Combined Epiregulin and Amphiregulin Expression Levels as a Predictive Biomarker for Panitumumab Therapy Benefit or Lack of Benefit in Patients With RAS Wild-Type Advanced Colorectal Cancer

Importance  RAS wild-type (wt) status is necessary but not sufficient for response to anti-epidermal growth factor receptor (EGFR) agents in advanced colorectal cancer (aCRC). RNA expression of EGFR ligands epiregulin (EREG) and amphiregulin (AREG) may correlate with EGFR-targeted therapy efficacy in aCRC, so may represent a much-needed additional predictive marker for these drugs.

Objective  To examine a novel ligand model in a randomized clinical trial of panitumumab, irinotecan, and ciclosporin in colorectal cancer (PICCOLO) with with the a priori hypothesis that high tumor expression of either AREG or EREG would predict panitumumab therapy benefit in RAS-wt patients; and low expression, lack of efficacy.

Design, Setting, and Participants  Prospectively planned retrospective biomarker study from the PICCOLO trial, which tested the addition of panitumumab to irinotecan therapy in patients with KRAS wt aCRC who experienced failure with prior fluoropyrimidine treatment. The analysis was conducted between 2012 and 2014. A predefined dichotomous model classified tumors as “high expressor” (either EREG or AREG in top tertile for messenger RNA level) or “low expressor” (neither EREG nor AREG in top tertile). Ligand expression was assessed as a prognostic and predictive biomarker. Expression of AREG/EREG and RAS and BRAF mutations were assessed in archival tumor tissue.

Main Outcomes and Measures  Primary end point was progression-free survival (PFS); secondary end points were response rate and overall survival (OS).

Results  Of the 696 PICCOLO trial patients in the irinotecan–vs–irinotecan with panitumumab randomization, 331 had sufficient tumor tissue available and measurement of ligand expression was successful in 323. High ligand expression was not prognostic for OS (hazard ratio [HR], 0.79 [95% CI, 0.58-1.09]; P = .15) or PFS (HR, 0.93 [95% CI, 0.68-1.27]; P = .64). The primary population had RAS wt aCRC (n = 220); for RAS wt patients with high ligand expression, median (interquartile range [IQR]) PFS was 8.3 [4.0-11.0] months (irinotecan with panitumumab) vs 4.4 [2.8-6.7] months (irinotecan alone); HR, 0.38 [95% CI, 0.24-0.61]; P < .001). In RAS wt patients with low ligand expression, median (IQR) PFS was 3.2 [2.7-8.1] months (irinotecan with panitumumab) vs 4.0 [2.7-7.5] months (irinotecan); HR, 0.93 [95% CI, 0.64-1.37]; P = .73; interaction test results were significant [P = .01]). Less marked effects were seen for response rate (interaction P = .17) and OS (interaction P = .11).

Conclusions and Relevance  High ligand expression is a predictive marker for panitumumab therapy benefit on PFS in RAS wt patients; conversely, patients with low ligand expression gained no benefit. The current “opt-in” strategy for anti-EGFR therapy in all patients with RAS wt aCRC should be questioned. Expression of EREG/AREG is a useful biomarker for anti-EGFR therapy; optimization for clinical use is indicated.

Trial Registration  isrctn Identifier: ISRCTN93248876

Clinical trials of the anti–epidermal growth factor receptor (EGFR) therapeutic antibodies panitumumab and cetuximab in colorectal cancer have yielded inconsistent results, varying among trials and patient subpopulations from worthwhile benefit to significant harm.^1,2 Activating mutations in Kirsten rat sarcoma viral oncogene (KRAS) exon 2 consistently confer lack of benefit,^1,3 and extended RAS mutation testing identifies further nonresponders.^4,5 However, an unmutated RAS pathway, although necessary, is not sufficient for response to anti-EGFR agents because many patients whose tumors are wild type (wt) for both KRAS and NRAS do not respond.^2,5

For EGFR blockade to be effective, upregulation of EGFR signaling may be required in addition to an intact signal transduction pathway. The EGFR ligands epiregulin (EREG) and amphiregulin (AREG) are commonly overexpressed in colorectal cancer^6,7 and are highly co-expressed at the transcriptional^7,8 and protein levels.⁶ In cell line models, binding of either ligand to EGFR leads to autocrine activation^9,10; furthermore, knockdown of the AREG or EREG genes significantly abrogates cetuximab-induced reduction in tumor proliferation.⁸ Thus, ligand overexpression may be an indicator of tumor EGFR dependence.^9,10 This provides the scientific rationale to investigate EREG/AREG as an additional predictive biomarker for anti-EGFR therapies.

High messenger RNA (mRNA) expression of EREG or AREG has been reported to correlate with the efficacy of EGFR-targeted agents in nonrandomized series of KRAS wt advanced colorectal cancer (aCRC)^7,11-13 and in preclinical studies in squamous cancers.^14,15 In biomarker analysis of the CO.17 trial (supportive care with or without cetuximab), KRAS wt patients with high-EREG RNA expression had improved survival with cetuximab therapy, while KRAS wt, low-EREG patients did not.¹⁶ However, a positive biomarker/treatment interaction in KRAS wt patients was not convincingly demonstrated.

Thus, while there are signals that either AREG or EREG may identify patients who will or will not benefit from anti-EGFR therapy, studies to date have fallen short of validating their role or providing a much-needed clinically applicable patient selection strategy. To move the field forward, we herein comprehensively examine the interactions between ligand expression and treatment benefit in a large randomized trial of irinotecan, alone or with panitumumab, as second-line treatment of aCRC.² Because AREG and EREG are commonly but not consistently co-expressed, either may mediate EGFR activation, and both have shown association with anti-EGFR efficacy in previous work,^7,11-13,16 we elected a priori to combine data from both ligands as a single putative predictive biomarker in an “either/or” model. We hypothesized that the benefit of panitumumab therapy would be confined to patients with both wt RAS (KRAS c.12,13,61 and NRAS c.12,13,61 wt) and high expression of either AREG or EREG. Conversely, RAS wt patients with low ligand expression would derive no benefit from the addition of panitumumab to chemotherapy. Secondary analyses addressed single-ligand effects and interactions with other mutations in the EGFR signal transduction pathway and primary tumor location (PTL).

Box Section Ref ID

The randomized clinical trial of panitumumab, irinotecan, and ciclosporin in colorectal cancer (PICCOLO) (ISRCTN93248876) has been reported previously.^2,17 Following United Kingdom national ethics committee review and approval, 1196 consenting patients entered, including 696 randomized to irinotecan vs irinotecan with panitumumab. From June 2008, this randomization was restricted to patients with KRAS c.12,13,61 wt cancers. The University of Leeds Clinical Trials Research Unit managed the trial, overseen by an independent trial steering committee.

Written consent was obtained for the retrieval of stored tumor blocks, and ethical approval was obtained for translational studies. The present study includes all available patients with adequate stored pathological material, recruited both before and after randomization was restricted to KRAS c.12,13,61 wt patients. Tumor tissue used was from the primary tumor collected prior to first-line chemotherapy for aCRC. This translational study was performed between August 2012 and February 2014.

Genotyping for KRAS c.12,13,61, KRAS c.146, NRAS c.12,13,61, BRAF V600E, and PIK3CA c.542,545-6,1047 was previously performed.² RNA was extracted from formalin-fixed, paraffin-embedded (FFPE) tissue sections (6 sections at 5 μm) (RNeasy FFPE kit; Qiagen). The expression of AREG and EREG was measured by reverse transcription polymerase chain reaction (RT-PCR) (eMethods and eTable 1 in the Supplement). Laboratory staff were blind to the patients’ treatment allocation, mutation profile, and clinical outcomes.

Stata was used for all statistical analyses (Stata Statistical Software, Release 12 [2011]; StataCorp). Baseline patient characteristics were compared between treatment arms using 2-tailed t tests, Wilcoxon rank sum tests (for variables with nonnormally distributed frequency distributions), and Pearson χ² tests (for categorical variables). Patient characteristics were compared with the whole trial population using the same tests.

Box plots were produced for raw AREG and EREG expression, and their correlation estimated using the Spearman coefficient. Data from both were combined to give a clinically usable single dichotomous classifier by taking the upper/middle tertile cut point for each ligand and dividing the population into “high expressors” (either EREG or AREG in top tertile) or “low expressors” (neither EREG nor AREG in top tertile). This cutoff was chosen pragmatically, to give high and low groups of similar size.

Three clinical end points were used: The primary end point was progression-free survival (PFS); secondary end points were overall survival (OS) and Response Evaluation Criteria in Solid Tumors (RECIST) response rate (RR). Progression-free survival and RR data were unchanged from the primary trial analysis, but updated 2-year OS data were used in this analysis.

Ligand expression was first assessed as a prognostic marker in all patients treated with irinotecan alone, both using the dichotomous classifier (high expressors vs low expressors) and using each ligand separately as a continuous variable (log-transformed to base 2), in Cox proportional hazards models.

Ligand expression was then assessed as a predictive marker for panitumumab therapy benefit by testing for interaction between the effects of expressor status (high vs low) and treatment (irinotecan with panitumumab vs irinotecan alone) on PFS and OS using the likelihood ratio test. The concordance probability was calculated using the Harrell C index. Adjustment was performed for significant prognostic factors in the trial population (World Health Organization performance status [WHO PS], response to previous therapy). Response to previous therapy was unknown in 37 patients and multiple imputation was used to impute values for these 37 patients. Multiple logistic regression was performed using “previous oxaliplatin therapy” and “previous chemotherapy” as predictors of previous response based on 20 imputed data sets. Secondary analysis of predictive effects was performed in patients with RAS or BRAF mutations. Relative risks were estimated from generalized linear models (with a log link) for the RR outcome. BRAF mutation and PTL (right colon vs left colon or rectum) were identified as possible confounding factors^2,18; therefore, survival models were estimated for the joint effects of BRAF and ligand, then PTL status and ligand, for the dichotomous classifier and each ligand separately.

Of the 696 PICCOLO trial patients in the irinotecan–vs–irinotecan with panitumumab randomization, 331 had sufficient tumor tissue available and measurement of ligand expression was successful in 323. Baseline characteristics were well balanced (Table 1), and no significant differences existed between the biomarker and main trial population. Survival data were available for all patients. Of these, 299 had a disease progression event (92.6%). In updated data used for OS analysis, 303 (93.8%) patients had died.

Complete KRAS c.12,13,61,146, NRAS c.12,13,61, BRAF V600E, and PIK3CA c.542,545-6,1047 genotypes were available in all cases. A total of 220 of 323 (68.1%) patients were wt across all KRAS and NRAS codons (“RAS wt”). This is a higher proportion than seen in an unselected aCRC population and reflects the fact that from June 2008 patients with KRAS c.12,13,61 mutations were not recruited to this randomization, thus enriching the population for KRAS c.12,13,61 wt patients. Of the 220 RAS wt patients, 47 (21.4%) had a BRAF mutation, also reflecting the study’s enrichment for KRAS wt patients (Figure 1).

Nineteen (8.6%) RAS wt patients had a PIK3CA mutation. There were no interactions between PIK3CA mutations and the prognostic and predictive role of ligands and these results are not further reported here.

Both ligands showed skewed distributions of RNA expression and were log-transformed for further analyses. Epiregulin and amphiregulin were highly co-expressed (Spearman correlation coefficient, 0.78; P < .001). Using the combined dichotomous classifier, 140 of 323 (43.3%) patients were “high expressors” (either ligand in the top tertile), and 183 of 323 (56.7%) “low expressors” (neither ligand in the top tertile) (Figure 2).

Ligand expression was not significantly associated with RAS mutation status (median log AREG RAS wt = −3.18 vs RAS mutated = −3.19, median log EREG RAS wt = −2.78 vs RAS mutated = −2.91; Wilcoxon rank sum tests, P = .41 and P = .31, respectively), but both ligands were significantly higher in BRAF wt than BRAF-mutated groups (P < .001, both comparisons). Both ligands were significantly higher in patients with PTL on the left than on the right side (P < .001, both comparisons) (eFigure 1 in the Supplement).

In patients treated with irinotecan alone, we saw no evidence for a prognostic effect of high vs low expressor status on PFS (hazard ratio [HR], 0.93 [95% CI, 0.68-1.27]; P = .64) or OS (HR, 0.79 [95% CI, 0.58-1.09]; P = .15) (eTable 2 in the Supplement).

The primary hypothesis was that high ligand expression would be predictive of the benefit of panitumumab on PFS in RAS wt patients. This hypothesis was supported by the data, as shown in Table 2 and Figure 3. In RAS wt patients with high ligand expression, panitumumab therapy had a significant effect on PFS (median, 8.3 [interquartile range {IQR}, 4.0-11.0] months for irinotecan with panitumumab vs 4.4 [IQR, 2.8-6.7] months for irinotecan alone; HR, 0.38 [95% CI, 0.24-0.61]; P < .001). Conversely, panitumumab therapy was ineffective in RAS wt patients with low ligand expression (median PFS, 3.2 [IQR, 2.7-8.1] months for irinotecan with panitumumab vs 4.0 [IQR, 2.7-7.5] months for irinotecan alone; HR, 0.93 [95% CI, 0.64-1.37]; P = .73). The ligand-treatment interaction was significant whether unadjusted (P = .01, Harrell C index 0.57) or after adjustment for other prognostic factors (P = .009, adjusted for WHO PS and response to previous therapy [imputed for 37 cases]; data not shown). The results were similar when including missing data for previous response as a separate category (P = .01, C index 0.59) or when excluding the missing data (P = .004, C index 0.61). The predictive effect of the ligand model on OS was less marked and was not statistically significant (interaction P = .11) (Table 2).

In RAS wt patients with high ligand expression, the RR was 47.7% (irinotecan with panitumumab) vs 11.3% (irinotecan) (relative risk, 4.22 [95% CI, 1.87-9.52]; P = .001), while in low ligand expressors RR was 18.9% (irinotecan with panitumumab) vs 10.5% (irinotecan) (relative risk, 1.81 [95% CI, 0.74-4.43]; P = .20) (interaction P = .17).

Within the RAS and BRAF wt subpopulation, the low ligand expressors did not gain benefit with irinotecan with panitumumab therapy (HR, 0.70 [95% CI, 0.43-1.16]; P = .17) compared with irinotecan alone, but the ligand-treatment interaction was not significant (interaction P = .16) (Table 2).

Patients in the RAS-mutated or BRAF-mutated groups gained no panitumumab therapy benefit by any end point, regardless of ligand status (Table 2).

As a secondary analysis, AREG and EREG were examined separately as continuous variables. In prognostic analyses, EREG was prognostic for OS (HR, 0.87 [95% CI, 0.80-0.94]; P = .001) but not for PFS (HR, 0.94 [95% CI, 0.86-1.02]; P = .16), while AREG was not prognostic for either (OS HR, 0.94 [95% CI, 0.85-1.03]; P = .18; PFS HR, 0.97 [95% CI, 0.87-1.07]; P = .50) (eTable 2 in the Supplement). Conversely, in the predictive analysis in RAS wt patients, AREG was predictive for the effect of panitumumab therapy on PFS (interaction P = .008), but results were less clear for EREG (interaction P = .08). Neither EREG nor AREG was predictive of panitumumab therapy OS benefit (eTable 3 in the Supplement).

BRAF mutations and PTL on the right side were associated with low ligand expression (eFigure 1 in the Supplement). We therefore explored whether the differential treatment effects by ligand level seen for PFS in the RAS wt patients were driven by these factors (eTable 4 in the Supplement). The dichotomous classifier remained a significant predictor of panitumumab therapy benefit after adjustment for BRAF (interaction P = .005) and for PTL (interaction P = .01). Similarly, continuous AREG remained significant following adjustment for BRAF (interaction P = .008) and for PTL (interaction P = .02), although, as before adjustment, EREG was not. Thus, the ligand-treatment effect appears to be independent of these potential confounders.

The dichotomous ligand model was decided a priori on the basis of 2 decisions: to combine data from both ligands because either can activate EGFR, and to use the middle/upper tertile boundary for each, so generating similarly sized high and low patient groups.

For exploratory purposes, we looked separately at patients classified on the basis of both ligands, or just one falling within the top tertile (eFigure 2 in the Supplement). Patients with both ligands in the top tertile had marked PFS benefit from panitumumab therapy (HR, 0.28 [95% CI, 0.15-0.53]; P < .001, n = 56). Among the 43 patients with just one ligand in the top tertile, the PFS benefit from panitumumab therapy seemed to be less, but with a wide confidence interval (HR, 0.60 [95% CI, 0.30-1.22]; P = .16).

We examined the choice of the middle/upper tertile boundary by reanalysis using the 50th, 80th, and 90th centiles (eTable 5 in the Supplement). None of these alternative cut points provided greater separation of benefiting/nonbenefiting populations, as determined from the P values for interaction and the HRs.

Continuous AREG and EREG, when analyzed separately, each acted as a significant predictive biomarker for panitumumab therapy benefit on PFS in RAS wt patients (P < .001 for both) (eTable 3 in the Supplement). When both were then entered into a multivariable model, only log AREG retained significance (AREG HR, 0.84 [95% CI, 0.71-0.99]; P = .04; EREG HR, 0.95 [95% CI, 0.83-1.08]; P = .43), suggesting that if just one ligand were to be selected for development, it should be AREG. We observed marked benefit in the high-AREG group (HR, 0.30 [95% CI, 0.17- 0.52]; P < .001, n = 76). However, this would mean withholding treatment from the small group of patients with isolated high EREG, for whom treatment benefit is uncertain (HR, 0.82 [95% CI, 0.32- 2.09]; P = .68, n = 23, data not shown).

This study confirms for the first time the utility of EREG and AREG mRNA expression as a predictive biomarker for anti-EGFR therapy in RAS wt patients, and proposes a simple model—either high vs neither high—to combine information from both ligands. Using this model, RAS wt patients with neither high gained no PFS benefit from the addition of panitumumab to chemotherapy, while those with either high gained marked benefit.

Although predictive, the combined ligand model was not significantly prognostic. However, consistent with previous findings,^11,12,19,20 EREG alone was prognostic for improved survival.

There are many requirements for development of a predictive biomarker for clinical application. They include a sound scientific basis, a validated assay method, optimized cut points, a model to combine data from different biomarkers, and retesting in prospective hypothesis-led studies in large randomized data sets powered to prove the biomarker-treatment interaction. The prospective hypothesis, randomized design, and model optimization included in this study move us closer to clinical application.

In recent analysis from the CO.17 trial, KRAS wt patients with high EREG expression had significant benefit with cetuximab therapy over supportive care alone (OS HR, 0.46, P < .001; PFS HR, 0.33, P < .001); conversely, no benefit was seen in KRAS wt low-EREG expressors (OS HR, 0.93, P = .81; PFS HR, 0.7, P = .21); the unadjusted tests for interaction were not statistically significant (OS P = .08; PFS P = .07).¹⁶ This study was exploratory in nature: It tested multiple EREG models but without AREG, did not correct for confounding factors, and included KRAS-mutated patients in the main analysis; however, it provides important corroboration for the findings reported here.

One novel aspect of this study was the a priori decision to combine data from both AREG and EREG in the simple either high vs neither high model, while others have measured only one, or have entered both into a multivariable model. There is preclinical evidence that either ligand can activate EGFR.^21-23 However, they are highly co-expressed so are affected by multicollinearity if treated competitively in a multivariable model. In previous studies using multivariable models, EREG outperformed AREG in some^7,11,16,19 while AREG outperformed EREG in others.^12,24 Thus, selection of a single ligand would be a poor basis for clinical decision making, leading to wrong decisions in that minority of patients whose ligands are not co-expressed. In our data set, AREG was the better predictive biomarker if taken in isolation; however, we propose continuing development of our combined ligand model incorporating data from both ligands because this better satisfies the scientific hypothesis, preclinical data, and totality of clinical data now available.

Amphiregulin and epiregulin are distributed continuously over a wide range, with no natural cut point. We chose the upper/middle tertile boundaries for a pragmatic reason: to give similar numbers of patients in the high (either/or) or low (neither/nor) populations for analysis. However, our exploratory analysis of other cut points suggests that fortuitously this may be the best discriminator in terms of strength of marker-treatment interaction (P = .01). A lower threshold (50th centile) would have assigned more patients to the high category while still retaining PFS benefit (interaction P = .02), an important consideration for its use as a potential negative predictive marker.

We elected not to test other known ligands of EGFR (EGF, TGF-α, HB-EGF) because they have not previously been consistently linked with anti-EGFR efficacy.^7,12,13 The hypothesis-based approach focused on AREG/EREG, avoiding multiple testing, retained adequate power to detect a ligand-treatment interaction.

We considered the effect of BRAF and PTL status on the ligand-treatment predictive model because both interact with anti-EGFR agent effect.^2,18 In PICCOLO, BRAF was a negative predictive marker for panitumumab therapy benefit in KRAS wt patients,² but PTL was not.²⁵ As expected, ligand expression varied significantly according to BRAF and PTL status; however, an independent predictive effect of ligands was demonstrated.

Importantly, there is now evidence of the utility of ligands in the first-line,¹⁹ third-line,¹⁶ and now second-line treatment of aCRC, and with both cetuximab^16,19and now panitumumab. Additionally, high expression of ligands correlates with anti-EGFR benefit in preclinical studies in squamous cancers.²⁶ Thus, there is strong impetus to further develop AREG and EREG into a clinically applicable test.

The further development of AREG and EREG for clinical decision making now requires clinical consensus. In health care systems in which all RAS wt patients currently receive anti-EGFR therapy, AREG and EREG will be of interest as a negative selection biomarker; consensus is therefore required on the level of certainly for lack of benefit in low expressors to justify withholding these drugs. Conversely, where anti-EGFR use is currently restricted, AREG and EREG will be of interest as for positive selection, so the degree of benefit required to justify treatment must be quantified. These subtly different scenarios require different positive and negative predictive values, which can then be calculated using independent data sets to establish the optimum cut point for clinical use.

The translation of a quantitative RNA biomarker into routine care presents special challenges. This study used routinely stored FFPE tumor material and validates the RT-PCR technique for the measurement of EREG and AREG mRNA expression.¹¹ However, it was necessary to extract both DNA (for RAS and other mutations) and mRNA (for AREG and EREG assays). This requires sufficient preserved tissue and incurs costs. However, given the mean drug acquisition cost of panitumumab of approximately $8000 per month for a 60-kg patient, it is highly likely that use of 2 RT-PCR assays costing approximately $100 to $300 for all RAS wt patients would be cost-effective. A protein-based ligand biomarker assay such as immunohistochemical analysis would be an attractive alternative to mRNA measurement, but as yet there is only preliminary evidence that ligand protein expression correlates with anti-EGFR treatment activity.²⁷

There were limitations of this study. Sufficient archived tissue was available for only 331 (47.6%) of the 696 patients in the irinotecan with panitumumab vs irinotecan alone randomization in PICCOLO. Reassuringly, the demographic characteristics and outcomes in the study population were consistent with the main trial analysis. As in other aCRC second-line trials of EGFR therapies,^28,29 the marked PFS benefit seen with panitumumab therapy in KRAS wt patients in PICCOLO did not translate into improved OS, driven in part by shorter survival after progression for patients who had received panitumumab.² Progression-free survival was therefore chosen as the primary end point of this study; and it is perhaps no surprise that even though RAS wt high-ligand–expressing patients had a marked improvement in PFS with panitumumab, this did not translate into longer OS.

Although responses to the irinotecan with panitumumab combination therapy were commoner in patients with high rather than low ligand expression (47.7% vs 18.9%), the low rates of RECIST response events in this second-line setting mean that this comparison is underpowered, and the interaction test was not statistically significant for this secondary end point.

This study serves as clinical validation of AREG and EREG mRNA as a predictive marker for anti-EGFR agents. Further development is now required to bring this potentially important biomarker rapidly into routine clinical use in colorectal cancer, and to test it in other cancers treated with anti-EGFR agents. With alternative new options emerging for patients with aCRC, it becomes ever more important to ensure that the use of anti-EGFR therapy is confined to those who will benefit from it, so avoiding the costs and lost opportunities of futile treatment. This is particularly pertinent with ongoing controversy regarding the optimal first-line treatment of RAS wt aCRC patients.^30,31 We propose this combined model as a clinically usable tool and would urge further revalidation using existing and new randomized trial biobanks.

Accepted for Publication: December 8, 2015.

Corresponding Author: Matthew T. Seymour, MD, St James’s Institute of Oncology, St James’s University Hospital, Beckett St, Leeds LS9 7FT, UK (matt.seymour@nihr.ac.uk).

Published Online: February 11, 2016. doi:10.1001/jamaoncol.2015.6065.

Author Contributions: Dr Seymour had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Drs Quirke and Seymour served as co–senior authors.

Study concept and design: Seligmann, Quirke, Seymour.

Acquisition, analysis, or interpretation of data: All authors.

Drafting of the manuscript: Seligmann, Elliott, Jacobs, Barrett, Seymour.

Critical revision of the manuscript for important intellectual content: Seligmann, Richman, Jacobs, Hemmings, Brown, Barrett, Tejpar, Quirke, Seymour.

Statistical analysis: Seligmann, Elliott, Jacobs, Barrett, Seymour.

Obtained funding: Seligmann, Quirke, Seymour.

Administrative, technical, or material support: Richman, Jacobs, Hemmings, Tejpar.

Study supervision: Barrett, Quirke, Seymour.

Conflict of Interest Disclosures: Dr Quirke receives research funding from Amgen and Affymetrix and travel expenses from Amgen and Ventana. Dr Seymour receives research funding from Amgen and Integragen. No other disclosures are reported.

Funding/Support: This study was supported by Cancer Research UK (CRUK) and Yorkshire Cancer Research. Dr Tejpar is supported by the Fund for Scientific Research Flanders (FWO-Vlaanderen). The PICCOLO trial was developed through the National Cancer Research Institute Colorectal Clinical Studies Group, funded by CRUK and supported by an unrestricted educational grant from Amgen Inc. It was conducted within the UK National Health Service with the support of the National Institute for Health Research (NIHR) Cancer Research Network. Molecular testing for RAS and BRAF mutations was supported by Amgen, NIHR/CRUK Experimental Cancer Medicine Centre, and Yorkshire Cancer Research.

Role of the Funder/Sponsor: No funding organization had any role in design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Additional Contributions: We are deeply indebted to the patients who participated in this study and to their families and carers. We would like to thank all staff, past and present, at the University of Leeds Clinical Trials Unit. We are grateful to staff at all the centers that contributed to the PICCOLO trial, including clinical management of patients, data collection, and contribution of tumor blocks to the PICCOLO trial biobank. The following participating centers contributed tumor blocks: Addenbrookes, Cambridge (PI: Charles Wilson); Airedale, Keighley (PI Michael Crawford); Bradford Royal Infirmary (PI: Andy Conn); Bristol (PI: Stephen Falk); Huddersfield & Halifax (PI: Jo Dent); Charing Cross & Hammersmith, London (PI: Charles Lowdell); Cheltenham & Gloucester (PI: Kim Benstead); Clatterbridge (PI: Sun Myint); Edinburgh Cancer Centre, Western General, Edinburgh (PI: Lesley Dawson); Hereford County (PI: Nick Reed); James Cook, Middlesbrough (PI: Nicholas Wadd); Kent and Canterbury Hospitals (PI: Catherine Harper-Wynne); Mount Vernon Hospital, Middlesex (PI: Rob Glynne-Jones); New Cross Hospital (PI: Mark Churn); Nottingham City (PI: Vanessa Potter); Peterborough (PI: Karen McAdam); Poole & Bournemouth (PI: Tamas Hickish); Burton (PI: Prabir Chakraborti); Royal Free Hospital, London (PI: Astrid Mayer); Royal Marsden (PI: Ian Chau); Leeds (PI: Matt Seymour); South Tyneside (PI: Ashraf Azzabi); St Luke’s Guildford (PI: Gary Middleton); St Mary’s, London (PI: Susan Cleator); St Mary’s/QA Portsmouth (PI: Ann O'Callaghan); St Thomas’s/QE London (PI: Nick Maisey); Swindon (PI: Claire Blesing); Torbay (PI: Nangi Lo); UCL/North Middlesex/Harlow (PI: John Bridgewater); Coventry & Warwickshire (PI: Robert Grieve); Velindre, Cardiff (PI: Tim Maughan); Wansbeck, Northumberland (PI: Werner Dobrowsky); West Middlesex (PI: Pippa Riddle); Weston Park, Sheffield (PI: Jonathan Wadsley); Worthing/West Sussex/Royal Sussex (PI: Andrew Webb); Yeovil, Somerset (PI: Clare Barlow); Ysbyty Gwynedd, Bangor (PI: Catherine Bale); Ysbyty Maelor & Glan Clwyd (PI: Simon Gollins). Molecular testing was carried out with the assistance of Morag Taylor, BSc, and Philip Chambers, PhD, Leeds Institute of Cancer and Pathology, University of Leeds. All the contributors to the study made their contributions within their regular employment roles and none were additionally compensated for their roles in this study.