Ad hoc Analysis of the Phase III ENGOT-OV16/NOVA Study: Niraparib Efficacy in Germline BRCA Wild-type Recurrent Ovarian Cancer with Homologous Recombination Repair Defects

The PARP family of nuclear proteins is recruited to DNA repair complexes and activated on sensing DNA single-strand breaks (SSB), playing a crucial role in SSB repair (1). In the presence of a PARP inhibitor (PARPi), unrepaired SSBs lead to stalled replication forks and accumulation of DNA double-strand breaks (DSB; ref. 2). In normal cells, DSBs are repaired effectively by a high-fidelity, error-free DNA repair process called homologous recombination repair (HRR; ref. 3). In cells with faulty HRR, called homologous recombination deficient (HRd), such as those bearing BRCA mutations (BRCAm), PARP inhibition induces accumulation of DNA DSBs, leading to the activation of nonhomologous end-joining pathway, an error-prone process to repair DNA DSBs; this process results in chromosomal instability, cell-cycle arrest, and subsequent apoptotic cell death (3). PARP inhibition also results in PARP–DNA complexes by trapping PARP1/2 protein on the DNA, which will further intensify the DNA replication fork damage. This synthetic lethality between PARP inhibition and homologous recombination defects has served as the basis of PARPi therapy development in multiple solid tumors and is the best-characterized mechanism of action for these agents (4, 5). In addition, preclinical and clinical studies have demonstrated that tumor cells that are homologous recombination proficient (HRp) may also respond to PARPi, suggesting the utility of PARPi beyond HRd tumors (6–9).

In addition to BRCA1 and BRCA2, other genes play critical roles in orchestrating the HRR process, including genes involved in DNA DSB recognition (7), initiation of HRR (7), DNA resection (10), RAD51 filament strand invasion (11, 12), DNA synthesis (12), and Holliday junction resolution (13). Defects in expressing these HRR genes will impair the integrity of HRR and may confer sensitivity to PARP inhibition. The sensitivity of tumors with HRR mutations (HRRm), including BRCA1, BRCA2, PALB2, and RAD51C, to PARPi has been reported in preclinical research (14), as well as clinically (15, 16), across tumor types (Supplementary Table S1). In ovarian cancer, analysis of samples from the ARIEL2 trial of rucaparib maintenance therapy found RAD51C and RAD51D mutations as well as high-level BRCA1 promoter methylation to be associated with PARPi sensitivity (17). However, the reported clinical evidence in ovarian maintenance with niraparib is limited.

The ENGOT-OV16/NOVA trial (NCT01847274) of niraparib maintenance enrolled 553 patients with platinum-sensitive recurrent ovarian cancer who responded to the penultimate platinum-based chemotherapy: 203 in the germline BRCAm (gBRCAm) (niraparib, n  =  138; placebo, n  =  65) and 350 in the non-gBRCAm (niraparib, n  =  234; placebo, n  =  116) cohorts (7). Patients in both cohorts experienced a statistically significant improvement in median progression-free survival (mPFS) compared with those in the placebo arm [21.0 vs. 5.5 months in the gBRCAm cohort (HR, 0.27; 95% confidence interval, CI, 0.17--0.41) and 9.3 vs. 3.9 months in the overall non-gBRCAm cohort (HR, 0.45; 95% CI, 0.34--0.61); P < 0.001 for both comparisons] (7). On the basis of the results of the ENGOT-OV16/NOVA trial, niraparib became the first PARPi approved for maintenance treatment of platinum-sensitive recurrent ovarian cancer regardless of biomarker status.

Whereas the NOVA study used the BRACAnalysis assay (Myriad Genetics, Inc.) to determine gBRCAm status for patient enrollment, the tissue-based myChoice CDx (Myriad Genetics, Inc.) was used to determine tumor, or somatic, BRCAm status and genomic instability score (GIS). Patient tumors were identified as HRd [somatic BRCAm (sBRCAm) and/or GIS ≥  42] or HRp (non-sBRCAm and GIS < 42). Patients whose tumors were identified as HRd in the non-gBRCAm cohort had longer mPFS in the niraparib arm than in the placebo arm [12.9 vs. 3.8 months (HR, 0.38; 95% CI, 0.24--0.59)]. In addition, patients in the niraparib arm who were identified as HRp had a significantly longer mPFS than those in the placebo arm [6.9 vs. 3.8 months (HR, 0.58; 95% CI, 0.36--0.92)] (7). These results potentially reflect the limitations of current tests to reliably capture patients with genomic scarring and HR-deficient tumors who could potentially benefit from therapy. In addition, they also suggest that mechanisms independent of BRCAm or HRd may confer clinical benefit with niraparib PARPi in ovarian cancer.

We performed a comprehensive retrospective analysis using tumor samples collected from the non-gBRCAm cohort in the ENGOT-OV16/NOVA trial to explore additional biomarkers or mechanisms that may predict sensitivity to niraparib. The mutation status of the 18 HRR genes, including BRCA1/2, within the Myriad HRD research assay (Myriad Genetics, Inc.) was evaluated as a biomarker for niraparib.

Of the 553 patients enrolled in the ENGOT-OV16/NOVA trial, 350 patients were in the non-gBRCAm cohort as determined by BRACAnalysis (Myriad Genetics, Inc.), which uses blood or saliva samples to test for the presence of deleterious or suspected deleterious germline BRCA1/2 mutations. Of the 350 patients in the non-gBRCAm cohort, 331 had archival tumor samples available for additional tumor biomarker testing and were included in this analysis.

As part of the ENGOT-OV16/NOVA trial, myChoice CDx was performed on tumor samples prior to database lock. The myChoice CDx test is an integrated next-generation sequencing test assessing sBRCAm status and measuring tumor genomic instability (18). Three algorithms were used to assess genomic instability—loss of heterozygosity (LOH) profiles, telomeric allelic imbalance, and large-scale state transitions—resulting in the myChoice CDx GIS, which is the sum of the three individual scores (Supplementary Fig. S1). Although the myChoice CDx GISs distribute along a continuum, patients were categorized as either myChoice CDx HRD-positive (now called HRd) or HRD-negative (now called HRp) according to the prespecified cut-off score of 42 and/or sBRCAm presence (19). This analysis excludes patients who were enrolled in the gBRCAm cohort of the trial, as determined by BRACAnalysis, regardless of HR status.

The Myriad research-grade assay (Myriad Genetics, Inc.) was performed on the 331 available patient tumor samples from the 350 patients in the non-gBRCAm cohort to determine the mutation status of 18 HRR genes (ATM, ATR, BAP1, BARD1, BRCA1, BRCA2, BRIP1, MRE11A, NBN, PALB2, RAD50, RAD51B, RAD51C, RAD51D, RAD54B, RAD54L, XRCC2, and XRCC3). Deleterious or suspected deleterious mutations were defined by Myriad Genetics based on review of multiple lines of evidence.

All statistical analyses in this article were post hoc. An exploratory analysis was performed on the 331 patients from the non-gBRCAm cohort with available tumor samples to determine HRR gene mutation status. For the subgroup analyses, we performed a two-sided log-rank test using the stratification factors from randomization (best response to last platinum-based therapy, HRD status, and time from penultimate platinum-based therapy to progression) to analyze PFS, which was summarized using Kaplan--Meier methods. Patients were censored according to the methods used in the primary analysis (7). We estimated HRs with two-sided 95% CIs using a stratified Cox proportional hazards model with the stratification factors used in randomization.

HRs refer to the comparison of the niraparib arm with the placebo control arm. Formal P-value correction for multiple testing was not applied, but the multiplicity was accounted for in the interpretation of results. This approach was considered the most suitable given the exploratory nature of the analyses and that the measured biomarkers (BRCAm, HRRm, and myChoice CDx GIS) were selected on the basis of biological relevance.

All patients provided written informed consent as approved by an Institutional Review Board, in accordance with ethical guidelines as described in the U.S. Common Rule. As patients were not specifically consented for open-access genomic data, complete sequencing data, such as BAM files, cannot be shared publicly. Anonymized individual participant data and study documents can be requested for further research from www.clinicalstudydatarequest.com.

Baseline characteristics of the 331 patients from non-gBRCAm cohort with known HRR status were included in this analysis are shown in Supplementary Table S2. Demographic and clinical characteristics were well balanced in the niraparib and placebo arms. For exploratory purposes, patients whose tumor contained a loss-of-function (LOF) mutation in at least one of the 18 HRR genes including BRCA1/2 were considered as HRRm (Supplementary Fig. S1). Of the 331 tumor samples analyzed from the non-gBRCAm cohort, 283 (85.5%, 283/331) were BRCA wild type (BRCAwt) with no detectable deleterious or suspected deleterious mutation in BRCA1/2, and 48 (14.5%, 48/331) carried deleterious or suspected deleterious somatic BRCA1/2 mutations (sBRCAm; Supplementary Fig. S2A). Of the 48 patients with sBRCAm disease, 43 carried biallelic sBRCAm (BRCAm and LOH event), 1 carried a monoallelic BRCAm, and 4 carried BRCAm with unknown allelic status of the tumor (Supplementary Fig. S2B). Of the 283 patients in the BRCAwt subgroup, 41 (14.5%, 41/283) had an LOF mutation in at least 1 non-BRCA HRR gene (non-BRCA HRRm; Supplementary Fig. S2A). Among the 16 non-BRCA HRR genes evaluated in this cohort, RAD51C was the most commonly mutated (n =  9), followed by BRIP1 (n =  7), and the remaining genes had mutations observed in 5 or fewer patients. None of the tested samples had mutated RAD54B or XRCC2 (Supplementary Fig. S2A). When using OncoPrint plot to illustrate the mutational spectrum at the patient level by different treatment groups, LOF mutations in any of the other 16 non-BRCA HRR genes were rarely detected (2/48) in the sBRCAm tumor samples (Supplementary Fig. S2B). One patient had both BRCA1 and RAD54B mutations, and 1 patient had both BRCA2 and RAD54 L mutations. Among patients with non-BRCA HRRm tumors, only 1 patient had both RAD51C and BRIP1 mutation (Supplementary Fig. S2B).

The GIS histogram of the non-gBRCAm cohort showed a bimodal distribution, and the current Myriad GIS cutoff of 42 separated the two modes (Fig. 1A). All 43 sBRCAm samples with myChoice CDx GISs available for analysis had scores ≥33, and 37 had a GIS ≥ 42 (Fig. 1A). myChoice CDx GISs for non-BRCA HRRm samples ranged from 2 to 79. Of the 36 patients with non-BRCA HRRm tumors with available GIS, 25 had a GIS ≥ 42 (Fig. 1A). sBRCAm tumors had the highest median GIS (Fig. 1B), non-BRCA HRRm tumors had an intermediate median GIS, and BRCAwt/HRR wild-type (HRRwt) tumors had the lowest median GIS (Fig. 1B).

Exploratory post hoc analysis of mPFS was performed to evaluate the benefit of niraparib in patients with sBRCAm (Figs. 2 and 3). The HR of niraparib versus placebo was 0.32 (95% CI, 0.09--1.11; mPFS, 20.9 vs. 11.0 months, Δ9.9 months) in the 43 patients with biallelic sBRCAm (Figs. 2 and 3A), 0.27 (95% CI, 0.08--0.88; mPFS, 20.9 vs. 11.0 months; Δ9.9 months) in the 48 patients with sBRCAm (Figs. 2 and 3B), and 0.47 (95% CI, 0.34--0.64; mPFS, 7.4 vs. 3.9 months, Δ3.5 months) in the 283 patients with BRCAwt tumors (Figs. 2 and 3C).

To evaluate whether other HRRms contributed to the clinical benefit of niraparib observed in patients with BRCAwt tumors, an exploratory post hoc analysis of mPFS was performed in subgroups of patients with or without other HRRms in BRCAwt tumors. The HR of niraparib versus placebo was 0.31 (95% CI, 0.13--0.77; mPFS, 6.2 vs. 3.8 months, Δ2.4 months) in the 41 patients with non-BRCAm HRRm tumors (Fig. 4A) and 0.49 (95% CI, 0.35--0.70; 7.4 vs. 4.2 months, Δ3.2 months) in the 242 patients with BRCAwt/HRRwt tumors (Fig. 4B). This result suggests that mutations in non-BRCA HRR genes may impair HRR and sensitize patients to niraparib beyond BRCA1/2, demonstrating similar positive predictive value to that of sBRCA mutation. Nevertheless, patients with BRCAwt tumors, regardless of their HRR gene mutation status, also benefit from treatment with niraparib. The positive predictive value of mutations in at least one of the 16 HRR genes is similar to that of sBRCAm, as suggested by HRs of biomarker-positive cohorts, BRCAwt/HRRm (HR, 0.31; 95% CI, 0.13--0.77) and sBRCAm (HR, 0.27; 95% CI, 0.08--0.88), when comparing between the niraparib and placebo arms (Figs. 2–4); however, neither biomarker demonstrates optimal negative predictive value, as suggested by HRs of biomarker negative cohorts, BRCAwt/HRRwt (HR, 0.49; 95% CI, 0.35--0.70) and BRCAwt (HR, 0.47; 95% CI, 0.34--0.64) when comparing between the niraparib and placebo arms (Figs. 2–4). When comparing niraparib and placebo, HRs of the biomarker-positive cohorts (BRCAwt/HRRm and sBRCAm, respectively) are smaller than HRs of the biomarker-negative cohorts (BRCAwt/HRRwt and BRCAwt, respectively), but HRs of both biomarker-negative cohorts are still statistically significant and less than 1.

In a subanalysis of the 41 patients with non-BRCA HRRm tumors, the HR of niraparib versus placebo was 0.22 (95% CI, 0.05--0.89; mPFS, 15.7 vs. 4.5 months, Δ11.2 months) in the 18 patients with biallelic HRRm (Fig. 4C) and 0.91 (95% CI, 0.35--2.34; mPFS, 4.8 vs. 3.7 months, Δ1.2 months) in the 23 patients whose HRRms were monoallelic or had unknown allelic status (Fig. 4D). Because non-BRCA HRR genes function differently within the HRR pathway, niraparib treatment response was also evaluated in patients with tumor mutations in the more well-characterized HRR genes known to contribute to HRD and PARPi sensitivity (i.e., RAD51C, RAD51D, BRIP1, and PALB2) and patients with other, less well-studied HRR genes. Although the data should be approached with caution because of the small sample size, the benefit of niraparib treatment was more apparent in patients with mutations in well-known HRR genes than in the other group (Supplementary Table S3).

A further exploratory post hoc analysis of mPFS was performed to evaluate the predictive value of GIS, with 42 as the cutoff, in patients with BRCAwt/HRRwt tumors. Niraparib treatment improved mPFS in the 88 patients with a GIS ≥ 42 (HR, 0.33; 95% CI, 0.18--0.61; mPFS, 9.3 vs. 3.8 months, Δ5.5 months; Fig. 5A). Similarly, an improvement in mPFS was seen with niraparib treatment in the 123 patients whose GIS was <42 [niraparib vs. placebo: HR, 0.60 (95% CI, 0.36--0.99); mPFS, 7.2 vs. 4.2 months, Δ3.0 months; Fig. 5B]. Outcomes were also examined by GIS in patients with non-BRCA HRRm tumors; however, the sample sizes for each subgroup were too small to draw conclusions (Supplementary Table S4).

We conducted a comprehensive retrospective analysis of the phase III NOVA/ENGOT-OV16 non-gBRCAm cohort, evaluating niraparib efficacy in subgroups of patients across the spectrum of biomarker status, including all combinations of sBRCA status, with or without a HRRm, and myChoice CDx--identified GIS status. We observed a statistically significant and clinically meaningful benefit of niraparib treatment in the broad NOVA patient population regardless of the biomarker status. These results are consistent with results from the NOVA clinical trial and demonstrate a continuum of benefits across biomarkers, with the highest sensitivity in patients with deleterious BRCAm, followed by those with myChoice CDx HRd tumors, and those with myChoice CDx HRp tumors (7, 20). Although HRRm in BRCAwt tumors was associated with higher sensitivity to niraparib, we were unable to determine any biomarker that could identify a patient subset that did not show a clinical benefit. Given these results, although the myChoice CDx GIS may help to estimate the magnitude of benefit from maintenance treatment with niraparib, the benefit-risk ratio of this testing is low in this patient population, especially when the failure rate of the test is high (17% in the ENGOT-OV16/NOVA trial).

In this analysis, assessment of the HRR gene mutation status in NOVA revealed several interesting findings. At the patient level, evaluation of the mutational spectrum found that LOF mutations in any of the other 16 non-BRCA HRR genes were rare in sBRCAm tumor samples, potentially indicating a mutual exclusivity between BRCA and other HRRms in some cases. In terms of the relationship between HRRm status and GIS, 25 of 36 patients with non-BRCA HRRm tumors had a GIS ≥ 42. This observation is consistent with previous findings from Study 19 (21) and might be explained by the myChoice CDx GIS cutoff of 42, which captures 95% of BRCA1/2 mutations in breast and ovarian cancer. In patients with non-BRCA HRRm tumors, the results suggest that patients with biallelic HRRms may show higher sensitivity to niraparib. This finding is consistent with the hypothesis that most non-BRCA HRR genes would require only one intact copy to be functional, which is similar to BRCA1/2.

Mechanistically, the responses seen in patients with BRCAwt/HRRwt tumors could be potentially explained in several ways. BRCA and HRR mutational profiling does not capture promoter methylation. Promoters of BRCA1 and RAD51C are frequently methylated in ovarian cancer, which results in a “BRCAness” phenotype that confers sensitivity to PARPis. BRCA1 and RAD51C promoter methylations are often reported to be mutually exclusive with mutation events (22). Alternatively, the efficacy of niraparib could be explained via a DNA repair--independent effect. PARP1/2 are known to have pleotropic effects that extend beyond DNA repair, and there is growing evidence that PARP-mediated actions impact tumor cell proliferation and viability via alternative mechanisms of action such as PARP-regulated gene transcription (23), ribosome biogenesis (24), and immune activation (25). The efficacy seen with niraparib in BRCAwt/HRRwt/HRp tumors could be the result of a functional HR deficiency not identified by either the BRCA/HRR mutational analysis or the myChoice CDx HRD test. The myChoice CDx assay is based on capturing genomic scarring resulting from past HR deficiency events, which may not always reflect the current homologous recombination status of the tumor (26). In this analysis, the inability to discern between past genomic scarring and current functional homologous recombination deficiency status could also have been exacerbated by the archival nature of the tumor samples used for testing. The archival samples may not have been reflective of the homologous recombination status of the tumor at the time of niraparib treatment. Moreover, myChoice CDx genomic scarring is based on large structural chromosomal instabilities that do not include the additional genomic features associated with homologous recombination deficiency, such as mutational signatures and microhomology deletion.

When the NOVA trial was designed, the myChoice CDx test was not an approved diagnostic test, and BRACAnalysis was used to assign cohorts during randomization. A key difference between BRACAnalysis and myChoice CDx is that BRACAnalysis only detects germline mutations in BRCA, whereas the BRCA portion of myChoice CDx detects both germline and somatic mutations (in addition to the genomic scarring score). Therefore, it was not surprising that a number of patients with sBRCAm were detected in this post hoc analysis. The results from a randomized phase III trial of niraparib in patients with newly diagnosed ovarian cancer (PRIMA/ENGOT-OV26/GOG-3012) were published recently (6). PRIMA used the myChoice CDx test during randomization, and therefore patients with sBRCAm have been identified and classified as BRCAm and included in the HRd subgroup of that trial. We would expect the results from this analysis—that even patients with GIS < 42 and no known HRR gene mutations benefit from niraparib as maintenance treatment—to be true for PRIMA, and that <1% of patients classified as HRp in the PRIMA trial would have a biallelic BRCAm (germline or somatic), consistent with the known parameters of the myChoice CDx test and the results from this analysis. In PRIMA, patients with HRd disease had a HR of 0.43 (95% CI, 0.31--0.59), and patients with HRp disease had a HR of 0.68 (95% CI, 0.49--0.94; ref. 6).

There are some important limitations of these results. This was an exploratory post hoc analysis and was not designed or powered to draw definitive conclusions on this topic. All HRR gene mutations are not expected to contribute equally to HR pathway deficiency. However, because the number of patients with any given HRR gene mutation was small, it was not feasible to assess the potential impact of each gene individually. The heterogeneous nature of non-BRCA HRR genes also may have contributed to the large 95% CI observed for HRs for patients with non-BRCA HRRm tumors. In addition, the myChoice CDx also has its limitations and may have failed to identify all patients with HR-deficient status. Small sample sizes must also be taken into consideration for the different subgroup analyses. Because the number of patients in each subgroup was small and not always evenly distributed between treatment arms, the 95% CIs for the HRs for several subgroups were quite large. Accordingly, caution should be used when extrapolating results to other patient populations. In addition, NOVA only enrolled patients who were platinum sensitive to their penultimate platinum-containing regimen. Because the platinum-free treatment interval following first-line treatment is an important predictor of responsiveness to subsequent treatment (27, 28), the selection of likely responders in NOVA may limit the generalizability of the HRR findings. Future studies to prospectively test the impact of HRR gene mutations on PARPi efficacy (generally and on a per-gene basis) will be important to validate these results.

The results presented here demonstrate the continuum of niraparib efficacy across the different biomarkers in the NOVA study. Patients with sBRCAm, other BRCA HRRms, or HRD score ≥42 benefitted the most from niraparib treatment. However, significant PFS benefit was also seen in HRp (HRD score <42) patients without HRRms. The results presented here support the use of niraparib in patients with recurrent ovarian cancer regardless of BRCA/HRRm status or myChoice CDx GIS, as all studied subgroups demonstrated a clinical benefit with niraparib when compared with placebo.

G. Lindahl reports personal fees from Honoraria for lectures outside the submitted work. S. Mahner reports grants, personal fees, and other from AbbVie, AstraZeneca, Clovis, Eisai, GSK, Hubro, Medac, MSD, Novartis, Nykode, Olympus, PharmaMar, Pfizer, Roche, Sensor Kinesis, Teva, and Tesaro outside the submitted work. A. Redondo reports personal fees from GSK, MSD; personal fees and other from AstraZeneca; other from Clovis; grants and personal fees from Pharmamar; and grants from Eisai outside the submitted work. M. Fabbro reports personal fees from GSK and AstraZeneca outside the submitted work. B.J. Rimel reports other from GSK, Merck, Immunogen, and personal fees from Deep6AI outside the submitted work. A.M. Oza reports PI and Steering Committees with AstraZeneca, GSK, and Clovis; advisory Board member with AstraZeneca and Morphosys. U. Canzler reports personal fees from AstraZeneca, Lilly, and Roche outside the submitted work. J.S. Berek reports grants from Tesaro during the conduct of the study. A. González-Martín reports personal fees from Alkermes, Amgen, AstraZeneca, Clovis Oncology, Genmab, GSK, ImmunoGen, Merck Sharp & Dohme, MacroGenics, Novartis, Oncoinvent, Pfizer/Merck, PharmaMar, Roche, Sotio, Sutro and grants from GSK and Roche outside the submitted work. P. Follana reports personal fees from GSK, AstraZeneca, and Clovis outside the submitted work. R. Lord reports personal fees from GSK outside the submitted work. Z. Wang reports other from GSK during the conduct of the study; other from GSK outside the submitted work. D. Gupta reports other from GSK during the conduct of the study; other from GSK outside the submitted work; and D. Gupta is an employee of GSK which sponsored the NOVA trial. U. Matulonis reports personal fees from GSK, AstraZeneca, Merck, Novartis, Next Cure, Agenus, 2X oncology, Symphogen, Alkermes, and Morphosys during the conduct of the study; personal fees from Clearity Foundation and Ovarian Cancer Research Alliance outside the submitted work. B. Feng is an employee of GSK. No disclosures were reported by the other authors.

M.R. Mirza: Conceptualization, supervision, investigation, writing-review and editing. G. Lindahl: Conceptualization, supervision, investigation, writing-review and editing. S. Mahner: Conceptualization, supervision, investigation, writing-review and editing. A. Redondo: Conceptualization, supervision, investigation, writing-review and editing. M. Fabbro: Conceptualization, supervision, investigation, writing-review and editing. B.J. Rimel: Conceptualization, supervision, investigation, writing-review and editing. J. Herrstedt: Conceptualization, supervision, investigation, writing-review and editing. A.M. Oza: Conceptualization, supervision, investigation, writing-review and editing. U. Canzler: Conceptualization, supervision, investigation, writing-review and editing. J.S. Berek: conceptualization, supervision, investigation, writing-review and editing. A. González-Martín: Conceptualization, supervision, investigation, writing-review and editing. P. Follana: Conceptualization, supervision, investigation, writing-review and editing. R. Lord: Conceptualization, supervision, investigation, writing-review and editing. M. Azodi: Conceptualization, supervision, investigation, writing-review and editing. K. Estenson: Resources, supervision, visualization, project administration, writing-review and editing. Z. Wang: Data curation, formal analysis, validation, visualization, methodology, writing-review and editing. Y. Li: Data curation, software, formal analysis, validation, visualization, methodology, writing-review and editing. D. Gupta: Conceptualization, resources, formal analysis, supervision, validation, visualization, methodology, project administration, writing-review and editing. U. Matulonis: Conceptualization, supervision, investigation, writing-review and editing. B. Feng: Conceptualization, data curation, formal analysis, validation, visualization, methodology, writing-original draft, writing-review and editing.

Medical writing and editorial support, funded by GSK, was provided by Nicole Renner, PhD, Betsy C. Taylor, PhD, CMPP, and Jennifer Robertson, PhD, of Ashfield MedComms, an Inizio company. The authors wish to thank Ilker Yalcin, Wei Guo, and Jing Wang for their support in analyzing the data.