Surrogate and Intermediate Endpoints in Randomized Trials: What's the Goal?

In this issue of Clinical Cancer Research, Mushti and colleagues evaluated objective response rate (ORR), progression-free survival (PFS), and a modified PFS as possible surrogate endpoints for overall survival (OS) using data from 13 randomized trials of anti–PD-1/PD-L1 therapies (1). When discussing surrogate endpoints, which are intermediate between the treatment and definitive endpoint, it is important to consider the goals of using an intermediate endpoint, which could be for drug development or patient management (Table 1). Evaluation of an intermediate endpoint for drug development requires results from multiple randomized trials, for example, OS and PFS HRs (Table 1; ref. 1). The most ambitious goal for an intermediate endpoint is to be a trial-level surrogate (goal 1), for example, establishing that PFS results will reliably predict OS results in future trials, making effective experimental therapies available sooner.

A key motivation frequently given for considering PFS as a trial-level surrogate for OS is that effective subsequent therapies (after progression) or long-term survival will cause OS differences between the treatment arms to be small or nonexistent. However, this is illogical as an argument for the need for an OS surrogate: If one expects OS differences to be clinically insignificant and believes OS differences are the true measure of clinical benefit to patients, then there is no role for an OS surrogate (2). A specific example is when an immunotherapy is being assessed in a first-line setting after it has been shown to be effective (and approved) in the second-line setting; the relevant clinical question is whether it is better to give the therapy early versus late, which would require showing a direct OS benefit to giving the therapy early.

There are several limitations of the trial-level analyses by Mushti and colleagues (1), some of which are noted by the authors. The pooling of diseases in a surrogacy analysis is problematic, especially with therapies that have different disease-specific efficacies. It should be noted that only trial-level results are required for assessing trial-level surrogacy (3), so the analyses presented for PFS and ORR could have been expanded to include all published trial results and disease-specific analyses (although this would not have allowed the modified PFS analyses). A second limitation is the inclusion of trials in the first and second line, where PFS/OS relationships would be expected to be quite different. A third limitation is the inclusion of only positive trials. A trial-level surrogate needs to be able to predict both positive and negative definitive endpoint results, and therefore it cannot be properly evaluated on only positive trials; ideally, the analysis should include all available trials to avoid potential bias (4). Finally, the use of R² to measure surrogacy is not ideal (especially with a restricted set of trials), with better approaches directly measuring the predictive ability of the surrogate endpoint (3).

The main challenge for demonstrating trial-level surrogacy is generalizing the surrogate/definitive outcome relationship to future trials. One needs intermediate and definitive endpoint results from a set of trials with agents with a similar mechanism of action in the same clinical setting to the future trials. However, if one had such a set of trials, then the important clinical questions of interest involving these agents may have already been answered. Using the surrogate endpoint as a surrogate for a different type of agent or in a different setting is a large extrapolation. In addition, the results of a surrogacy analysis would no longer apply if the clinical landscape (including available life-extending salvage therapies) has changed from when the surrogacy analysis was performed. For example, the trial-level surrogacy of PFS demonstrated for OS using trials of 5FU-based regimens for advanced colon cancer may not apply to trials of targeted drugs tested against updated standards of care with improved salvage therapies (5).

Given the above challenges, the interest in showing PFS as a trial-level surrogate should perhaps be directed to more modest goals in using an intermediate endpoint for drug development (goals 2–5; Table 1). For example, to use PFS as the endpoint in a screening randomized, phase II trial (or as the phase II endpoint in a phase II/III trial; goal 2) does not require PFS to be a surrogate for OS but just the weaker requirement that a clinically interesting between-arm difference in OS should be reflected in an observable difference in PFS. A related screening application (goal 3) of PFS is to use interesting PFS activity of an experimental agent in an advanced-disease setting for moving the agent forward to a first-line or adjuvant setting with a definitive endpoint evaluation (without making a definitive statement about the agent's clinical utility in the advanced-disease setting). Goal 4 covers the special circumstance where the experimental therapy and the standard-agent control arm have similar mechanisms of action, for example, the experimental therapy is the standard agent with a modified schedule. In this situation, showing an improved PFS (and no demonstrated decrement in OS) may be enough to have this experimental therapy available to patients as another option. Unfortunately, the trial of nivolumab versus docetaxel in advanced nonsquamous non–small cell lung cancer (6) shows that PFS may not be useful for even these limited goals for immunotherapies (possibly due to a delayed treatment effect; ref. 7). Finally, improvement in non-OS endpoints could represent direct patient benefit in some settings (goal 5) regardless of whether OS is improved, for example, a large improvement in disease-free survival for a relatively nontoxic therapy in an adjuvant breast cancer trial (2, 8); this is generally not true for PFS improvement without OS improvement (9).

Mushti and colleagues appropriately distinguish trial-level and individual-level surrogacy (1). An individual-level surrogate is an intermediate endpoint that can predict who will do well or poorly on a treatment; it need not be a trial-level surrogate (3, 10). An individual-level surrogate has the potential for improving patient care (goal 6) by suggesting additional (less) treatment when the patient is predicted to do poorly (well). Immunotherapies are known to work well for a subset of patients, leading to much interest in finding predictive biomarkers (7). Barring that, an individual-level surrogate predicting OS available shortly after the commencement of immunotherapy would be very helpful for patient management.

The evaluation of the individual-level surrogacy can be done on a cohort of similarly treated patients (the modeling of trial-level data from multiple trials is not required). For example, Mushti and colleagues show that responders live longer than nonresponders in the trials they considered (Fig. 1A in ref. 1), demonstrating that response is an individual-level surrogate for OS. Whether response would have any utility in patient management in the immunotherapy setting would depend on whether the difference in OS curves for responders versus nonresponders would be enough to change the way the patient was treated. For evaluating improved patient management, the analysis of the individual-level surrogacy of PFS for OS with the rank correlation used by Mushti and colleagues could be supplemented by a landmark analysis of OS curves for patients who did or did not progress by a fixed time point, say 6 months.

In many cancer settings, it may not be realistic to expect PFS to predict trial-level effects on OS due to the changing therapeutic landscape. However, there are several important drug development and patient management applications for intermediate endpoints.

No potential conflicts of interest were disclosed.

Conception and design: E.L. Korn, B. Freidlin

Development of methodology: E.L. Korn, B. Freidlin

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): E.L. Korn, B. Freidlin

Writing, review, and/or revision of the manuscript: E.L. Korn, B. Freidlin