Prediction of survival in patients with advanced, refractory colorectal cancer in treatment with trifluridine/tipiracil: real-world vs clinical trial data

Trifluridine/tipiracil increases overall survival (OS) in patients with refractory, metastatic colorectal cancer (mCRC). A post hoc exploratory analysis of the RECOURSE randomized clinical trial (RCT) established two categories, a good prognosis corresponding to subjects having a low tumor burden and indolent disease. Other models in refractory mCRC are the FAS-CORRECT and Colon Life nomogram. The main objective was to externally validate the prognostic factors of the RECOURSE and FAS-CORRECT trials, and the Colon Life nomogram in a multicenter, real-world series of mCRC treated in 3rd and successive lines with trifluridine/tipiracil. The secondary aim was to develop an OS predictive model, TAS-RECOSMO. Between 2016 and 2019, 244 patients were recruited. Median OS was 8.15 vs 8.12 months for the poor (85% of the subjects) and good (15%) prognosis groups from the RESOURCE trial, respectively, log-rank p = 0.9. The most common grade 3–4 toxicities were neutropenia (17%), asthenia (6%), and anemia (5%). The AFT lognormal model TAS-RECOSMO included six variables: ECOG-PS, KRAS/NRAS/BRAF mutation status, time between diagnosis of metastasis and beginning of trifluridine/tipiracil, NLR, CEA, and alkaline phosphatase. The model’s bootstrapped bias-corrected c-index was 0.682 (95% CI, 0.636–0.722). The factors from the Colon Life model, FAS-CORRECT, and RECOURSE displayed a c-index of 0.690, 0.630, and 0.507, respectively. TAS-RECOSMO, FAS-CORRECT, and the Colon Life nomogram appear to predict OS in patients with refractory mCCR who begin trifluridine/tipiracil treatment in the real world. The prognostic groups of the RECOURCE RCT were unable to capture the situation of real-world subjects treated with trifluridine/tipiracil in this series.

www.nature.com/scientificreports/ Despite the inroads in targeted therapies made in recent years, individuals with refractory metastatic colorectal cancer (mCRC) have poor long-term survival. After progressing to second or third lines of chemotherapy, some patients maintain an acceptable functional situation and are eligible for successive therapies that prolong overall survival (OS) and preserve quality of life 1 3 . Nevertheless, the OS benefit does not project equally to all groups, casting doubt on the use of aggressive therapies in subjects with an expectation of limited survival or at the end of life when the foreseeable benefit is diminished 4 . Following this line of reasoning, Tabernero et al. evaluated the prognostic factors in the RECOURSE trial, concluding that OS was independent of age, Eastern Cooperative Oncology Group Performance Status (ECOG-PS), KRAS mutational status, and site of metastases at randomization 5 . According to this analysis, factors of good prognosis were low tumor burden and indolent disease when initiating trifluridine/tipiracil. However, several aspects call into question the external validity and applicability of these results in the real world. Pooled estimates from real-life studies reveal several differences in the baseline characteristics of individuals treated in clinical practice vis-à-vis the ideal population of the RECOURSE RCT 6 . Thus, subjects with worse ECOG-PS or more aggressive tumors are routinely treated in the real world. For example, in a series from the Netherlands, patients treated with trifluridine/tipiracil had worse functional status (ECOG PS of 1 or 2) in 57% and 9%, compared to 44% and 0% in the RECOURSE study population, respectively 7 . Similarly, there were more pretreated patients, as well as a greater proportion of KRAS mutated cancers, in comparison with the RECOURSE RCT; both variables were associated with impacting prognosis. Furthermore, this classification is based on bivariate analyses (log rank tests) 5 . Consequently, the additive contribution of multiple variables or information sources on prognosis (e.g., an individual may have begun trifluridin-tipiracil > 18 months, a favorable factor, yet also have a high tumor burden, an unfavorable factor) has not been contemplated. The Colon Life nomogram, a tool to predict prognosis in mCRC 8 has recently been developed and subsequently validated in the RECOURSE RCT cohort 9 . In addition, the FAS-CORRECT model was devised on the basis of the compassionate use of regorafenib (REBECCA) program in refractory mCRC 10 . These models might help to enhance patient classification.
In this sense, we have sought to externally validate the RECOURSE RCT prognostic factors, as well as the Colon Life nomogram and FAS-CORRECT, in a multicenter, real-world series. Secondarily, we have elaborated the TAS-RECOSMO (TAS-102-trifluridine-tipiracil-in REfractory COlorectal cancer Spanish MOdel) model that makes individualized prediction possible in this population.

Method
Patients and study design. The study population proceed from a database to which 12 Spanish hospitals have contributed. The design was a retrospective. Eligibility criteria included age ≥ 18 years, presence of histologically confirmed mCRC, administration of at least one cycle of trifluridine/tipiracil in third or successive lines, and treatment initiation between June 2016 and June 2019. Centers were asked to collect all consecutive cases meeting eligibility criteria. The study was performed in accordance with Good Clinical Practice guidelines and the Declaration of Helsinki. This observational, non-interventional trial was approved by the Research Ethics Committee of all centers that includes Ethics Committee of Galicia, Hospital General Universitario José María Morales Meseguer and Hospital Central de Asturias. All participants still alive at the time of data collection provided written, signed, informed consent. Informed consent and approval by the competent national authorities includes permission for publication and dissemination of the data. The protocol is shown in Supplementary File 1.

Selection of variables.
The primary endpoint principal was OS defined as the time between commencement of trifluridine/tipiracil until death or loss to follow-up. Progression-free survival (PFS) was defined as the interval between beginning trifluridine-tipiracil until progression or demise, right-censoring event-free subjects at the time of last follow-up. Factors for the predictive model were selected after comprehensively reviewing previously published literature 5,11 . The covariates chosen were neutrophil-lymphocyte ratio (NLR, continuous variable), CEA (continuous, non-linear variable), ECOG-PS, number of metastatic sites (organs involved, dichotomized as in Tabernero et al., < 3 vs ≥ 3) 5 , time since diagnosis of metastasis until starting TAS-102 (evaluated continuously and dichotomized as in Tabernero et al., < 18 vs ≥ 18 months), and alkaline phosphatase (continuous, non-linear variable). Tabernero et al. established two groups: cancers with good prognostic characteristics (GPC) defined as neoplasms having a low tumor burden (< 3 metastatic sites) and indolent disease (≥ 18 months from diagnosis of metastatic disease to trifluridine/tipiracil) (RECOURSE groups); the rest were deemed to have poor prognostic characteristics (PPC) 5 . The Colon Life nomogram comprises four variables (ECOG-PS, resection primary tumor, LDH value, and peritoneal involvement); the model was assessed as per the original description 8,9 . The FAS-CORRECT model consists of four variables: ECOG PS (0, 1, ≥ 2), time since initial diagnosis (≥ 18, < 18 months), number of metastatic sites (< 3, 3 +), and liver metastases 10 .

Statistics.
A log-normal accelerated failure time (AFT) model was used, given that several variables exhibited a dynamic effect. This model assumes that the effect of the covariates is to accelerate or decelerate the course of illness, making it suitable when the assumption of proportional hazards is not met 12 . Survival times in AFT models are multiplied by a constant effect under this formulation, such that the exponential coefficients of the model are denominated time ratios (TR). A TR > 1 implies a longer time to event, whereas a TR < 1 means that the events occurred sooner. Thus, a TR equal to log(0.5) represents that the median time to event is halved in its presence. Since this is a non-interventional, fixed sample size study, inferences should be interpreted accord- www.nature.com/scientificreports/ ing to the magnitude of the CI with a descriptive purpose. The strategy to specify the model was to adopt one degree of freedom for every 15 events available until the highest number of variables ran out, basing decisions on the correlation of the variables with OS (with Somers' Dxy rank correlations) and comparing nested models by applying the Akaike information criterion (AIC) 13 . Non-linear effects were visually inspected and, when necessary, continuous variables were modelled using restricted cubic splines. Discrimination was evaluated by means of bootstrapped bias-corrected Harrel's c-index, while 6-and 12-month calibration was evaluated visually. Analyses were performed in R v4.0.4 with the rms, Hmisc, and visreg software packages [14][15][16] .
AFT model. We then fitted an AFT lognormal model for OS. TAS-RECOSMO included 6 variables: ECOG-PS, KRAS/NRAS/BRAF mutation status, time since diagnosis of metastasis to initiation of trifluridine/tipiracil, NLR (Fig. 5D), CEA, and alkaline phosphatase (Figs. 4A, B). The prognostic effect of laboratory values, alkaline phosphatase and CEA, was clearly non-linear (Fig. 5B, C), and they turned out to be the variables that most closely correlated with OS (Somers' Dxy rank correlations in the Data Supplement) (Annex Fig. 1). The indolent vs aggressive course variable quantified by time to trifluridine-tipiracil displayed a slight biphasic pattern so that the initial protective effect was diluted and counteracted at late timepoints (Fig. 5A). Tumor burden, defined as number of metastatic sites, revealed a weak correlation with OS; its inclusion elevated the model's AIC, and was therefore excluded from the final model. The contrasts for this model are shown in Table 2. TAS-RECOSMO has acceptable discriminatory capacity with bootstrapped bias-corrected c-index of 0.682 (95% CI, 0.636-0.722). The 6-and 12-month calibration plots are illustrated in Annex Fig. 2. The model is well calibrated, except for the range of lowest expected survivals, where OS was slightly overestimated.

Comparison with the colon life model, FAS-CORRECT, and RECOURSE. The Colon Life Model,
formulated as a proportional hazards model, demonstrated poor calibration, associated with PH assumption violation. Once reformulated as an AFT model, the Colon Life Model yielded a c-index of 0.690 in this series and displayed excellent calibration (Annex Fig. 3), while the FAS-CORRECT model, reformulated as an AFT model, revealed a moderate discriminatory capacity with a c-index of 0.630. By comparison, the AFT model constructed on the RECOURSE groups performed poorly, with a c-index of 0.507, consistent with the absence of discrimination in Kaplan-Meier estimations (Fig. 2).

Discussion
External validity is the dimension that is overlooked in ranking evidence, since some RCTs may not be representative of the target population or exclude types of patients who do receive the therapy in the real world 17,18 . This pertains to average effects and safety concerns, but is key, inasmuch as it also impacts the capacity to generalize the analyses of prognostic factors to specific populations. In this work, we have evaluated the prognostic factors of patients with mCRC treated with trifluridine/tipiracil. www.nature.com/scientificreports/ Our data from clinical practice uphold the external validity and applicability of the outcomes of the RECOURSE RCT, albeit with several nuances. First of all, the survival endpoints are comparable, with a median OS of 7.1 and 8.1 months, and median PFS of 2.0 and 2.7, in the RECOURSE trial and in this study, respectively 2 . Likewise, these results are consistent with other reports of real-world observational studies [19][20][21] . The response rate is similar, despite differences in the time pattern in chemotherapy administration -every 8 weeks (RECOURSE RCT) and a median of 11 weeks in this series. In the RECOURSE study, 2% partial response and 16% stabilization rates were observed, while in our study, 81% progressed. The administration of trifluridine/tipiracil was likewise feasible; in the RECOURSE RCT it was administered over 12.7 ± 12.0 weeks (median, 6.7; range, 0.1-78.0), whereas here, a median of 3 cycles (range, 1-16) was administered. Our study saw less toxicity compared to the RECOURSE RCT, with less grade 3-4 neutropenia (17% vs 28%), grade 3-4 anemia (5% vs 18%), and less grade 3-4 thrombopenia (1% vs 5%). This appears to be contingent on the use of lower doses to treat unfit individuals and is on a par with other real-world results 19 .
However, the differences are remarkable as they refer to the analysis of prognostic factors. In the RECOURSE RCT, the protracted time to initiation of trifluridine/tipiracil evaluated dichotomously constituted a protective factor, as it was associated with indolent tumors 5 . This was also seen in the Regotas study, where treatment administered < 18 months since diagnosis of metastasis correlated with worse outcomes 22 , albeit not in other studies 23 . www.nature.com/scientificreports/ Nonetheless, the dichotomization of continuous variables, in this case < 18 vs ≥ 18 months, entails bias and loss of information 24 . In fact, in our study, the timing of administration was similar to that of the RECOURSE RCT, but the results point toward a slight non-linear, continuous effect, with initial protection that is quickly offset by the greater risk with therapies administered after extended periods of time, in advanced phases and cancers. Likewise, tumor burden is a prognostic factor in the RECOURSE RCT and in other series 25 , although it does not appear to be relevant in our study, possibly because trifluridine/tipiracil was administered to a series in which there were many subjects with a high tumor burden, where this variable failed to discriminate between some subjects and others. All this explains why the GPC and PPC groups in the RECOURSE RCT that combine these two factors do not have discriminatory capacity in our clinical practice series. In contrast, TAS-RECOSMO consists of 6 variables with known prognostic effect. ECOG-PS was not a prognostic factor in the RECOURSE RCT, although it is in the Colon Life nomogram, as it is in this and other series 9,10,22 . Similarly, ECOG-PS, RAS/BRAF mutations, and CEA emerge in a prognostic model for mCRC 11 . NLR is a marker of a pro-inflammatory state and known to be prognostic in colon cancer and other tumors 26,27 . Other predictive factors in the literature are bone metastases, albumin or AST, or platelets 22,25 . Our study yields no evidence that the number of previous lines contributes to prognosis, unlike other series 20,28 , perhaps due to the predominance of 3th or 4th line treatments in our sample.
Overall, the performance of the RECOURSE RCT groups was low in our dataset, possibly because their discriminatory capacity is tied to the population in which the model was elaborated. Their baseline characteristics are more homogenous than those of real-world series, as they are selected according to the strict inclusion criteria of RCTs with ECOG 0-1, laboratory variables within a pre-established range, and good liver and kidney function. This suggests that part of the information about prognosis in clinical practice is found in subjects who were not recruited in the RECOURSE RCT 2 . The percentage of patients with ECOG-PS 0 and 2-3 is 57% and 0 in the RECOURSE RCT, compared to 13% and 32% in our series. Likewise, 48% of the RECOURSE RCT population were included in the GPC group, versus 15% in the TAS-RECOSMO. Likewise, KRAS mutations occurred in 51% and 63%, respectively.
In contrast, the Colon Life nomogram is a valid model in our series, with the distinction that the dynamic effect of its variables required that it be reformulated as an AFT model to conserve suitable calibration over time. The FAS-CORRECT is also a valid model, regardless of it having originally been fitted for a cohort of individuals treated with regorafenib 10 . This is not surprising in that the covariates that comprise it are not specific to antiangiogenic therapies.
Readers must be aware of the limitations of our study, the most salient being that TAS-RECOSMO must be externally validated by other groups before it can be recommended for widespread use, although the internal validation suggests that it performs comparably to the Colon Life nomogram and FAS-CORRECT. As far as causal inference is concerned, the small sample size precludes the analysis of multiple variables (restricted by the effective sample size in 15 events per degree of freedom 13 ) and increases the uncertainty of some estimations. For instance, KRAS/NRAS/BRAF mutations appear to lower OS by 16%, although the CI is broad, which is also compatible with a smaller magnitude of effect. Finally, the discrepancies in toxicity may be accounted for by the retrospective nature of data collection in this registry, unlike the RECOURSE RCT.
In conclusion, we have developed and internally validated a model, TAS-RECOSMO, that predicts prognosis on the basis of six clinical-pathological and laboratory variables (general status; neutrophil-lymphocyte ratio; www.nature.com/scientificreports/ KRAS, NRAS and BRAF mutational status; CEA; alkaline phosphatase, and time between diagnosis of metastases until start of trifluridine/tipiracil). Additionally, we have externally validated another two models, the Colon life nomogram and FAS-CORRECT, that could predict OS in individuals with mCCR initiating trifluridine/tipiracil in the real world. The RECOURCE RCT prognostic groups failed to exhibit validity in this series, given that the two variables (tumor burden and indolent/aggressive disease) do not appear to capture the situation of patients treated with trifluridine/tipiracil in the real world. Our data speak to the importance of externally validating the prognostic outcomes obtained in RCTs in patient populations who receive the treatments and scantly resemble those of the trial.