Clinical and Functional Characterization of Atypical KRAS/NRAS Mutations in Metastatic Colorectal Cancer

Mutations in KRAS/NRAS (RAS) are important biomarkers in metastatic colorectal cancer (mCRC) that predict lack of benefit from anti-EGFR antibodies and occur in approximately 60% of mCRCs (1–3). Identification of these alterations is important not only to avoid potential toxicity of ineffective therapy, but RAS mutations may also predict worse outcome following anti-EGFR treatment. In the PRIME study, patients with RAS-mutant mCRC treated with panitumumab + FOLFOX4 had a worse median progression-free survival than patients who received FOLFOX4 alone [HR, 1.31; 95% confidence interval (CI), 1.07–1.60; P = 0.008; refs. 4, 5). Due to the strength of RAS mutations as a predictive biomarker, testing for these alterations has become essential prior to treatment with cetuximab or panitumumab (6).

Although the evidence supporting KRAS exon 2 mutations as predictive is robust, with a positive interaction test in placebo-controlled trials, the predictive nature of many of the less common variants remains less clear (2–4, 7). For example, only 7 patients with codon 59 mutations were identified in the PRIME trial that demonstrated extended RAS variants had clinical relevance (4). These mutations were not part of the extended RAS mutation analysis, but rather were assessed in a post hoc analysis that showed removing them from the wild-type population resulted in decreasing the HR in favor of adding anti-EGFR therapy. Even among guideline-cited variants, significant work went into evaluating whether KRAS G13D mutations may still benefit from anti-EGFR therapy based on retrospective evidence (8, 9). This led to the prospective ICECREAM trial which showed a 0% response rate following single-agent cetuximab among KRAS G13D-mutant mCRC (10).

Beyond extended RAS mutations, other “atypical” variants in RAS have been noted, and their clinical relevance remains unclear. As we move beyond hot-spot to full gene coverage, these mutations are increasingly observed. Despite G13D being one of the most common RAS mutations, the ICECREAM study took over 2 years to recruit 53 patients and demonstrated the difficulty in studying rare variants prospectively. If a similar prospective strategy is utilized to validate other less common atypical variants, it is unlikely to succeed, and alternative strategies are required. With this in mind, we aimed to describe the prevalence of atypical RAS variants across a pooled cohort of 9,485 patients and explore their functional and clinical significance using in vitro functional data with cross-platform validation, in vivo mouse experiments, and retrospective clinical data. We hope these data will provide guidance to clinicians treating patients with uncommon alterations, and although this work does not confirm the predictive nature of each individual variant, it provides best available evidence to guide patient care.

This study was completed after receiving Institutional Review Board approval and performed according to the Declaration of Helsinki and following institutional guidance for the care of animals. A waiver of consent was obtained for the retrospective review of patient records.

KRAS exon 2 mutations were defined as KRAS codon 12 and 13 mutations, extended RAS mutations were defined as KRAS codon 59, 61, 117, 146 and NRAS codon 12, 13, 59, 61, 117, 146 mutations, and atypical RAS mutations were defined as all other mutations not currently included in standard guidelines in KRAS/NRAS (6). For patients with more than one mutation in KRAS/NRAS, variants were considered independently for calculating the specific mutations' prevalence. These patients were categorized into KRAS exon 2, extended, or atypical with preference given in descending order from KRAS exon 2 to extended to atypical for subsequent analysis.

The eight cohorts studied are summarized in Table 1 and consisted of seven tissue-based and one cell-free DNA (cfDNA) cohort (Guardant Health). A total of 9,485 patients were assessed. The MD Anderson (MDA) CMS 46, MDA T200, a portion (377/1,078 patients) of the Mayo cohort, and a portion of the Caris Life Sciences cohort (62/2200 patients) had clinical annotation available for comparison of baseline characteristics and clinical outcomes. Mayo patients with clinical outcomes included all RAS-mutant cases treated at the Mayo Clinic, Rochester, MN. Caris Life Sciences patients with clinical outcomes included patients with extended and atypical RAS mutations treated at Georgetown University Medical Center, the West Cancer Center, and Karmanos Cancer Institute.

Microsatellite instability (MSI) status was retrospectively reviewed from patient's charts where access to charts was available and only evaluated in patients with testing performed as part of their standard care. Testing consisted of a mixture of IHC staining for mismatch repair protein deficiency (MLH1, MSH2, PM2, and MSH6) and PCR microsatellite assessment.

Categorical characteristics were compared using χ² test or Fisher exact tests as appropriate, whereas continuous variables were compared with the Mann–Whitney or Kruskal–Wallis tests when a median is reported and the Student t test or ANOVA when averages are shown. P < 0.05 was considered significant for all analyses. Right-sided tumors were defined based on pathology and surgical reports as those occurring from the cecum up to but not including the splenic flexure. Left-sided tumors were defined as those occurring from the splenic flexure to the rectum. Relative variant allele frequency (rVAF) was defined by dividing the allele frequency of a mutation by the maximum allele frequency of any somatic mutation detected in the same sample. Overall survival (OS) was defined as the time from diagnosis with stage IV colorectal cancer until death or last follow-up. Patients alive at the time of last follow-up were censored. OS was summarized using Kaplan–Meier curves and compared using the log-rank test and Cox regression. Where multivariate models were performed, a forward likelihood ratio selection was used. Variables with P < 0.05 were included, and P > 0.1 were excluded during stepwise assessment. All variables met the proportional hazards assumption and were chosen based on differences in baseline characteristics between groups or known prognostic features in colorectal cancer. Age was entered as a dichotomy (<60 vs. ≥60). Analysis was performed using Graph Pad Prism software version 5.0, SPSS version 22.0, and R studio version 3.30.

Functional significance was assessed for all RAS variants (i) detected at MDA among patients who received a CMS 46 next-generation sequencing (NGS) assay for any malignancy, (ii) present in a patient with colorectal cancer in the CARIS Life Sciences Molecular Diagnostics database, or (iii) noted to be of clinical significance or with prior functional annotation in PubMed or COSMIC (11). For example, KRAS P34R is associated with cardiofaciocutaneous syndrome, and KRAS T58I is associated with Noonan syndrome. They have both been well characterized to increase cellular proliferation, decrease KRAS GTPase activity, and stimulate downstream phosphorylation of MEK, serving as reasonable controls for activating atypical alterations (12, 13)

Variants were functionally characterized using an in vitro cell‐based assay (functional annotation for cancer treatment, FACT) designed to analyze oncogenic activity based on signaling activation (Novellus). Variants were generated on a wild‐type expression vector then transfected into a live-cell assay with fluorescently tagged ERK2 that is part of the MAPK/ERK pathway and which shuttles from the cytoplasm to the nucleus upon pathway activation (14). Cells were then fixed and scanned by a fluorescent microscope to detect reporter localization that generated nuclear‐to‐cytoplasmic ratios (NCR) to provide comparisons between signaling activity for each variant. Further details are available in the Supplementary Methods.

NCR values were normalized and scored according to the activation levels of wild-type and KRAS G13D mutations, so that 0% represents wild-type activity and 100% is the activity of the KRAS G13D. This was achieved using standard rescaling methods: score = (MT – KRAS wild-type)/(KRAS G13D - KRAS wild-type), where MT is the reported NCR of the studied mutation, and wild-type is the reported NCR of the wild‐type condition (15, 16). Graphical presentation represents mean and 95% CI.

Analysis comparing OS based on functional activity as defined by the FACT assay was performed using the clinical cohorts and by defining a cut point using histograms demonstrating the distribution of patients with mutations at each activity level (Supplementary Fig. S1). This approach was used as there were very few patients with nonactivating mutations. As such, the cut points still represent variants that have functional activity significantly above wild-type. Cut point selection and sensitivity analysis are outlined in the results section.

In order to validate findings of the Novellus FACT assay, we utilized a Ba/F3 transformation assay. A selection of 17 RAS mutations representing KRAS exon 2, extended, atypical, activating, and nonactivating (per the FACT assay) mutations were assessed using the previously described Ba/F3 transformation assay (17–19). Wild-type cell viability was assessed with four technical repeats, and each mutation was assessed with two technical repeats. Cell viability was compared with wild-type construct using an unpaired t test. Full details of the methodology for the Ba/F3 transformation assay are available in the Supplementary Methods.

We created stable RAS-mutant SW48 cell lines that were used in mouse xenograft models to assess whether functional annotation was concordant with response to cetuximab in animal models. Full details of cell line generation are available in Supplementary Methods. BALB/c nu/nu 6- to 8-week-old athymic nude mice were maintained in the MDA animal facilities under Institutional Animal Care and Use Committee–approved protocols. Wild-type SW48 and mutant KRAS/NRAS-transduced isogenic clones were used to establish xenografts. Approximately 3 × 10⁶ cells/mouse were injected s.c. into right posterior flanks of 5 mice per group. Tumor establishment was monitored twice/week, and when tumor volume reached 100 to 200 mm³, mice were randomized into treatment groups. The treatment groups received either 250 μL saline or cetuximab (0.5 mg/mouse) twice a week via i.p. injection for 21 days at which point mice were euthanized for tumor collection.

The prevalence of missense, nonsense, and indel mutations in RAS across the eight cohorts is demonstrated in Fig. 1. RAS mutations were noted in 4,596 of 9485 patients (48.5%) and varied between cohorts (P < 0.0001) with a range in prevalence from 32.1% to 53.3%. Data from the Project Genie collaboration include patients with a variety of NGS platforms. Given the heterogeneity in techniques used, data are presented for the entire available version 1 Genie cohort (N = 1,879) and for only those patients who utilized assays that would cover all exons of KRAS/NRAS (depicted with a *, N = 1,149). There was no statistically significant difference in RAS mutation frequency (P = 0.65) or distribution of RAS mutation class (P = 0.93) between versions of the cohort.

A total of 3,582 out of 9,485 (37.8%) patients had KRAS exon 2 mutations, 898 of 9,485 (9.5%) had extended RAS mutations, and 116 of 9,485 (1.2%) had atypical RAS mutations. Thirteen of 129 atypical variants (10.1%) occurred in patients who had a co-occurring typical or extended mutation, and they were categorized according to their more common variant, whereas only 37 of 4,518 (0.8%) extended or exon 2 variants occurred in patients with co-occurring RAS alterations (P < 0.0001). Of atypical variants detected, 90 occurred in KRAS and 39 in NRAS. Prevalence of individual variants is summarized in Supplementary Tables S1 and S2, and although most atypical variants occur at low frequencies, certain variants are more common than guideline-cited variants. For example, KRAS Q22K occurred in 16 of 9,485 (0.2%) patients, and yet not a single NRAS codon 117 or 146 variant was detected, and only 1 NRAS codon 59 variant was detected. Other atypical variants occurring at frequencies ≥0.1% include KRAS L19F (7/9,485) and KRAS D33E (7/9,485).

Data on VAF were available in the MDA CM 46, MDA T200, Project Genie, and cfDNA cohorts (N = 5,360). Tissue and cfDNA results are provided separately due to inherent differences in allele frequency distribution (Fig. 1B and C). In both tissue and cfDNA, we noted that KRAS exon 2 and extended RAS mutations had significantly higher VAFs than atypical RAS mutations (P < 0.0001 and P = 0.0015, respectively), even after correcting for tumor content using the rVAF (both P < 0.0001). In addition, atypical mutations demonstrated a more diffuse distribution of clonality, whereas KRAS exon 2 and extended RAS mutations appeared highly clonal.

Among these same cohorts, we assessed prevalence and clonality of concurrent RAS/BRAF V600 mutations. We defined mutations as subclonal if they occurred at rVAF<10%. In tissue, there were 8 concurrent RAS/BRAF V600 mutations, all of which were clonal, and 3 (37.5%) of which were atypical mutations, a significantly higher prevalence of atypical mutation than in BRAF V600 wild-type colorectal cancer (P < 0.0001). In cfDNA, there were 10 concurrent RAS/BRAF V600 mutations, of which 7 were subclonal for both the RAS and BRAF partner. Atypical mutations accounted for 2 of 10 cases (20%), again significantly more prevalent than in BRAF V600 wild-type colorectal cancer (P = 0.037).

We reviewed baseline characteristics and clinical outcomes among 2,581 patients with available clinical data. As seen in Table 2, BRAF V600 mutations were associated with older age than all other groups, whereas KRAS, NRAS, and wild-type cancers did not differ in age distribution (all pairwise P > 0.05). KRAS, NRAS, and BRAF V600 mutations more commonly associated with female gender (all pairwise P < 0.002) and right-sided tumors (all pairwise P < 0.004), whereas BRAF V600 mutations were more commonly associated with MSI-H status (P < 0.0001), but RAS-mutant tumors did not differ from wild-type tumors regarding MSI status (all pairwise P > 0.15). NRAS mutations more commonly co-occurred with additional RAS alterations than KRAS or BRAF V600 alterations (P < 0.0001).

Patients with RAS/BRAF V600 wild-type colorectal cancer had better OS than patients with KRAS (P < 0.0001), NRAS (P < 0.0001), or BRAF V600 mutant tumors (P < 0.0001; Fig. 2A). KRAS mutations were associated with a better prognosis than NRAS (HR, 0.75; 95% CI, 0.58–0.97, P = 0.012) or BRAF V600 mutations (HR, 0.55; 95% CI, 0.43–0.70, P < 0.0001). NRAS mutations were also associated with better prognosis than BRAF V600 mutations (HR, 0.75; 95% CI, 0.57–1.00, P = 0.047). In a multivariate model controlling for age, gender, MSI status, and primary tumor location, KRAS (HR, 1.41; 95% CI, 1.23–1.62), NRAS (HR, 1.83; 95% CI, 1.40–2.40, P < 0.0001), and BRAF V600 (HR, 2.26; 95% CI, 1.75–2.91, P < 0.0001) had worse prognosis than wild-type tumors with only right-sided location remaining significant in the model for controlled covariates (HR, 1.34; 95% CI, 1.17–1.53, P < 0.0001). When directly comparing KRAS with NRAS, there was no statistical difference between genes after controlling for covariates.

Extended RAS mutations were associated with older age than wild-type tumors (P = 0.010), but other RAS mutation classes did not differ (P > 0.14). KRAS exon 2 (P < 0.0001) and extended (P = 0.0018) RAS mutations were more common in women, whereas atypical mutations did not have different sex distribution from wild-type tumors. KRAS exon 2 (P < 0.0001), extended (P < 0.0001), and atypical (P = 0.0022) RAS mutations were more likely to be right sided than wild-type tumors. There was no difference in MSI status based on RAS mutation class (P = 0.25).

KRAS exon 2 (HR, 1.48; 95% CI, 1.32–1.67, P < 0.0001), extended (HR, 1.59; 95% CI, 1.31–1.93, P < 0.0001), and atypical (HR, 2.07; 95% CI, 0.89–4.85, P = 0.014) RAS mutations had worse OS than RAS/BRAF V600 wild-type tumors (Fig. 2B). There was no difference in survival among RAS mutation classes (all pairwise P > 0.28). In a multivariate model controlling for age, sex, MSI status, and primary tumor location, we found that KRAS exon 2 (HR, 1.44; 95% CI, 1.24–1.66, P < 0.0001), extended (HR, 1.46; 95% CI, 1.19–1.78, P < 0.0001), and atypical mutations (HR, 2.27; 95% CI, 1.21–4.27, P < 0.0001) all had worse prognosis than wild-type tumors, but there was no statistically significant difference between class when assessing prognosis among only patients with RAS mutations. In fact, if we included class of mutation and whether the mutation occurred in KRAS or NRAS in the model as two separate covariates, we found that class of mutation was not retained in the model, and only the presence of the mutation in KRAS (HR, 1.41; 95% CI, 1.23–1.62, P < 0.0001) or NRAS (HR, 1.84; 95% CI, 1.40–2.41, P < 0.0001) remained in the model. This suggests that differences in prognosis based on mutation class may be defined more by distribution of these variants across the two genes rather than class; however, a test of interaction was not significant.

We characterized 62 KRAS and 52 NRAS variants using the Novellus FACT assay. As seen in Fig. 3, KRAS and NRAS showed a large dynamic range of functional activity. Values represent relative signaling compared with wild-type (0%) and a known activating control mutation for KRAS (G13D = 100%) and NRAS (Q61R = 100%). Mean activity on this relative scale ranged from 26% below wild-type signaling for a KRAS E76G mutation to 313% for a KRAS Q22K mutation. All exon 2 and extended mutations were more active than wild-type RAS, indicating the utility of the assay (Fig. 3). KRAS alterations had a higher median score relative to NRAS variants (P = 0.0002). Of 57 atypical RAS variants, 18 resulted in signaling below wild-type, 23 had signaling between wild-type and activating control, and 16 were more active than the activating control. KRAS atypical variants (23/31) and NRAS atypical variants (16/26) showed similar ratios of variants that were more activating than wild-type (OR, 1.80; 95% CI, 0.57–5.07, P = 0.39) but atypical KRAS variants were more likely to be hyperactivating with signaling above the KRAS-activating control (15/31) than NRAS variants (1/26; OR, 23.44; 95% CI, 3.45–255.80, P = 0.0002). There was no difference between KRAS and NRAS in terms of the number of atypical variants with signaling below wild-type (P = 0.31).

Using a Ba/F3 transformation assay, we selected a subset of variants for validation from the FACT assay results (19). Variants chosen for validation focused on atypical mutations that were highly prevalent (KRAS L19F, Q22K, and D33E) and a sampling of exon 2, extended, atypical, activating, and inactivating mutations across KRAS and NRAS as comparators. As seen in Fig. 4, 17 of 18 tested variants were concordant between the FACT and Ba/F3 assay. Only KRAS L19F was discordant, lacking transformation activity in the Ba/F3 assay but showing increased signaling in the FACT assay. Importantly, concordance was shown for both the activating alterations in the FACT assay (15/16) and the nonactivating alterations (2/2). Using a large library of previously classified BRAF variants, we also compared protein expression levels using reverse phase protein arrays and the Ba/F3 viability and demonstrate that there is no correlation between relative protein level and cell viability (r = 0.054, P = 0.68), suggesting transformation is due to functional alterations rather than potential differences in protein expression (Supplementary Fig. S2).

Next, we further evaluated the well-characterized KRAS G12D and NRAS Q61K mutations and atypical variants with prevalence of ≥0.1% in Supplementary Tables S1 and S2 using in vivo xenograft models derived from transduced isogenic clones of the SW48 cell line. As none of the atypical variants for NRAS were present at this threshold, we randomly selected one atypical alteration (K135N) that was nonactivating for characterization. As shown in Fig. 5, the SW48 parental and wild-type transduced controls show tumor suppression following cetuximab in xenograft models. Similarly, the well-described KRAS G12D and NRAS Q61K mutant xenograft models show reduced growth suppression with cetuximab. The KRAS and NRAS mutations fell into three classes when mice were treated with cetuximab, those that completely blocked the effects of cetuximab (KRAS G12D, KRAS T50I, and NRAS Q61K), those that were associated with decreased activity of cetuximab (KRAS Q22K and KRAS D33E), and those that resulted in response to cetuximab that was similar to wild-type xenografts (KRAS L19F, KRAS D57N, and NRAS K135N). Among atypical variants assessed, only KRAS T50I showed a lack of tumor growth suppression with cetuximab, whereas KRAS Q22K and D33E showed proliferation that was intermediate between wild-type and known activating mutations. Of tested variants, seven of eight were concordant with the FACT assay if the intermediate category was considered similar to activating in the FACT assay, and eight of eight were concordant with the Ba/F3 assay. KRAS L19F was inactivating by the Ba/F3 assay and xenograft models but was considered activating with the FACT assay.

When we compared xenografts grown in the absence of cetuximab, we noted no difference in tumor volume change when comparing individual mutations with each other and wild-type–transduced controls with a one-way ANOVA with correction for multiple comparisons. We also performed a Western blot assessing RAS expression and a RAS-GTP pulldown of isogenic lines (Supplementary Fig. S3) which demonstrated comparable expression and that all lines were more active than negative control and showed a range of activity that reflected other functional assays except for the NRAS K135N line, which was activating in the pulldown but not in any other assay. Mutations with all four assays are summarized in Supplementary Table S3.

We next assessed the prevalence and clinical impact of FACT assay characterized variants among patients from our epidemiologic cohorts with clinical outcomes. As seen in Supplementary Fig. S1, most patients had a functionally active variant with activity similar to the KRAS G12D–activating mutant. OS did not differ in patients with a mutation occurring above versus below 60% relative activity (HR, 1.38; 95% CI, 0.84–2.27, P = 0.13) but showed trends toward being worse in the less active variants. We reviewed the variants making up this group and noted that all but one of the patients included in the group with RAS mutation activity <60% of control were NRAS, which likely drove this difference in outcome. We chose 60% based on the histogram in Supplementary Fig. S1A, but subsequently recreated histograms split by gene and created new cut points specific to KRAS (80%) and NRAS (60%) and found that neither KRAS (HR, 0.97; 95% CI, 0.83–1.14, P = 0.73) nor NRAS (HR, 1.14; 95% CI, 0.68–1.92, P = 0.60) signaling activity affected OS. Unfortunately, the population with minimally active RAS variants was small so cut points are quite high, and we were unable to determine if atypical variants with functional activity near wild-type had prognosis closer to wild-type patients. When assessing RAS activity as a linear prognostic variable in a Cox-regression model, there was no association with OS in univariate (P = 0.94) or multivariate models (P = 0.41) that controlled for age, sex, primary location, and MSI status.

With increased use of more comprehensive sequencing beyond hot-spot annotation, we increasingly find variants of unknown significance. These mutations pose a challenge for clinicians to interpret. Here, we present a landscape of RAS mutations in colorectal cancer across 9,485 patients with tissue and blood sequencing demonstrating a 1.2% prevalence of atypical RAS mutations and provide functional characterization across 4 orthogonal platforms. We demonstrate that NRAS mutations are associated with worse prognosis than KRAS mutations; however, atypical NRAS variants appear more likely to lack increased functional activity than atypical KRAS alterations. Though some atypical RAS variants do not appear to increase signaling activity (18/57), many resulted in signaling between wild-type and activating control (23/57) and a subset were more activating than known activating mutations, such as KRAS G12D (16/57; Fig. 3). In addition, the majority of highly recurrent atypical variants do activate RAS-dependent signaling. Thus, in addition to having functional annotation, the landscape and prevalence data in Supplementary Tables S1 and S2 provide further information about which variants may have clinical significance as recurrent alterations are more likely to have functional and clinical relevance.

Although we tried to provide a comprehensive annotation of the prevalence and functional significance of as many mutations as possible, there will always be newly discovered uncommon alterations, and it was not pragmatic to perform functional validation on all variants. We focused on recurrent alterations as they affect the most patients and, by nature of the fact that they are recurrent, are more likely to have oncogenic significance. The atypical variants KRAS L19F, Q22K, D33E, and T50I warrant special attention as they occur at frequencies greater than KRAS codon 59 or NRAS codon 117/146 variants which are currently in standard-of-care guidelines for clinical testing (6). Although KRAS L19F increased MAPK signaling in the FACT assay, it did not appear to transform cells in the Ba/F3 assay or xenograft models. Although categorizing variants as “active” or “inactive” is convenient, it overlooks the fact that variants, even in the same gene, may have different effects on cellular signaling and survival that manifest in different ways. Although KRAS L19F increased signaling, the presence of this variant alone may not cause resistance to anti-EGFR therapy or cause transformation in the Ba/F3 assay. Our results mirror prior work that showed increased RAS pathway signaling with L19F, however limited oncogenic transforming potential (20). This variant has not been previously assessed for effects on anti-EGFR therapy, and the observation that our work shows increased signaling but continued response to cetuximab would suggest a potential intermediate phenotype and highlights the benefit of orthogonal functional annotation. Using a single assay in isolation may not detect nuanced responses or differing signaling profiles. The FACT assay is focused on ERK translocation to the nucleus; however, this may not be sufficient to determine that a variant is clinically relevant if taken in isolation. Based on this, KRAS L19F mutant may represent a functional neomorph that retains the aspects of RAS activity assessed in the FACT assay (translocation of ERK to the nucleus) but does not retain aspects of KRAS signaling that may be required to mediate transformation (or cetuximab resistance). This may explain other discordant findings across assays, such as the discordant RAS-GTP pulldown for NRAS K135N suggesting increased signaling, despite the other three assays showing the variant does not cause ERK translocation, transformation, or cetuximab resistance.

This idea that atypical variants need to be considered within the context of other alterations is also seen in atypical BRAF mutations. The BRAF D594F mutation is a kinase-dead variant that can hyperactivate ERK phosphorylation in the presence of concurrent RAS alterations but lacks this activity in RAS wild-type models (21). A recent review of 163 patients with atypical RAS mutations by Pietrantonio and colleagues showed high rates of co-occurring RAS/BRAF and NF1 mutations in 30% and 12% (respectively) of cases with atypical RAS variants (22). Although we had a small number of patients with atypical variants in our clinical cohort, we did note a higher rate of atypical RAS variants co-occurring with BRAF alterations (P < 0.0001); however, this may be spurious due to sample size. We have seen similar findings when reviewing ERBB2 mutations in mCRC, where concurrent PIK3CA mutations were 2–4x as common in tumors with ERBB2 mutations than in tumors with ERBB2 wild-type tumors (23). Even among well-conserved and highly prevalent mutations in RAS, variants may have different evolutionary fitness. Winters and colleagues introduced a library of mutations in KRAS codons 12 and 13 and showed through barcoded sequencing that variants within the same codon had surprising diversity in fitness based on ability to establish tumors (24). These findings all suggest that both individual mutations and the genomic context within which they occur are important to understand when considering targeted therapeutic strategies.

Separately, our finding of worse outcomes in patients with NRAS mutations helps add further support to the prognostic relevance of these mutations in light of mixed prior findings. In the MRC COIN trial, a pooled analysis of five German trials, and an Italian retrospective cohort, there was no difference in outcome between KRAS and NRAS variants (25–27). There were only 38, 39, and 47 NRAS-mutant cancers noted in these studies, respectively, limiting statistical power. Contrary to this, Cercek and colleagues showed in a larger study (N = 87 NRAS mutant tumors) that NRAS mutations showed a trend toward worse survival than KRAS mutations (P = 0.05; ref. 28). Our study was larger than these prior reports and showed a statistically significant better OS in KRAS compared with NRAS alterations (Fig. 2; HR, 0.75; 95% CI, 0.58–0.97, P = 0.012). There are different isoform-specific signaling pathways of KRAS, NRAS, and HRAS that drive differential expression through Raf-1 and Pi3K, and their differing distribution for different cancer histologies suggests these genes affect biology differently (29, 30).

Our study must be interpreted in the context of several limitations. Although we aimed to provide orthogonal validation across Ba/F3 and xenograft models for a selection of RAS variants (Supplementary Table S3), it was not practical to functionally validate all 114 variants assessed with 3 assays as this would have required enormous expenditure beyond our resource capacity. Despite this, positive controls, such as the KRAS P34R and T58I which were chosen due to prior characterization in hereditary cancer syndromes, showed matched activity in the FACT assay, and most atypical variants were highly concordant across platforms. We also acknowledge that although we did see substantial differences in growth following cetuximab noted in the xenograft work performed, the SW48 cell lines using for our isogenic clone generation contain an MEK Q56P mutation that may affect sensitivity to EGFR inhibitors (31). Although the gold standard would be to treat patients with each atypical variant with anti-EGFR agents and assess response in prospective trials, the rarity of these variants makes that impractical. As well, most of our sequencing results came from large tertiary centers which may not be representative. Future efforts to support multi-institutional data sharing from community and tertiary centers could address this gap while also providing a large population to better describe prevalence and clinical implications of uncommon variants. This would help confirm the prognostic and clinical implications of individual atypical variants which was challenging in our study due to small sample size for individual variants.

Despite these limitations, this work demonstrated atypical variants are present in 1.2% of patients with mCRC, a rate higher than previously reported, and provides best available evidence to guide patient care when one of these variants is found (22). We demonstrate that NRAS mutations are associated with a worse prognosis than KRAS alterations and identify KRAS L19F, Q22K, D33E, and T50I as more prevalent than many guideline-included RAS variants. These variants should be considered for testing in patients with mCRC as part of standard care in future testing and alternate non–anti-EGFR antibody treatments should be prioritized where available if a KRAS Q22K, D33E, or T50I variant is identified.

J.M. Loree reports nonfinancial support from Novellus Dx during the conduct of the study, as well as grants and personal fees from Amgen and Ipsen; personal fees from Eisai, Bayer, Pfizer, and Merck; and grants from AstraZeneca outside the submitted work. J. Xiu reports other support from Caris Life Sciences during the conduct of the study. B.A. Weinberg reports personal fees from Taiho, Lilly, Sirtex, Bayer, and Daiichi Sankyo/AstraZeneca, and grants from Ipsen outside the submitted work. A.M. Vanderwalde reports personal fees from Bristol Myers Squibb, Genentech, Caris Life Sciences, Elsevier, George Clinical, Mirati, Pfizer, ConcertAI, Immunocore, and Compugen, and nonfinancial support from AstraZeneca outside the submitted work. A. Tesfaye reports other support from AstraZeneca outside the submitted work. V.M. Raymond is an employee and shareholder of Guardant Health, Inc. B. Miron reports personal fees from NovellusDx during the conduct of the study, as well as personal fees from Seagen outside the submitted work. G. Tarcic reports a patent for US10815477B2 issued to Novellus and is a full-time employee of Fore Biotherapeutics (previously Novellus). O. Zelichov reports personal fees from NovellusDx during the conduct of the study. P.K.S. Ng reports grants from NIH during the conduct of the study. G.B. Mills reports grants, personal fees, nonfinancial support, and other support from Amphista, AstraZeneca, Chrysallis Biotechnology, GSK, ImmunoMET, Ionis, Lilly, PDX Pharmaceuticals, Signalchem Lifesciences, Symphogen, Tarveda, Turbine, and Zentalis Pharmaceuticals; nonfinancial support and other support from Adelson Medical Research Foundation, Breast Cancer Research Foundation, Komen Research Foundation, Ovarian Cancer Research Foundation, Prospect Creek Foundation, and Nanostring Center of Excellence; and other support from HRD assay to Myriad Genetics DSP to Nanostring during the conduct of the study. A. Grothey reports grants and other support from Bayer, Merck, and Genentech/Roche, and other support from Caris and Foundation Medicine during the conduct of the study, as well as grants and other support from Boston Biomedicals and Natera outside the submitted work. J.L. Marshall reports personal fees from Caris Life Science and Indivumed outside the submitted work. S. Kopetz reports personal fees from Roche, Genentech, Merck, Karyopharm Therapeutics, Amal Therapeutics, Navire Pharma, Symphogen, Holy Stone, Biocartis, Amgen, Novartis, Lilly, Boehringer Ingelheim, Boston Biomedical, AstraZeneca/MedImmune, Bayer Health, Pierre Fabre, EMD Serono, Redx Pharma, Jacobio, Natera, Repare Therapeutics, Daiichi Sankyo, Lutris, Pfizer, Ipsen, and HalioDx outside the submitted work. No disclosures were reported by the other authors.

J.M. Loree: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. Y. Wang: Conceptualization, resources, data curation, formal analysis, writing–review and editing. M.A. Syed: Resources, data curation, formal analysis, investigation, visualization, methodology, writing–review and editing. A.V. Sorokin: Resources, data curation, formal analysis, investigation, methodology, writing–review and editing. O. Coker: Resources, formal analysis, investigation, visualization, methodology, writing–review and editing. J. Xiu: Resources, data curation, formal analysis, investigation, writing–review and editing. B.A. Weinberg: Resources, data curation, formal analysis, investigation, writing–review and editing. A.M. Vanderwalde: Resources, data curation, formal analysis, investigation, writing–review and editing. A. Tesfaye: Resources, data curation, formal analysis, investigation, writing–review and editing. V.M. Raymond: Resources, data curation, formal analysis, investigation, methodology, writing–review and editing. B. Miron: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. G. Tarcic: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. O. Zelichov: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. R.R. Broaddus: Conceptualization, resources, data curation, formal analysis, supervision, validation, investigation, visualization, methodology, writing–review and editing. P.K.S. Ng: Resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. K.J. Jeong: Resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. Y.H. Tsang: Resources, data curation, formal analysis, investigation, visualization, methodology, writing–review and editing. G.B. Mills: Resources, data curation, formal analysis, supervision, validation, investigation, visualization, methodology, writing–review and editing. M.J. Overman: Conceptualization, resources, data curation, formal analysis, supervision, validation, investigation, visualization, methodology, writing–review and editing. A. Grothey: Resources, data curation, formal analysis, supervision, investigation, visualization, methodology, writing–review and editing. J.L. Marshall: Conceptualization, resources, data curation, supervision, investigation, methodology, writing–review and editing. S. Kopetz: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing.

J.M. Loree was a member of the UBC Clinician Investigator program, the recipient of a CAMO Fellowship, the ASCO - J. Edward Mahoney Foundation Young Investigator Award, and the RCPSC Detweiler Fellowship. S. Kopetz is the recipient of NIH R01 (CA 1184843 and CA 187238) and CCSG grants (CA016672) that supported this research. We are grateful to the MDA Moon Shots program for research funding that supported this work.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.