The introduction of immune checkpoint inhibitors (ICIs) has produced a paradigm shift in the treatment of patients with deficient mismatch repair (dMMR)/microsatellite-instability (MSI)-high metastatic colorectal cancer (mCRC). Compared with the incremental gains achieved by several treatment options in the last 10 years, the efficacy of anti-programmed death (ligand)1 (PD-(L)1) ± anticytotoxic T-lymphocyte antigen 4 (CTLA-4) agents in this small subgroup of patients (~5%) is unprecedented. In fact, a relevant proportion of patients achieve long-term disease control, which may be the result of pharmacological tumor eradication.1–7

As reported in other immune-sensitive cancer types,8–10 a plateau of survival curves was observed in pivotal trials assessing the activity and efficacy of anti-PD-1 ± anti-CTLA-4 agents in patients with MSI-high mCRC, with reported long-term progression-free survival (PFS) rates of ~30% to 50% and~75% in patients receiving anti-PD-1 monotherapy and anti-CTLA-4 + anti-PD-1 combinations, respectively.3–7 In this scenario, a proportion of patients with MSI-high mCRC achieving durable responses and long-term survival outcomes may be considered potentially ‘cured’ by ICI treatment, as opposed to patients experiencing primary ICI refractoriness or short-term clinical benefit.

The development of validated, reproducible, and easy-to-use clinical tools able to discriminate between patients with MSI-high mCRC achieving long-term disease control and ICI-refractory patients represents an unmet clinical need. The early identification of potentially resistant patients with adverse clinical and biological features might be useful to timely implement more effective ICI combinations (eg, anti-CTLA-4+antiPD-1).

In the current study, we developed and externally validated a nomogram to estimate the PFS and event-free probability (EFP) in patients with MSI-high mCRC receiving ICIs and potentially assist clinicians in their clinical practice.

The developing set included 163 patients. The validating series originally included 161 consecutive patients; 15 patients had missing information on ECOG PS (8 patients) and/or NLR (8 patients); thus, the cure model could be validated on a subset of 146 patients. Main patients and disease characteristics in the developing and validating sets are listed in table 1. A higher proportion of patients in the developing set had an ECOG PS of 0 (60.7% vs 39.7%, p<0.001), a mucinous histotype (54.0% vs 34.2%, p<0.001), lymph nodal metastases (65.6% vs 54.1%, p=0.048), ≥2 metastatic sites (59.5% vs 44.5%, p=0.005); received ≥1 prior lines of systemic treatment (77.3% vs 52.7%, p<0.001); and underwent primary tumor resection (97.5% vs 63.7%, p<0.001).

Median follow-up, overall suvival (OS) and PFS estimates in the two sets, overall and according to regimen, are reported in online supplemental table 1; the corresponding OS and PFS Kaplan-Meier curves are shown in online supplemental figure 1. We observed that in both the development and validation series the greater proportion of PFS events occurred within 12 months (55 out of 73 overall and 48 out of 57 overall, respectively). Thus, to forecast patients’ prognosis as earlier as possible, we used 12 months as cut-point for PFS prediction. Among the 23 variables included in the RSF model, five were selected: ICI treatment type (anti-PD-1 mono vs anti-CTLA4 combo), ECOG PS at baseline (0 vs >0), NLR (modeled as binary variable ≤3 vs >3, using a literature cut-off14), PLT count (continuous linear term) and prior lines of treatment for metastatic disease.

Figure 1 represents the nomogram to predict the 12-month PFS probability according to the cure model including the selected variables. The calibration of the 12-month PFS predictions was good in both the developing and validating sets (figure 2). For comparison, we show in online supplemental figure 2 the calibration plot obtained fitting a Cox model on the developing set data including the same five covariates as the cure model; Cox model predictions were not well calibrated in the group of patients with predicted 12-month PFS equal to 0.76. The Harrell C statistics were 0.707 (95% CI 0.653 to 0.762) in the developing set and 0.641 (95% CI 0.567 to 0.714) in the validating set.

Table 2 shows the two sets of scores allowing determination of the 12-month PFS and the EFP based on the nomogram covariates. The median value of the EFP (equal to 46%) allowed segregation of two well-separated prognostic groups in both developing and validating sets. In detail, in the developing set, the PFS of 12 and 36 months was 43% and 31% vs 84% and 67% in patients with EFP of <46% or≥46%, respectively (HR=3.73, 95% CI 2.25 to 6.18; p<0.0001; figure 3A). In the validating set, the PFS of 12 and 36 months was 55% and 50% vs 74% and 64% in patients with EFP of <46% or ≥46%, respectively (HR=1.86, 95% CI 1.07 to 3.23; p=0.027; figure 3B). Regarding OS, in the developing set, the OS of 12 and 36 months was 57% and 42% vs 90% and 75% in patients with EFP of <46% or≥ 46%, respectively (HR=3.86, 95% CI 2.11 to 7.06; p<0.0001; figure 3C). In the validating set, the OS of 12 and 36 months was 74% and 54% vs 94% and 78% in patients with EFP of <46% or ≥46%, respectively (HR=2.99, 95% CI 1.37 to 6.55; p=0.006; figure 3D).

We also depicted the outcomes according to the quartiles of the EFP determined on the developing set, to investigate whether more homogeneous subgroups were still associated with differential outcomes in the context of the high efficacy of ICI in our patients’ population. Online supplemental figure 3A, C shows the PFS curves stratified according to the quartiles of the EFP in the developing set (p<0.0001) and in the validating set (p=0.024), respectively. Online supplemental figure 3B, D show the OS curves stratified according to the quartiles of the EFP in the developing set (p<0.0001) and in the validating set (p=0.004). Online supplemental table 2 shows the estimates for PFS and OS of 12 and 36 months in the subgroups identified by the median value or the quartiles of the EFP.

Online supplemental table 3 illustrates a simulation of clinical scenarios to highlight the weight of ICI treatment type on 12-month PFS and EFP estimates according to different values of cure model covariates.

In several advanced cancers, ICIs are associated with survival outcomes that are deeply different from those observed after chemotherapy or targeted therapy, with a plateau of the survival curves after a specific follow-up timepoint.8–10 A fraction of the patients’ population could be then considered potentially cured and, from a statistical point of view, may be subject to the same mortality of the cancer-free general population.15 Therefore, mixture cure models have been proposed as an alternative to traditional proportional hazard models, as the Cox model, to explore the association between survival endpoints and putative prognostic factors in a multivariable context. The advantage of cure models relies on the assumption of two different populations of patients: those who are cured and those who are ‘not cured’, thus allowing the identification of patients at high chance to be alive and event-free independently of time. Therefore, the concept of pharmacological eradication of MSI-high cancers may be better and properly investigated thanks to cure models, with relevance of this approach for trials’ design. As confirmation of the eradicating potential of immunotherapy in this molecular subgroup of patients, post-treatment surgery of metastases or early-stage primary tumors is associated with an extremely high rate of pathological complete responses, independent of the presence of radiological residuals.16 17 Also, patients with Response Evaluation Criteria in Solid Tumors(RECIST) very good partial responses≥50% or complete responses are usually cured, even if the best tumor response may be achieved after several months of treatment.18

Here we have used a multivariable mixture cure model to develop a tool for predicting the 12-month PFS probability of patients with MSI-high mCRC receiving ICIs. The timepoint for PFS prediction was chosen based on the observation that the greater proportion of PFS events occurred within 1 year, in order to allow predictions as early as possible, therefore, with specific focus on patients with low chance of being cured; this consideration is particularly true for the two groups with the worst prognosis, as appears from the PFS curves stratified according to the EFP quartiles (probability <24%, online supplemental figure 3). The EFP is focused instead on patients likely to be event-free at a certain point of their follow-up; high levels of such probability can identify those patients who will experience delayed or no progression (group with the best prognosis, probability ≥62%; online supplemental figure 3).

As regards the model covariates, we investigated the ICI treatment type and a multiplicity of variables potentially related to life expectancy and commonly used in clinical practice, and then used a random forest model leading to select five easy-to-collect variables: anti-PD-(L)1 monotherapy versus anti-CTLA-4 combination, number of prior treatment lines for metastatic disease, ECOG PS, NLR, and PLT.

ECOG PS and blood-based biomarkers are known independent prognostic variables identified by standard Cox models in available datasets of patients with MSI-high mCRC,19 20 and these parameters were validated also by our methodological approach. However, despite the presence of poor prognostic characteristics such poor PS and laboratory evidence of systemic inflammation, a relevant proportion of patients can still achieve long-term survival thanks to immunotherapy.21

ICI treatment type has a relevant impact on PFS and EFP estimates, especially in high-risk patients as identified by cure model covariates (online supplemental table 3, lines 7 and 8).

Anti-PD-(L)1 mono-immunotherapy has a manageable safety profile but is associated with a ~30% rate of primary refractoriness,7 whereas non-randomized trials showed that anti-CTLA-4-containing dual immunotherapy is associated with limited rates of primary resistance, at a price of higher risk of immune-related adverse events.5 6 The better outcomes achieved by an anti-CTLA-4-based dual regimen may be even more relevant for patients’ subgroups with poor prognosis. Therefore, a careful balance between the risk of disease progression or death and the potential occurrence of severe toxicities needs to be carried out in clinical practice. In this perspective, a nomogram encompassing both baseline variables and treatment type may be useful to inform therapeutic choices by weighing the predicted benefit of a given ICI regimen against the risk of toxicity on an individual basis. Moreover, in the perspective of future clinical trials, patients who are more likely to achieve long-term benefit from anti-PD-(L)1 monotherapies alone could be spared from unnecessary toxicity.

To date, ongoing clinical trials are aimed at investigating several novel approaches including immunotherapy combinations and immunotherapy in combination with chemotherapy.22 23

The clinical usefulness of our tool should be validated in ongoing randomized clinical trials investigating anti-CTLA-4-based dual immunotherapy versus single agent and could be tested in other randomized trials with combination strategies. A free app called MSI mCRC Cure has been developed as a web-based online system (freely downloadable online); it allows the user to calculate the 12-month PFS and the EFP based on the combination of the nomogram covariates in individual patients.

Our study has some clear limitations, including its retrospective nature and patients’ and treatment heterogeneity. From a statistical viewpoint, resorting to a global beforehand variable selection using the RSF model13 and not allowing covariate differentiation between the two cure and failure probability model components could be thought of as limitations. However, we think that this strategy is consistent with the ultimate end of the mixture cure model, that is, to estimate a survival probability. Moreover, using the same covariates in both the cure and failure probability models can be an advantage to generalize model estimates and to identify which model covariates are more associated to the cure component and less to the failure component and vice versa, which is useful from a clinical viewpoint for decision-making activities. Another potential limitation lies in that the 12-month timepoint may be early to consider a subject as cured, despite recent evidence on the maintenance of the plateau of the survival curves at the 4 year follow-up update of the ipilimumab plus nivolumab cohort of the CheckMate-142 trial.24 However, we selected the 12-month timepoint to allow predictions as early as possible, with a specific focus on patients with high chance of resistance to ICIs. Moreover, the cure model was subsequently fitted based on PFS independently of a specific timepoint.

In conclusion, thanks to a large multicenter collaborative effort, we developed a nomogram based on five easily assessable clinical variables to estimate 12-month PFS, giving both a scoring system to calculate the 12-month PFS and a scoring system to estimate the time-independent EFP of patients with MSI-high mCRC receiving ICIs. Prospective validation of the nomogram is required, particularly to assess its discriminative performance in selected patients’ subgroups.