Rethinking surrogate endpoints in metastatic colorectal cancer: counting chickens only when they hatch

Metastatic colon cancer is a challenging disease to study due to its highly heterogeneous clinical presentation and outcome. This heterogeneity is due to widely varying tumor biology, site, and burden of disease, numerous effective locoregional and systemic therapies, and widely disparate prognoses. The gold standard for assessing the efficacy of a novel treatment in cancer is overall survival (OS) with intention-to-treat analysis of patients enrolled into comparative randomized controlled trials (RCTs). This endpoint, however, requires lengthy follow-up, which can delay approvals and progress. As a result, surrogate endpoints, requiring shorter follow-up, are often used. Common surrogate endpoints include tissue and radiographic endpoints such as radiographic and pathologic response, and disease-specific endpoints such as progression-free survival (PFS) and recurrence-free survival (RFS). However, these endpoints can be misleading, in part due to the narrow scope of their assessment and their potential for poor correlation with OS. Optimal surrogates for OS fulfill Prentice’s criteria of capturing the relationship between the treatment and the true endpoint (1). As we will explore, PFS and RFS fall short of correlating with OS in the study of multimodal treatment for unresectable colorectal liver metastases (CRLM).

The post-hoc analysis of the CAIRO5 trial, recently reported in JAMA Oncology, offers a modern examination of long-term outcomes among patients treated with first-line systemic therapy for unresectable CRLM, subclassified by sidedness and driver oncogenes (2). The findings from this RCT reaffirm the central importance of hepatic resection after downstaging in this disease—but also highlight a more nuanced observation: in metastatic colorectal cancer, particularly liver-limited disease, PFS is a poor surrogate for OS. This discrepancy arises from the complex nature of CRLM disease progression, the high response rates to first-line therapy, the availability of effective locoregional strategies after downstaging, and the frequent decoupling of PFS from long-term outcomes. In light of these dynamics, this commentary challenges the relevance of PFS and RFS as primary endpoints in CRLM trials and advocates for trial designs that emphasize durable oncologic control and survival.

The CAIRO5 trial was a prospective, randomized study evaluating optimal first-line systemic therapy for patients with initially unresectable CRLM conducted through the Dutch Colorectal Cancer Group. The trial stratified patients by molecular subtype—RAS/BRAF^V600E mutation status—and by tumor sidedness, randomizing subjects to receive either doublet chemotherapy (FOLFOX or FOLFIRI) plus bevacizumab, or FOLFOXIRI plus bevacizumab for patients with mutations in RAS or BRAF^V600E or right-sided primary tumors. For subjects with left-sided primary tumors without RAS or BRAF^V600E mutations, patients were randomized to doublet chemotherapy regimens and compared based on the targeted agent (bevacizumab or panitumumab). The primary outcome was PFS, assessed in a modified intention-to-treat analysis.

The initially reported endpoint analyses showed an objective response rate of 54% with FOLFOXIRI plus bevacizumab compared to 33% with doublet chemotherapy plus bevacizumab (P<0.001). Correspondingly, the rate of local therapy—defined as resection and/or ablation—was numerically higher in the FOLFOXIRI arm at 57% vs. 46% with doublet chemotherapy, although this difference did not reach statistical significance (P=0.08) (3). When considering complete local therapy, defined in the protocol as R0–R1 resection or ablation of all visible liver metastases, 51% of patients in the FOLFOXIRI group and 37% in the doublet group achieved this outcome, with surgical resection comprising the vast majority of cases (3). These gains in tumor response, conversion to local therapy, and the 1.6-month improvement in the prespecified primary endpoint of PFS (P=0.03) supported the conclusion that triplet chemotherapy should be preferred for patients with right-sided or RAS/BRAF^V600E-mutated unresectable CRLM (3). For patients with left-sided primary tumors and without mutations in RAS/BRAF^V600E, the use of panitumumab was not superior to bevacizumab, but was associated with increased toxicities, including a significantly higher incidence of grade 3–4 skin toxicity (25% vs. 1%, P<0.001) and diarrhea (16% vs. 4%, P=0.007), and a lower incidence of hypertension (7% vs. 18%, P=0.02) (3).

The recently published post-hoc analysis offers a mature view of treatment trajectories and long-term outcomes (2). Overall survival was included as a secondary outcome in the CAIRO5 trial. Among the patients who were able to undergo complete local therapy for right-sided or RAS- or BRAF^V600E-mutated tumors, nearly half of the patients recurred within 6 months (49% in patients treated with FOLFOX/FOLFIRI + bevacizumab; compared with 39% in patients treated with FOLFOXIRI + bevacizumab) (2). However, in both groups, approximately two-thirds (66–68%) of patients with recurrence after conversion to local therapy were able to undergo subsequent repeat salvage local therapies (2). Those who were salvageable had a median OS of 58.5 months, nearly identical to those who never recurred (2). This finding challenges the notion that local recurrence necessarily equates to poor prognosis and underscores the role of salvage local therapies in impacting survival. As a result, according to the initial chemotherapy regimen, this trial found no significant OS difference between chemotherapy strategies in the compared cohorts (2). Thus, despite the initially reported increased response rates and PFS advantage observed with triplet chemotherapy plus bevacizumab for right-sided and RAS/BRAF^V600E-mutated tumors (3), the authors ultimately supported the use of doublet therapy with bevacizumab. This recommendation was based on the statistically indistinguishable OS between groups—23.6 months [95% confidence interval (CI), 20.1–27.5] in group 1 (doublet + bevacizumab) vs. 24.1 months (95% CI, 21.0–30.9) in group 2 (triplet + bevacizumab) [hazard ratio (HR) 0.90; 95% CI, 0.70–1.17; P=0.44] (2).

Given the high rate of conversion to resection in the CAIRO5 trial, a related and compelling example of the limitations of surrogate endpoints in CRLM can be gleaned from the JCOG0603 RCT assessing adjuvant chemotherapy with mFOLFOX6 vs. observation following hepatectomy for CRLM (4). Notably, the trial was terminated early due to a statistically significant improvement in RFS in the chemotherapy arm (5-year RFS: 49.8% vs. 38.7%, HR 0.67; P=0.006). However, the final analysis revealed a paradoxical finding: OS was statistically indistinguishable between the two arms, and numerically worse in the chemotherapy arm (5-year OS: 71.2%) compared to the surgery-alone arm (83.1%), with a HR of 1.25 (P=0.42) (4). This disconnect illustrates how a surrogate endpoint like RFS may be driven by relatively small differences in recurrence timing, but not necessarily by recurrence patterns and effective treatments for recurrent disease, which may ultimately dictate OS. Unlike some cancers, in CRLM, salvage local therapies and effective systemic options following recurrence often mitigate early failures. Consequently, decisions based solely on PFS or RFS may lead to overestimation of benefit and underappreciation of toxicity or long-term trade-offs. JCOG0603’s contradictory findings parallel those of CAIRO5 and underscore the need to prioritize OS until a surrogate endpoint that accurately mirrors OS is found.

The discordance between RFS/PFS and OS in resectable CRLM has been well documented, most comprehensively in an analysis of institutional data from Memorial Sloan Kettering Cancer Center and a meta-analysis of randomized trials (5). An analysis of nearly 3,000 patients treated with hepatic resection for CRLM revealed little correlation between RFS and OS, with correlation estimates ranging from 0.30 [standard deviation (SD) 0.17] to 0.56 (SD 0.13) (5). Additionally, a meta-analysis of phase III RCTs demonstrated a pooled HR for RFS benefit with adjuvant chemotherapy was 0.69, yet the corresponding HR for OS was only 0.95 and not statistically significant. As a result, there was minimal correlation between RFS and OS with a Spearman’s correlation of r=0.43 [standard error (SE) =0.39; P=0.22] (5). This underscores that early improvements in RFS do not necessarily translate into a survival advantage in settings where salvage strategies are effective. Given these findings in resectable CRLM, it is not surprising that a similar phenomenon is now emerging in trials involving initially unresectable CRLMs, especially in patients who experience downstaging to resection as was observed in the CAIRO5 trial (2,3).

Following hepatectomy for CRLM, approximately 60% to 70% of patients experience disease recurrence, with the liver and lungs being the most common sites (6). Notably, a substantial proportion of these recurrences—estimated between 25% and 35%—are amenable to salvage interventions such as repeat hepatectomy, pulmonary resection, or ablation of isolated sites of disease (6). The efficacy of these salvage therapies is well-documented; for instance, patients undergoing repeat hepatectomy for recurrent liver metastases have reported 5-year OS rates ranging from 20% to 65%, depending on various prognostic factors (6,7). Similarly, data from the CAIRO5 trial underscore that many recurrences are amenable to salvage local therapy, which is associated with OS comparable to patients without recurrence (2,3). These findings emphasize the heterogeneity of disease biology and therapeutic trajectories in CRLM. They also suggest that recurrence alone should not be viewed as a definitive trial endpoint—particularly when effective salvage remains feasible and is associated with long-term survival, comparable to those without recurrence.

While molecular profiling and sidedness can help refine treatment selection in metastatic colorectal cancer, the current staging and response frameworks remain overly simplistic when evaluating long-term outcomes. For example, patients with a solitary lung recurrence after resection are biologically and therapeutically distinct from those with liver recurrences or with multifocal peritoneal or systemic relapse (8). Although a full classification system accounting for recurrence kinetics, anatomic location, multifocality, and salvageability would provide clinical nuance, such granularity may not be feasible within the constraints of prospective trials. Instead, these considerations underscore the need to prioritize OS as the primary endpoint for future studies, with secondary and complementary endpoints that capture the use of additional therapies, treatment-related toxicity, and patient-reported quality of life. Given the potential for long-term survival through iterative multimodal treatments, trial designs and endpoints must reflect the complex trajectory of metastatic colorectal cancer rather than rely solely on early radiographic surrogates, especially when considering early stopping rules within protocols. Furthermore, as treatment paradigms grow complex with multimodality, the study of treatment “packages” capturing the synergistic or futile interaction between system, local, targeted, and immunotherapies becomes paramount in elucidating treatment sequencing for patients with CRLM.

Studying patients with CRLM presents a unique clinical and methodological challenge. There remains no universal consensus among surgeons on what constitutes resectability, leading to substantial variation in practice and trial eligibility criteria (9). Meanwhile, emerging loco-regional strategies—such as liver transplantation in highly selected patients, hepatic artery infusion pump chemotherapy, and ablative modalities—are redefining the treatment landscape (10-13). Added to this is the intrinsic biological heterogeneity of CRLM, driven by both tumor sidedness and key oncogenic mutations. Specific mutational constellations, such as TP53/KRAS co-mutations or BRAF^V600E, are associated with a propensity for disease recurrence and worse OS (14,15). Disease burden also plays a critical role, as illustrated by concepts that integrate tumor biology and volume to estimate post-hepatectomy prognosis (16,17). In this complex setting, understanding the impact of any new therapy—systemic, surgical, or otherwise—requires a holistic assessment of the patient’s clinical course, beyond a readout of radiographic response such as PFS/RFS. Instead, trial outcomes should prioritize OS and consider including detailed analyses of recurrence patterns, salvageability, and therapeutic sequencing and secondary endpoints. The disease trajectory in CRLM is not often defined by the first treatment, but rather by the impact of multiple sequential treatments—be it a salvage resection, a second-line systemic regimen, or ablative local therapy, let alone the emerging role of liver transplantation. As such, we must be disciplined in interpreting early trial readouts and cautious in overvaluing imperfect surrogates such as PFS, which may capture treatment timing but not necessarily therapeutic value.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.