Comparison of tumor-agnostic and tumor-specific clinical oncology trial designs: a systematic review and meta-analysis

Advancements in our understanding of the molecular basis of cancer have revealed new targets for cancer therapeutics. Relative to traditional cytotoxic agents that work to slow disease progression by killing cancerous cells in a nonspecific manner, targeted therapies act on specific cellular pathways to induce antitumor effects with greater efficacy and reduced toxicity. This has evolved into an interest in tumor-agnostic drug development, particularly for rare cancer mutations, where few patients meet both a histologic and molecular definition. In tumor-agnostic trials, classification by molecular alteration supersedes classification by tumor type, and patients are recruited into trials primarily by the presence of a specific molecular target [1]. Since 2017, three therapies have been approved for tumor-agnostic indications [2]. However, even for targeted therapy, the response may differ by histologic context [3,4]. For example, the efficacy of BRAF inhibitor monotherapy, such as vemurafenib, for BRAF-mutant melanoma is not consistent with the response seen in BRAF-mutant colorectal cancer for which efficacy is relatively poor [4]. Thus, given the importance of histological context in predicting the success of targeted therapy, the use of a tumor-agnostic approach in drug development has not been fully embraced, and the question of whether or not tumor-agnostic trials can offer comparable results to tumor-specific trials to support new drug development and approval has not been thoroughly investigated.

It is timely to understand the comparability of estimates of clinical efficacy for targeted agents using tumor-specific versus tumor-agnostic designs, and the relative importance of classifying patients by tumor type versus molecular alteration. Using two focused systematic reviews and meta-analyses (SRMA), the current work was designed to compare clinical outcomes of published tumor-specific and tumor-agnostic clinical oncology trials to better understand how results compare across study designs and the potential implications for research and development of pharmaceuticals. This analysis focused solely on clinical trials of tyrosine kinase inhibitors (TKI) in solid advanced cancers meeting one of two criteria: tumor-specific trials in BRAF-mutated cancer or tumor-agnostic trials in any molecular alteration. The comparison of tumor-specific and tumor-agnostic trials then focused on cohorts that had BRAF-mutated cancers, to allow for less heterogeneous comparisons. BRAF mutations were the focus of the analysis, as they occur across a broad group of human cancers and are common enough to provide sufficient sample sizes for data analysis.

Methods

The research question in this review differed from that of traditional SRMAs as it aimed to compare results from trial designs rather than clinical interventions. The comparison of interest in this meta-analysis occurred across trials, not within, and thus required some modifications to the traditional SRMA methodology. Nevertheless, the study followed the Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA) recommendations wherever possible and appropriate.

As the research question necessitated capturing two distinct trial designs, this qualitative comparison was conducted as two parallel SRMAs with independent eligibility criteria and search strategies for tumor-agnostic trials and tumor-specific trials. The search strategy and eligibility criteria for tumor-specific trials were limited to capturing only trials conducted in BRAF-mutated cancers. In comparison, the tumor-agnostic search strategy and criteria were broad enough to capture all tumor-agnostic trials across diverse molecular targets to provide a tumor-agnostic evidence base for future studies. However, this study focuses only on BRAF-mutated cohorts within tumor-agnostic trials to allow for less heterogenous comparisons between tumor-agnostic and tumor-specific trials. Thus, study selection occurred in two steps, with the first step being to screen for publications meeting overall eligibility criteria, and the second step being to identify publications that included BRAF-mutated trial populations or cohorts. This subset of trials containing BRAF-mutated cohorts was included in the descriptive analysis of trial/cohort characteristics and subsequent meta-analyses.

Eligibility criteria

To be eligible for inclusion, tumor-specific trials must have met the following criteria: phase Ib, phase I/II, phase II or phase III clinical trial; employed a tumor-specific trial design that recruited just one type of cancer; recruited patients with advanced cancer and a molecular alteration of the BRAF gene; experimental intervention included a TKI targeted to BRAF and/or MEK alterations; and had published results, including at least one of overall response rate (ORR), progression-free survival (PFS) or overall survival (OS).

Tumor-agnostic trials were eligible for inclusion if the publication met the following criteria: phase Ib, phase I/II, phase II or phase III clinical trial; employed a tumor-agnostic trial design that recruited across multiple cancer types; recruited patients with advanced cancer and a specific molecular alteration; experimental intervention was a TKI targeted to the molecular alteration for which patients were recruited and had published results, including at least one of ORR, PFS or OS.

Although not part of the screening criteria, the tumor-agnostic trials included in this analysis were limited to those that included BRAF-mutated cancer cohorts and reported those results separate from other mutations. Both tumor-agnostic and tumor-specific trials were also excluded if they met any of the following additional exclusion criteria: included pediatric tumors where results for adult tumors were not presented separately; did not include solid tumors; included only early-stage cancers; or if the experimental arm included treatments other than a TKI alone (e.g., TKI and monoclonal antibody combination therapy).

Information sources & search strategy

Electronic systematic literature searches were performed in PubMed/MEDLINE and EMBASE via the Ovid interface to identify trial publications from 1 January 2000, to 23 September 2021. The reference lists of key studies were reviewed for additional publications and hand searches of other search engines including ClinicalTrials.gov and Google Scholar were also conducted to identify potential additional studies not captured in the search strategies.

Separate search strategies were developed for tumor-specific and tumor-agnostic trials (Supplementary Tables 1 & 2). As the tumor-agnostic trial design is a relatively new phenomenon in the history of oncology clinical trials, the search was restricted to only trials published after the year 2000. Upon reviewing the distribution of publication years, trials published in 2010 or more recently were screened, to keep the time period as comparable as possible between the tumor-agnostic and tumor-specific trials. Abstracts were also excluded during study selection as they are not peer-reviewed publications and are unlikely to contain sufficient information to assess the risk of bias.

Study selection, evaluation & data extraction

Two independent reviewers conducted two levels of screening. Titles and abstracts for all trials were screened by both reviewers independently in level 1 to determine which articles potentially met eligibility for inclusion. Full-text articles for all trials screened for inclusion in level 1 were then screened in level 2 by both reviewers independently. A third reviewer was consulted as necessary to resolve any disagreements between reviewers and reach a consensus. Kappa statistics were calculated to determine the level of agreement between the two independent reviewers.

Data extraction followed a uniform data extraction form that was common to both the tumor-agnostic and tumor-specific publications. Data extracted for each study included: study details such as study name/clinical trial ID, full citation, study design, study phase, blinding, number of centers and number of treatment arms; baseline characteristics such as geographic trial location(s), primary patient definition (histological vs molecular), eligibility criteria, sample size, age, sex, solid cancer type(s), primary tumor location(s) and genetic alteration characteristics; intervention and comparator characteristics such as experimental and comparator treatment regimen(s) and efficacy outcomes including PFS, ORR and OS.

Study evaluation

In addition to the data items just described, the methodological quality of the studies was assessed using the Cochrane Collaboration's Risk of Bias tool. This tool was used to assess the risk of selection bias, performance bias, detection bias, attrition bias, reporting bias and other sources of bias. The risk of bias was rated as ‘high’, ‘low' or ‘uncertain'. Heterogeneity between studies was also assessed using the I² value, with 0–40% indicating low heterogeneity, 30–60% indicating moderate heterogeneity and 50% or greater indicating substantial heterogeneity [5].

Data analysis

Analyses were focused on BRAF-mutated cohorts only. Multiple cohorts could be included from a single trial (e.g., results reported separately for different TKI interventions or clinical subgroups). Some trials also had results published across multiple articles, with separate publications from the same trial typically being for separate cohorts (e.g., separate tumor cohorts from tumor-agnostic trials), separate clinical subgroups (e.g., patients with brain metastases) or different time points (e.g., extended follow-up). All available data for all eligible cohorts were extracted, ensuring that patients were not double-counted across multiple cohorts, subgroups or time points. Where the same results were published for the same patients across multiple articles, the most recent publication was used.

Multiple meta-analyses were conducted for tumor-agnostic and tumor-specific trials, pooling ORRs across cohorts from tumor-agnostic trials and tumor-specific trials, respectively. ORR was the only outcome meta-analyzed, as it was the most consistently and uniformly reported compared with PFS and OS, for which reporting was more variable. ORR was pooled across multiple tumor types, creating one multitumor pooled ORR each from tumor-agnostic trials and tumor-specific trials. Results were then pooled within tumor types, creating single-tumor pooled ORRs for both tumor-agnostic trials and tumor-specific trials. Finally, results were pooled by intervention type received (e.g., BRAF inhibitors and BRAF/MEK inhibitors). This stepwise approach to the analysis, with each step limiting meta-analyses to a certain factor, such as tumor type or intervention type, was intended to simulate controlling for extraneous variables beyond trial design that would differ between the two groups of trials, to increase confidence that observed differences in pooled effect sizes and heterogeneity were at least partly attributable to differences in trial design. The clinical characteristics of the patient populations and methodological differences in the trials were also described qualitatively to further contextualize the differences observed in pooled results from the respective meta-analyses.

As considerable between-study heterogeneity was expected due to variability in populations, intervention and time periods, a generalized linear mixed-effects model [6] was used to pool effect sizes, and a maximum likelihood estimator was used to calculate the heterogeneity variance τ². A Knapp-Hartung adjustment [7–9] was also applied to the mixed-effects model to calculate the 95% CIs around the pooled effect sizes. Analyses were conducted in R software, version 4.1.1 [10] using the ‘meta' [11] and ‘forestplot' [12] packages. For all meta-analyses, results are reported as pooled ORR (95% CI) and statistical heterogeneity of the estimate (I²). Comparison of pooled ORRs (i.e., tumor-specific vs tumor-agnostic) was performed using the χ2 test of independence, with exact p-values reported to describe the statistical significance of the comparison. Other descriptive characteristics are summarized as mean and standard deviation (SD) or median and interquartile range (IQR) for continuous variables and frequency and proportion (%) for categorical variables.

Results

Study selection

The electronic database search for tumor-specific trials returned a total of 1372 citations. Of these, 557 were peer-reviewed, full-text articles that underwent title and abstract screening. A total of 495 articles were excluded from level 1 screening and 62 articles underwent full-text screening. After the second-level screening, an additional 19 articles were excluded and 43 articles were considered eligible. An additional five articles were identified from hand-searching, resulting in a total of 48 tumor-specific publications eligible for inclusion, coming from 33 clinical trials (Figure 1). For tumor-agnostic trials, the electronic database search returned 8572 total citations, of which 2344 were peer-reviewed, full-text articles. Level 1 screening excluded 2300 of these articles, leaving 44 articles for level 2 screening. A total of 18 articles were excluded during level 2 screening and 26 were considered eligible. Hand-searching identified an additional six articles, resulting in 32 tumor-agnostic publications coming from 25 clinical trials (Figure 2). The vast majority of articles that were excluded at each phase of screening, for both tumor-agnostic and tumor-specific trials, met multiple exclusion criteria, the most common being that the experimental intervention included a non-TKI and/or the intervention was tested in an all-comer population (i.e., not targeted to a specific molecular alteration).

For both search strategies, the overall Kappa statistic for the inter-rater agreement was 0.57, indicating moderate agreement. However, this was mostly attributable to the iterative nature of refining inclusion criteria given the extensive scope of literature captured. That is, most disagreements between raters were due to novel publication or trial characteristics that had not been considered during protocol development. All disagreements were easily resolved through discussion of the specific citation, and any decisions made were then consistently applied to the screening of all other citations.

Study characteristics of publications included in meta-analyses

Publications from tumor-agnostic trials

Of the 25 tumor-agnostic trials identified in the initial step of study selection, 5 were identified as recruiting patients with targetable BRAF alterations and had published ORR results available. Four of the five trials were phase II trials, and one was a phase I/II trial. All trials were single-arm, multicenter and open-label. All five trials recruited patients from the USA, but three trials recruited patients internationally as well. All five trials were judged to have a high risk of bias.

Among the five tumor-agnostic trials, there were 22 distinct BRAF-mutated cohorts spanning the following tumor types: colorectal, non-small-cell lung (NSCLC), biliary tract, ovarian, bladder, pancreatic, uterine, salivary gland, small intestine, prostate, unknown primary and 21 other tumor types not further specified. Further cohorts were defined based on criteria such as prior treatment exposure. All TKI interventions were BRAF and/or MEK inhibitors, including vemurafenib, cobimetinib, trametinib, encorafenib, binimetinib and dabrafenib. Sample sizes of the cohorts ranged from 7 to 230. The proportion of males in each cohort ranged from 25 to 73%, and with the exception of one cohort, all cohorts had a median age over 50 years old. The vast majority included patients with any BRAF V600 mutation, though two cohorts recruited only V600E mutations and one cohort recruited only fusions in the BRAF gene or non-V600 mutations. Tumor-agnostic trial and cohort characteristics are described in detail in Supplementary Tables 3 & 4, respectively, and PFS and OS are described in Supplementary Table 5.

Publications from tumor-specific trials

Of the 33 tumor-specific clinical trials identified in step one of study selection, 27 were identified with published ORR results for meta-analysis. Of the trials, 14 (51.6%) were phase II; 7 (25.9%) were phase III; and the remaining six were phase I/II or phase Ib (22.2%). Three studies used a single-blind or double-blind design and the rest were open-label. 17 (63.0%) were single-arm trials, with the remaining ten (37.0%) testing two or more treatment arms. A total of 24 (88.9%) were multicenter, 16 of which recruited patients internationally. Three (11.1%) trials were single-center trials, one taking place in the USA and another in Australia, with no location specified for the last trial. 25 (92.6%) of the tumor-specific trials were conducted on melanoma, with the other two trials conducted in NSCLC and thyroid cancer, respectively. 21 trials were judged to have a high risk of bias, three were low risk and three had an unclear risk of bias.

Among the 27 tumor-specific trials conducted in BRAF-mutated cancers, there were 41 distinct BRAF-mutated cohorts. Although each tumor-specific trial was conducted in a prespecified tumor type, there could be multiple cohorts per trial for different interventions (e.g., monotherapy vs BRAF and MEK inhibitor combination therapy), different dosing (e.g., intermittent vs continuous) or different clinical subgroups (e.g., with vs without brain metastases or pretreated vs treatment-naive). TKI interventions included the following BRAF and/or MEK inhibitors: vemurafenib, cobimetinib, trametinib, encorafenib, binimetinib and dabrafenib. Sample sizes of the cohorts ranged from 9 to 856 patients, and the proportion of males in each cohort ranged from 35 to 75%. All cohorts had a median age between 52 and 66 years old. Of the 41 cohorts, 17 (41.5%) cohorts included V600E or K mutations only, 14 (34.1%) included any V600 mutation, six (14.6%) included V600E only, with the remaining four (9.8%) requiring other specifications of BRAF mutations or alterations. Tumor-specific trial and cohort characteristics are described in detail in Supplementary Tables 6 & 7, respectively, and PFS and OS are described in Supplementary Table 8.

Meta-analysis

Multitumor pooled ORRs

The multitumor meta-analysis for tumor-agnostic trials included 23 cohorts and 554 patients. The following cancer cohorts were represented: colorectal (n = 2), ovarian (n = 1), thyroid (n = 3), NSCLC (n = 4), melanoma (n = 3), head and neck (n = 1), Erdheim–Chester disease (ECD) and Langerhans cell histiocytosis (LCH; n = 2), cholangiocarcinoma (n = 2), glioma (n = 1), cancer of unknown primary (n = 1) and multicancer cohorts not further specified (n = 3). The pooled ORR was 37% (27–49%) with 55% heterogeneity. The multitumor meta-analysis for tumor-specific trials included 44 cohorts and 4891 patients. The meta-analysis included melanoma cohorts (n = 39), thyroid cancer (n = 2) and NSCLC (n = 3). There were no tumor-specific trials in colorectal cancer that met the eligibility criteria and reported ORR. The pooled ORR was 50% (43–57%) with 90% heterogeneity. The difference between the pooled ORRs for tumor-agnostic versus tumor-specific trials did not reach significance (p = 0.05; Figure 3).

Pooled ORRs by tumor type

Thyroid cancer

There were three cohorts totaling 24 patients from tumor-agnostic trials that reported results only for thyroid cancer. The pooled ORR from thyroid cohorts from tumor-agnostic trials (n = 3) was 57% (12–93%) with 32% heterogeneity. For tumor-specific trials, the pooled ORR from thyroid cohorts (n = 2) was 33% (1–96%) with 0% heterogeneity. There was no significant difference in effect size between the two pooled ORRs (p = 0.10; Figure 4).

Non-small-cell lung cancer

There were four NSCLC cohorts from tumor-agnostic trials with a pooled ORR of 39% (25–55%) and 0% heterogeneity. In comparison, the pooled ORR from tumor-specific trials (n = 3) was 53% (19–84%) with 85% heterogeneity. There was no significant difference in effect size between the two pooled ORRs (p = 0.18; Figure 5).

Melanoma

The pooled ORR for melanoma cohorts from tumor-agnostic trials (n = 3) was 55% (24–83%) with 49% heterogeneity. Most cohorts from tumor-specific trials were melanoma cohorts (n = 39), with a pooled ORR of 51% (43–58%) and 91% heterogeneity. There was no significant difference in effect size between the two pooled ORRs (p = 0.58; Figure 6).

Pooled ORRs by intervention type

BRAF inhibition

Nine cohorts from tumor-agnostic trials received treatment with BRAF inhibitor monotherapy, specifically vemurafenib. Cohorts treated with vemurafenib included the following tumor types: ECD and LCH (n = 2), cholangiocarcinoma (n = 1), NSCLC (n = 3), thyroid cancer (n = 1), glioma (n = 1) and multitumor not otherwise specified (n = 1). The pooled ORR was 36% (29–45%) with 29% heterogeneity. Among tumor-specific trials, 17 cohorts comprised of melanoma (n = 14), NSCLC (n = 1) or thyroid tumor types (n = 2) had received treatment with BRAF inhibitor monotherapy, either vemurafenib, encorafenib or dabrafenib. The pooled ORR was 44% (36–52%) with 84% heterogeneity. There was no significant difference in the two pooled effect sizes (p = 0.17; Supplementary Figure 1).

BRAF + MEK inhibition

13 cohorts from tumor-agnostic trials received treatment with a combination of a BRAF and a MEK inhibitor, specifically one of the following regimens: vemurafenib plus cobimetinib, dabrafenib plus trametinib or encorafenib plus binimetinib. The tumor cohorts treated included colorectal (n = 2), head and neck (n = 1), NSCLC (n = 1), ovarian (n = 1), thyroid (n = 2), melanoma (n = 3), cholangiocarcinoma (n = 1), multitumor (n = 1) and unknown primary tumor (n = 1). The pooled ORR was 46% (29–63%) with 42% heterogeneity. There were 22 cohorts from tumor-specific trials, all melanoma, that were treated with either dabrafenib plus trametinib or vemurafenib plus cobimetinib. The pooled ORR was 61% (52–70%) with 87% heterogeneity. There was no significant difference in the two pooled effect sizes (p = 0.10; Supplementary Figure 2).

Discussion

To our knowledge, this is the first study to assess the comparability of tumor-agnostic and tumor-specific methodologies and outcomes, with implications for future trial design and targeted therapy development. To create a focused, less heterogenous comparison, we isolated cohorts of patients with BRAF alterations receiving treatment with TKIs, pooled their ORRs by trial design (tumor-specific vs tumor-agnostic), then applied stratifications for tumor type and intervention type to qualitatively understand how clinical results compare between the two trial design groups. We also aimed to understand the extent to which potential observed differences in clinical results can be attributed to trial design versus other clinical factors.

We captured a larger number of cohorts from tumor-specific trials than tumor-agnostic trials and with a larger number of patients. This is not surprising given the relatively recent advent of tumor-agnostic trials, and that a larger proportion of the tumor-specific trials were phase III. Comparisons of the two trial groups were unable to detect a statistically significant difference in pooled effect sizes for the comparison pooling across multiple tumor cohorts, or when stratified by tumor type and intervention types. Tumor-specific trials showed numerically higher ORRs for most meta-analyses, with the exception of those for thyroid cancer and melanoma. However, there was almost always more heterogeneity in the estimates from tumor-specific trials than tumor-agnostic trials, and often by a considerable margin. While this study limited its scope to BRAF-mutated cancers, this novel application of the SRMA methodology sets the stage for similar comparisons in other molecular cancer subtypes.

The results of these meta-analyses highlight important considerations in evaluating tumor-agnostic versus tumor-specific trial designs. Although both designs yielded consistent estimates of the outcome, there was typically less heterogeneity for meta-analyses of cohorts from tumor-agnostic trials than those of tumor-specific trials. This could be interpreted as more consistency in trial design, as cohorts of different tumor types are all exposed to the same overall trial design in a tumor-agnostic trial, while variability between trials conducted in separate tumor types introduces increased heterogeneity. Some of this heterogeneity was reduced when the meta-analyses were conducted within a single tumor type, suggesting some heterogeneity in the multitumor analyses is attributable to the range of tumor types included. However, if tumor type explained most of the heterogeneity observed, one would expect heterogeneity to be greater for the analysis of tumor-agnostic cohorts, which pooled at least 31 tumor types compared with just three tumor types from tumor-specific trials. Further, this pattern of lower heterogeneity across tumor-agnostic trials was also consistent within the meta-analyses for NSCLC and melanoma, increasing confidence that variability in trial design between tumor-specific trials is at least partly responsible for heterogeneity in corresponding effect sizes. Although differences in patient characteristics could create heterogeneity, the characteristics of patients in the two groups of trials were ultimately very similar: cohorts from both groups recruited patients of similar age, with comparable variability in trial location and, by extension, patient ethnicity. Both sets of trials included mostly patients with advanced cancer, with comparable proportions of males versus females. Both tested the same TKIs and combination therapies, mostly targeted to V600 mutations. Though trials showed differences in the range of tumor types that were recruited, with most tumor-specific trials being in melanoma relative to a more diverse mix from tumor-agnostic trials, results were still consistent when meta-analyses were stratified by tumor type.

While it has already been shown that not all genomic and immunological targets are created equally, and histological context will be more important in evaluating efficacy for some targets than others, it is not well understood how trial design itself may contribute to trial outcomes, particularly in this new age of tumor-agnostic trial design. While the current results indicated that tumor-specific trials may show slightly increased effect sizes, they also showed substantial heterogeneity across trials that was not observed to the same extent across tumor-agnostic trials, potentially attributable to inconsistent trial design. In comparison, tumor-agnostic trials offer the advantage of applying a single, consistent trial design to multiple tumor cohorts simultaneously, which may serve to increase confidence that variability in results between cohorts is due to true differences in efficacy rather than artifacts of trial design. Subsequently, a tumor-agnostic trial design need not imply a tumor-agnostic interpretation of results. As demonstrated in this analysis, effect sizes from both tumor-specific and tumor-agnostic studies may vary by tumor type, with some tumor types showing better efficacy than others despite a common mutation. However, it is with a tumor-agnostic trial design that one can best check for this complexity in certain molecular alterations. That is, the ability to compare efficacy across multiple tumor types with a common target, within a single trial rather than across multiple independent trials, may enable researchers to recognize the importance of histological context for a specific target much earlier and with greater confidence. This advantage goes beyond that of tumor-agnostic trials that have already been documented, including the ability to expand clinical trials and streamline regulatory approvals more efficiently, particularly for rarer cancers [1,3].

Limitations

This study has some limitations. First, the scope of this analysis was limited to statistical comparisons of separate meta-analyses conducted within the tumor-agnostic and tumor-specific trial groups, rather than tests of noninferiority or equivalence, limiting the ability to make definitive claims that results from the two trial designs were equivalent. However, the aim of this study was not necessarily to determine equivalence, but rather to explore the comparability of estimates produced from tumor-agnostic and tumor-specific trial designs, and perhaps develop hypotheses for further research. This analysis was also limited to ORR as it was the most uniformly reported variable across cohorts. However, ORR is not necessarily correlated with survival and thus may not be as important to patients and clinicians as outcomes such as PFS, OS or health-related quality of life. That said, the objective of this analysis was not to assess the efficacy of any given treatment regimen but rather to comment on the comparability of outcomes between trials of different designs and inform clinical trial design. While comparability of outcomes may differ for other survival-based outcomes and may represent an important area for future research, this analysis of ORR still provides a first look at the variability in outcomes between and across tumor-agnostic and tumor-specific trials.

The lack of statistical significance between pooled ORR estimates from tumor-agnostic and tumor-specific trials does not necessarily imply clinically meaningful equivalence. We recognize that the p-value depends on various factors such as sample size and contend that the results should be interpreted with ample consideration of these elements, (e.g., small sample size for some analyses, underpowered study). In the future, as tumor-agnostic trials become more common and bolster sample sizes for analysis, this will be an important area for future research. That is, will more favorably powered analyses, or alternative tests such as noninferiority and equivalence tests, show consistent results with those observed in this analysis?

Further, although we attempted to simulate control of factors other than trial design, either through stratified analysis or qualitative review of patient characteristics, it is still possible that between-study heterogeneity beyond overall trial design could have produced biased pooled results. However, we argue that the degree of heterogeneity is itself an important finding, as we saw notable patterns in the degree of heterogeneity across tumor-specific trials versus tumor-agnostic trials, and this may have implications for how the results of these trials can be generalized to the real world. Nonetheless, different baseline characteristics of patients and trial design could still contribute to different baseline probabilities of response, which may partly explain observed differences or lack thereof.

Lastly, this study was restricted to trials in BRAF-altered cancers, and findings on the comparability of outcomes between trial designs may not generalize to other populations. That said, the framework developed in this study provides a template for future comparisons of trial design in oncology clinical trials of targeted therapies, where the analysis could be further expanded to additional molecularly defined populations. The results of this analysis suggest it is plausible that ORRs yielded from tumor-agnostic trials may not be different from their tumor-specific counterparts. Future research that tests this comparison between trials of more comparable designs and patient populations would lend additional confidence to these conclusions.

Conclusion

The results of these meta-analyses indicate that, for BRAF-altered advanced cancers, tumor-agnostic clinical trials do not yield substantially different results from tumor-specific studies and may offer a more efficient method of studying efficacy across multiple tumor types simultaneously with less between-study heterogeneity.