Comparing efficacy and safety of first-line treatment of metastatic renal cell carcinoma: A Bayesian network meta-regression analysis

Background This Bayesian network meta-regression analysis provides a head-to-head comparison of first-line therapeutic immune checkpoint inhibitors (ICI) and tyrosine kinase inhibitors (TKI) combinations for metastatic renal cell carcinoma (mRCC) using median follow-up time as covariate. Methods We searched Six databases for a comprehensive analysis of randomised clinical trials (RCTs). Comparing progression free survival (PFS) and overall survival (OS) of different interventions at the same time node by Bayesian network meta-analysis. Bayesian network meta-regression analysis was performed on objective response rate (ORR), adverse events (AEs) (grade ≥ 3) and the hazard ratios (HR) associated with PFS and OS, with the median follow-up time as the covariate. Results Eventually a total of 22 RCTs reporting 11,090 patients with 19 interventions. Lenvatinib plus Pembrolizumab (LenPem) shows dominance of PFS, and Pembrolizumab plus Axitinib (PemAxi) shows superiority in OS at each time point. After meta-regression analysis, for HRs of PFS, LenPem shows advantages; for HRs of OS, PemAxi shows superiority; For ORR, LenPem provides better results. For AEs (grade ≥ 3), Atezolizumab plus Bevacizumab (AtezoBev) is better. Conclusion Considering the lower toxicity and the higher quality of life, PemAxi should be recommended as the optimal therapy in treating mRCC. Systematic review registration https://www.crd.york.ac.uk/prospero/, identifier CRD4202236775.


Introduction
According to the latest statistics, renal cancer is the sixth most common malignant tumor in men and the ninth in women. It is estimated that 79,000 new diagnoses will be determined in the United States in 2022, resulting in 13,920 deaths (1). Among them, 80% are renal cell carcinoma (RCC) (2). In the past 20 years, the incidence rate of RCC has continued to rise (3). In addition, in approximately 35% of cases, metastatic RCC (mRCC) was first diagnosed (4).
Since 2006, treatment of mRCC in the first instance has gradually changed from interleukin-2 (IL-2) and interferon-a (IFN-a) with serious toxic and side effects to the treatment of tyrosine kinase inhibitors (TKI) Sunitinib (Suni) and Pazopanib (Pazo) (5). Over the ensuing years, with the advent of TKI and mamman target of rapamycin (mTOR) inhibitors, survival rates for those with mRCC have improved dramatically (6). Simultaneously in 2016, the immune checkpoint inhibitors (ICI) has achieved significant effect in the treatment of mRCC, which could be regarded as a milestone (7). Recently, ICI-TKI and ICI-ICI have demonstrated remarkable efficacy in patients with mRCC. Phase 3 CLEAR Trial demonstrated that the objective response rate (ORR), OS and progression-free survival (PFS) of first-line treatment Lenvatinib plus Pembrolizumab (LenPem) were significantly higher than that of Suni (8). First-line treatment Nivolumab plus Cabozantinib (NivoCabo), when compared with Suni, showed significantly improved OS, PFS, and ORR in phase 3 CheckMate 9ER (9). A four-year-long study indicated that in intermediate/poor-risk patients with IMDC, Nivolumab plus lpilimumab (Nivolpi) had better OS, PFS, and ORR than Suni, according to the phase 3 CheckMate 214 trial (10). According to the phase 3 KEYNOTE-426 trial, significant improvement in PFS, OS and ORR for all IMDC risk group patient treated with Pembrolizumab plus Axitinib (PemAxi) versus Suni (11). The phase 3 JAVELIN Renal 101 trial shows that PFS was significantly higher in Avelumab plus Axitinib (AveAxi) than in Suni among all patients in the IMDC risk group (12).
Notwithstanding, there is still no direct comparison between ICI-ICI and ICI-TKI for certain reasons, a head-to-head comparison of different ICI and TKI combinations remains paucity. Naturally, network meta-analysis (NMA) acts as an indispensable bridge to materialize the indirect comparisons. However, previous NMAs did not compare PFS and OS of different interventions at the same time node, resulting potential bias as treatment period might be the confounding factor. Moreover, no studies have focused on the effect of different median follow-up times on ORR and AEs (grade ≥ 3) of different interventions.
Hence, based on this study, we performed a Bayesian NMA to investigate the effectiveness of different combinations of ICI and TKI at each time node and Bayesian network meta-regression analysis adjusting follow-up time using hazard ratios from kaplan meier curve as primitive data to provide more precise evidence for practice in clinical settings.

Methods
This NMA was guided by the PRISMA guideline (Preferred Reporting Items for Systematic Reviews and Meta-analysis) (13).
Below are the exclusion criteria: (1) observational studies, letters, review, or conference abstract; (2) single-arm design studies; (3) animal studies or research in vitro; (4) interferon as control arm (in light of the widespread acceptance of TKI as a standard of care); (5) non-Chinese and non-English literature.

Data extraction and quality assessment
In the included studies, data were independently extracted by two investigators (SQ and ZX) and used the Cochrane Risk of Bias 2.0 tool, assessed the risk of bias for each included RCT by Review Manager 5.3, any discrepancy was arbitrated by the senior reviewer (XC). Variables recorded include: name of the first author, country, publication year, number of patients, condition, therapeutic drugs, treatment dosage, median follow-up time, treatment level, ORR, AEs (grade ≥ 3), and the hazard ratios (HR) and 95% confidence intervals (CI) associated with PFS and OS. Subsequently, the data regarding to PFS and OS at 3, 6, 12, 18, 24, 30, 36 month were harvested from kaplan meier curve by Getdata 2.26.

Data analysis
For PFS and OS at each time point, the Bayesian NMA was conducted with STATA 17.0 MP to directly and indirectly compare multiple treatments. After evaluating OS and PFS at each time point with odds ratio (OR) and 95% Cl, treatment ranking was performed conducting the surface under the cumulative ranking curve (SUCRA) values, however, whether the effect size between any pair reached the significance was determined by net-league table, which was also called matrix in algebra. Inconsistency and consistency tests were performed to examine the existence of inconsistency. Publication bias was assessed by funnel plots as well.
HRs for OS and PFS, Napierian Logarithm HR (lnHR) and standard error of lnHR (selnHR) for each study were calculated by STATA 17.0 MP. For the ORR and AEs (grade ≥ 3), conventional meta-analyses were conducted by STATA 17.0 MP to generate Napierian Logarithm odds ratios (lnOR) and standard error of lnOR (selnOR) for each study. Subsequently these data (lnHR and selnHR for OS and PFS, lnOR and selnOR for ORR and AEs, respectively) were input into Rstudio 4.1.2 by "gemtc" package to conduct Bayesian NMA. if I 2 <50% and p>0.01, fixed effect model would be implemented; if 50%<I 2 <75%, random effect model was carried out; if I 2 >75%, Galbraith plot would be drawn to preclude the studies outside the outlines. Markov-chain Monte Carlo (MCMC) was used to obtain posterior distributions, with 20,000 burn-ins and 150,000 iterations of 4 each chain and a thinning interval of 10 for each outcome. Brooks-Gelman-Rubin diagnostics and Trace plots were used to evaluate and visualize the convergence of the model over iterations. Matrices were also generated by Rstudio 4.1.2.
Finally, we conducted sensitivity analyses, using median followup time as a covariate to perform meta-regression analyses to eliminate potential confounding factors.

PFS at each time point
In PFS, compared with Suni, the intervention measures with significant effect from 3 to 36 months were LenPem, Axi, NivoCabo and LenEvero in order from high to low. We summarize the details of the interventions with significant results compared with Suni in Table 2.

OS at each time point
18 out of 26 articles reported outcomes related to OS. In this study, adequate data were available at 3, 6, 12, 24 and 30 month to conduct NMA. An analysis of pairwise comparison of regimens on every OS time point is shown in Figure 2B. As for agents, Suni was most commonly used, with Axi and Sora being most commonly At 3 rd month, two interventions were significantly compared with Suni, but were not significantly compared with placebo. This diagram shows the PSRISMA flow diagram for study search and selection (updated in 2020). PSRISMA, Preferred Reporting Items for Systematic Reviews and Meta-Analyses; CENTRAL, Cochrane Central Register of Controlled Trials; Embase, Excerpta Medica database.  In OS, compared with Suni, PemAxi and NivoCabo were significant from 6 to 36 months; Nivolpi significant from 12 to 36 months; LenPem significant from 6 to 24 months. We summarize the details of the interventions with significant results compared with Suni in Table 3.

Survival Analysis of PFS, OS, ORR and AEs (grade ≥ 3)
Twenty-two of the 26 articles reported outcomes related to the HRs of PFS. We compared the 19 interventions included directly and indirectly. The network graph is shown in Figure 3A. The results related to OS were reported in 18 of 26 articles. A direct and indirect comparison was made between the included 15 interventions. The network graph is shown in Figure 3A. The interventions with significant differences from Suni are LenPem (HR=2.27, 95%Crl 1.28 to 4.03), PemAxi (HR=2.08, 95%Crl 1.07 to 4.03), and AveAxi (HR=1.93, 95%Crl 1.04 to 3.56). Among them, the highest SUCRA ranking is LenPem, followed by PemAxi. Detailed results are shown in Supplementary Table 4B.
Results related to ORR were reported in 20 out of 26 articles. The 18 interventions were compared directly and indirectly. The network graph is shown in Figure 3A. Regarding AEs (grade ≥ 3), indirect and direct comparisons were conducted between 16 interventions. The network graph is shown in Figure 3A. Nivolumab, Sora, Atezolizumab, Anlotinib, Savo, Nivolpi and Atezolizumab plus bevacizumab are less toxic than Suni. Cabo, PemAxi, NivoCabo, Axi, Pazo had no significant

Heterogeneity and network metaregression analysis
In order to better explain the heterogeneity, we performed sensitivity analysis on the four outcomes. Meta-regression analysis was performed on PFS, OS, ORR and AEs (grade ≥ 3) with the median follow-up time as the covariate. The network graphs are shown in Figure 3B. Detailed results are shown in Supplementary Table 5A-D. For PFS, after meta-regression analysis, PemAxi, NivoCabo are not significantly compared with Suni, LenPem, Cabo and AveAxi are consistently better than Suni. For OS, PemAxi and LenPem are still better than standard Suni. For ORR, in comparison to Suni, Cabo, LenPem, PemAxi, NivoCabo and LenEvero consistently provide better results. For AEs (grade ≥ 3), after meta-regression analysis, the serious adverse events of LenPem and LenEvero were not different from that of Suni. The detailed results before and after meta-regression are shown in the Table 4.
For meta-regression, as shown by Brooks-GelmanRubin diagnostic, inferential iterations were reproducible and stable for each MCMC. Additionally, we used the history feature to confirm the model's convergence in all outcomes. Detailed results are presented in Supplementary Figures 7A-D

Principal findings
This is the first Bayesian NMA investigating the pairwise effect of regimens on OS and PFS at each time node; meanwhile, the prominent innovativeness is the implementation of network meta- regression analysis adjusting confounding factor, which is the Qomolangma in NMA. There are the following findings.
Regarding PFS, compared to Suni, the interventions with significant effects were LenPem, Axi, NivoCabo and LenEvero from high to low from 3 to 36 months. PemAxi and AveAxi also showed good results compared to Suni, but due to the lack of data, it was not possible to tell whether the significance persisted until month 36. Based on Bayesian NMA of HRs of PFS and Bayesian network meta-regression analysis with median follow-up time as a covariate, the comparisons with Suni were significant in descending order of LenPem, Cabo and AveAxi. In summary, LenPem is the first choice for improving PFS.
Regarding OS, from 6 to 36 months, PemAxi and NivoCabo were significantly superior to Suni. Bayesian NMA of HRs of OS and Bayesian network regression analysis with median follow-up time as covariate showed that LenPem and PemAxi were significantly different from Suni. Considering LenPem significant only from 6 to 24 months, PemAxi is the first choice for improving OS.
Regarding ORR, in comparison to standard chemotherapy, Cabo, LenPem, PemAxi, NivoCabo and LenEvero consistently provide better results. Notably, LenEvero needs to be excluded because according to the Meta-regression analysis, LenEvero is inferior to D in both primary endpoints PFS and OS although the ORR results are significant.
Regarding AEs (grade ≥ 3), after Bayesian network metaregression analysis with median follow-up time as a covariate, Atezo and Savo were significantly less toxic than Suni, and the rest were not significant with Suni. However, both Atezo and Savo were inferior to Suni in primary endpoints PFS and OS, so they were not considered as first-line therapeutic agents for mRCC.
These results demonstrate that the combination of ICI-TKI has significant OS, PFS and ORR benefits in patients with mRCC. Interestingly, clinical studies show the separating survival benefit of ICI-TKI much earlier than ICI-ICI in first-line treatment of mRCC (9,35). As well, some investigations showed that the OS of ICI-TKI combination treatment effect is more favorable than dual combination immunotherapy (10, 11). Moreover, the comparison of efficacy results and tumors with sarcomatoid differentiation in clinical trials concluded that the ORR and PFS of PemAxi were superior to those of Nivolpi (36, 37). Therefore, ICI-TKI combination therapy is also the preferred therapy for aggressive, rapidly progressive renal cancer.
RCC is a highly vascularized tumor, and the expression level of VEGF-A is significantly higher in RCC patients than in patients with other types of cancer (38). In addition, TKI can increase immune infiltration directly or indirectly while improving vascularity (39, 40). Studies have shown that both Cabo and Len have modulating and immune-promoting properties (41, 42). Thus, the combination of TKI and ICI has a synergistic anti- tumor effect. Treatment-related toxicity should also be considered when TKI and ICI are used in combination. ICI produces immune-related adverse events, while TKI is chronically toxic. The toxicity of the combination, although greater than that of monotherapy such as Atezo, was not significantly compared to standard treatment. Its toxicity is within the acceptable range, only the superposition of dual adverse effects will increase the difficulty of clinical management. In our research, although both LenPem and PemAxi showed significant advantages in terms of PFS and ORR, the toxicity of LenPem cannot be ignored, and in this clinical trial, the dose of Lenvatinib was consistently reduced by constant reductions to reduce adverse events that could discontinue treatment. In terms of OS, the remarkable performance of PemAxi lasts up to 36 months, while LenPem lasts only up to 24 months. In addition, the toxicity of PemAxi is lower and the quality of life of patients is higher. In summary, our study shows that for mRCC patients PemAxi can have better survival outcomes, lower toxicity, and higher quality of life. Therefore PemAxi appears to be the superior first-line TKI-ICI combination.

Previous network meta-analyses
Treatments for mRCC in the first instance has been changing rapidly in recent years, and the earlier NMAs did not incorporate the multiple ICI-TKI interventions recommended by the mRCC first-line treatment guidelines in recent years (43,44). In 2019, Wang et al (45) published a NMA that focused only on the analysis of ICI and included many interventions that were completely withdrawn from the clinic due to high toxicity, such as IFN-a and IL-2. Manz et al (46) published a NMA in 2020, Focusing solely on TKI. Other studies have focused only on immune-based interventions (47,48). In two recent studies, they included only a few treatment nodes and used a frequentist NMA rather than a Bayesian framework (49, 50).

Strengths and limitations
We evaluated 19 first-line interventions using 26 high-quality studies that were screened. First, we used a Bayesian framework that is more flexible relative to the frequentist, describing pairwise comparisons in terms of probabilistically distributed random variables (51). Second, for the analysis of PFS and OS, it lasted until 36 months. In addition, based on PFS, OS, ORR and AEs (grade ≥ 3), we performed a sensitivity analysis. We performed a network meta-regression analysis with median follow-up time as a covariate. Third, due to the inclusion of a sufficient number of studies, we performed a paired meta-analysis. Closed loops existed in the network, so heterogeneity was also evaluated and the results showed good agreement between the trials included in the study. Fourth, the stability and replicability of each MCMC chain iteration was demonstrated using Brooks-GelmanRubin diagnostics, as well as the convergence of the model was estimated.
A number of limitations are associated with this NMA. First, we have compared ICI and TKI combinations directly or indirectly; however, this approach cannot fully replace a head-to-head comparison. Therefore direct comparative clinical trials are still indispensable. Second, the quality of the trials included in this analysis may have been affected by several types of bias, which could have some impact on the validity of the overall outcomes. Third, the study population included patients with clear cell histology, so the final results are not applicable to patients with non-clear cell histology. Fourth, only trials with standard dosing regimens were included in this study, and the doses and schedules administered in actual clinical settings may differ from those of the included studies; consequently, efficacy and tolerability may be affected to some extent. Some investigations have demonstrated that modifications to the dose and schedule pattern of Suni administration may improve its efficacy and enhance tolerability (52,53). Fifth, there was a large variation in median follow-up time across studies, and although this influence was corrected using meta-regression, the results need to be further investigated in the clinic. In addition, another part of confounding factors (e.g., PD-L1 status, number of focal metastases, patient risk class, etc.) had missing data in some trials, and we could not correct for these factors using metaregression; therefore, the results of this Bayesian NMA need to be treated with caution.

Future research
We hope that future clinical studies will be more precise and focus more on the outcomes of ORR and AEs at each time point. According to the studies, the median follow-up period ranged from 6.4 to 55 months, and the wide variation in follow-up time can have an impact on outcome indicators. Although we corrected for this with meta-regression using time as a covariate, the results this result cannot be used as a proxy for accurate clinical studies. The findings would be more convincing if the ORR or AEs of different interventions were compared at the same time points.

Conclusion
Considering the lower toxicity and the higher quality of life, PemAxi should be recommended as the optimal therapy in treating mRCC. Certainly, it is necessary to conduct more head-to-head comparisons in order to confirm these findings.

Data availability statement
The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding author.