Predictive Biomarkers of Overall Survival in Patients with Metastatic Renal Cell Carcinoma Treated with IFNα ± Bevacizumab: Results from CALGB 90206 (Alliance)

Renal cell carcinoma (RCC) accounts for 4% to 5% of all new cancers diagnosed each year and is the most common type of kidney cancer in the United States (1). Prior to the 1990s, cytokines such as IFNα and IL2 represented the main therapeutic choices for RCC (2). In 2005, the first VEGF inhibitor sorafenib was approved by FDA to treat RCC (3). The median overall survival (OS) for untreated patients with metastatic RCC (mRCC) improved from under 1 year to more than 33 months with the introduction of VEGF targeted therapy (4, 5). This initial approval was followed by the approval of six additional VEGF inhibitors for use in patients with mRCC, including sunitinib, pazopanib, axitinib, bevacizumab, cabozantinib, and lenvantinib (6).

In mRCC, the most widely used prognostic factor model is from Memorial Sloan Kettering Cancer Center (MSKCC), which includes five negative predictive factors for short survival: low serum hemoglobin, high corrected serum calcium level, Karnofsky performance status less than 80%, high lactate dehydrogenase level, and time from diagnosis to treatment of less than 1 year (7). Motzer and colleagues validated the MSKCC model and added platelet and neutrophil counts as additional markers (8, 9). Mekhail and colleagues further expanded the model by accommodating two prognostic factors: prior radiotherapy, as well as presence of hepatic, lung, and retroperitoneal nodal metastases (10). Currently, the most adopted system for OS prediction is the International Metastatic Renal Cell Carcinoma Database Consortium (IMDC) criteria, consisting of two clinical variables (Karnofsky performance status and months from diagnosis to treatment start) and four laboratory variables (serum levels of calcium, hemoglobin, platelets, and neutrophils; refs. 11, 12).

In contrast to these extensively optimized clinical factors, circulating biomarkers are less understood in mRCC. VEGF inhibitors represent an important option for the treatment of first-line patients with mRCC, but no predictive biomarkers exist for this class of drugs. Since the general response rate to VEGF inhibitors is less than 50% (13), it remains pivotal to identify predictive biomarkers to enable patient selection and to improve clinical outcomes (14). To this end, we have developed and optimized a protein multiplex array termed the Angiome, a panel of circulating biomarkers crucial in tumor angiogenesis, inflammation, and immune modulation (15–17). The Angiome biomarker platform is currently approved for use as an integrated biomarker by the NCI Biomarker Review Committee in several ongoing National Clinical Trials Network studies.

In CALGB 90206, 732 patients with mRCC were randomized with equal probability to receive IFNα alone (IFN) or bevacizumab plus IFNα (BEV + IFN; ref. 18). Bevacizumab (Avastin, Genentech/Roche Inc.) is a mAb against VEGF, and has been approved for use in many cancers, including metastatic colorectal, non–small cell lung, renal cell, and glioblastoma (19). In CALGB 90206, bevacizumab significantly prolonged the progression-free survival (PFS) time, but no OS benefit was observed (18). Recognizing the potential value of biomarkers that might predict for sensitivity and resistance to bevacizumab, EDTA plasma samples were collected prospectively at baseline and after each restaging during treatment on CALGB 90206. Here, we report the Angiome analysis focusing on the pretreatment (baseline) plasma samples from patients on CALGB 90206. Patients were randomly split into training and testing sets, and prognostic and predictive biomarkers for bevacizumab were first identified in the training set, and then validated in the testing set.

The design and results of CALGB 90206 have been described previously (18) and the trial was conducted in accordance with the Declaration of Helsinki. Briefly, eligible patients had mRCC with a clear cell histologic component, Karnofsky performance status of ≥70%, adequate end organ function, blood pressure less than 160/90 mmHg, and lack of central nerve system metastases, significant comorbidity, or recent history of bleeding or clotting. Prior chemotherapy for metastatic disease was not permitted. Institutional review board (IRB) approved, written informed consent was obtained from all the patients who opted to participate in this correlative analysis of CALGB 90206.

Peripheral venous blood was collected into EDTA anticoagulant vacutainers. The tubes were centrifuged within 30 minutes of collection at 2,500 g for 15 minutes. Plasma was aliquoted into cryovials and snap frozen, and samples were shipped on dry ice for centralized storage at -80°C at the CALGB Pathology Coordinating Office in Ohio State University (Columbus, OH). Before analysis, all patients' samples were shipped to Duke Phase I Biomarker Laboratory, thawed on ice, realiquoted based on specific assay requirements, and stored at -80°C. All assays were performed in duplicate at Duke University Medical Center (Durham, NC), limited to two freeze-thaw cycles only, and all analyses were conducted while blinded to clinical outcomes.

All biomarkers [Ang-2, BMP-9, CRP, Endoglin, Gro-α, hepatocyte growth factor (HGF), ICAM-1, IGFBP-1, IGFBP-2, IGFBP-3, IL6, IL8, MCP-1, OPN, PAI-1 total, PAI-1 active, PDGF-AA, PDGF-BB, PEDF, PlGF, P-selectin, SDF-1, TGF-β1, TGF-β2, TSP-2, VCAM-1, VEGF, VEGF-C, VEGF-D, VEGF-R1, and VEGF-R2) were measured using the SearchLight multiplex platform (Aushon Biosystems, Inc.; now Quanterix), except for TGFβ-R3 which was tested as previously described (16).

The primary endpoint used for the analysis was OS, which was also the primary endpoint in CALGB 90206 and was defined as the interval from the date of random assignment to the date of death from any cause. PFS was a secondary endpoint and was measured from the date of random assignment to the date of progression or death, whichever occurred first. A total of 498 CALGB 90206 patients (68%) consented to this correlative study and had baseline plasma available for analysis. Data were averaged among the duplicate measures of the markers and then transformed using the natural logarithm function. Furthermore, the data were randomly split into a 0.56:0.44 allocation ratio with 279 patients and 219 patients assigned to the training and testing sets, respectively. Spearman correlations coefficients were computed among the biomarkers and are presented visually in a dendrogram. This exploratory retrospective analysis had a prespecified analysis and conforms to the reporting guidelines established by the REMARK criteria (20).

The objectives of this analysis were to identify prognostic and predictive factors of OS and PFS. A two-step procedure was used to identify and test for important prognostic markers for OS and PFS. In the first step, the proportional hazards models were utilized to test for the prognostic importance of the 32 markers in predicting OS and PFS. Biomarkers that had FDR (Benjamini–Hockberg method) < 0.05 in the univariate analysis were selected for the multivariable models (21). In the second step, the adaptive least absolute shrinkage and selection operator (ALASSO) (22) penalty was used to select important markers for the multivariable models of OS adjusting for the stratification factors (nephrectomy and Motzer criteria). The regularization parameter was chosen to minimize the Schwarz information criterion. The 95% confidence interval (CI) for the ALASSO was derived by adopting the perturbation method (23).

The proportional hazards model was used to identify predictive biomarkers of OS and PFS. In each model, the main effect for each biomarker, treatment arm, and biomarker–treatment arm interaction terms were evaluated. Biomarkers–treatment arm interaction terms that had FDR < 0.05 were considered statistically significant.

The final model of OS was built with the selected biomarkers and clinical variables. A risk score was computed for each patient in the training set from the estimated predictive regression coefficients.

The estimated regression coefficients from the training set were used to compute risk scores for each patient in the testing set. The performance of the final predictive model of OS was evaluated for its discriminative ability utilizing the integrated time-dependent area under the receiver operating characteristics curve (tAUROC) with a follow-up period of up to 92 months (24). The final model of OS was validated using the testing set with the risk score as a continuous parameter. In addition, in order to assess the robustness of the final model of OS, the combined data (n = 498) was randomly split into 100 training/testing sets of 0.56:0.44 allocation ratio. The final model of OS was fitted 100 times using the training sets and the tAUROC was computed from each testing set and averaged over the 100 testing sets. We report the first and third quartile for the tAUROC across the 100 randomly split testing sets.

In addition, the risk score based on the second quartile (median) was used to group patients into low or high-risk groups. The Kaplan–Meier product-limit method was utilized to estimate the PFS and OS distributions by the different groups and the log-rank statistic was used to test if the two risk groups have different survival outcomes. All statistical analyses for model development and validation were performed using the R package (25). Data collection and statistical analyses were conducted by the Alliance Statistics and Data Center. Data quality was ensured by review of data by the Alliance Statistics and Data Center following Alliance policies. All analyses were based on the study database frozen on March 24, 2009.

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.

Of the 732 patients with mRCC accrued to CALGB90206, a total of 498 patients had baseline plasma available for analysis. The training set consisted of 279 patients, 135 patients in the IFN arm and 144 patients in the BEV + IFN arm; while the testing test consisted of 219 patients; 106 patients in the IFN arm and 113 patients in the BEV + IFN arm. The CONSORT diagram is shown in Fig. 1. The baseline clinical characteristics of the training and testing sets were similar and representative of the total patient population on the parent study (Table 1). The cut-off date is March 24, 2009 and the median duration of follow-up among surviving patients is 83.8 months (18).

In order to investigate prognostic and predictive biomarkers in this disease and treatment setting, we assessed 32 circulating protein biomarkers at baseline. Multiplex analyses demonstrated good sensitivity and reproducibility across all of the 32 markers with coefficients of variation (CV) ranging from 5% to 20%. The median baseline levels and ranges for all markers were shown in Supplementary Table S1. Spearman-based correlation analyses identified five distinct clusters among the measured markers (Supplementary Fig. S1).

Twelve biomarkers were identified as prognostic of OS with a FDR < 0.05, including IL6, Ang-2, OPN, IL8, CRP, IGFBP-1, HGF, IGFBP-2, PIGF, TSP-2, VEGFR-1, and VEGF. All of these markers have HR >1, indicating higher levels of these markers were associated with shorter survival duration (i.e., negative prognostic markers). The HRs and 95% CIs for all 32 markers are shown in Supplementary Table S2.

Similarly, nine biomarkers were identified as prognostic of PFS with statistically significance (FDR < 0.05). These were OPN, PlGF, IL6, IGFBP-1, CRP, Ang-2, IL8, HGF, and VEGF. All biomarkers identified as prognostic for PFS were also prognostic for the OS endpoint. The complete listing of HRs for the PFS endpoint and 95% CIs for all markers are presented in Supplementary Table S3.

In multivariable analyses of OS, we identified three prognostic markers: HGF, IL6, and Ang-2. Higher levels of these markers were associated with greater risk of death (Table 2). Adaptive LASSO identified four prognostic biomarkers of PFS: OPN, IL6, PlGF, and Ang-2 (Table 2). A high degree of consistence was observed between the univariate and the multivariable analysis for prognostic markers.

Our next goal was to identify potential predictive biomarkers of bevacizumab benefit. From the training set, we identified two predictive markers for OS. IL6 was identified as a predictive marker (FDR = 0.0497; Supplementary Table S4). The Kaplan–Meier plots demonstrated that patients with higher than median IL6 levels (>15.9 pg/mL) benefitted from BEV + IFN compared with those receiving IFN alone (Fig. 2A). The median OS was 14.4 months (95% CI = 9.8–20.0) and 10.1 months (95% CI = 6.7–13.0) in patients with high levels of IL6 treated with BEV + IFN and IFN alone, respectively. In contrast, patients with low levels of IL6 levels (≤15.9 pg/mL) treated with BEV + IFN had similar median OS [31.4 months (95% CI, 24.3–43.3)] as that in patients treated with IFN alone [median OS 31.6 months (95% CI, 25.1–41.6)] (Fig. 2A).

HGF was identified as a potential predictive marker for OS, but it did not meet the statistically significant threshold after multiple testing adjustment (FDR = 0.3267; Supplementary Table S4). In contrast to IL6, patients with lower than median HGF levels (≤173.6 pg/mL) benefitted from BEV + IFN treatment (median OS = 40.8 months; 95% CI, 26.9–55.1) compared with those treated with IFN alone (median OS = 26.1 months; 95% CI, 18.1–33.5). No benefit was seen with BEV + IFN (median OS = 11.5; 95% CI, 9.2–16.4) compared with IFN alone (median OS = 11.0; 95% CI, 8.6–18.4) in patients with high HGF levels (>173.6 pg/mL; Fig. 2B). A complete list of predictive values for all markers is shown in Supplementary Table S4.

None of the 32 biomarkers were predictive of PFS, as shown in Supplementary Table S5. Interestingly, IL6 was the top candidate as a predictive biomarker for PFS (P = 0.0735). No further data analyses were conducted on IL6 as it did not meet the statistical significance threshold.

In multivariable analyses, a three-way interaction was detected between IL6, HGF, and treatment in the training set. For simplicity, the three-way interaction is presented in Table 3 with the IL6 and HGF as binary variables dichotomized by the observed median levels. Evaluation of HGF and IL6 stratified by arm revealed that patients with high IL6 levels benefitted from the addition of bevacizumab in both the low (median OS = 33.8 in BEV + IFN vs. 13.9 months in IFN) and high HGF groups (median OS = 9.8 in BEV + IFN vs. 6.7 months in IFN; Table 3). However, in patients with low IL6, only those patients with low HGF levels benefitted from the addition of bevacizumab (median OS = 43.0 months in BEV + IFN vs. 33.5 months in IFN); patients with low IL6 and high HGF did not benefit from the addition of bevacizumab (median OS = 17.1 months in BEV + IFN vs. 27.6 months in IFN) and trended toward a worse survival (Table 3). This model identifies a subgroup of patients (IL6 high/HGF low) who derived the most benefit from the addition of bevacizumab, while the group of patients (IL6 low/HGF high) are predicted for lack of benefit from the addition of bevacizumab.

An alternate view of the three-way interaction model is presented with Kaplan–Meier plots with the markers as binary in Fig. 3. Based on the median levels of IL6 and HGF, patients are grouped into the same four subgroups as described above, including IL6 low/HGF low; IL6 low/HGF high; IL6 high/HGF low; IL6 high/HGF high. In both the IFN and BEV + IFN arms, patients with IL6 low/HGF low exhibited superior OS while patients with IL6 high/HGF high exhibited worse outcomes. Interestingly, in the IFN arm, patients with IL6 low/HGF high had longer survival than patients with IL6 high/HGF low (Fig. 3A). This situation is reversed in the BEV + IFN arm, where patients with IL6 high/HGF low demonstrated longer survival compared with patients with IL6 low/HGF high (Fig. 3B).

To validate these findings from the training set, we tested the predictive value of IL6 and HGF using biomarker levels and OS durations in the testing set. No significant differences were noted in patient demographics or biomarker baseline levels in both the training and testing sets (Table 1; Supplementary Table S1). For each patient in the testing set, a risk score (RS) was computed using the estimated regression coefficients from the three-way interaction (IL6, HGF, treatment arm), along with MSKCC risk score and prior nephrectomy. RS was highly predictive of OS with tAUROC of 0.80 (95% CI, 0.76–0.88) based on one-split testing set. The average tAUROC based on 100 randomly splits testing sets was 0.78 (1^st and 3^rd quartile, 0.77–0.79). Patients were classified into either high-risk or low-risk groups based on the median RS (Fig. 4). The HR for death in the high-risk patients was 2.8 (95% CI, 2.1–3.8; P < 0.0001) compared with low-risk patients. The high and low risk groups had median OS of 10.2 (95% CI, 8.0–13.8) months and 34.3 (95% CI, 28.5–40.5) months, respectively. This model based on the three-way interaction of HGF and IL6 and treatment identified patients who benefitted most from addition of bevacizumab to IFN treatment.

In this retrospective biomarker study, IL6 and HGF were identified as potential predictors of OS benefit from bevacizumab in untreated patients with mRCC enrolled in CALGB 90206 in univariate analysis. These findings represent a comprehensive circulating biomarker analysis performed in first-line patients with mRCC and have the potential to guide current therapeutic approaches in mRCC treatment.

In the past 5 years, the RCC field has undergone a paradigm shift with immune checkpoint blockage becoming the first-line treatment of choice, marked by the FDA approval of nivolumab in 2015 and ipilimumab in 2018 (26, 27). It has been shown that coadministration of antiangiogenic drugs and immune checkpoint inhibitors synergistically increase T-cell infiltration into the tumor (28) and combinations of these two categories of drugs have been rigorously tested in patients with RCC. As a result of this work, the FDA has approved the use of both pembrolizumab and axitinib as well as avelumab and axitinib in the treatment of RCC, based on the superior clinical outcomes revealed in two phase III trials, KEYNOTE 426 (29) and JAVELIN Renal 101 (30). In 2021, the FDA approved the combination of cabozantinib and nivolumab (31), lenvantinib and pembrolizumab (32) in advanced RCC. Even in this era of immuno-oncology, antiangiogenic drugs remain a valid option with the potential to enhance immunotherapy and to modulate immune responses. Our biomarker findings have the potential to improve the application of not only antiangiogenic agents, also immune checkpoint inhibitors in mRCC.

The phase III trial of CALGB 90206 was conducted before immune therapy was introduced to mRCC field, testing the efficacy of bevacizumab, an mAb against VEGF. We identified nine prognostic markers negatively associated with OS and PFS. Five of these markers clustered as two closely related biomarker groups in Dendrogram. One consisted of IL6, IL8, PlGF; the other consisted of HGF and VEGF (Supplementary Fig. S1). The clustering suggested that baseline expression of these markers were very similar. Prognostic value of these markers is consistent with the findings of Tran and colleagues where they reported that serum concentrations of IL6, IL8, HGF were negatively associated with prognosis in a phase III trial with 344 patients with RCC randomized to pazopanib or placebo (33). Escudier and colleagues reported baseline VEGF level associated with both PFS and OS in a phase III trial with 903 patients with RCC randomized to sorafenib or placebo (34). Several groups have identified IL6 as a prognostic marker in patients with mRCC (35–37). These findings all align and confirm the prognostic value of circulating biomarkers as a complement to clinical parameters (38).

In addition to being significantly prognostic (Supplementary Tables S2 and S3), IL6 is the only significant predictive marker after multiple testing adjustment (FDR = 0.0497). We observed that high baseline level of IL6 was associated with improved OS with bevacizumab treatment over IFN alone (Supplement Supplementary Table S4). Using a similar approach, Tran and colleagues identified high IL6 levels as a predictive marker of improved PFS from a phase II trial where 215 patients with RCC received pazopanib (39), then validated the predictive role of IL6 in 344 patients with RCC randomized to placebo or pazopanib in a phase III trial (40). Tran and colleagues showed that high expression level of IL6 was predictive of PFS benefit from pazopanib compared with placebo with a P_interaction of 0.009 (33). Together, these data revealed the predictive value of IL6 for bevacizumab, an mAb, and pazopanib, a multikinase inhibitor of VEGFR, raising the possibility that the finding may be applicable to general VEGF inhibitors in the disease setting of mRCC. Furthermore, we have recently reported a similar predictive role of IL6 in ovarian cancer, where women with high IL6 derived the most benefit from bevacizumab (41). Together these findings suggest that IL6 may assist in directing front-line bevacizumab therapy to maximize benefit and minimize toxicity. Further testing to validate the role of IL6 as a predictive biomarker is warranted.

Our analysis identified a predictive role of HGF that was IL6 dependent (i.e., three-way interaction). By itself, HGF showed modest predictive value (P = 0.0203), but did not reach statistical significance after multiple testing adjustment (Supplement Supplementary Table S4). Upon evaluation of HGF-IL6 and treatment arm for patients, we found the subgroup of patients with high HGF and low IL6 levels derived no benefit from bevacizumab. Therefore, high HGF and low IL6 may be used as a predictor of lack of benefit from bevacizumab. In addition, the predictive importance of HGF has potential therapeutic implications. HGF is the ligand for c-MET, a receptor playing key roles in tumor angiogenesis, metastasis, and plasticity (42). Recently FDA approved cabozantinib, a novel antitumor drug that inhibits VEGFR2, MET, and other kinases (43). Cabozantinib has demonstrated clinical activity in patients with RCC previously treated with VEGF inhibitors, as well in the first-line setting as revealed in the METEOR and CABOSUN trials (44–46). At this time, no predictive biomarker for cabozantinib has been identified. Given the biological relevance, the predictive value of HGF in the context of VEGFR and MET dual inhibition should be investigated.

Lastly, the model combined key clinical risk factors (MSKCC RS and history of nephrectomy) and biomarkers for bevacizumab, demonstrating a robust and reproducible association with OS that was validated in the testing set of CALGB 90206 patients (Fig. 4). This model demonstrated that a combination of clinical and plasma biomarkers represent a stronger predictive marker than either one alone. Our findings of predictive markers and the resulted model have the potential to select patients who have the best chance to benefit from antiangiogenic therapy, as well as those more likely to respond to other therapies. The application of our findings needs to be validated prospectively in future trials.

Finally, we acknowledge several limitations to our study. Although bevacizumab has been tested in combination with immunotherapeutic drugs (47), it is not a commonly used regimen for RCC. Axitinib is more commonly used in combination with immune therapy, likely due to its high specificity as a VEGF inhibitor (48). Rather than targeting the soluble VEGF ligand, small-molecule, VEGFR-targeted therapies (sunitinib and pazopanib) remain the standard first-line treatment in combination with ICI for RCC (49). The biomarker findings reported here for bevacizumab may not be applicable to anti-VEGFR tyrosine kinase inhibitor drugs. Another limitation of our study is that our statistical analysis focused on OS, and the risk factor model was based on OS only. This may impede the application of our findings to other trials which use PFS as the primary endpoint.

In summary, we identified IL6 and HGF, two factors in tumorigenesis and progression, as negative prognostic and predictive biomarkers in mRCC. Combined IL6 and HGF together, a subgroup of patients (IL6 high/HGF low) benefited the most from BEV + IFN, while the other group (IL6 low/HGF high) benefited more from IFN alone. Application of these findings awaits further validation in prospective trials.

A.B. Nixon reports grants from Alliance Foundation during the conduct of the study as well as grants from Genentech, HTG Molecular Diagnostics, MedImmune/AstraZeneca, Medpacto, and Seattle Genetics; grants and personal fees from Promega Corporation; and personal fees from Eli Lilly and Company, GlaxoSmithKline, Leap Therapeutics, AdjuVolt Theapeutics, and NCI outside the submitted work; in addition, A.B. Nixon has a patent for Methods of developing a prognosis for pancreatic cancer and predicting responsiveness to cancer therapeutics issued, and a patent for Blood-based biomarkers for colon cancer pending. I. Shterev reports a patent for Methods of predicting responsiveness of a cancer to an agent and methods of determining a prognosis for a cancer patient issued. H.I. Hurwitz reports Roche stock. P.G. Febbo reports other support from Illumina Inc. outside the submitted work. B.I. Rini reports grants and personal fees from Roche outside the submitted work as well as research funding from Pfizer, Hoffman-La Roche, Incyte, AstraZeneca, Seattle Genetics, Arrowhead Pharmaceuticals, Immunomedics, BMS, Mirati Therapeutics, Merck, Surface Oncology, Dragonfly Therapeutics, Aravive, Exelixis, and Janssen and consulting from BMS, Pfizer, Genentech/Roche, Aveo, Synthorx, Compugen, Merck, Corvus, Surface Oncology, 3DMedicines, Aravive, Alkermes, Arrowhead, GSK, Shionogi, Eisai, and Nikang Therapeutics; in addition, B.I. Rini reports stock from PTC Therapeutics. E.J. Small reports personal fees and other support from Fortis, Harpoon, and Teon and personal fees from Janssen, Johnson and Johnson and Ultragenyx during the conduct of the study. M.J. Morris reports personal fees from ORIC, Curium, Athenex, Exelixis, AstraZeneca, and Amgen outside the submitted work. D.J. George reports personal fees from American Association for Cancer Research (AACR), Axess Oncology, Capio Biosciences, Constellation Pharma, EMD Serono, Flatiron, IdeoOncology, Ipsen, Merck Sharp & Dohme, Michael J. Hennessey Associates, Inc., Millennium Med Pub, Modra Pharmaceuticals B.V., Myovant Sciences, NCI GU Steering Committee, Nektar Therapeutics, Physician Education Resource, PlatformQ, Propella Therapeutics, RevHealth, LLC, Seattle Genetics, UroGPO, and WebMD; grants and personal fees from Astellas, AstraZeneca, BMS, Janssen Pharmaceuticals, and Pfizer; grants, personal fees, and nonfinancial support from Bayer, Exelixis, and Sanofi-Aventis; grants from Calithera and Novartis; personal fees and nonfinancial support from UroToday; and personal fees from Xcures outside the submitted work. No disclosures were reported by the other authors.

A.B. Nixon: Conceptualization, data curation, supervision, funding acquisition, investigation, visualization, methodology, writing–original draft, writing–review and editing. S. Halabi: Formal analysis, validation, visualization, methodology, writing–original draft, writing–review and editing. Y. Liu: Data curation, investigation, visualization, methodology, writing–original draft, writing–review and editing. M.D. Starr: Data curation, methodology, writing–review and editing. J.C. Brady: Data curation, methodology, writing–review and editing. I. Shterev: Formal analysis, validation, visualization, methodology, writing–review and editing. B. Luo: Formal analysis. H.I. Hurwitz: Conceptualization, data curation, supervision, funding acquisition, investigation, methodology, writing–review and editing. P.G. Febbo: Investigation, writing–review and editing. B.I. Rini: Investigation, writing–review and editing. H. Beltran: Investigation, writing–review and editing. E.J. Small: Investigation, writing–review and editing. M.J. Morris: Investigation, writing–review and editing. D.J. George: Conceptualization, data curation, supervision, funding acquisition, investigation, methodology, writing–original draft, writing–review and editing.

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 and UG1CA233253. Detailed information is listed in https://acknowledgments.alliancefound.org. This work is also supported in part by funds from Genentech. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. We gratefully acknowledge the invaluable contributions of the patients, their families, and the staff who participated in this study. The research for CALGB 90206 (Alliance) was supported, in part, by grants from the NIH and by grants from NCI to the Alliance for Clinical Trials in Oncology (Monica M. Bertagnolli, MD, Chair) and to the Alliance Statistics and Data Center (Sumithra J. Mandrekar, PhD). The content of this manuscript is solely the responsibility of the authors and does not necessarily represent the official views of the NIH or NCI. More than 400 institutions participated in this study. Detailed information can be found at NCT00072046.

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.