Impact of KRAS Mutations and Co-mutations on Clinical Outcomes in Pancreatic Ductal Adenocarcinoma

The relevance of KRAS mutation alleles to clinical outcome remains inconclusive in pancreatic adenocarcinoma (PDAC). We conducted a retrospective study of 803 PDAC patients (42% with metastatic disease) at MD Anderson Cancer Center. Overall survival (OS) analysis demonstrated that KRAS mutation status and subtypes were prognostic (p<0.001). Relative to patients with KRAS wildtype tumors (median OS 38 months), patients with KRASG12R had a similar OS (median 34 months), while patients with KRASQ61 and KRASG12D mutated tumors had shorter OS (median 20 months [HR: 1.9, 95% CI 1.2–3.0, p=0.006] and 22 months [HR: 1.7, 95% CI 1.3–2.3, p<0.001], respectively). There was enrichment of KRASG12D mutation in metastatic tumors (34% vs 24%, OR: 1.7, 95% CI 1.2–2.4, p=0.001) and enrichment of KRASG12R in well and moderately differentiated tumors (14% vs 9%, OR: 1.7, 95% CI 1.05–2.99, p=0.04). Similar findings were observed in the external validation cohort (PanCAN’s Know Your Tumor® dataset, n=408).


Introduction
Pancreatic ductal adenocarcinoma (PDAC) is projected to be the second leading cause of cancer death in US by 2040; with limited available treatment options for metastatic PDAC, the 5-year survival rate is less than 5% 1,2 . The median overall survival (OS) for the current standard of care chemotherapy (oxaliplatin, irinotecan, fluorouracil, and leucovorin [FOLFIRINOX]) is 11.1 months in the first line treatment of metastatic disease, with an objective response rate (ORR) of 31.6% and median progression-free survival (PFS) of 6.4 months 3,4 . The median OS for the other available first line chemotherapy regimen, gemcitabine/nab-paclitaxel, is 8.5 months with an ORR of 23% and median PFS of 5.5 months 5 . In the setting of second line treatment, the median OS with chemotherapy (liposomal irinotecan, fluorouracil and leucovorin) is only 6.1 months, with an ORR of 16% and median PFS of 3.1 months 6 . Better therapy for PDAC is urgently needed.
Among the identified genomic alterations (GAs) in PDAC, oncogenic KRAS mutations are the most common, occurring in close to 90% of patients, followed by TP53, CDKN2A, and SMAD4 7,8 . The majority of KRAS mutations are at codon 12, with the highest prevalence of G12D mutation (35%), followed by G12V (20-30%), G12R (10-20%), Q61H (~5%), 1%-2% G12C and other rare muations [9][10][11][12] . Targeting KRAS has been challenging for decades until the allosteric KRAS G12C mutant-specific inhibition by covalent binding to the mutant cysteine beneath the switch-II region, which locks it in the inactive GDP bound form, has been discovered 13 . Exciting results from clinical trials of the KRAS G12C inhibitors sotorasib (AMG510) and adagrasib (MRTX849) have been reported, and both have been approved by the US FDA for previously treated KRAS G12C -mutated advanced lung cancer. Moreover, efficacy of both sotorasib and adagrasib against PDAC has also been observed [14][15][16][17][18] . Sotorasib had a 21% ORR with a median PFS of 4.0 months in patients with pancreatic cancer who had received chemotherapy previously 19 . Adagrasib monotherapy had an ORR of 33.3% with a median PFS of 5.4 months (95% CI 3.9-8.2) and a median OS of 8.0 months ( 95% CI 5. 2-11.8) in pancreatic cancer patients refractory to chemotherapy (n=21) 20 . More excitingly, preclinical development of a KRAS G12D inhibitor (MRTX 1133) has shown promising results and MRTX 1133 is currently in phase 1 clinical trial 21 . Pan KRAS inhibitor RMC-6236, which binds to the chaperone protein, cyclophilin A, and active GTP-bound RAS (RAS ON inhibitor) are also being tested in patients with KRAS G12 mutations, including G12D, G12V, G12R, G12A, or G12S mutations (NCT05379985). Moreover, T cell therapy with KRAS G12D -targeting T cell receptors (TCRs) caused tumor regression in a pancreatic cancer patient, and T cells with TCRs targeting other KRAS mutations, such KRAS G12V , are under development 22,23 . We are at a breakthrough point in attempts to target KRAS in pancreatic cancer. The remaining challenges include the short duration of response and primary/secondary resistance to KRAS inhibition. Additionally, while multiple genomic and non-genomic factors have been associated with resistance to KRAS inhibitors, such as co-mutations of KEAP1/STK11 with KRAS as observed in patients with lung cancer, comutations KEAP1/STK11 mutations are rare in pancreatic cancer; little is known about the landscape of KRAS mutations and co-mutations in pancreatic cancer or their impact on clinical outcomes 12,24,25 .
KRAS-mutated cancers are heterogeneous with different mutation allele subtypes and co-mutations [26][27][28] . Each KRAS mutation allele subtype has unique biochemical and clinicopathological features, and the differences between the mutation subtypes and commutations in pancreatic cancer have not been well studied [26][27][28][29] . The KRAS G12D mutation has an intrinsic wildtype and SOS1 guanine exchange activities while the KRAS Q61 mutation has deficiencies in GTP hydrolysis 27,30 . The KRAS G12R mutation, which accounts for approximately 15% in pancreatic cancer but less than 1% in lung cancer, was reported to be associated with different downstream signaling pathways from other KRAS mutations 27 . The KRAS G12D mutation was reported to be more immune suppressive with shorter survival in lung cancer and pancreatic cancer 31,32 . Moreover, it has been reported that the mutations that co-mutations with KRAS vary with the KRAS mutation alleles in patients with lung cancer, and these different patterns of co-mutation with KRAS differentially affect clinical outcomes 33 . For example, co-mutation of KEAP1/STK11 was more common in patients with KRAS G13 -mutated lung cancer than KRAS G12D -mutated lung cancer, and the co-mutation of KEAP1/STK11 with KRAS G13 was associated with poor prognosis and treatment resistance 28 .
Research to date on the impact of KRAS allele subtypes and co-mutations on clinical outcome in PDAC has been limited, and the conclusions remain controversial. KRAS G12D was reported to be associated with worse OS compared that among with patients with KRAS G12R -mutated pancreatic cancer in a single institutional study (n=126); however, within the KRAS G12R -mutated PDAC group, those with co-occurring PI3K pathway mutations experienced worse OS 34 . Meanwhile, there was no statistically significant difference in OS between different KRAS mutation alleles in another study 12 . Our institution has collaborated with the data science firm Syntropy to deploy the Palantir Foundry software platform for extraction and analysis of merged clinical data and laboratory data across a variety of platforms, including the Electronic Health Record (EHR); molecular testing/next generation sequencing (NGS), pathology, and radiology results; and tumor registry data [35][36][37] . Together with the development of data science tools such as natural language processing (NLP) and the increased use of NGS in pancreatic cancer, the Foundry platform now gives us the ability to analyze large datasets comprising real-world clinical and molecular information to dissect the heterogeneity of KRASmutated pancreatic cancer. In this study, we illustrate the co-mutation landscape of KRAS mutations and the allele specific differences of KRAS-mutated pancreatic cancer with clinical outcome in our institution. In addition, we validated our findings in an external cohort from the Pancreatic Cancer Action Network (PanCAN)'s Know Your Tumor® (KYT) Dataset 38 .

Patient Characteristics
A total of 803 patients with PDAC who had tumor tissue somatic mutation testing at MD Anderson were identified (Fig 1); the demographic and clinical characteristics of this cohort are summarized in Table 1. The median age was 63 years (range 26-86), 43% were female, and 29.3% had a smoking history (current or former). A total of 336 (42%) patients had documented stage IV disease at the time of their initial diagnosis, and 321 (40%) had poorly differentiated tumors. KRAS gene mutation status was tested in 703 patients, including 302 with stage IV disease; 578 (82%) were positive for mutated KRAS. In addition to KRAS, TP53 was tested in 604 patients, 418 (69%) of whom were positive; TP52. CDKN2A was tested in 509 patients, 102 (20%) of whom were positive; and SMAD4 was tested in 536 patients, 68 (13%) of whom were positive. The median followup time from initial diagnosis was 41 months. Median OS of the entire cohort of 803 patients at MD Anderson was 19 months (range 0.07-348).
In patients with metastatic disease and known KRAS, TP53, and CDKN2A mutation status (n=232), we classified four distinct molecular subtypes of metastatic PDAC

PanCAN's Know Your Tumor® Dataset
To validate our findings, an external cohort from PanCAN's KYT dataset (n=408) was analyzed. Baseline characteristics of patients in the KYT cohort are summarized in Table  2. The median age at the time of diagnosis was 65 years (range 36-88). 46% were female and 54% were male. The median follow-up time from diagnosis was 15 months. Disease staging information was not available in majority of the patients in this cohort (59.8%). 23.8% (n=97) patients had documented stage IV disease at the time of diagnosis. Median overall survival in all the patients was 22 months (range 0.2-93 months). KRAS (92%), TP53 (77%), SMAD4 (24%), CDKN2A (21%), and ARID1A (5%) were the most commonly mutated genes in the PanCAN cohort (Fig 6A).

Discussion
In this study, we analyzed the impact of KRAS mutation status, KRAS allele subtypes, and co-occurring mutations on clinical outcome of patients with PDAC in two real-world datasets. The study included 803 patients who had been tested for somatic tumor mutations at MD Anderson Cancer Center and an external cohort (n=408 of patients with pancreatic cancer from the PanCAN KYT® dataset. We found that KRAS mutation status and allele subtypes were associated with OS; median OS was longer in patients with KRAS wildtype and KRAS G12R -mutated tumors compared to median OS in patients with KRAS G12D or KRAS Q61 -mutated tumors. We illustrated the co-mutation landscape with KRAS mutation. We also found that ARID1A mutation was associated with worse OS and SMAD4 was associated with better OS. We found TP53 and ATM mutation were mutually exclusive. There was a higher rate of ARID1A mutation in KRAS G12D compared with KRAS G12R patients. We also found enrichment of KRAS G12D in metastatic disease and enrichment of KRAS G12R in well to moderately differentiated tumors. Among the 803 patients with PDAC analyzed, 703 were tested for KRAS mutation at MD Anderson (Fig 1). The overall positive rate for KRAS mutation was 82% (n=578) with the most common mutation of KRAS G12D (39%), followed by KRAS G12V (31%), KRAS G12R (14%), KRAS Q61 (6%), and other uncommon KRAS variants (9%) (Fig 2D). There were differences in OS with KRAS mutation status and allele subtypes in the overall population (Fig 2A) and stage IV patients who had tested for KRAS (n=302) (Fig 2B).  (Fig 2C). The external cohort of PanCAN KYT® dataset (n=408) validated that KRAS G12R was associated with best median OS (32 months), while KRAS Q61 (16 months, HR: 2.6, 95% CI 0.88-7.8, p=0.02) and KRAS G12D (23 months, HR: 1.68, 95% CI 1.06-2.65, p=0.04) were associated with shorter median OS (Fig 7B). Our results were consistent with the previous report of significantly longer OS (HR 0.55) in patients with KRAS G12R -mutated PDAC (n=23) compared with those with non-KRAS G12R PDAC (n=88) 34 . Another study comparing KRAS G12C (n=30) and other KRAS mutations reported longer median OS (p=0.03) for KRAS wildtype patients (n=91) which was consistent with our findings of better survival in KRAS wildtype patients 12 . The previously reported study analyzed the OS of metastatic PDAC from starting the first line therapy and did not show statistically significant difference between other KRAS alleles while compared against KRAS G12C patients 12 . Due to the low frequency of KRAS G12C mutation, we grouped the KRAS G12C with other uncommon mutations. In our cohort, OS was defined from initial diagnosis and there was enrichment of KRAS G12D mutation in metastatic disease (stage IV) (OR: 1.7, 95% CI 1.2-2.4, p=0.001) (Fig 3C). Our data suggested worse outcome of KRAS G12D tumors. This is consistent with the study of 356 resected patients with PDAC, which reported that KRAS mutation had worse diseasefree survival (DFS) (median 12.3 months) and OS (median 20.3 months) compared with wildtype (DFS 16.2 months and OS 38.6 months) and poor outcome in KRAS G12D patients (median OS 15.3 months) 39 . The mechanisms of why KRAS G12D had worse prognosis is not fully understood beyond the comutations. More immunosuppressive tumor microenvironment (TME) with KRAS G12D tumors was found in lung cancer 28,31 . In KRAS G12D mutation driven PDAC mice model, there were immune suppressive cytokines IL-4 and IL-13 and remodeling of myeloid cell composition in TME 40,41 . In PDAC mice models treated with the KRAS G12D inhibitor MRTX1133, increased macrophages (CD11b and F4/80+) in the TME with decreased the total myeloid cells was observed 42 . Correlative tissue and blood samples for potential KRAS mutation allele specific immune features were not included in this project and could be a future direction in patients with PDAC.
KRAS G12R existed most commonly in PDAC (~15%) with low frequency in other cancer types 12 . It has distinct biochemical features from KRAS G12D/V with altered switch-II structure that could not activate p110α/PI3K directly 43 . We found the median OS of KRAS G12R mutated patients was comparable to wildtype patients and longer than KRAS G12D or KRAS Q61 mutated patients. There was enrichment of KRAS G12R mutation in well and moderately differentiated tumors vs poorly differentiated/anaplastic tumors (OR: 1.7, 95% CI 1.05-2.99, p=0.04) (Fig 3D), which suggested less aggressive biology and better outcome for the KRAS G12R -mutated tumors. On the other hand, KRAS Q61 mutants had decreased GTP hydrolysis rate with high RAF-dependent MEK phosphorylation and it did not response to SOS1 inhibition 29,44 . While KRAS Q61 mutants had shorter median OS in our cohort, little is known about the clinical features of KRAS Q61 mutants. To our best knowledge, this is the first study to report worse OS with KRAS Q61 which could be consistent with its biochemical features. Due to the rarity of KRAS Q61 mutations, we grouped different KRAS Q61 mutations together, which may be mutant specific 45 . The clinical and molecular features of KRAS G12R and KRAS Q61 mutated PDAC warrant further and larger studies which could help the development of KRAS allele specific inhibitors such as the KRAS G12R inhibitor 46 .
Co-mutations with KRAS could be one of the contributing factors for the allele specific clinical outcome in PDAC. KEAP1 co-mutation with KRAS in lung cancer was associated with early progression on KRAS G12C inhibitor sotorasib 25 . Co-occurrence of other mutations in PDAC were common and the disease progression model was proposed with early KRAS mutation followed by CDKN2A then loss of TP53 and SMAD4 47,48 . Our data was consistent with previous reports that TP53 (67%) was the most common co-mutation with KRAS followed by CDKN2A (17%), SMAD4 (11%), and ARID1A (6%) (Fig 4B) 12 . We tested the KRAS/CDKN2A/TP53 disease progression model by classified four distinct molecular subtypes of metastatic patients in patients who had tested for KRAS, TP53, and CDKN2A (n=232). We found patients with triple negative (KRAS-/TP53-/CDKN2A-) tumors demonstrated better OS (median 28 months) with CDKN2A predominant type had the worst OS (median OS12 months, p=0.014) (Fig 5C). In our study, CDKN2A mutation included any mutation either missense or deletion of CDKN2A. Germline CDKN2A mutation increased the risks of melanoma and pancreatic cancer and somatic CDKN2A loss was common in pancreatic cancer [49][50][51] . CDKN2A loss had worse survival (median DFS 11.5 and OS 19.7) in patients with resected PDAC compared with wildtype patients (median DFS 14.8 and median OS 24.6) 39 . In another study of 100 patients with both metastatic and nonmetastatic PDAC, CDKN2A mutations were also associated with shorter OS (22 months vs 35 months; P = 0.018) 52 . In KRAS-mutated lung cancer, CDKN2A mutation was associated with worse survival on imunotherapy 53 . In mice model, CDKN2A loss accelerated KRAS G12D driven tumor 54 . Targeting CDKN2A in KRASmutated PDAC is under investigation yet clinical activities of CDK4/6 inhibitors in early phase trials was not seen 55,56 . The location of the methylthioadenosine phosphorylase gene (MTAP) is adjacent to CDKN2A and majority of PDAC with CDKN2A loss also had MTAP loss [57][58][59] . The surrogate role of CDKN2A is not clear with low reported rate of MTAP loss in our cohort, due to the detection method for MTAP loss is not currently validated by comparative genomic hybridization for pancreatic cancer in our NGS testing panel.
Univariate OS analysis in our study did not show statistically significant association of comutation with TP53 or CDKN2A mutation with OS but revealed that ARID1A mutant was associated with poor OS (median 18 vs 31 months, HR: 1.6, 95% CI 0.99-2.6, p=0.05) and SMAD4 mutant was associated with better OS (median 35 vs 27 months, HR: 0.67, 95% CI 0.46-0.99, p=0.046) (Fig 5B). SMAD4 is a tumor suppressor gene and inconsistent results were reported about the prognostic value of SMAD4 39,60-62 . SMAD4 inactivation in resected PDAC was associated with poor prognosis while the metaanalysis did not show its association with OS 61,62 . Our data showed 13% SMAD4 mutation rate and it was associated with better OS. Further studies with larger sample size and different populations are needed to understand the different results. ARID1A was found to be significantly co-mutated with CDKN2A (OR: 2.7, 95% CI 1.18-6.02, FDR-corrected p=0.095), and with SMARCA4 (OR: 5.17, 95% CI 1.15-18.44, FDR-corrected p=0.1). KRAS G12R mutated patients had lower ARID1A compared with KRAS G12D (0% vs 8% in p=0.02) (Fig 4B). Similar findings were also observed in the validation cohort of the PanCAN KYT® dataset. Both ARID1A and SMARCA4 are Switch/Sucrose Nonfermentable (SWI/SNF) chromatin remodeling complex genes which are important in epigenetic reprogramming in PDAC 63 . Context specific tumor suppressive or oncogenic function of SWI/SNF chromatin regulation was noticed in PDAC 64,65 . In mice models, disrupted ARID1A promoted the carcinogenesis from KRAS-mutated premalignant intraductal papillary mucinous neoplasms (IPMN) to PDAC 44 . In KRAS-mutated colon cancer, a similar tumor supporting role of ARID1A was required for MEK/ERK signaling 66 . Our results of worse OS with ARID1A support the oncogenic role of ARID1A and targeting ARID1A in PDAC. ARID1A regulates DNA damage checkpoints and sensitizes cells to DNA damage response (DDR) targeting agents [67][68][69] . ATM-TP53 signaling pathway is critical in DDR targeting in pancreatic cancer 70 . Interestingly, in both of our cohorts, TP53 mutation was mutually exclusive with ATM mutation. TP53 and ATM mutation exclusivity was reported in mantle cell lymphoma with distinct clinical impact and sensitive to the protein arginine methyltransferase 5 (PRMT5) inhibitor 71,72 . Our findings of worse OS with ARID1A mutation and mutual exclusivity of TP53 and ATM mutation in PDAC provided insights on therapeutic vulnerabilities of PDAC.
In summary, we reported the KRAS mutation allele specific clinical outcome in PDAC using a single institution retrospective study and an external validate cohort. Our findings suggested that KRAS targeting, and combination strategies may warrant the mutant allele specific approaches with consideration of the co-occurring mutations with KRAS.

Limitations
The limitations of this study are heterogeneities in both patient populations and tumor mutation testing methods and gene panels. Only patients who had tissue molecular testing done at MD Anderson were included in this study while patients who had tested by other panels were not included. It is a retrospective study in a single tertiary cancer institution with ascertainment bias. The external validation cohort had limited clinical information. Treatment information was not available. Tumor genomic factors may not be the main contributor for KRAS mutation allele specificities. Correlative tissue and blood samples from patients for other non-genomic factors accounting for the differences in clinical outcome are not included in this study.

Conclusion
In our analysis of 803 patients with PDAC, we found that KRAS mutation status and mutation allele subtypes were associated with OS. Patients with KRAS wildtype and KRAS G12R -mutated tumors survived longer than patients with KRAS G12D or KRAS Q61 -mutated tumors, and this observation was confirmed in an external validation cohort. We also found enrichment of KRAS G12D mutations in patients with metastatic disease and KRAS G12R mutations in patients with well to moderately differentiated tumors. We found co-mutations could contribute to the KRAS allele specific outcome. We found worse OS in ARID1A mutated patients and lower co-mutation rate of ARID1A in KRAS G12R . Our findings of different clinical outcomes by KRAS mutation subtypes and co-mutations status suggest allele-and co-mutation-specific impact of KRAS mutations on pancreatic cancer outcome and provide guidance in improving approaches to target KRAS in pancreatic cancer.

Patients and Methods
The MD Anderson Cancer Center Institutional Review Board (IRB) approved the collection of demographics, clinical, and pathological information under IRB protocol 09-0373 and 2023-0091. Informed consent was waived, as per the IRB guidelines for retrospective studies of previously collected clinical and molecular information. The Palantir Foundry software system (Palantir, Denver, CO) was used to query the MD Anderson EHR to identify patients with a confirmed diagnosis of PDAC who underwent somatic tumor tissue mutation testing at MD Anderson from 3/14/1997 to 4/27/2023 for inclusion in the study.
Patient demographics, histopathology, tumor grade, surgical history, and mutational profiles were collected from the MD Anderson EHR and tumor registry data using the Foundry system. Histologic classification and grade were collected from the patients' pathology reports. Molecular testing was performed at MD Anderson's molecular diagnostics laboratory, which is College of American Pathologists (CAP) accredited and Clinical Laboratory Improvement Amendments (CLIA) certified. The gene panels used evolved during the study inclusion period, with expanding lists of genes over time. The Page 13 of 28 information on tumor genomic alterations (GAs) was extracted from the available clinical and molecular data. Deidentified information was used for analysis.
For the co-mutation analysis, only patients who were tested with multigene panels were included (n=513). The Oncoplot function within MAFtools was used to visualize the somatic mutation distribution. The function performs pair-wise Fisher's exact test to uncover mutually exclusive or co-occurring gene sets and an FDR -corrected p<0.1 was considered significant. To better understand the co-mutations patterns with KRAS and the rest of the genes, a heatmap was constructed to demonstrate the co-mutation landscape of KRAS mutation status, as well as the status of the different KRAS alleles, and the rest of the genes (Fig 4B). The percentage of co-occurrence between KRAS alleles and pathogenic mutations in the genes listed in the heatmap were determined using in-house R scripts. Fisher's exact test was used to test for significance in cooccurrence between KRAS alleles and pathogenic mutations. Based on the co-mutation patterns observed, we divided patients into 4 molecularly distinct PDAC co-mutation subtypes to visualize and test the relationship between co-mutation pattern and OS.

Statistical Analysis
Differences in disease stage and tumor grade between patients with different KRAS mutations were assessed using Chi-square and Fischer's exact test. Overall survival (OS, from the time of initial diagnosis) was calculated from the date of initial diagnosis until death or last known contact. OS curves were estimated using the Kaplan-Meier method, and the difference in survival curves was tested using the log-rank test. Univariate Cox proportional hazards models were used to estimate hazard ratios (HRs) and test the associations of KRAS mutation status, KRAS mutation allele subtypes, and other driver mutations with OS.
In the co-mutation analysis, the somatic interactions function within MAFtools was used to detect mutually exclusive or co-occurring mutation events. Pair-wise Fisher's exact test was used to uncover mutually exclusive or co-occurring gene sets with Benjamini-Hochberg multiplicity correction, and a false discovery rate (FDR)-corrected p<0.1 was considered significant. The OS curves for the 4 co-mutation subtypes were estimated with the Kaplan-Meier method and compared using the log-rank test.
GraphPad Prism version 9 (GraphPad Software, San Diego, California USA) and Rstudio 2020 (RStudio, PBC. Boston, MA) were used for the statistical analyses and data visualization 73 . All tests were two-sided, and statistical significance was identified by a pvalue < 0.05.

PanCAN's Know Your Tumor® Program and Dataset
PanCAN, in partnership with Tempus (Tempus Labs Inc., Chicago, IL), offers the Know Your Tumor® (KYT) precision medicine service to patients with pancreatic cancer. KYT data is available through the PanCAN SPARK platform (www.pancan.org/spark). Tempus processes, sequences and conducts group-level bioinformatics analyses on tumor biopsy samples. Data is derived from the Tempus xT NGS panel that covers 648 genes with actionable oncologic mutations. Variants are called from the resulting alignment files using an analysis pipeline that detects SNPs and indels using Freebayes and Pindel 74,75 . A filtered variant file which contains biologically relevant DNA variants, as determined by the Tempus pipeline, were used for all KYT related analyses. Patients with PDAC who had their tumor sequenced by Tempus were included in the analysis. Pathogenic or likely pathogenic mutations were determined by Tempus' proprietary Knowledge Database which is based on the American College of Medical Genetics and Genomics (ACMG) and Association for Molecular Pathology (AMP) guidelines for variant classification. All mutation data was converted to Mutation Annotation Format (MAF) to enable use of the functions in the Bioconductor R package, MAFtools 76 . The Oncoplot function within MAFtools was used to visualize the somatic mutation distribution across the KYT cohort. The somatic interactions function within MAFtools was used to detect mutually exclusive or co-occurring mutation events. The function performs pair-wise Fisher's exact test to uncover mutually exclusive or co-occurring gene sets with Benjamini-Hochberg multiplicity correctio and an FDR-corrected p<0.1 was considered significant. The percentage of co-occurrence between KRAS alleles and pathogenic mutations in the genes listed in the heatmap in Figure 6 were determined using in-house R scripts. Fisher's exact test was used to test for significance in co-occurrence between KRAS alleles and other pathogenic mutations. Overall survival (OS, from the time of initial diagnosis) was calculated from the date of initial diagnosis until death or last known contact. OS curves by KRAS mutation and subtype status were estimated using the Kaplan-Meier method, and the difference in survival curves was tested using the log-rank test.