Plasma Protein Biomarkers in Advanced or Metastatic Colorectal Cancer Patients Receiving Chemotherapy With Bevacizumab or Cetuximab: Results from CALGB 80405 (Alliance)

Colorectal cancer is a leading cause of cancer mortality in the United States. Significant progress has been made in the past 20 years in both detection and treatment of colorectal cancer (1). In 1990s, fluorouracil became the only cytotoxic drug in colorectal cancer as continuous infusion of fluorouracil improved the overall survival (OS) from 12 to 15 months (2). Since then, combination therapies of fluorouracil with leucovorin and either irinotecan [FOLFIRI regimen (3)] or oxaliplatin [FOLFOX regimen (4)] became the first-line chemotherapeutic treatment for colorectal cancer. In 2000s, targeted therapies were added to chemotherapy, including mAbs targeting EGFR and VEGF (5, 6).

Two anti-EGFR antibodies, panitumumab and cetuximab, have been approved by FDA to treat patients with metastatic CRC (mCRC) as monotherapies and combined with chemotherapy (7, 8). EGFR binding to its ligands results in the activation of the mitogenic MAPK/ERK signaling cascade via RAS GTPases (9). Patients whose tumors harbor activating RAS mutations do not respond to EGFR-targeting agents (10–12). Current guidelines recommend the use of panitumumab and cetuximab only for patients with CRC with wild-type RAS (13).

There are currently four FDA approved biological agents for targeting the VEGF-pathway in patients with colorectal cancer: first-line bevacizumab which binds VEGF-A (5); second-line ramucirumab which binds VEGFR-2 (14), and ziv-aflibercept, which binds VEGFs and PlGF (15); and third-line regorafenib which blocks activity of all VEGF receptors and other tyrosine kinases (16, 17). Extensive effort has been made to identify biomarkers to guide anti-VEGF drugs, including markers at the molecular, cellular, and tissue levels (18).

Circulating protein biomarkers have gained increasing attention given the advantages of ease to obtain, cost-effectiveness, and amenability to repeated sampling. We have developed and optimized the Angiome, a panel of circulating protein biomarkers and analyzed their predictive potential in multiple trials involving thousands of patients in different disease settings (19–22). Ang-2, SDF-1, and VEGF-D plasma levels were identified as predictive biomarkers for bevacizumab in patients with pancreatic cancer treated with gemcitabine ± bevacizumab in a study conducted by Cancer and Leukemia Group B (CALGB), now part of the Alliance for Clinical Trials in Oncology cooperative group, CALGB 80303 (20). Benefit from bevacizumab was associated with high levels of Ang-2 and SDF-1 and low levels of VEGF-D. The VEGF-D effect was strongest in patients in the lowest quartile of plasma VEGF-D (20). Consistently, negative predictive effect of VEGF-D has been reported in patients with mCRC treated in the Australian Gastro-Intestinal Trials Group testing Mitomycin, Avastin, and Xeloda (AGITG MAX trial; ref. 23). Despite a number of promising candidates including VEGF-D, currently there are no validated biomarkers to guide the applications of anti-angiogenic agents in any disease settings (24).

CALGB and the Southwest Oncology Group (SWOG), conducted the intergroup study, CALGB/SWOG 80405, to better guide clinical decision making for patients with advanced or mCRC (25). CALGB 80405 was initially designed to compare bevacizumab, cetuximab, and the combination of both agents in addition to FOLFIRI or FOLFOX chemotherapy regimen chosen by the physician and patient. Early in the course of the study, the combination regimen was discontinued due to results from the CAIRO2 and PACCE studies showing increased toxicity and a lack of increased survival benefit in the arms with dual VEGF and EGFR inhibition (26, 27). CALGB80405 was negative for overall survival (OS), the primary endpoint for this study. The clinical results of CALGB 80405 showed similar progression-free survival (PFS) and OS benefits from either bevacizumab-chemotherapy or cetuximab-chemotherapy groups (25).

To improve outcomes in future clinical trials of bevacizumab in mCRC, a homogenous selected patient population would be pivotal. Disease parameters and circulating biomarkers would greatly facilitate patient selection to enrich patient deriving the most benefit from anti-angiogenic drugs. As stated earlier, our Angiome analysis has successfully identified prognostic and predictive biomarkers for anti-angiogenic agents, bevacizumab included. Although prognostic biomarkers reflect how well a patient perform, independent of treatments; predictive biomarkers have the potential to identify patients benefitting the most from a specific treatment.

In this study, we assessed the Angiome expression in plasma samples collected at baseline in the primary cohort of CALGB 80405. These patients habored KRAS wild-type tumors, received either bevacizumab or cetuximab in addition to FOLFIRI or FOLFOX. Angiome analysis provided a snapshot of tumor angiogenesis, inflammation, and immune status at baseline. Potential prognostic and predictive biomarkers were identified.

The study design of CALGB 80405 and clinical results have been reported (25) and detailed information can be found at ClinicalTrials.gov (NCT00265850). Briefly, patients with locally advanced or mCRC were randomized to receive either bevacizumab, cetuximab, or the combination of both in conjunction with chemotherapy (FOLFIRI or FOLFOX) at the discretion of the treating physician. The study was later amended to stop enrollment in the combination arm and to implement KRAS mutation testing. The samples used in this study were collected at baseline from the primary cohort of CALGB 80405. The primary cohort consists of 1137 KRAS wild-type patients randomized to the bevacizumab + chemotherapy or cetuximab + chemotherapy treatment groups. This correlative study of CALGB 80405 was approved by the Institutional Review Board at all participating centers, adhered to Guidelines for Good Clinical Practice and guiding principles laid out in the Declaration of Helsinki. Written informed consent was obtained from patients who opted to participate in this analysis.

Peripheral venous blood was collected at baseline from consenting patients into vacutainers containing EDTA anticoagulant. Samples were centrifuged at 2,500 g for 15 minutes within 30 minutes of collection. The resulting plasma was spun again at 2,500 g for 15 minutes. Double-spun, platelet-poor plasma was aliquoted, frozen, and shipped on dry ice to the Alliance Pathology Coordinating Office for centralized storage. Prior to analyses, samples were shipped to Duke Phase I Biomarker Laboratory, thawed once on ice, re-aliquoted, and stored at −80°C until assayed.

Levels of 24 soluble protein biomarkers were measured in duplicates with the personnel blinded to the clinical results. Plasma samples from patients in the primary cohort were analyzed using the multiplex ELISA techniques on the Quanterix and Meso Scale Discovery platforms as described previously (20, 28). The analysis conforms to the guidelines established by the REMARK criteria.

All biomarker levels were log-transformed prior to analysis, to de-emphasize outliers and better approximate a normal distribution. All analyses were performed using baseline data from patients in the primary cohort with continuous values for the protein analytes. Biomarkers prognostic of clinical outcome [overall survival (OS) or progression-free survival (PFS)] were determined using univariate Cox proportional hazards models independent of treatment arm, and the resulting HR, 95% confidence intervals (CI), and asymptotic P values based on the score test were reported. Multivariable Cox regression models were used to test for interactions between biomarker levels and treatment arms to identify biomarkers predictive of benefit from bevacizumab or cetuximab and asymptotic P values based on the Wald test for the interaction terms were reported. All analyses were stratified by type of chemotherapy (FOLFIRI or FOLFOX) to control for any differences in baseline survival rate among these two groups. Biomarker levels were also tested for interaction with treatment arm separately within each chemotherapy subpopulation. The reported P values were not adjusted for multiple testing.

Hierarchical agglomerative clustering, using Euclidean distance and complete linkage, was used to identify groupings among protein biomarkers based on their baseline levels. The results were illustrated using a dendrogram. Forest plots were created as illustrations of prognostic effect sizes (HRs and corresponding 95% CIs). For the purposes of illustrating interaction effects, biomarker levels were dichotomized as “high” or “low” relative to the median, and Kaplan–Meier plots of OS and PFS were created, with separate curves for each combination of treatment group and dichotomized biomarker level. It has previously been shown that low levels of VEGF-D (first quartile) were predictive of benefit from bevacizumab in advanced pancreatic cancer and colon cancer, respectively (20, 23), so we also considered VEGF-D levels separated by quartiles. Unless otherwise stated, the results reported pertain to the analysis of biomarkers as continuous measures.

The Alliance Statistics and Data Center conducted data collection and statistical analyses on the basis of clinical data locked as of December 15, 2015. The R software environment for statistical computing and graphics and the survival extension package were used to conduct the statistical analyses and to generate the figures (29, 30). The analyses were carried out with adherence to the principles of reproducible analysis using the knitr package (31) for generation of dynamic reports and GitLab for source code management.

The code for replicating the statistical analysis has been made accessible through a public source code repository (https://gitlab.oit.duke.edu/dcibioinformatics/pubs/calgb-80405-plasma).

De-identified patient data may be requested from Alliance for Clinical Trials in Oncology via [email protected] if data are not publicly available. A formal review process includes verifying the availability of data, conducting a review of any existing agreements that may have implications for the project, and ensuring that any transfer is in compliance with the IRB. The investigator will be required to sign a data release form prior to transfer.

In CALGB 80405 primary cohort, patients with mCRC with KRAS wild type were randomized to receive bevacizumab (BEV) or cetuximab (Cetux), in addition to either FOLFIRI or FOLFOX as chemotherapeutic agents per physician choice. The last date of follow up was December 15, 2015, and the median follow up for surviving patients was 47.4 months (25). The demographic and survival outcomes for the total primary cohort population (n = 1,137) and the biomarker population (n = 715) are shown in Fig. 1 and Supplementary Table S1. No significant differences were noticed regarding age, race, gender, or clinical outcomes between the biomarker population and the total primary cohort. To better present the four patient cohorts receiving different treatments, a detailed demographic table was generated. As reported in the clinical paper, there is no statistically significant difference in OS and PFS in the bevacizumab-chemotherapy group and the cetuximab-chemotherapy group (25). All four patient cohorts showed similar baseline characteristics (Table 1).

A total of 24 circulating protein biomarkers were assessed using baseline plasma samples for 715 patients with mCRC enrolled in CALGB 80405. The median levels and ranges for all biomarkers are shown in Supplementary Table S2. A dendrogram illustrating the associations among these 24 biomarkers is shown in Supplementary Fig. S1. As observed in previous analyses, we noted several consistent relationships among the protein biomarkers. PlGF was tightly associated with general angiogenic biomarkers such as Ang-2 and TSP-2, whereas VEGF-A, -C, -D were more clustered with factors in TGFβ signaling pathway. CD73 and HER-3, two previously identified biomarkers predictive of benefit from cetuximab (28), were tightly clustered with each other.

The clinical report showed similar OS and PFS across patients in all treatment groups, therefore, prognostic analysis was conducted with all patients combined. We identified ten biomarkers (Ang-2, CD73, HGF, ICAM-1, IL6, OPN, TIMP-1, TSP-2, VCAM-1, and VEGFR-3) to be prognostic for both OS and PFS in the overall population, as illustrated by the Forrest plots (Fig. 2). All of the biomarkers identified were negative prognostic biomarkers, that is, higher expression levels were associated with shorter times to event (OS or PFS). The strongest prognostic effects with respect to HR for OS and PFS were observed for VCAM-1 (for OS, HR, 2.49; 95% CI, 2.01–3.09; P < 0.0001; for PFS, HR, 1.55; 95% CI, 1.28–1.88; P < 0.0001). A complete list of the prognostic value for all 24 biomarkers for OS and PFS was shown in Supplementary Tables S3 and S4, respectively. All biomarkers tested were independently prognostic from one another. PlGF and VEGF-A were prognostic for OS (PlGF, HR, 1.31; 95% CI, 1.05–1.63; P = 0.0166; VEGF-A, HR, 1.10; 95% CI, 1.02–1.19; P = 0.0143; Supplementary Table S3), but not for PFS (PlGF, HR, 1.16; 95% CI, 0.95–1.41; P = 0.1473; VEGF-A, HR, 1.06; 95% CI, 0.99–1.14; P = 0.0811; Supplementary Table S4).

We identified PlGF as the only predictive biomarker of PFS benefit in the overall population. A high baseline level of PlGF was associated with better PFS when treated by cetuximab (BEV arm: HR, 1.51; 95% CI, 1.10–2.06; Cetux arm: HR, 0.94; 95% CI, 0.71–1.25; P_intx, 0.0298; Table 2). No other biomarkers were observed to be predictive of OS or PFS in the chemotherapy-pooled populations (Supplementary Tables S5 and S6). A Kaplan–Meier curve illustrating the predictive effect of PlGF is shown in Fig. 3. Patients with lower than median PlGF exhibited longer PFS and benefited more from bevacizumab than cetuximab.

To control for the possibility of the interaction effects across chemotherapies used in CALGB 80405, the biomarkers were re-analyzed for predictive associations within chemotherapy subpopulations. In patients receiving FOLFOX, low VEGF-D was associated with better PFS and benefit from bevacizumab (BEV arm: HR, 1.70; 95% CI, 1.19–2.42; Cetux arm: HR, 0.92; 95% CI, 0.68–1.24; P_intx = 0.0097, Table 2). Although a borderline trend was noted for PlGF, the PFS benefit did not reach statistical significance (P_intx = 0.0656, Table 2). Kaplan–Meier curves showed the predictive effects of VEGF-D in chemotherapy combined, as well as in FOLFOX and FOLFIRI subpopulation, respectively (Fig. 4). No other biomarkers were observed to be predictive for OS or PFS when patients were stratified by chemotherapy (Supplementary Tables S7 and S8).

To optimize the cut-point of VEGF-D level, Forest plots were generated to illustrate the effects of VEGF-D by quartile on OS and PFS in patients receiving FOLFOX (Fig. 5). Patients with VEGF-D levels in the lowest quartile had better OS (HR, 0.62; 95% CI, 0.41–0.92) and PFS (HR, 0.59; 95% CI, 0.41–0.85).

CALGB 80405 is the largest phase III trial to directly compare bevacizumab and cetuximab, two of the most commonly used targeted agents, in addition to FOLFIRI or FOLFOX as first-line treatment in advanced or mCRC. Blood samples collected during CALGB 80405 provide a unique and valuable source of materials for biomarker studies that may help guide the use of these therapeutics. Our retrospective biomarker analysis identified PlGF and VEGF-D, two chemo-dependent predictive biomarkers, as well as numerous prognostic biomarkers.

We identified 10 biomarkers with highly consistent prognostic impact across both OS and PFS. These biomarkers play important roles in angiogenesis (Ang-2, HGF, TSP-2, VEGF-R3), inflammatory and immune modulating (CD73, IL6, ICAM-1, VCAM-1), and extracellular matrix remodeling (OPN, TIMP-1). Intrinsic high expression of all these biomarkers was associated with worse outcomes, reflecting the coordinated interplay of angiogenesis, inflammation, and immune modulation. Interestingly, PlGF and VEGF-A were found to be prognostic for OS, but not for PFS. This is likely due to the manifest of stronger prognostic effect at later stage of the disease or over longer time, as the median OS was approximately 20 months longer than median PFS (25).

With respect to EGFR inhibition, we previously identified two predictive biomarkers for cetuximab benefit (HER-3 and CD73) across all patients (KRAS mutant and wild-type) in CALGB 80203 (28), a randomized study evaluating the efficacy of cetuximab in combination with FOLFOX or FOLFIRI in mCRC. However, neither biomarker predicted benefit from cetuximab here in CALGB 80405 (Supplementary Tables S5–S8). HER-3 belongs to the same family of receptor tyrosine kinases as EGFR, with the potential to serve as a resistance mechanism to cetuximab (32). CD73, an immune modulatory AMP phosphatase, was associated with cetuximab benefit in previous studies (28, 33). The lack of predictive value in this study may be due to the potential association between CD73 and RAS mutational status (34). In this correlative study, all patients included were KRAS wild type at codons 12 and 13. Exclusion of patients with RAS mutations may confound the analysis. Although RAS wild-type has been widely applied as a prerequisite for patients to receive anti-EGFR inhibitors, not all patients with wild-type RAS respond to EGFR inhibitors. Our Angiome analysis could provide novel insight regarding the patient population most likely to derive benefit from the EGFR-targeting drugs.

With respect to VEGF inhibition, VEGF-A has long been considered a candidate for predictive biomarker for anti-angiogenic therapies, and high levels of VEGF-A may indicate increased biological dependency on this angiogenic factor. Yet, a meta-analysis of more than 1,800 patients participating in phase III trials in colorectal cancer, non–small cell lung cancer, and renal cell carcinoma confirmed that pretreatment VEGF-A levels have only prognostic, but not predictive value (35). In the MERiDiAN trial, preselecting metastatic breast cancer patients based on high plasma levels of short isoforms of VEGF-A for bevacizumab treatment revealed no significant association of PFS by VEGF-A interaction, failing to support using baseline VEGF-A to identify patients benefitting most from bevacizumab (36). Similarly, PlGF has prognostic value, but has not been shown to be predictive for anti-angiogenic therapy (37). Our data provide novel insight into the potential predictive role of these VEGF family members.

In this study, we found PlGF was predictive of lack of PFS benefit from bevacizumab independent of the type of chemotherapy used (Table 2; Fig. 3). PlGF belongs to the VEGF family; it binds and activates signaling through VEGF-R1, but does not directly signal through VEGF-R2 (38). PlGF modulates VEGF-A signaling both at the ligand level (by forming heterodimers with VEGF-A) and the receptor level (by engaging VEGF-R1; ref. 39). PlGF has been noted as a pharmacodynamic biomarker with circulating levels consistently increasing in response to anti-VEGF agents (22, 40, 41), but this is the first report of a potential predictive association between baseline PlGF levels and outcome in colorectal cancer. Patients with lower than median levels of PlGF derived more PFS benefit from bevacizumab (P_intx = 0.0298, Fig. 3), independent of chemotherapy received. It is important to note that the comparative arm used in this study is an active therapy (cetuximab) and not a true placebo. This may explain why we observed PlGF as a predictive biomarker only in this specific study population.

VEGF-D was identified as the strongest predictive biomarker for lack of PFS benefit from bevacizumab; however, this was only observed in the FOLFOX treated patients (P_intx = 0.0097, Table 2). VEGF-D is another VEGF family member; it binds VEGF-R2 promoting angiogenesis, as well as VEGF-R3 where it promotes lymphangiogenesis (42). The predictive effect was largest in patients with VEGF-D levels in the lowest quartile, where patients derived the most benefit from bevacizumab (Fig. 5). Importantly, far more patients in CALGB 80405 study received FOLFOX (n = 523) than FOLFIRI (n = 192). It is possible that additional clinical factors related to the use of FOLFOX over FOLFIRI may confound the VEGF-D effect observed here. The exact mechanism for this chemotherapy effect remains unclear. The VEGF-D effect is consistent with results seen previously in patients with pancreatic cancer treated with bevacizumab and gemcitabine in CALGB 80303 (20), and in patients with colorectal cancer treated in the Australian Gastro-Intestinal Trials Group testing Mitomycin, Avastin, and Xeloda (AGITG MAX trial; ref. 23). In all cases, low VEGF-D predicts benefit from bevacizumab, while high VEGF-D predicts for lack of benefit from bevacizumab. Excessive VEGF-D can provide compensatory signaling in the VEGF axis through VEGF-R2, serving as a resistance mechanism upon VEGF-A blockage by bevacizumab. Therefore, patients are less likely to benefit from bevacizumab given the presence of a readily available compensatory mechanism. The predictive roles of PlGF and VEGF-D noted here in CALGB 80405 may suggest that alternate VEGF family members may mediate intrinsic resistance mechanisms to bevacizumab, which selectively binds VEGF-A.

Although no other biomarkers reach statistical significance, a few interesting trends were noted. First, VEGF-C, an additional VEGF family member with potent angiogenic and lymphangiogenic properties, appears to predict for OS benefit from cetuximab in FOLFIRI subgroup (P_intx = 0.0791, Supplementary Table S7). In addition, in the subgroup of patients receiving FOLFIRI, we identified three TGFβ-related biomarkers to be potentially predictive of PFS benefit, including TGFβ1 (P_intx = 0.0524), PDGF-AA (P_intx = 0.0511), and SDF-1 (P_intx = 0.0688; Supplementary Table S8). Higher levels of TGFβ1 and PDGF-AA potentially predict for benefit from cetuximab. These biomarkers may reflect overlapping biologies contributing to cetuximab sensitivity. Interplay between EGFR and TGFβ signaling pathways have been reviewed extensively (43).

Despite extensive effort, no predictive biomarker has been identified that would enable a more personalized use of bevacizumab. Angiogenesis is a complex process involving diverse cell types, cytokines, and growth factors signaling through multiple pathways. The biological redundancies eventually lead to anti-angiogenic drug resistance. As the biology of angiogenesis and its role in tumor development varies across tumor types, suitable biomarkers may be tumor type specific. Although angiogenic factors (PlGF and VEGF-D) were noted to be associated with PFS benefit in patients with mCRC in CALGB80405, other studies noted the predictive potential of the inflammatory cytokine IL6 in the context of anti-angiogenic therapy in renal cancer and ovarian cancer (21, 44).

This analysis has several limitations. The patients receiving FOLFIRI were much fewer than those receiving FOLFOX, therefore underpowered to explore key subgroups. The unbalanced use of FOLFIRI and FOLFOX may introduce uncontrolled bias into these analyses. Although all patients are considered KRAS WT, KRAS testing was only performed at codons 12 and 13. It is possible that some patients with KRAS mutant were included in our analysis. Genetic information such as extended RAS and RAF mutations (45), microsatellites instability and tumor mutational burdens (46), consensus molecular subtypes (CMS), representing intrinsic heterogeneity of colorectal cancer at the genetic expression level (47) will be further investigated in subsequent analysis. In addition, this retrospective biomarker analysis was intended to be exploratory and hypothesis-generating in nature, hence the P values presented here were not adjusted for multiple testing. The current findings are promising and biologically plausible but should be considered preliminary and not ready to guide patient care at this time. To advance any of these putative biomarkers along the validation path, a predefined cut-point would need to be established and tested. Ideally, any biomarker cut-point would be tested prospectively as a stratification factor or for entry onto the trial. The establishment of a cut-point is a crucial step for clinical decision making when considering the prospective use of any prognostic or predictive biomarker. As such, validation of VEGF-D as a biomarker with a specific cut-point awaits further investigation.

Discovery and validation of predictive biomarkers is challenging and requires a detailed understanding of the target expression and biology in patients, technical optimization of biomarker assays, and clinical validation of any findings. This analysis in CALGB 80405 is the culmination for nearly two decades of work by this group. The findings of predictive biomarkers may be applied to other FDA-approved drugs for mCRC. Notably, ziv-aflibercept, a recombinant receptor containing the ligand-binding domains of VEGF-R1 and VEGF-R2, binds both VEGFs and PlGF (15, 48); while ramucirumab blocks the binding of both VEGF-A, -C, and -D to VEGF-R2 (49, 50). It is not known whether these agents may be more active in patients with higher levels of plasma PlGF and VEGF-D. Nevertheless, our data support the rationale for biomarker-driven clinical studies to address this possibility.

In conclusion, our biomarker analysis of CALGB 80405 identified two candidate predictive biomarkers for lack of benefit from bevacizumab: PlGF and VEGF-D. Future studies to validate and refine these findings are warranted.

A.B. Nixon reports personal fees from Consulting and NCI and personal fees and other support from Scientific Advisory Board during the conduct of the study as well as grants from Industry Funding; personal fees from Consulting, Scientific Advisory Board, and NCI; and personal fees and other support from Scientific Advisory Board outside the submitted work; in addition, A.B. Nixon has a patent for U.S. Patent No. 9255927 issued and a patent for U.S. Patent No. 62/337,633 pending. A.J. Hatch reports other support from Owlstone Medical, Inc. outside the submitted work. A.P. Venook reports grants from Genentech/Roche, Merck, and NIH during the conduct of the study. H.-J. Lenz reports personal fees from Merck KG, Bayer, Roche, Fulgent, and Oncocyte and grants from NCI during the conduct of the study. B.H. O'Neil reports other support from Eli Lilly and Company outside the submitted work. F. Innocenti reports other support from AbbVie during the conduct of the study; in addition, F. Innocenti has a U.S. Patent No. 63/169,301 pending and a patent for U.S. Patent No. 16/932,002 issued, licensed, and with royalties paid from Emerald Lake Safety. J.A. Meyerhardt reports personal fees from COTA Healthcare, Taiho Pharmaceutical, and Merck outside the submitted work. E.M. O'Reilly reports grants from Genentech, Celgene, BioNTech, AstraZeneca, Arcus, Elicio, and Parker Institute; personal fees from Cytomx Therapeutics, Rafael Therapeutics, Seagen, Boehringer Ingelheim, BioNTech, Ipsen, Merck, AstraZeneca, Noxxon, BioSapien, Cend Therapeutics, and Thetis; and other support from Agios, Genentech, and Eisai outside the submitted work. H.I. Hurwitz reports grants, personal fees, and other support from Genentech-Roche outside the submitted work. No disclosures were reported by the other authors.

A.B. Nixon: Conceptualization, formal analysis, supervision, validation, investigation, writing–original draft, writing–review and editing. A.B. Sibley: Formal analysis, validation, writing–review and editing. Y. Liu: Data curation, writing–original draft, writing-review and editing. A.J. Hatch: Writing-original draft, writing–review and editing. C. Jiang: Formal analysis, writing–review and editing. F. Mulkey: Writing–review and editing. M.D. Starr: Data curation. J.C. Brady: Data curation. D. Niedzwiecki: Formal analysis, writing–review and editing. A.P. Venook: Conceptualization, supervision, writing–review and editing. L. Baez-Diaz: Conceptualization, supervision, writing–review and editing. H.-J. Lenz: Supervision, writing–review and editing. B.H. O'Neil: Supervision, writing–review and editing. F. Innocenti: Conceptualization, supervision, writing–review and editing. J.A. Meyerhardt: Supervision, writing–review and editing. E.M. O'Reilly: Supervision, writing–review and editing. K. Owzar: Conceptualization, formal analysis, methodology, writing–review and editing. H.I. Hurwitz: Conceptualization, resources, supervision, funding acquisition, methodology, writing–original draft, writing–review and editing.

We gratefully acknowledge the invaluable contributions of the patients, their families, and the staff who participated in this study. The authors would also like to thank the Alliance Biorepository at the Ohio State University and administrative staff. More than 680 institutions participated in this study. The content of this manuscript is solely the responsibility of the authors and does not necessarily represent the official views of NCI. Research reported in this publication was supported by the NCI of the NIH under Award Numbers U10CA180821, U10CA180882, and U24CA196171 (to the Alliance for Clinical Trials in Oncology), UG1CA233180, UG1CA233253, UG1CA233290, UG1CA233337, UG1CA233373, U10CA180820 (ECOG-ACRIN), U10CA180888 (SWOG). The following have provided support to Alliance for Clinical Trials in Oncology and Alliance Foundation Trials programs (https://acknowledgments.alliancefound.org). This study was also supported in part by funds from Bristol-Myers, Genentech, Pfizer, and Sanofi. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

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.