3. DISCUSSION
The inability to select an appropriate dose in pivotal trials has been shown to contribute to the declining success rates of drug development programs.41, 42 A study examining FDA approval packages between 2015 and 2017 found that in a third of development programs no E-R analysis was reported.43 The expanded use of E-R analysis in more drug development programs may serve as a solution for declining success rates. E-R analysis is particularly useful in early clinical trials where multiple doses are administered to inform dose selection and optimization. It is often repeated in phase 3 trials given the meaningful sample size for efficacy and safety interpretation. Rather than assuming one dose fits all patients this approach identifies whether specific patient subgroups would benefit from alternative doses. A successful example of E-R application was shown for the exposure-survival analysis of ipilimumab. Ipilimumab was originally approved in several countries at a dose of 3 mg/kg for the treatment of advanced melanoma. A phase 2 dose-ranging study, however, suggested improvement in OS with the 10 mg/kg dose.44While this study was not statistically powered to detect differences in survival an E-R analysis pooling data from four phase 2 trials demonstrated that OS improved with increasing exposure. In the CPH model results patients in the 5th and 95thpercentiles of steady state trough concentration (Cminss) had an OS HR of 1.52 and 0.552, respectively, relative to patients with median Cminss.45 This suggested that OS improved with increased ipilimumab doses. In the post-marketing trial conducted with the 3 mg/kg and 10 mg/kg doses this relationship was confirmed. Median OS was 15.7 months for the 10 mg/kg dose group, and 11.5 months for the 3 mg/kg dose group (HR 0.84, p=0.04).46 The results of this phase 3 trial are included in the ipilimumab label, and demonstrate that E-R analyses could identify potential survival benefits gained from increased doses and exposures. While the 10 mg/kg dose provided a survival benefit it was also associated with increased treatment-related adverse events. The 3 mg/kg dose was selected as the labeled dose after accounting for efficacy benefit and safety risk.
While the utility of E-R analyses applies across a variety of therapeutic areas additional considerations are needed for oncology due to the impact of prognostic factors on both exposure and outcome. Performance status, clinical symptoms (dyspnea, appetite loss, cognitive function), primary tumor site, and c-reactive protein (CRP) concentration are examples of prognostic factors used to predict outcome in clinical settings.47, 48 In oncology, E-R is more than just considering the unidirectional relationship where the dose affects exposure which subsequently affects response. OS is often the primary response endpoint for oncology trials, and its relationship with exposure is confounded by prognostic factors. Recognizing and accounting for the impact of time-varying clinical response and prognostic factors on exposure are critical for accurate E-R interpretations.49 This relationship is illustrated by the findings for nivolumab, avelumab, durvalumab, and pembrolizumab where patients with improved post-treatment disease status showed greater time-dependent decreases in drug CL.7-10 The mechanism is not fully understood, but there is an interaction between clinical response, prognostic factors, and exposure. When patients respond to treatment their prognostic factors improve, which in turn decrease the drug CL and increase drug exposure (Figure 2) . In an E-R analysis this could lead to incorrect conclusions that higher drug exposure caused clinical response when in fact the E-R relationship is confounded by the effect of changing prognostic factors on exposure. This may be caused by the unique nature of disease progression in oncology. As a patient’s disease status declines clinical changes such as cachexia and inflammation can increase the catabolism and clearance of both endogenous and therapeutic proteins.50, 51This is supported by the significance of tumor burden and albumin as a covariate on CL in the population PK analyses for nivolumab, avelumab, durvalumab, and pembrolizumab.8-10, 12 Clearance increased with higher tumor burden and lower albumin concentrations. In addition, time-dependent PK was observed for nivolumab in advanced malignancies, but not in patients with resected melanoma.52 The latter had tumors surgically resected prior to adjuvant treatment with nivolumab and were overall healthier than patients with advanced malignancies. This further supports the impact of disease status and prognostic factors on exposure. Looking prospectively, these collective observations also suggest that the presence of time-dependent PK, and the significance of albumin as a covariate on CL would indicate the risk of a confounded E-R analysis.
A confounded E-R analysis may result in false positive E-R relationships which may lead to the wrong conclusion that the dose for patients with lower exposure is suboptimal. As seen in the ToGA/HELOISE example it may lead to the initiation of a new trial in an attempt to rescue patients who failed treatment. Considering these risks three mitigation strategies have been in this review: CPH modeling and case matching analysis, TGI-OS modeling, and multiple dose study design. Studying multiple dose levels in randomized, balanced groups appears to be an effective approach that can distinguish with certainty between a true positive E-R relationship versus a false positive relationship with hidden confounders. This strategy, however, is impractical in most oncology indications, and may offer limited value for monoclonal antibodies with a wide therapeutic window. TGI-OS modeling explicitly separates the drug-specific and disease-specific effect on OS when evaluating the E-R relationship. It also incorporates an estimate of tumor dynamics which serves as an informative biomarker of disease status. CPH modeling and case-matching analysis lack this separation between drug and disease-specific effects. This makes it more challenging to consistently distinguish between exposure- and prognostic factor-driven changes in OS. Case-matching may be preferred over CPH modeling due to the assumptions involved in CPH modeling regarding the relationship between predictors and outcome. While a suggestion of the relative utility of each approach is made here the unique limitations of each methodology should be considered.
While the effect of changing prognostic factors on exposure and OS can confound exposure-efficacy relationships it does not appear to significantly impact exposure-safety relationships. Among molecules discussed here it appears that exposure-safety analyses for pembrolizumab, nivolumab, durvalumab, and T-DM1 have not faced the same confounding issues as exposure-efficacy analyses.15, 20, 53-58 If patients with worsening disease status and prognostic factors are more likely to experience adverse events, and have a decreased drug exposure it could be thought that differences in prognostic factors can confound the exposure-safety relationship. The confounding, however, would contribute to an inverse E-R relationship rather than a positive relationship. In addition, because safety endpoints in E-R analyses are usually drug-related adverse events rather than disease-related adverse events exposure-safety relationships may be less likely to be influenced by differences in prognostic factors.
Current oncology drug development is rapid and aggressive. Many recent development programs bypass a dose-ranging phase 2 trial and go directly from phase 1 to phase 3 trials with a single dose level. In some programs, a phase 2 trial is done, but only with a single dose level or a limited efficacy endpoint. This severely limits the range of exposure data available for an exposure-survival analysis and increases the risk of a confounded E-R analysis. The FDA’s E-R Guidance has previously described the risk of characterizing E-R relationships based on data from single dose levels.59 Sponsors should consider conducting expanded dose-ranging trials early in development programs to better inform dose selection and potentially avoid the need to study multiple dose levels in late phase trials. Despite the current limitations of E-R analyses they are required to be included as part of a filing package. The HELOISE trial is an example of the potential risks of confounded E-R analyses. The E-R analysis performed using data from the ToGA trial supported the conduct of the HELOISE trial. No dose-response relationship was observed, however, and patients did not benefit from higher doses of trastuzumab. If the E-R analysis is limited by the range of available exposure data, and could be confounded any observed E-R relationship should be interpreted with great caution.
Drug development must not only focus on developing novel treatment modalities, but also on selecting the optimal dose for patients. E-R analysis is a useful tool for dose optimization in a variety of therapeutic areas, and also has many applications to support modern drug development. Despite its wide utility E-R analysis in oncology faces unique challenges when applied to monoclonal antibodies tested at a single dose level. E-R analyses in oncology are susceptible to confounding from unique, disease-related factors. Mitigation strategies presented in the current paper can be employed to account for confounding factors and elucidate the true E-R relationship. In a broader scope, the design of oncology drug development programs may be structured to more effectively inform dose-response and E-R relationships for dose optimization. Once an E-R analysis is performed its application in decision-making must be carefully considered based on the methodology and the data used in the analysis. The improvement and effective use of E-R analysis is an effort that must be addressed on multiple fronts of oncology drug development with the common goal of maximizing benefit to the patient and minimizing toxicities.