Codon-specific KRAS mutations predict survival benefit of trifluridine/tipiracil in metastatic colorectal cancer

Genomics has greatly improved how patients with cancer are being treated; however, clinical-grade genomic biomarkers for chemotherapies are currently lacking. Using whole-genome analysis of 37 patients with metastatic colorectal cancer (mCRC) treated with the chemotherapy trifluridine/tipiracil (FTD/TPI), we identified KRAS codon G12 (KRASG12) mutations as a potential biomarker of resistance. Next, we collected real-world data of 960 patients with mCRC receiving FTD/TPI and validated that KRASG12 mutations were significantly associated with poor survival, also in analyses restricted to the RAS/RAF mutant subgroup. We next analyzed the data of the global, double-blind, placebo-controlled, phase 3 RECOURSE trial (n = 800 patients) and found that KRASG12 mutations (n = 279) were predictive biomarkers for reduced overall survival (OS) benefit of FTD/TPI versus placebo (unadjusted interaction P = 0.0031, adjusted interaction P = 0.015). For patients with KRASG12 mutations in the RECOURSE trial, OS was not prolonged with FTD/TPI versus placebo (n = 279; hazard ratio (HR) = 0.97; 95% confidence interval (CI) = 0.73–1.20; P = 0.85). In contrast, patients with KRASG13 mutant tumors showed significantly improved OS with FTD/TPI versus placebo (n = 60; HR = 0.29; 95% CI = 0.15–0.55; P < 0.001). In isogenic cell lines and patient-derived organoids, KRASG12 mutations were associated with increased resistance to FTD-based genotoxicity. In conclusion, these data show that KRASG12 mutations are biomarkers for reduced OS benefit of FTD/TPI treatment, with potential implications for approximately 28% of patients with mCRC under consideration for treatment with FTD/TPI. Furthermore, our data suggest that genomics-based precision medicine may be possible for a subset of chemotherapies.

Genomics has greatly improved how patients with cancer are being treated; however, clinical-grade genomic biomarkers for chemotherapies are currently lacking. Using whole-genome analysis of 37 patients with metastatic colorectal cancer (mCRC) treated with the chemotherapy trifluridine/tipiracil (FTD/TPI), we identified KRAS codon G12 (KRAS G12 ) mutations as a potential biomarker of resistance. Next, we collected real-world data of 960 patients with mCRC receiving FTD/TPI and validated that KRAS G12 mutations were significantly associated with poor survival, also in analyses restricted to the RAS/RAF mutant subgroup. We next analyzed the data of the global, double-blind, placebo-controlled, phase 3 RECOURSE trial (n = 800 patients) and found that KRAS G12 mutations (n = 279) were predictive biomarkers for reduced overall survival (OS) benefit of FTD/TPI versus placebo (unadjusted interaction P = 0.0031, adjusted interaction P = 0.015). For patients with KRAS G12 mutations in the RECOURSE trial, OS was not prolonged with FTD/TPI versus placebo (n = 279; hazard ratio (HR) = 0.97; 95% confidence interval (CI) = 0.73-1.20; P = 0.85). In contrast, patients with KRAS G13 mutant tumors showed significantly improved OS with FTD/TPI versus placebo (n = 60; HR = 0.29; 95% CI = 0. 15-0.55; P < 0.001). In isogenic cell lines and patient-derived organoids, KRAS G12 mutations were associated with increased resistance to FTD-based genotoxicity. In conclusion, these data show that KRAS G12 mutations are biomarkers for reduced OS benefit of FTD/TPI treatment, with potential implications for approximately 28% of patients with mCRC under consideration for treatment with FTD/TPI. Furthermore, our data suggest that genomics-based precision medicine may be possible for a subset of chemotherapies.
Systemic anticancer therapy based on the chemotherapeutic agents 5-fluorouracil (5-FU)/capecitabine, oxaliplatin and irinotecan in combination with epidermal growth factor receptor (EGFR) or vascular endothelial growth factor inhibitors are the cornerstone of the treatment of metastatic colorectal cancer (mCRC) 1 . More recently, the chemotherapeutic drug trifluridine/tipiracil (FTD/TPI), a combination of trifluridine (FTD), a nucleoside analog, and tipiracil (TPI), a thymidine phosphorylase inhibitor, has been approved for patients with advanced, heavily pretreated mCRC [2][3][4] . Although durable responses to FTD/TPI have been observed in some patients with mCRC, the median overall survival (OS) benefit in the general population with mCRC is modest (1.8 months), highlighting the unmet need for patient selection 1,3,5-9 . Article https://doi.org/10.1038/s41591-023-02240-8 a codon-agnostic manner diluted the observed effect ( Fig. 1a,b and Supplementary Table 4). Similar results were obtained when time on FTD/TPI treatment was used as the end point (Extended Data Fig. 3 and Supplementary Table 5). Based on these hypothesis-generating results, we wondered if KRAS G12 mutation status could be a determinant of FTD/ TPI treatment outcome in mCRC.

KRAS mutations and real-world survival on FTD/TPI treatment
We next collected real-world data of 960 patients with mCRC who were treated with FTD/TPI in 36 centers across Italy and the UK (Supplementary Tables 6 and 7). Based on routine diagnostics (largely performed at diagnosis), the cohort contained 385 patients with RAS/RAF wild-type (WT) tumors, 343 patients with KRAS G12 mutations, 86 patients with KRAS G13 mutations, 53 patients with KRAS mutations at codons other than G12 or G13 (KRAS other ), 32 patients with BRAF mutations and 61 patients with NRAS mutations. In the full population, patients with KRAS G12 mutations had more frequent right-sided disease and more recent diagnoses of metastatic disease (Table 1). Importantly, these factors were well balanced when patients with KRAS G12 mutations were compared to patients with other RAS/RAF mutations, or specifically to those with hotspot mutations affecting the directly adjacent codon KRAS G13 , whereas this latter subgroup had relatively good performance status (Table 1).
In the real-world validation cohort, codon-specific RAS/RAF mutations were significantly associated with clear differences in OS on treatment with FTD/TPI (log-rank P < 0.001; Fig. 2a). Again, KRAS G12 mutations were significantly associated with poor OS, with a similar effect in the population as a whole (unadjusted hazard ratio (HR) for death = 1.31; 95% CI = 1.11-1.55; P = 0.0017; adjusted HR for death = 1.24; 95% CI = 1.04-1.47, P = 0.016; Fig. 2b, left panel), as in the RAS/RAF mutant subpopulation (unadjusted HR for death = 1.30; 95% CI, 1.04-1.61, P = 0.019; adjusted HR for death, 1.28; 95% CI 1.03-1.60, P = 0.027; Fig. 2b, middle panel). Notably, the OS of patients with KRAS G12 mutations was also poor as compared with patients with KRAS G13 mutations (unadjusted HR for death = 1.79; 95% CI = 1.29-2.48; P < 0.001; adjusted HR for death = 1.61; 95% CI = 1. P = 0.0061;Fig. 2b,right panel). Similar results were obtained when the analysis was based on progression-free survival (PFS) (Extended Data Fig. 4). Patients with KRAS G12 mutations did not show significantly shorter OS than patients in any of the other, smaller RAS/RAF mutant subgroups (those with Precision medicine is widely used to select patients for targeted therapies and immunotherapies for mCRC according to the presence or absence of genomic biomarkers. As such, the detection of KRAS hotspot mutations is a critical step in the diagnostic workup of mCRC as RAS/RAF mutations predict clinical resistance to EGFR-targeting antibodies [10][11][12] . Such KRAS mutations are found in 44% of patients with mCRC; these most frequently occur at codon G12 (KRAS G12 ; 28% of patients) or codon G13 (KRAS G13 ; 8% of patients) (Extended Data Fig. 1 and Supplementary Table 1) 13 . Although KRAS G12 and KRAS G13 mutations are regarded as a single entity in clinical practice guidelines, they have different biochemical properties 14,15 and display tissue-and treatment-specific mutational patterns 16 .
In this study, given the lack of genomic biomarkers and the limited clinical benefit of FTD/TPI in unselected patients with mCRC, we harnessed the power of whole-genome somatic profiles coupled with patient outcomes to identify biomarkers of response and resistance to FTD/TPI. Key findings were then validated in a real-world cohort of FTD/TPI-treated patients with mCRC (n = 960) and in the double-blind, placebo-controlled, phase 3 RECOURSE trial (n = 800; study overview shown in Extended Data Fig. 2).

KRAS G12 mutations as potential biomarkers for FTD/TPI treatment
We first performed whole-genome analysis of a real-world discovery cohort that consisted of 37 patients with mCRC from the publicly available Hartwig Medical Foundation database 17 , who received FTD/TPI treatment in a standard-of-care setting in 13 hospitals across the Netherlands (Hartwig Medical Foundation (HMF) cohort; Supplementary Table 2). In accordance with late-stage disease, the median OS was relatively short, that is, 6.1 months (95% confidence interval (CI) = 4.2- 8.3). Ten genomic drivers occurred in at least five patients and were tested as candidate biomarkers for OS (Supplementary Table 3 and Methods). After correction for multiple-hypothesis testing, KRAS G12 status was most significantly associated with reduced OS (exact log-rank test-based two-sided P = 0.0016; Benjamini-Hochberg false discovery rate (FDR) = 0.016; threshold for significance, FDR < 0.05; Fig. 1a

KRAS mutations and survival in the RECOURSE trial
To further strengthen our findings and investigate if our observations were based on prognostic or predictive effects, we analyzed the data of a large, independent, placebo-controlled clinical cohort, the RECOURSE trial 3 . Briefly, this international, randomized, double-blind, placebo-controlled, phase 3 study assigned 800 heavily pretreated patients with mCRC to receive either FTD/TPI or placebo in a 2:1 ratio. Based on routine diagnostics (largely performed at diagnosis), approximately half of the patients (n = 393) were KRAS WT , whereas the other half (n = 407) were KRAS mutant. In this study, KRAS mutation status (mutated yes/no) was not significantly associated with reduced OS or PFS benefit of FTD/TPI versus placebo; however, codon-specific analyses were not performed 3,4 .  Codon-specific mutational status was available for 367 out of 407 (90%) patients with KRAS-mutated tumors in the RECOURSE trial. Of these, 279 (76%) had KRAS G12 mutations, 60 (16%) had KRAS G13 mutations, 21 (5.7%) were reported to have KRAS G12/G13 double mutations (largely due to the use of analytical methods that could not discriminate between the two codons) and 7 (1.9%) had other mutations. (The true percentage of patients with other mutations was probably higher because their assessment was only broadly implemented later 11 .) Throughout our analyses, we considered patients with KRAS G12/G13 double mutations as a distinct subgroup.

MMR status
The prespecified baseline characteristics of the RECOURSE trial were well balanced between the FTD/TPI and placebo arms in KRAS G12 mutant, KRAS G13 mutant and KRAS WT subgroups (Table 2), with some exceptions; patients whose tumors harbored a KRAS mutation generally had more recent diagnoses of metastatic disease, were less heavily pretreated and were more frequently refractory to fluoropyrimidine  as part of the last previous regimen (Table 2). Importantly, all these factors were balanced between the KRAS G12 and KRAS G13 mutant populations. Between these two groups, the only significant difference was that patients with KRAS G13 mutations originated less often from Japan (Table 2). To understand the prognostic effects of codon-specific KRAS mutations in the trial population, we first analyzed OS in the placebo arm (Extended Data Fig. 5). This showed that patients with the KRAS G12 and KRAS WT mutants had similar OS. Interestingly, placebo-treated patients with KRAS G13 mutations (the other main KRAS mutant subgroup in the study) had a remarkably shorter OS than those with KRAS G12 mutations (median OS KRAS G13 mutants: 2.9 months, 95% CI = 2.1-6.1 months versus median OS KRAS G12 mutants: 5.8 months, 95% CI = 4.7-7.3; HR = 2.20; 95% CI = 1.25-3.86; P = 0.0060; Extended Data Fig. 5), which held after adjustment for the ten baseline characteristics (HR = 2.46; 95% CI = 1.33-4.57; P = 0.0043; Extended Data Fig. 5). In the placebo arm, patients with KRAS G13 mutant tumors also had shorter OS than those with KRAS WT tumors, which was statistically significant in unadjusted analysis (HR = 1.95; 95% CI = 1. 13-3.36; P = 0.017; Extended Data Fig. 5), but did not attain statistical significance in the adjusted analysis (HR = 1.79; 95% CI = 0.96-3.32; P = 0.065; Extended Data Fig. 5). Taken together, these analyses indicate that KRAS G12 mutations are not associated with poor prognosis in late-stage mCRC.
We then studied if KRAS G12 mutations were predictive biomarkers for reduced OS benefit of FTD/TPI in the RECOURSE trial. In the KRAS G12 mutant population (n = 279 patients), OS was not prolonged with FTD/TPI versus placebo (HR = 0.96; 95% CI = 0.71-1.29; P = 0.78; Fig. 3a, upper left). Within the full study population (n = 800 patients), KRAS G12 mutations were significantly associated with a reduced OS benefit of FTD/TPI versus placebo (unadjusted interaction P = 0.0031;   Fig. 6). Taken together, these data demonstrate that FTD/TPI treatment did not lead to a clinically relevant prolongation of OS in patients with mCRC with KRAS G12 mutations in the RECOURSE trial. When patients whose tumors harbored a KRAS G12 mutation were excluded from the analysis, FTD/TPI resulted in a pronounced OS benefit over placebo (n = 521; HR = 0.55; 95% CI = 0.45-0.69; P < 0.001; Fig. 3a, upper right), with a median OS benefit of 2.7 months in this subgroup (versus 1.8 months in the full population, as reported by Mayer et al. 3 ).
Next, we analyzed the treatment effect of FTD/TPI in patients with KRAS G13 mutant tumors. In sharp contrast to the KRAS G12 mutant population, patients with the KRAS G13 mutation showed a clear OS benefit in the FTD/TPI arm versus the placebo arm (HR = 0.34; 95% CI = 0.17-0.67; P = 0.0018; Fig. 3a, lower left). This remained significant in the adjusted analysis (HR = 0.21; 95% CI = 0.090-0.48; P < 0.001; Fig. 3b). The median OS was three times longer in the FTD/TPI arm versus the placebo arm (8.7 versus 2.9 months; Fig. 3). The OS benefit of FTD/TPI treatment was significantly more pronounced in patients with KRAS G13 -mutated mCRC versus those with KRAS G12 -mutated disease (unadjusted interaction P = 0.0026; adjusted interaction P = 0.0023; Fig. 3b; the full regression model fits are shown in Supplementary Table 8). Thus, KRAS G13 mutations marked patients with clear OS benefit from FTD/TPI treatment.
We then assessed PFS in KRAS codon-specific subgroups of the RECOURSE trial. A minimal PFS benefit of FTD/TPI versus placebo was observed in all three subgroups (median PFS benefit 0.1, 0.3 and 0.3 months for patients with KRAS G12 , KRAS G13 and KRAS WT mutations, respectively), which did not significantly differ among these subpopulations (interactions nonsignificant for all pairwise comparisons; Extended Data Fig. 7).

KRAS G12 mutations and FTD/TPI resistance in vitro
Finally, we aimed to replicate these findings in vitro using isogenic cell lines and mCRC patient-derived organoids (PDOs) (n = 7; Supplementary Table 9). KRAS G12 mutation knock in significantly reduced responsiveness to FTD (the cytotoxic component of FTD/TPI) in two colorectal cancer cell line models, SW48 and Colo320 (two-sided Wilcoxon rank-sum-based P = 0.029 for both models; Fig. 4a-d). The parental models are KRAS WT and do not harbor other frequent mCRC oncogenic drivers like mutations in NRAS, BRAF, PTEN or PIK3CA. Similar results were obtained with PDOs, with KRAS G12 -mutated lines consistently showing reduced FTD responsiveness (two-sided Wilcoxon rank-sum-based P = 0.034; Fig. 4e,f). Notably, the presence of a KRAS G12 mutation was associated with suppression of FTD-induced DNA damage (as measured by γH2AX) in both isogenic cell lines and PDOs (Fig. 4g,h). We next tested in vitro sensitivity to 5-FU because this chemotherapeutic is closely related to FTD/TPI but exerts its main effect through damaging RNA rather than DNA. In all models, KRAS G12 mutations did not significantly reduce in vitro sensitivity to 5-FU ( Fig. 4i-l). Of note, the higher sensitivity to FTD in KRAS WT models could not be explained by higher baseline proliferation rates, as the (untreated) KRAS WT PDOs demonstrated lower proliferation rates than (untreated) KRAS G12 PDOs (Extended Data Fig. 8). Taken together, these results show that KRAS G12 mutation-based resistance to FTD can be modeled in vitro and is characterized by limited FTD-induced DNA damage.

Discussion
Using two independent real-world datasets from three different countries and an independent validation cohort based on the global, double-blind, placebo-controlled, phase 3 RECOURSE trial, we demonstrate that codon-specific KRAS mutations predict OS benefit for patients treated with the chemotherapeutic agent FTD/TPI in late-stage mCRC. Specifically, KRAS G12 mutations identify patients who experience no clinically relevant 18 survival benefit from FTD/TPI, while the remaining population-including KRAS G13 -mutated patients-benefits substantially. The RECOURSE trial showed only a modest OS benefit of FTD/TPI versus placebo in the general, unselected population with mCRC. In this context, our results offer a framework to (re)assess the risk-benefit profile of FTD/TPI according to codon-specific KRAS mutations. Given that KRAS testing is routinely performed in the molecular workup of all patients with CRC to guide treatment with EGFR-targeting agents 1,11 , our findings can be readily adopted in the clinic.
In line with previous clinical and preclinical evidence [14][15][16] , our data demonstrate that KRAS G12 and KRAS G13 mutated mCRC are different clinical entities. The former disease is characterized by better prognosis but shows no clinically relevant OS benefit of FTD/TPI treatment (predictive effect), whereas the latter disease behaves aggressively when treated with placebo but can be more effectively managed with FTD/TPI treatment. These data caution against lumping together KRAS mutations at different codons in biomarker analyses and clinical trial designs because different biological and biochemical properties may be associated with different clinical outcomes.
The primary objective of the RECOURSE trial was to detect differences in OS between FTD/TPI and placebo, the gold standard outcome measure for regulatory approval studies for new treatments for metastatic cancer 19,20 . One of the reasons for this is that marginal improvements in PFS may not translate into an OS benefit 21 as we observe in the subpopulation of the RECOURSE trial with KRAS G12 mutant tumors. The main caveat of OS is that lines of treatment administered after progression on the study drug might bias the conclusions. Notably, information on 5-FU-based rechallenges was not collected in the RECOURSE trial but are unlikely to underlie the reduced OS benefit of FTD/TPI in the population with KRAS G12 mutations. The reason is that this would require that placebo-treated patients with KRAS G12 mutations received considerably more treatments after progression in the study than FTD/ TPI-treated patients with KRAS G12 mutations. Nevertheless, even in this unlikely scenario the conclusion would still be that, in terms of OS, the treatment with FTD/TPI has not been a useful intervention because it did not provide a relevant OS benefit over treatment with placebo.
Analysis of the real-world validation cohort showed that mismatch repair (MMR) deficiency was rare (Table 1) and was not associated with KRAS G12 status nor with OS of patients treated with FTD/TPI (data not shown). Furthermore, tumor sidedness was well balanced among all RAS/RAF-based subgroups of the real-world cohort and adjustment for this covariate in multivariate models did not affect our conclusions. In addition, pretreatment variables, such as the number of previous regimens, refractoriness to fluoropyrimidine or previous use of regorafenib were not responsible for our results. Namely, our RECOURSE trial-based analyses showed that these (1) were well balanced between the populations with KRAS G12 and KRAS G13 mutant tumors, (2) were not associated with OS benefit of FTD/TPI versus placebo and (3) did not alter our conclusions when incorporated into multivariate models.
While all RAS/RAF-based subgroups were molecularly well defined in our real-world datasets, this classification was not as complete in the RECOURSE trial. Indeed, KRAS hotspot mutations outside codons G12 and G13 were only tested in a small fraction of the RECOURSE trial population and data on NRAS and BRAF mutations were (largely) missing. Given the results of our real-world analyses, patients with KRAS mutations outside of codons G12 and G13 or BRAF mutations may more closely resemble patients with KRAS G12 mutations; inclusions Article https://doi.org/10.1038/s41591-023-02240-8 of these cases in the KRAS WT group might have underestimated the survival benefit conferred by FTD/TPI in the 'real' KRAS WT population. A further observation with potential clinical implications relates to the fact that virtually all KRAS WT patients in our cohorts were pretreated with anti-EGFR therapeutics, while their RAS (and RAF) status was determined before any therapy. Given that RAS mutations can emerge as drivers of acquired resistance in this scenario 22 , some patients might have been misclassified. Although the above considerations are important to keep in mind when interpreting the results, our conclusions hold regardless because such misclassifications may only have diluted the differences between the analyzed subgroups.
A potential limitation of our study is that this investigator-initiated reanalysis of the RECOURSE trial was not predefined in the original trial protocol. However, based on our findings, this reanalysis was hypothesis-driven and prespecified in a formal data request before access to the RECOURSE trial data was granted.
Several clinico-pathological and molecular biomarkers of benefit to FTD/TPI have been tested but none has reached clinical application 23 . colorectal cancer cell lines after 2 weeks' exposure to a concentration range of FTD in vitro. b, As in a, but for KRAS WT and KRAS G12D isogenic Colo320 CRC cell lines. c, Dose-response curves of KRAS WT (black) and KRAS G12 (red) isogenic SW48 (dots) or Colo320 (diamonds) CRC cell lines exposed to a concentration range of FTD in vitro. The dots and error bars represent the mean and s.d. among four biological replicates at the tested concentrations, respectively. d, Half-maximal inhibitory concentrations (IC 50 ; log 2 ) for FTD of KRAS WT and KRAS G12 isogenic SW48 or Colo320 CRC cell lines, as indicated on the x axis. Data are plotted for four biological replicates. The box center lines, box ranges, whiskers and dots indicate the medians, quartiles, 1.5 times the IQR and data points of individual experiments (biological replicates; n = 4 for each line), respectively. The twosided Wilcoxon rank-sum test-based P value is shown. e, Dose-response curves of mCRC PDOs harboring WT KRAS (black; n = 3) or different KRAS G12 mutations (red; n = 4) exposed to FTD in vitro. The dots and error bars represent the mean and s.d. at the tested concentrations, respectively. f, IC 50 (log 2 ) for FTD of KRAS WT (black; n = 3) and KRAS G12 (red; n = 4) mCRC PDOs. The box center lines, box ranges, whiskers and dots indicate the medians, quartiles, 1.5 times the IQR and data points of individual organoid lines (see legend), respectively. The two-sided Wilcoxon rank-sum test-based P value is shown. g, Representative western blot of the DNA damage marker γH2AX on treatment of KRAS WT (black) and KRAS G12 (red) SW48 (left) and Colo320 (right) cells with FTD at increasing concentrations. Hsp90 was used as a loading control. Data were confirmed in three and two biological replicates for SW48 and Colo320, respectively. h, As in g, but for mCRC PDOs harboring KRAS WT (black, left) or different KRAS G12 mutations (red, right). The left and right panels were exposed together (Source Data 1). Data were confirmed in three biological replicates. i, As in c but for 5-FU. j, As in d but for 5-FU. k, As in e but for 5-FU. l, As in f but for 5-FU.
Article https://doi.org/10.1038/s41591-023-02240-8 Our results show that KRAS mutational analysis, a standard-of-care test already implemented worldwide, can identify patients with KRAS G12 mutant mCRC who are unlikely to benefit from FTD/TPI treatment, avoiding unnecessary toxicities to patients and rationalizing the use of resources for healthcare systems. Thus, we report the first proof of genomics-based precision medicine for a chemotherapy in mCRC, which has the potential to substantially improve patient selection for FTD/TPI treatment.

Online content
Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41591-023-02240-8.

Study participants
Discovery cohort. The large, publicly available, real-world dataset with clinical annotation and whole-genome sequencing (WGS) by the HMF was used as the discovery cohort 17 . All patients who received FTD/TPI as part of their standard-of-care treatment for mCRC were identified in May 2018 (Supplementary Table 2). These patients were included in 13 academic, teaching, and regional hospitals in the Netherlands. The study was approved by the Medical Ethical Committee of the University Medical Center Utrecht and was conducted in accordance with the Declaration of Helsinki (fourth edition). All patients provided written informed consent for the collection, analysis and pseudonymized sharing of paired tumor-normal WGS data and clinical characteristics for research purposes. RECOURSE trial cohort. The RECOURSE trial design has been previously described in detail 3 . Briefly, the RECOURSE trial (NCT01607957) was an international double-blind, randomized, placebo-controlled, phase 3 trial comparing FTD/TPI plus best supportive care to placebo plus best supportive care. Heavily pretreated patients with refractory mCRC (n = 800) were randomly assigned in a 2:1 ratio to receive FTD/ TPI or placebo. Within this process, patients were stratified based on KRAS status (mutant yes/no), time between first diagnosis of metastases and randomization (<18 versus ≥18 months) and geographical region ( Japan or USA, Europe and Australia). The data cutoff was at 571 deaths, in accordance with the cutoff of the primary analysis. All patients in the study provided written informed consent, as stated in the original publication 3 .  Table 9). The study was conducted in accordance with the Declaration of Helsinki. All patients provided written informed consent for organoid culture and collection, analysis and pseudonymized sharing of clinical characteristics for research purposes.

End points and study objectives
In the real-world discovery analysis, we searched for genome-wide somatic variants associated with OS and time on FTD/TPI treatment as end points. In the real-world validation analysis, the primary and secondary objectives were to assess the association of KRAS G12 mutation status with OS and PFS, respectively, both in the population as a whole and in RAS/RAF mutation-based subpopulations. All end points used in the real-life analyses were measured from the start of FTD/TPI treatment and evaluated at participating institutions over the treatment course according to local practice. In our reanalysis of the RECOURSE trial, we tested OS benefit and PFS benefit of FTD/TPI versus placebo as the primary and secondary end points, respectively, in subgroups defined by codon-specific KRAS mutation status. This was in accordance with the hierarchy of end points prespecified in the RECOURSE trial protocol; these reanalyses were prespecified in a formal data request to the sponsor of the RECOURSE study before access to the data was granted.

Bioinformatics analysis
All genomics data of the discovery cohort was publicly available and provided by the HMF under the approved data request DR-015. WGS (median depths approximately 100 and approximately 40 for tumor and normal, respectively) and bioinformatics analysis of the discovery cohort were performed by the HMF as described previously 17 , with an optimized pipeline based on open-source tools freely available on GitHub (https:// github.com/hartwigmedical/pipeline5). Somatic genomic drivers were identified as an integrated functionality of PURPLE v.2.43 (ref. 17 ). Briefly, somatic mutations were considered drivers if they fulfilled one of the following criteria: (1) mutations in oncogenes located at-or within five bases of-known hotspots; (2) inframe indels in oncogenes with repeat count <8 repeats; (3) biallelic (that is, the WT allele is lost) nonsense, splice or indel variants in tumor suppressor genes (TSGs); and (4) mutations in oncogenes or TSGs with a sample-specific driver likelihood >80%, as calculated by PURPLE as described previously 17 . For this manuscript, we only considered TSG mutations to be drivers if (1) they were biallelic or (2) in the case of multiple mutations in the gene for which the summed variant ploidies exceeded the gene ploidy within the sample −0.5 (for example, the classical APC two-hit hypothesis). Amplifications were considered to be drivers if (1) they affected an oncogene with pan-cancer evidence for recurrent amplification 17 and (2) this oncogene had a copy number exceeding three times the sample ploidy. Deletions were considered to be drivers if (1) they affected TSGs with pan-cancer evidence for recurrent deletion 17 and (2) they were homozygous (absolute gene copy number <0.5).

Statistical analysis
Median time on treatment, OS and PFS were calculated using the Kaplan-Meier method. OS and time on treatment were compared between biomarker positive versus negative patients in the discovery cohort using the exact log-rank test. In this analysis, multiple-hypothesis correction was performed using the Benjamini-Hochberg procedure. HRs and corresponding 95% CIs, and Wald test-based two-sided P values, were estimated from Cox regression models. The proportional-hazards assumption was tested using the methodology developed by Grambsch and Therneau 25 , with a significance threshold of P = 0.05; categorical covariates were modeled as stratification factors (rather than covariates) where appropriate to prevent assumption violations. 'Unadjusted' Cox regression analyses of the real-world validation cohort were performed in a univariate manner. 'Adjusted' Cox regression analyses of the real-world validation cohort were stratified for ECOG performance status ( . RECOURSE trial data-based survival analyses were performed on the intention-to-treat population and were prespecified in a formal data request before access to the data was granted.

Candidate biomarker selection for the discovery cohort
The procedure for the selection of candidate biomarkers was as follows. Somatic genomic driver alterations (mutations and copy number alterations) were included as candidate biomarkers at increasingly specific 'levels': (1) gene-level biomarkers, for example, 'APC alteration', which could either be by mutation or copy number alteration; (2) variant class-level biomarkers, for example, 'APC mutation' or 'APC deletion'; (3) codon-level biomarkers, for example, 'APC codon 1450 mutation'; and (4) amino acid change-specific biomarkers, for example, 'APC p.Thr562Met mutation'. In cases where biomarkers of different levels showed complete redundancy, only the most specific level was included. For example, all KRAS alterations in the cohort were mutations leading to complete redundancy between 'KRAS alteration' and 'KRAS mutation'. Hence, KRAS mutation was selected as the most specific level and included as a candidate biomarker, whereas KRAS alteration was excluded. All candidate biomarkers occurring in at least five patients in the discovery cohort were tested for association with treatment outcomes. Supplementary Table 3 provides a comprehensive overview of the frequencies of all candidate biomarkers identified in our cohort.

Variable selection for multivariate Cox regression
Real-world validation cohort. We selected eight variables for multivariate (adjusted) Cox proportional-hazards modeling of the real-world validation cohort. In this process, we aimed to harmonize the selection as much as possible to the variables used in the Cox regression modeling of the RECOURSE trial-based data (see below), with some alterations. The ECOG performance status (0-1 versus ≥2) was used as a stratification factor because this variable violated the proportional-hazards assumption when modeled as a covariate. Furthermore, we adjusted for seven additional covariates: time since diagnosis of first metastases (<18 versus ≥18 months); geographical region (UK versus Italy); age (<65 versus ≥65 years); sex; sidedness (left versus right); previous surgery (yes versus no); and peritoneal disease at the start of FTD/ TPI treatment (yes versus no). Sidedness was used instead of primary site of the disease (colon versus rectum, as used in the RECOURSE trial-based analyses), because sidedness was most strongly associated with OS and only one of these two variables could be included due to high collinearity. Due to data unavailability for the real-world validation cohort, we were unable to factor in if the disease was refractory to fluoropyrimidine as part of the last previous regimen, previous use of regorafenib, the number of previous regimens and the number of metastatic sites in the analyses. In RECOURSE trial-based analyses, none of these factors were predictive and only the latter variable was prognostic for OS. Instead, based on significant (univariate) associations with OS in the real-world validation cohort, we decided to add the two variables 'previous surgery' and 'peritoneal disease at the start of FTD/TPI treatment' to our selection of covariates, although these data were unavailable for the RECOURSE trial dataset.

RECOURSE trial-based analyses.
We selected ten variables for multivariate (adjusted) Cox proportional-hazards modeling of RECOURSE trial data.
This selection included all factors prespecified in the RECOURSE trial study protocol, except KRAS status and ethnicity, totaling eight prespecified factors: time since diagnosis of first metastases (<18 versus ≥18 months (stratification factor of the study); geographical region ( Japan versus the USA, Europe and Australia; stratification factor of the study); age (<65 versus ≥65 years); sex; ECOG performance status (0 versus 1); primary site of the disease (colon versus rectum); number of previous regimens (2, 3 or ≥4); and number of metastatic sites (1-2 versus ≥3). KRAS status was excluded because of collinearity with our variables of interest (KRAS G12 mutation, KRAS G13 mutation, KRAS WT ). Ethnicity was excluded for two reasons. First, the sponsor of the RECOURSE trial could not share the original ethnicity data for privacy reasons because the number of Black participants (nine patients) was below a predefined threshold put in place to prevent patient reidentification. For this reason, the ethnicity item has been modified to a quasi-identifier of 'Asian' versus 'Other' (White or Black). In the RECOURSE trial, the original ethnicity variable was not significantly prognostic or predictive for OS 3 . Second, the modified ethnicity variable showed high collinearity (and hence redundancy) with the included factor 'geographical region' because 266 out of 266 (100%) of participants from the 'Asia' region had the 'Asian' ethnicity and 522 out of 534 (98%) of participants from the USA, Europe and Australia regions had the 'Other' (which included Black and White) ethnicity.
Next, we included two additional factors in our multivariate models: (1) disease refractory to fluoropyrimidine as part of the last previous regimen; and (2) previous use of regorafenib. These factors were not prespecified in the RECOURSE trial protocol for multivariate analyses but were used for the subgroup analyses reported by Mayer et al. 3 . We decided to include these pretreatment-related factors in our multivariate models because patients with KRAS mutant tumors showed significant differences regarding their pretreatment profiles as compared to patients with KRAS WT tumors. Patients with KRAS mutant tumors were more often refractory to fluoropyrimidine as part of their last previous regimen and were less heavily pretreated than patients with KRAS WT (Table 2).
BRAF mutation status was not included in our selection because this information was missing for 676 out of 800 (85%) patients. Subgroup analysis of the population with BRAF mutant tumors was not possible because BRAF mutations were detected in only eight patients.

Western blot analysis
Western blot analysis was performed on the isogenic cell lines and PDOs treated with FTD or 5-FU, at different concentrations, for 24 h (cell lines) or 48 h (PDOs). For PDOs (but not for the isogenic cell lines), the extracellular matrix was removed by incubating with 2 mg ml −1 type II dispase (catalog no. D4693, Sigma-Aldrich) for 10 min at 37 °C.

Colony formation assay
Cells were seeded into six-well plates (1.5-2 × 10 4 cells per well) and cultured in the presence of drugs at the indicated concentrations. For each cell line, cells cultured using different conditions were fixed in methanol (catalog no. 32213, Honeywell) and stained with 0.1% crystal violet solution (catalog no. V5265, Sigma-Aldrich).

Role of the funding source
The funders of the study had no role in study design, data collection, data analysis, data interpretation or writing of the article. Authors had full access to all the data and had the final responsibility to submit for publication.

Statistical methods and associated software
The statistical methods and associated packages used in this study are summarized in Table 3.

Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability
The somatic mutation data of the MSKCC cohort are freely available via the cBioPortal for cancer genomics (http://cbioportal.org/ msk-impact); patient identifiers are provided in Supplementary  Table 7. The RECOURSE trial data can be accessed upon approval of a data request at Servier (https://clinicaltrials.servier. com/data-request-portal/). Source data are provided with this paper.

Code availability
The bioinformatics for the WGS data of the discovery cohort were performed with an optimized pipeline based on open-source tools; this is freely available on GitHub (https://github.com/hartwigmedical/ pipeline5).  Last updated by author(s): Jan 12, 2023 Reporting Summary Nature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency in reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly The statistical test(s) used AND whether they are one-or two-sided Only common tests should be described solely by name; describe more complex techniques in the Methods section.
A description of all covariates tested A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals) For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted Data Policy information about availability of data All manuscripts must include a data availability statement. This statement should provide the following information, where applicable: -Accession codes, unique identifiers, or web links for publicly available datasets -A description of any restrictions on data availability -For clinical datasets or third party data, please ensure that the statement adheres to our policy Discovery cohort HMF Sample identifiers and patient-level biomarker and clinical outcome data are available in Supplementary Table 2. Raw and processed genomics data are freely available at Hartwig Medical Foundation through standardized procedures and request forms (https://www.hartwigmedicalfoundation.nl/en/data/data-accesrequest).

Real-World validation cohort
The original data used in all analyses of the real-world validation cohort can be found Supplementary Table 7.

RECOURSE trial
The RECOURSE trial data can be accessed upon approval of a data request at Servier (https://clinicaltrials.servier.com/data-request-portal/).

MSKCC Colorectal Cancer Cohort
Somatic mutation data of the MSKCC cohort are freely available via the cBioportal for cancer genomics (http://cbioportal.org/msk-impact) and patient identifiers are provided in Supplementary Table 1.

Human research participants
Policy information about studies involving human research participants and Sex and Gender in Research.

Reporting on sex and gender
Throughout the manuscript, we report sex as an biological attribute, which was collected through self reporting. Sharing of patient-level, pseudonymized clinical data, including sex, for research purposes was covered by the informed consents of the individual studies and local legislation.

Population characteristics
Discovery cohort HMF Patients with advanced or metastatic cancer for whom there is an indication for any systemic treatment were included in the HMF database as part of the CPCT-02 (NCT01855477) clinical study. We included all patients in the HMF database who received FTD/TPI as part of their standard-of-care treatment for advanced/metastatic, histologically proven, colorectal carcinoma, which were identified in the HMF database in May 2018. The median age was 62 years (IQR 59-68), 38% was (self-reported) female and 62% male, and the percentage of patients with KRAS G12 and other KRAS mutations were 54% and 11%, respectively. All patients received FTD/TPI treatment as part of standard-of-care. As FTD/TPI is approved as final line therapy, these were all late-stage patients.
Real-World Cohort (UK and Italy) Retrospective collection of all sequential patients with advanced/metastatic, histologically proven, colorectal carcinoma undergoing treatment with FTD/TPI following progression to prior lines of standard chemotherapy, with known codonspecific RAS/RAF status known between November 2016 and March 2022. As reported in Table 1, the median age was 64 years (IQR 56-71), 41% was (self-reported) female and 59% male, 72% and 28% were diagnosed with colon and rectum cancer, respectively, and the percentage of patients with KRAS G12, KRAS G13, KRAS other, NRAS and BRAF mutations were 36%, 9.0%, 5.5%, 6.4%, and 3.3%, respectively. All patients received FTD/TPI treatment as part of standard-of-care. As FTD/ TPI is approved as final line therapy, these were all late-stage patients.
RECOURSE trial (population description from the original publication, by Mayer et al, NEJM 2015) "Baseline demographic and disease characteristics were well balanced between the two study groups (Table 1). All the patients had received prior chemotherapy regimens containing a fluoropyrimidine, oxaliplatin, and irinotecan; all but one patient (in the placebo group) had received bevacizumab. All but two patients (one patient in each study group) with KRAS wild-type tumors had received cetuximab or panitumumab. Regorafenib, an oral multikinase inhibitor, became available for the management of previously treated colorectal cancer during the course of the study; 17% of the patients in the TAS-102 group, as compared with 20% of those in the placebo group, had received this drug. A large percentage of patients in both study groups -93% of patients receiving TAS-102 and 90% of those receiving placebo -had disease that had been refractory to fluoropyrimidines when they were last exposed to this class of drugs. Moreover, 58% of the patients receiving TAS-102 and 54% of the patients receiving placebo had disease that had been refractory to fluoropyrimidine when that drug was administered as part of their last treatment regimen before study entry." "Patients with biopsy-documented adenocarcinoma of the colon or rectum were eligible for participation in the study if they had received at least two prior regimens of standard chemotherapies, which could have included adjuvant chemotherapy if a tumor had recurred within 6 months after the last administration of this therapy; if they had either tumor progression within 3 months after the last administration of chemotherapy; or if they had had clinically significant adverse events from standard chemotherapies that precluded the readministration of those therapies. Eligibility also required knowledge of tumor status with regard to KRAS (i.e., wild-type or mutant), as reported by investigators. Patients were also required to have received chemotherapy with each of the following agents: a fluoropyrimidine, oxaliplatin, irinotecan, bevacizumab, and -for patients with KRAS wild-type tumors -cetuximab or panitumumab. In addition, patients had to be 18 years of age or older; have adequate bone-marrow, liver, and renal function; and have an Eastern Cooperative Oncology Group (ECOG) performance status of 0 or 1 (on a scale of 0 to 5, with 0 indicating no symptoms, 1 indicating mild symptoms, and higher numbers indicating increasing degrees of disability)."

Recruitment
Discovery cohort HMF Patients with advanced or metastatic cancer for whom there is an indication for any systemic treatment were included in the HMF database as part of the CPCT-02 (NCT01855477) clinical study. The CPCT-02 study, patients were included by 41 academic, teaching and general hospitals across The Netherlands and collected material and clinical data by standardized protocols. Metastatic cancer patients were asked to participate in the studies in any of the 41 participating hospitals. Recruitment involved hundreds of medical specialists and research nurses which minimizes self-selection biases. Recruitment was independent on tumor type. An important requirement for participation was the ability to safely undergo a tumor biopsy. Health conditions and lesion site related risk could therefore have resulted in exclusion of patients.
Real-world validation cohort As per above RECOURSE trial (recruitment details from the original publication, by Mayer et al, NEJM 2015) "Between June 17, 2012, and October 8, 2013, a total of 1002 patients were screened for eligibility, of whom 800 underwent randomization, with 534 assigned to receive TAS-102 and 266 assigned to receive placebo (intention-to-treat population) (details regarding the disposition of patients are provided in Fig. S1 in the Supplementary Appendix, available at NEJM.org). Treatment was initiated in 798 patients, with 533 receiving TAS-102 and 265 receiving placebo (safety-analysis population). All treated patients received their assigned study drug according to the randomization schema, and 760 could be evaluated for assessment of tumor response (tumor-response population)." Ethics oversight HMF discovery cohort The study was approved by the Medical Ethical Committee of the University Medical Center Utrecht and was conducted in accordance with the Declaration of Helsinki. All patients provided written informed consent for collection, analysis and pseudonymized sharing paired tumor-normal whole genome sequencing data and clinical characteristics for research purposes. Patient-derived organoid (PDO) cohort The study was approved by the Medical Ethical Committee of the Netherlands Cancer Institute and was conducted in accordance with the Declaration of Helsinki. All patients provided written informed consent for organoid culture and collection, analysis and pseudonymized sharing of clinical characteristics for research purposes.

Real-world validation cohort
Note that full information on the approval of the study protocol must also be provided in the manuscript.

Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.

Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design
All studies must disclose on these points even when the disclosure is negative.

Sample size
A priori sample size calculations were not performed. Sample sizes were determined by the max amount of patients/samples with available data.
Discovery cohort: All patients in the HMF database who received FTD/TPI as part of their standard-of-care treatment for mCRC were identified in May 2018.
Real-World Cohort (UK and Italy): Clinical pathological and molecular data of all the sequential patient treated with FTD/TPI at 35 centers in Italy and UK between between November 2016 and March 2022 were retrieved accessing electronic patients records (n=1012). 52 patients were excluded from the analysis as explained below, leading the final number of analyzed cases to n=960.

Blinding
Discovery cohort HMF Collection of genomics data and bioinformatics analysis were performed prospectively and hence blinded for clinical outcomes. Collection of clinical outcomes was performed by trained research nurses and medical doctors at the sites of inclusion, who were blinded, as they lacked knowledge about the research question of our study. Data analysis happened in an unblinded fashion.
Real-world Cohort (UK and Italy) Collection of clinical outcomes was performed by trained research nurses and medical doctors at the sites of inclusion, who were blinded, as they lacked knowledge about the research question of our study. Data analysis happened in an unblinded fashion.

RECOURSE trial
The RECOURSE trial was a double-blind study. Details on this process can be found in the original publication (Mayer et al, NEJM 2015). In this re-analysis, data analysis was prespecified in a formal and detailed data request before access to the data was granted, but happened in an unblinded fashion.
In vitro experimentation occurred in an unblinded fashion, as the experimenters generated the appropriate models for their experiments, but given standardized procedures for data readout (automated cell viability readouts and Western blotting) the risk of potential biases was limited.
Reporting for specific materials, systems and methods We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. The following primary antibodies were used: phospho-Histone H2A.X (Ser139) (#05-636) was purchased from Sigma-Aldrich; HSP 90α/β (#sc-13119) was purchased from Santa Cruz Biotechnology.

Eukaryotic cell lines Policy information about cell lines and Sex and Gender in Research
Cell line source(s)