamic (PK-PD) models serves to estimate the impact of the time-course of therapy on the occurrence of the AR both within and between treatment cycles. For example, PK-PD time-course models were developed to estimate the timing of the myelosuppressive effects of the chemotherapy agent CPX-351 as compared to current standard of care (22). Critically, these PK-PD time-course models not only captured the onset of response with respect to reduction in platelet and neutrophil counts, but also accounted for the effects of platelet transfusions and leukocyte growth factors on these counts. These risk mitigation interventions can dramatically affect the exposure-response relationship for thrombocytopenia and neutropenia, and failure to appropriately account for their use during treatment may underestimate the true underlying impact of chemotherapy on myelosuppression (23,24). Accounting for use of these concomitant medications and interventions in the PK-PD time-course model led to more realistic simulations of safety outcomes for various dosing regimens, as well as to the identification of potential critical moments in clinical monitoring schemes which may require dosage reduction or a delay in the start of the next treatment cycle. Such a technique is potentially broadly applicable to diverse time-dependent risks and intervention strategies. This longitudinal PK-PD modeling approach can also be applied to biologics as well, with key modifications. For example, in the case of antibody drug conjugates (ADC), population PK models would typically be developed to include characterization of multiple analytes in serum including both total and unconjugated antibody, as well as the unconjugated payload, which is most closely associated with their myelosuppressive effects. Development of longitudinal PK-PD time-course models characterizing the effect of the unconjugated payload on platelets and neutrophils can be used to select ADC dosage amounts that can minimize the risk of a Grade 3/4 AR occurring as well as identify optimal time between treatment cycles (e.g., every three versus every four weeks).

Model-based approaches can also be used to assess the impact of drug combinations on the probability of ARs and assist selection of starting doses for the individual drugs in the evaluated combination therapy. For development of the combination of venetoclax with CPX-351, the same semi-mechanistic PK-PD model for myelosuppression created for CPX-351 was used to recommend starting dosages (25). The myelosuppressive effects of venetoclax and CPX-351 were modeled individually, then simulations were performed based on the assumption that the effects on neutrophils and platelets were either additive or synergistic. These simulations were subsequently completed to determine dosing regimens that targeted ≥ 50% of patients to recover their neutrophils and platelets at the end of each cycle to support and recommend a safe starting dose for a FIH trial of their combination therapy (26).

While many approaches utilize models of safety or efficacy in parallel, the integration of multiple modeled factors into a clinical utility index (CUI) can facilitate their simultaneous consideration (27,28). For example, the dosage of ipatasertib with abiraterone selected for the registration trial in metastatic castration-resistant prostate cancer was supported by CUIs in conjunction with multiple exposure-response models of efficacy and safety based on available clinical data (29). This combined approach supported the in silico exploration of clinically evaluated and unevaluated dosages across various endpoints. A caveat and strength of this method is that it requires determination of important efficacy and safety endpoints for inclusion, their relative weights, and clinically meaningful thresholds through multistakeholder discussions and the testing of multiple scenarios. Clinical data from the subsequent registrational trial showed that the models adequately captured dose intensity and safety findings, although the trial failed to show significant improvement in survival (30). However, because of this thorough analysis, the investigators could be confident in their dose selection and the decision to discontinue development of ipatasertib in prostate cancer (31).

These model-based approaches demonstrate how diverse methods can be implemented to understand the probability of adverse reactions and efficacy, thereby providing key information to help select optimized dosage(s) for late-stage clinical trials and beyond. As demonstrated by the above examples, diverse nonclinical and clinical data, including continuous, categorial, and time to event data from the same drug or drug class, can be incorporated in these models. In general, utilizing the totality of data available will lead to more robust assessment and thereby the selection of optimized dosages for the registration trial. However, as illustrated by each example using a different evidence package to support their dosing decisions, considerations for relevant pharmacology mechanism and modeling approaches in alignment with the context of use are critical to success. Other considerations include understanding how well the selected safety or efficacy endpoints predict clinical outcomes and if the analytical methods allow for consistent and accurate measurement of the endpoints. In short, there is no appropriate one-size-fits-all or algorithmic approach to leveraging models to aid in dosage selection, as specific drug classes in specific disease contexts will require individually reasoned decision-making.

Implementing Seamless and Adaptive Registrational Trial Designs

Aside from modeling approaches, an evolving collection of tools is available to help sponsors select dosages for registrational development, including seamless trial design for the investigation of alternate dosages in either early or late-stage trials (Fig. 1). While development programs such as that of pembrolizumab were able to evaluate multiple dosages in the FIH trial and take two to the end of the initial confirmatory trial, thereby collecting a considerable amount of data to support the selection of the dosages to be evaluated in multiple trials, this successful strategy was associated with a large sample size and substantial economic cost (32). Alternatively, seamless designs can yield similar benefits with reduced operational burden by combining clinical trial phases into one protocol (33). Two examples include operationally seamless designs, where one trial protocol spans multiple development phases, and inferentially seamless designs, where one overarching statistical design allows for consideration of multiple developmental phases in a single analysis. Benefits of these designs include reduced enrollment pauses before and after a go/no go decision, increased patient access to innovative investigational medicines, accelerated development timelines, reduced total sample size, increased design flexibility within pre-designed parameters, and better estimated risks of ARs. However, seamless trials are complex, requiring thoughtful and intentional design. Amidst an evolving dosage optimization paradigm in the US, it is important to consider how such trial designs may be best employed.

Figure 1.

Seamless trial approaches for aiding dosage optimization. Traditionally, oncology drugs progress through multiple trials to support FDA approval, with dose finding trials evaluating the safety of the drug at escalating dosages in small cohorts, proof of concept trials investigating the efficacy and refining to specific cohorts or smaller sets of dosages in expanded cohorts, and confirmatory trials evaluating a specific dosage in a specific population against standard of care therapies in large cohorts. This approach can be associated with gaps in enrollment as new protocols are designed and approved. Seamless trial designs minimize such gaps by organizing multiple phases of development under one protocol, represented by the large colored boxes, providing opportunities to expedite development. Operationally seamless trials combine multiple phases under one trial protocol, but only leverage distinct phases in planning for the next phase. Inferentially seamless trials allow earlier phases to impact both planning and analysis of later phases by continuing evaluation of chosen cohorts/dose-levels of interest. While this can lead to the beneficial accumulation of long-term safety data and a decrease in required sample size, statistical penalties may need to be applied due to re-analysis. (Created with BioRender. AACR Science [2025]. https://BioRender.com/8×6vr9k.)

Depending on the specific needs of the development program, dosage optimization may be incorporated as an operationally seamless component of either a FIH trial or as an initial lead-in for a registrational design. For example, the design of the FIH trial of osimertinib (AURA) in EGFR-mutant non-small cell lung cancer (NSCLC) pre-specified that if preliminary activity was demonstrated during the dose escalation phase the trial would seamlessly expand into relevant biomarker-driven expansion cohorts across multiple doses (34). Similar strategies have been used to enable efficient enrollment of other trial cohorts such as for pembrolizumab in PD-L1 high NSCLC (KEYNOTE-001 Part F) and selpercatinib in RET positive NSCLC (LIBRETTO-001) (35,36). Alternatively, the operationally seamless approach can also be used in later stage trials, such as in the development of a second-generation kinase inhibitor plus pembrolizumab in participants with KRAS G12C NSCLC (SUNRAY-01), where an initial dosage optimization and safety lead-in seamlessly transitioned to a placebo-controlled registration trial, allowing for rapid activation of the registration trial across global trial sites (37,38). Considerations for such an approach include more extensive up-front IRB review, iterative amendments to the protocol based on interim learnings, and adoption of built-in checkpoints for safety and efficacy after the lead-in but before fully opening a registration trial.

Inferentially seamless adaptive confirmatory trial designs are also potentially attractive options to answer questions about dosage optimization. By pre-specifying the inclusion of phase II dose expansion/proof of concept trial patients in the registrational phase III trial statistical analysis and requiring only a minimal operational pause after phase II, inferentially seamless trials can expedite a development program by generating high-quality, long-term, multi-dose randomized controlled data while requiring lesser sample size than a fully seamed program. This inclusion of multiple doses across trial stages provides drug developers and regulators with more data to consider and may facilitate better dosing but requires a nuanced approach to statistical analysis. As a result of the multiple analysis of data from the phase II trial, both at the end of phase II to help determine phase III registrational dosage and together with phase III, an alpha penalty is typically required to control type I error rates in final efficacy analyses. The size of penalty depends on dose selection criteria, the amount of data used for dose selection and various other statistical factors. Different statistical designs may be applied to account for this issue, and in specific situations the penalty can be negligible (39,40). Following FDA’s guidance for adaptive designs, simulation studies may be conducted to assist with the assessment (41). Inferentially seamless approaches in the early and later setting may be most appropriately applied in scenarios where the disease context has high unmet need, given their potential benefit-risk tradeoff. Inferential seamless trials carry many of the challenges of operationally seamless trials but in addition may rely on rapid decision-making based on emerging data, leading to additional patient risk due to individual investigators potentially being less familiar with the safety signals associated with the investigational drug.

Regardless of planned trial approach, sponsors are encouraged to meet with FDA to discuss how the dosages will be selected prior to initiating the registration trial, including but not limited to what nonclinical and clinical data will be considered and who is the target patient population. Appropriate venues for such conversations may include regulatory milestone meetings, clinical pharmacology-led type C meetings, and other newer meeting types such as type D (42).

Discussion & Conclusion

Final dosage selection for registrational trials must no longer remain confined to MTD, both because the pharmacology of modern oncology drugs no longer conforms to this approach and because patient advocates are calling for change in dosing. The development programs summarized here demonstrate that model-based approaches and innovative clinical trial designs can play an integral role in the identification of an optimized dosing regimen but must be designed on a case-by-case basis for the specific development program.

Model-based approaches can effectively integrate the totality of the data available for a new drug, including key nonclinical data and emerging clinical data, as well as incorporate knowledge from similar oncology drugs. The models can thereby provide mechanistic insight, project optimized dosing regimens, and inform go/no-go decisions. The benefits of seamless trial designs can also be brought to bear under the dosage optimization paradigm when thoughtfully planned, including greater trial uptime and the derivation of additional potentially dosage decision-altering data. In turn, it is necessary for sponsors to be prepared to engage with FDA as early as possible regarding how decisions about dosage will be made, including the principles for that decision-making, safety signals, and other information that will be considered.

The FDA–AACR workshop that inspired this manuscript, the final in a three manuscript series on dosage optimization in oncology drug development, discussed in greater depth the nuances in using models and seamless trials in late-stage drug development as well as efforts that can be undertaken at earlier stages to facilitate dosage optimization for oncology products and alignment with FDA’s Project Optimus and other dosage optimization initiatives; recordings are available (6). Of particular interest for future development in dosage optimization is the inclusion of patient reported outcomes in decision-making, as highlighted in subsequent workshops (43), and the handling of combination therapies. Whether thinking ahead to plan a molecule’s path to FDA approval prior to FIH dose administration or considering what parameters to use to in modeling, it is never too early to consider how best to incorporate dosage optimization strategies in order to better serve both patients and the drug development program itself.

Abbreviations:

ADC

Antibody Drug Conjugates

AR

Adverse Reactions

BiTE

Bispecific T-cell Engager

CUI

Clinical Utility Index

FIH

First-in-Human

MTD

Maximum Tolerated Dose

NSCLC

Non-Small Cell Lung Cancer

PK

Pharmacokinetic

PK-PD

Pharmacokinetic-Pharmacodynamic

QSP

Quantitative Systems Pharmacology