RAS–Pathway Mutation Profiles and Clinical Outcomes Among Patients With Colorectal Cancer

2.1. Study Population

This study integrates patient data and biospecimens from multiple contributing U.S. studies [19], including the Hispanic Colorectal Cancer Study (HCCS) [20], the Puerto Rico Biobank (PRBB) [21], the Colon Cancer Family Registry [22], and the Total Cancer Care (TCC) Protocol at H. Lee Moffitt Cancer Center and Research Institute [23]. In addition, data from The Cancer Genome Atlas (TCGA) was included to enhance the generalizability of findings. All participants provided written informed consent, or a waiver of authorization under the Health Insurance Portability and Accountability Act (HIPAA) was granted by the Institutional Review Board (IRB), as applicable. The study received approval from the IRBs at Cleveland Clinic (central coordinating center) (IRB protocol no. 20–1140), Cedars‐Sinai Medical Center (HCCS) (IRB approval nos. Pro00046506, STUDY00003167, STUDY00001918, and Pro00056359, as applicable), H. Lee Moffitt Cancer Center and Research Institute (TCC and PRBB) (IRB approval no. MCC 20004/IRB 33379), and Ponce Research Institute (PRBB) (IRB approval no. PRI‐IRB 080121‐IF). All procedures were conducted in accordance with established ethical standards and regulations under the U.S. Common Rule.

2.2. Data Collection

Socio‐demographic variables included sex, ethnicity (Hispanic or non‐Hispanic), and race (American Indian, Asian, Black, White, or Other). Race and ethnicity are social, rather than biological, constructs representing shared cultural background, nativity, language, and/or religious practices [8, 24]. Hispanic individuals were classified within the Hispanic ethnic group regardless of their race. We analyzed race and ethnicity as separate variables. For combined race/ethnicity categories, we used both race and ethnicity data; patients missing either were classified as Unknown/Missing. Hispanic versus non‐Hispanic analyses used ethnicity alone, regardless of race availability. These categories were used for descriptive and regression analyses and should not be interpreted as proxies for genetic ancestry. For PRBB participants without self‐reported ethnicity, all patients were classified as Hispanic, and potential misclassification cannot be excluded. Demographic and clinical information were obtained from surveys, medical records, and cancer registry data, with additional variables including age at diagnosis, vital status, tumor site, stage at diagnosis, and microsatellite instability (MSI) status. Age was analyzed as a continuous variable and dichotomized at 50 years for subgroup analyses. Tumor location was categorized as right colon (cecum, ascending colon, hepatic flexure, transverse colon, and ileo‐cecal region), left colon (splenic flexure, descending colon, and sigmoid colon), and rectum (including rectum, rectosigmoid, and rectosigmoid junction). We categorized disease stage as AJCC stage I, II, III, or IV for survival analyses; we grouped stages I/II for early‐stage subgroup analyses, while analyzing stages III and IV separately [25]. Microsatellite instability (MSI) status was determined using MSIsensor, with MSI‐High (MSI‐H) defined as a score ≥ 10 and microsatellite stable (MSS) as a score < 10 [26].

2.3. Sequencing Analysis and Quality Control

Tumor DNA was extracted from either fresh frozen specimens or, more commonly, formalin‐fixed, paraffin‐embedded (FFPE) tissue. Germline DNA was obtained from blood, saliva, or normal colorectal tissue samples. Whole‐exome sequencing (WES) was conducted on tumor samples and their matched normal counterparts, using study‐specific protocols for DNA extraction, library preparation, exome capture, and sequencing [26]. Sequence reads were aligned to the GRCh38 human reference genome using Burrows–Wheeler Aligner (BWA v0.7.17‐r1188) [27]. Output from BWA was converted to BAM files using SAMtools (v1.9) [28]. Duplicate read identification and base quality recalibration were performed using GATK (v4.1.2) [29], and indel realignment was performed using GATK (v3.8.1).

Somatic mutations were identified from matched tumor–normal pairs using a five‐method somatic mutation‐calling pipeline. Single‐nucleotide variants (SNVs) were detected using Mutect (v1.1.7) [30], Strelka (v1.0.15) [31], MuSE (v1.0rc) [32], SomaticSniper (v1.0.5) [33], and FreeBayes (v1.3.1) [34]. Small insertions and deletions were detected using Mutect2 (GATK v4.1.2) [35], Strelka (v1.0.15) [31], Strelka2 (v2.9.10) [36], FreeBayes (v1.3.1), and Lancet (v1.1.0) [37]. The GQ field in the variant call format (VCF) file represented the number of mutation‐calling methods that detected a given variant, and only variants detected by at least three of the five mutation‐calling methods were retained for downstream analyses. To reduce inclusion of common population polymorphisms, variants with an allele frequency greater than or equal to 1% in the 1000 Genomes Project Phase 3 dataset (https://www.internationalgenome.org/) were excluded. Both somatic and germline variants were restricted to the intersection of targeted regions, including padding around intron–exon boundaries. Accurate tumor–normal matching was verified by comparing germline variant calls at common single‐nucleotide polymorphism loci using TimeAttackGenComp [38]. To reduce potential artifact calls, mutation allele frequency skew was evaluated between matched tumor and normal samples. Variants were retained when the tumor‐versus‐normal allele frequency skew was ≥ 2 at positions with sequencing depth ≥ 10 in at least 80% of both tumor and normal samples. Sequencing depth was calculated across targeted regions after excluding duplicate reads. Median on‐target depth was 179× for tumor and 139× for normal samples. By contributing study, median tumor/normal depths were 346×/512× (CC), 311×/383× (HCCS), 298×/143× (PRBB), 226×/139× (TCC), and 110×/110× (TCGA), respectively. Tumor mutation burden (TMB) was calculated as the number of non‐synonymous somatic mutations divided by megabases of target region (38,094,056 bases). TMB distributions were summarized overall and evaluated across contributing studies and by ethnicity.

2.4. Mutation Annotation

Mutations were annotated using ANNOVAR (2017‐07‐17) with the Gencode v30 gene model. Only mutations resulting in a change in protein sequence and located in gene bodies based on the GRCh38 assembly were considered in the association analyses (synonymous, intronic, 5′ UTR, and 3′ UTR mutations were excluded). For lollipop‐style mutation diagrams, only mutations with protein change predictions were presented.

The study cohort comprised 1243 patients with available tumor‐normal sequencing data. Samples were classified as mutant if at least one protein‐truncating, splice, or non‐synonymous coding mutation was detected in the respective gene. We defined RAS mutation status by the presence of mutations in KRAS, NRAS, or HRAS, classifying samples as RAS mutant or RAS wild‐type; RAS‐mutant tumors were further categorized by the specific gene mutated (KRAS, NRAS, or HRAS). Gene‐level frequencies represented unique patients and gene‐specific RAS categories were not mutually exclusive. Patients with multiple mutations could contribute to more than one codon or mutation‐type category. For detailed analysis, the amino acid change and codon position were extracted for KRAS, NRAS, HRAS, BRAF, and PIK3CA, representing key genes in the RAS–RAF–MAPK and PI3K signaling pathways. Mutations were classified into functional categories (Missense, Frameshift, Nonsense, In‐frame deletion, In‐frame insertion, Delins, Splice site, Stop‐loss, or Other) using HGVS nomenclature. As a sensitivity analysis, we also classified variants using ClinVar somatic oncogenicity annotations; only variants classified as oncogenic or likely oncogenic were considered driver‐positive for this sensitivity analysis [39, 40]. For survival analyses, BRAF was evaluated specifically as V600E; non‐V600E BRAF alterations were excluded from BRAF prognostic analyses.

2.5. Statistical Analysis

Patient characteristics were summarized overall and by RAS mutation status. Continuous variables are presented as median (interquartile range [IQR]) and compared using Wilcoxon rank‐sum tests; categorical variables are presented as n (%) and compared using chi‐square or Fisher exact tests, as appropriate. Missing data were retained as separate categories in descriptive tables but excluded from statistical comparisons.

Somatic mutation frequencies were calculated for KRAS, NRAS, HRAS, BRAF, and PIK3CA. Mutation classes and codon‐level distributions were summarized as counts and percentages. RAS pathway and KRAS codon‐specific mutation frequencies were compared descriptively by sex, age at diagnosis (< 50 vs. ≥ 50 years), ethnicity (Hispanic vs. non‐Hispanic), and tumor location (right colon, left colon, and rectum) using chi‐square or Fisher exact tests. Ethnicity‐associated mutation frequencies were additionally evaluated using multivariable logistic regression adjusted for age at diagnosis, sex, tumor location (right colon, left colon, and rectum), and MSI status. Pairwise mutation co‐occurrence was evaluated using Fisher exact tests applied to 2 × 2 contingency tables, with Benjamini‐Hochberg false‐discovery‐rate adjustment. Mutation patterns were visualized using Oncoprint, lollipop plots, heatmaps, and subgroup bar plots.

Overall survival (OS) was defined as time from diagnosis to death from any cause; patients alive at last follow‐up were censored. All survival analyses were restricted to 5 years after diagnosis, with follow‐up administratively censored at 60 months. We estimated median potential follow‐up using the reverse Kaplan–Meier method. Kaplan–Meier curves were compared using log‐rank tests, and five‐year OS estimates with 95% CIs were calculated. Cox proportional hazards models estimated hazard ratios (HRs) and 95% confidence intervals (CIs) for associations of RAS, BRAF V600E, and PIK3CA mutation status with OS within the 5‐year follow‐up period. As a sensitivity analysis, we assessed 5‐year CRC‐specific mortality using competing‐risk and cause‐specific Cox regression analyses, excluding deaths of unknown cause. Multivariable models adjusted for age at diagnosis, sex, disease stage (I, II, III, or IV), MSI status (MSI‐H vs. MSS), year of diagnosis, and primary tumor location. The overall complete‐case multivariable cohort included 822 patients for RAS and PIK3CA analyses; the corresponding BRAF V600E model included 807 patients after exclusion of non‐V600E BRAF alterations. Subgroup models used complete cases within each stratum. Patients with missing or nonpositive survival time or missing vital status were excluded from survival analyses. The proportional hazards assumption was evaluated using Schoenfeld residuals [41]. Stratified analyses by ethnicity and stage were performed to assess potential effect modification. We additionally stratified BRAF V600E survival analyses by MSI status. We also repeated overall‐survival analyses for RAS and PIK3CA using the ClinVar driver‐restricted definition as a sensitivity analysis, with the same covariates as the corresponding broad mutation analyses.

All analyses were performed using R version 4.2.1 (R Foundation for Statistical Computing). Two‐sided p values < 0.05 were considered statistically significant. Benjamini‐Hochberg false‐discovery‐rate adjustment was applied to pairwise mutation co‐occurrence tests and the race/ethnicity mutation‐frequency comparisons presented in Table S2 and Figure S3; other subgroup comparisons in Table 3 were prespecified descriptive analyses and were not multiplicity‐adjusted.

3.1. Cohort Characteristics

The study cohort included 1243 patients with a median age of 65 years (IQR, 54.2–74.0); 592 patients (47.6%) were female (Table 1). Diagnosis years ranged from 1997 to 2023, with a median of 2010 (IQR, 2009–2013); therefore, most were diagnosed before the contemporary immunotherapy era. Overall, 507 patients (40.8%) had tumors harboring at least one RAS mutation. Age at diagnosis, sex, race/ethnicity, stage distribution, observed follow‐up time, and vital status did not differ significantly between patients with and without a RAS mutation. Years of diagnosis differed modestly between the groups (p = 0.04), although the median year of diagnosis was 2010 in both groups. RAS‐mutant tumors were more frequently right‐sided (39.8% vs. 28.4%, p < 0.001) and microsatellite stable (88.2% vs. 77.2%, p < 0.001).

3.2. Mutation Spectrum and Co‐Occurrence Patterns

Across RAS–RAF–MAPK and PI3K pathways, KRAS mutations were the most common, occurring in 439 tumors (35.3%), followed by PIK3CA in 221 (17.8%), BRAF in 107 (8.6%), NRAS in 67 (5.4%), and HRAS in 8 tumors (0.6%) (Table 2; Figure 1). Co‐occurring KRAS and NRAS mutations were rare, observed in 6 tumors (0.5%), and no tumors harbored mutations in all three RAS genes.

Missense alterations dominated the mutation spectrum. KRAS mutations were nearly all missense variants (438 of 439, 99.8%). At the codon level, codon 12 was the predominant hotspot (284, 64.7%), followed by codons 13 (83, 18.9%), 146 (31, 7.1%), 61 (21, 4.8%), and 117 (8, 1.8%). At the variant level, the most frequent amino acid substitution was p.G12D (25.3% of all KRAS mutations), followed by p.G12V (21.0%) and p.G13D (17.8%). NRAS mutations were exclusively missense and concentrated at codons 12 (29 of 67, 43.3%) and 61 (27 of 67, 40.3%), with similar frequencies at these two codons. At the variant level, the most common NRAS mutation was p.G12D, accounting for 25.4% of all NRAS mutations. PIK3CA mutations were primarily missense (212 of 221, 95.9%). At the codon level, mutations were most frequently located at codon 545 (64 of 221, 29.0%) and codon 1047 (35 of 221, 15.8%), followed by codons 542 (19, 8.6%), 546 (15, 6.8%), and 1043 (8, 3.6%). At the variant level, the most common specific amino acid substitutions were p.E545K (26.2% of all PIK3CA mutations) and p.H1047R (13.1% of all PIK3CA mutations). HRAS mutations were rare and were predominantly missense (5 of 8, 62.5%). The most frequent alteration was p.A134T at codon 134, accounting for 25.0% (2 of 8) of all HRAS mutations. None of the detected HRAS variants involved the canonical activating hotspots G12, G13, or Q61, and none met the ClinVar driver‐restricted oncogenicity definition. BRAF mutations were detected in 8.6% of tumors and were also predominantly missense variants (95.3%, 102/107). At the codon level, the transcript‐equivalent codon group 600 was the predominant hotspot (85 of 107, 79.4%). At the variant level, the most common mutation was BRAF p.V600E/p.V640E, representing the same genomic event reported on different transcript isoforms (85 of 107, 79.4%).

Compared with missense mutations, other mutation types were uncommon across the genes examined. Nonsense mutations were identified in KRAS (0.2%), HRAS (37.5%), BRAF (2.8%), and PIK3CA (0.9%), whereas splice‐site variants were observed only in KRAS (0.2%) and BRAF (1.9%). Frameshift mutations were detected only in BRAF (2.8%), in‐frame deletions only in PIK3CA (3.6%), and a stop‐loss mutation only in PIK3CA (0.5%). Table S1 provides the complete mutation spectrum for all five genes.

3.3. Co‐Occurrence of RAS/MAPK and PI3K Pathway Mutations

Most tumors with RAS/MAPK or PI3K pathway alterations harbored a mutation in only one of the examined genes. Pairwise co‐mutations were less common. KRAS and PIK3CA co‐mutations were observed most frequently (126 samples), followed by BRAF and PIK3CA (31 samples). KRAS‐PIK3CA and BRAF‐PIK3CA co‐mutations occurred more often than expected, whereas KRAS‐BRAF and KRAS‐NRAS co‐mutations were uncommon (8 and 6 samples, respectively), consistent with mutual exclusivity (adjusted p < 0.05; Figure 2).

3.4. Mutation Frequency by Clinical Characteristics

In unadjusted comparisons, mutation frequencies varied by selected clinical and demographic characteristics (Table 3; Figure S3; Table S2). By sex, KRAS, NRAS, and PIK3CA mutation frequencies were similar in male and female patients. BRAF V600E was more frequent in female than male patients (10.0% vs. 4.1%; p < 0.001). BRAF V600E was also less frequent in patients with early‐onset disease than in those diagnosed at age ≥ 50 years (1.2% vs. 7.9%; p = 0.002). KRAS mutations showed a nonsignificant trend toward higher frequency in early‐onset patients (39.3% vs. 34.9%, p = 0.27). Median TMB was 2.23 mutations/Mb overall and was similar across contributing studies (CC, 2.22; HCCS, 2.05; PRBB, 2.34; TCC, 2.31; TCGA, 2.23). Median TMB was 2.15 mutations/Mb among Hispanic/Latino patients and 2.39 mutations/Mb among non‐Hispanic patients. Table S2 reports exact counts and denominators for Hispanic/Latino (n = 542), White non‐Hispanic (n = 406), Black non‐Hispanic (n = 58), and Asian non‐Hispanic (n = 10) groups. KRAS mutations were significantly more frequent in right‐sided than left‐sided tumors (44.3% vs. 25.7%; p < 0.001), including enrichment of KRAS G12D in right‐sided disease (10.5% vs. 5.4%; p = 0.01). BRAF V600E mutations were also significantly enriched in right‐sided tumors (15.6% vs. 3.5%; p < 0.001), as were PIK3CA mutations (26.8% vs. 14.6%; p < 0.001). Hispanic patients had lower observed frequencies of KRAS mutations than non‐Hispanic patients (30.6% vs. 40.6%; unadjusted p < 0.001), but this association was attenuated after adjustment for age, sex, tumor location, and MSI status (adjusted OR, 0.76; 95% CI, 0.56–1.02; p = 0.07). For KRAS G12V, the point estimate was marginally significant (adjusted OR, 0.57; 95% CI, 0.33–0.99; p = 0.05), whereas G12D, G12C, and G12A were not significant. PIK3CA frequency among Hispanic patients (14.0% vs. 21.3%; p < 0.01) was also attenuated after adjustment (adjusted OR, 0.75; 95% CI, 0.51–1.09; p = 0.14; Table 3). BRAF V600E remained less frequent among Hispanic patients after adjustment (adjusted OR, 0.40; 95% CI, 0.20–0.78; p < 0.01).

3.5. Survival Analysis by Mutation Status and Ethnicity

Overall survival was evaluable in 1017 patients; Figure S4 shows the derivation of the analytic population. The median potential follow‐up, estimated using the reverse Kaplan–Meier method, was 87 months. All survival analyses were restricted to 5 years after diagnosis, with follow‐up administratively censored at 60 months. RAS mutation status was not significantly associated with 5‐year OS in univariable analysis (HR, 1.26; 95% CI, 0.95–1.67; p = 0.10) or after multivariable adjustment (HR, 1.16; 95% CI, 0.83–1.60; p = 0.38; Table 4). Five‐year OS was 76.3% (95% CI, 71.9%–80.2%) in RAS wild‐type tumors and 73.0% (95% CI, 67.9%–77.5%) in RAS‐mutant tumors.

When stratified by stage, RAS mutation status was not significantly associated with 5‐year OS in stage I/II (adjusted HR, 1.02; 95% CI, 0.49–2.11; p = 0.96), stage III (adjusted HR, 1.17; 95% CI, 0.71–1.93; p = 0.54), or stage IV disease (adjusted HR, 0.99; 95% CI, 0.55–1.75; p = 0.96). In the overall cohort, neither BRAF V600E (adjusted HR, 1.32; 95% CI, 0.62–2.79; p = 0.47) nor PIK3CA mutation status (adjusted HR, 1.06; 95% CI, 0.68–1.63; p = 0.80) was significantly associated with 5‐year OS (Table 4). The overall adjusted analyses included 822 patients for RAS and PIK3CA and 807 patients for BRAF V600E. In an exploratory stage‐specific analysis, BRAF V600E was associated with worse 5‐year OS in stage IV disease (adjusted HR, 4.82; 95% CI, 1.36–17.17; p = 0.02); however, this estimate was based on only five BRAF V600E‐positive patients, and the global BRAF V600E‐by‐stage interaction was not significant (p = 0.14). No significant mutation‐by‐stage or mutation‐by‐ethnicity interactions were observed for RAS, BRAF V600E, or PIK3CA (all p > 0.05; Table 4). Among samples with known MSI status, BRAF V600E was strongly associated with MSI‐H status: 61 of 83 (73.5%) BRAF V600E tumors were MSI‐H, compared with 100 of 1093 (9.1%) BRAF V600E‐negative tumors (Fisher's exact p < 0.001). In MSI‐stratified analyses, BRAF V600E was associated with worse unadjusted 5‐year OS in MSI‐H tumors (HR, 3.31; 95% CI, 1.15–9.55; p = 0.03), but the adjusted estimate was imprecise and not statistically significant (adjusted HR, 2.51; 95% CI, 0.47–13.32; p = 0.28). No significant association was observed in MSS tumors (adjusted HR, 1.15; 95% CI, 0.42–3.16; p = 0.78), and the BRAF V600E × MSI interaction was not significant (p = 0.69; Table S3). Restricting mutation status to ClinVar‐classified oncogenic or likely oncogenic variants did not materially alter survival findings for RAS (adjusted HR, 1.11; 95% CI, 0.80–1.55; p = 0.53) or PIK3CA (adjusted HR, 1.10; 95% CI, 0.67–1.79; p = 0.71; Table S4). The proportional‐hazards assumption was not satisfied for RAS mutation status (Schoenfeld‐residual p < 0.05). In a sensitivity analysis of 5‐year CRC‐specific mortality among patients with classifiable cause of death, no significant associations were observed for RAS (adjusted HR, 1.37; 95% CI, 0.82–2.30; p = 0.23), BRAF V600E (adjusted HR, 2.13; 95% CI, 0.73–6.24; p = 0.17), or PIK3CA (adjusted HR, 0.88; 95% CI, 0.41–1.88; p = 0.74; Table S5).

4.1. Study Limitations

This study has limitations. Some subgroup analyses, particularly those involving HRAS and BRAF mutations, included small numbers of mutated tumors and events and should be interpreted as exploratory. Treatment information was available for a subset of patients but was incomplete and lacked detailed targeted‐therapy information. We included year of diagnosis to partially account for temporal changes in treatment, but residual confounding by treatment exposure cannot be excluded. This is a substantive limitation for the RAS and BRAF findings specifically, as RAS mutation status is a key determinant of eligibility for anti‐EGFR therapy, and management of BRAF V600E‐mutated metastatic CRC evolved during the follow‐up period, including the introduction of encorafenib plus cetuximab after prior therapy [14]. Ethnicity was based on self‐report, medical records, or cancer registry data and may not fully capture genetic ancestry, lifestyle or environmental exposures, or health care access. The contributing study was strongly associated with ethnicity and could not be incorporated into the primary models; residual confounding by other study‐level factors may therefore remain. The RAS Cox model should be interpreted cautiously because the proportional‐hazards assumption was not satisfied. Mechanistic studies are needed to deepen our understanding of the associations between mutation status and clinical outcomes found in this study.

4.2. Conclusions

RAS/MAPK‐ and PI3K‐pathway mutations in CRC varied by tumor location, MSI status, sex, and age at diagnosis. The mutation distributions observed here support comprehensive molecular profiling and provide a reference for biomarker‐driven translational studies and clinical trial design.