Our protocol was prospectively registered in PROSPERO (CRD42023406401). We adapted and expanded methods from our previous analyses of risk and benefit in oncology clinical trials [11, 19, 20]. We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 reporting guidelines (see the PRISMA checklist in the Electronic Supplementary Material [ESM]) [21].

We included original reports (abstracts, articles, and summary reports on ClinicalTrials.gov) with results of interventional clinical trials with a basket design in oncology. Specifically, we looked for trials recruiting adult patients with different types of cancer (at least two cancer types) that tested an intervention based on the molecular profiling of cancer, rather than the location or histology of the cancer. We included both matched studies in which a mutation in a specific biomarker was confirmed and non-matched studies, i.e., studies that confirmed a lack of presence of a mutation in a specific biomarker. Both single-arm and multi-arm trials of all phases that met our criteria were included. We also considered parts of complex trials, for example, platform studies or multi-stage trials, where only a selected study part met the criteria for a basket trial.

Both solid tumors and hematological malignancies were eligible. In terms of interventions, we included targeted therapy, immunotherapy, or combinations of these, evaluated for cancer treatment. We had intended to include studies testing chemotherapy, but none of these studies met the other eligibility criteria. We considered only studies that reported results on treatment-related toxicity (e.g., number of patients with grade 3 and/or 4 and/or 5 treatment-related AEs) and/or treatment response (e.g., objective response rate, number of patients with partial responses).

We excluded studies that recruited pediatric populations, healthy volunteers, or patients with only one cancer type. We also excluded trials without a basket design and studies that did not perform molecular profiling to find a specific molecular alteration in tumors. The full eligibility criteria relevant to the Population, Intervention, Comparator, Outcomes and Study Type (PICOS) framework are listed in Table 1 of the ESM.

We systematically searched Embase, PubMed, and ClinicalTrials.gov for interventional cancer BCTs published between 1 January, 2001 and 14 June, 2023 (see our search strategies in Table 2 of the ESM). We did not apply language restrictions. We checked our strategies using the Canadian Agency for Drugs and Technologies in Health peer-review checklist for search strategies [22].

We performed additional searches of conference proceedings found in the records retrieved from Embase and PubMed. We also searched for other eligible trials in the reference lists of included articles and relevant review papers. For all studies included in the full-text screening, we searched for the ClinicalTrials.gov ID (NCT number) to check the trial description and results on ClinicalTrials.gov.

We selected trials in a stepwise process. Three review authors (KK, KS, LZ) screened the records for the initial study inclusion. Then, six reviewers (KK, KS, LZ, AW, TK, UBK) performed the screening of full texts and checked other resources (conference proceedings, reference lists, and records on ClinicalTrials.gov). At each step, the records were merged, and duplicates removed. Each study report was assessed for inclusion by two researchers independently. Disagreements were resolved by discussion and, if necessary, a third person, an arbiter (MW), was involved.

The specific unit for data extraction was an “arm” of a BCT. The arm referred to each subgroup of patients with different cancer types who received a specific intervention (or combination therapy) that was matched or not matched (non-match arm) to the specific biomarker (or group of biomarkers). A basket trial could have more than one arm (e.g., more than one intervention matched to a biomarker, or a specific intervention matched to more than one specific biomarker, or both). For the purposes of this article, the term “arm” is used interchangeably with “study” when reporting results.

The information provided on ClinicalTrials.gov supported data extraction. Each arm of the basket trial could have a separate registration and individual NCT number in the ClinicalTrials.gov registry, or there was one registration number for all arms of a basket trial. We extracted data from each arm separately (regardless of the registration type).

For each basket study arm, we merged information from all available sources (abstract, article and/or record(s) on ClinicalTrials.gov). As each source could report different results, we extracted data from (whichever came first): (a) the latest or the final report or (b) we chose the source with results reported for the highest number of patients or (c) with the most detailed information about the outcomes of interest.

We created and piloted our data extraction form. Its final version is available from the Open Science Framework, https://osf.io/w4ke9/. For each study arm, we extracted data related to study characteristics (e.g., phase, funding, location, study status), patient characteristics (e.g., number of enrolled participants, age, cancer type and stage of disease), intervention (e.g., agent names, type of therapy), and outcomes (e.g., objective responses, drug-related grade 3–5 AEs). Data were extracted independently by experienced reviewers working in pairs (KK, KS, LZ, AW, TK, or UBK). We resolved the discrepancies by discussion and, if necessary, an arbiter (MW) was involved. An experienced oncologist had a supervisory role (SVW).

We measured surrogate benefit using the objective response rate (ORR). We defined the ORR as the proportion of participants with a partial or complete response as defined by the study authors. We calculated the ORR as the sum of both partial and complete responses, or the ORR was reported directly in the study. For other benefit measures, we collected median progression-free survival (PFS) and overall survival (OS) data.

To assess risk, we used data on grade 3, 4, or 5 drug-related AEs as defined by the Common Terminology Criteria for Adverse Events version 5.0 (or earlier versions) [23]. We assessed separately the number of treatment-related AEs and the proportion of patients who experienced them. We considered an AE as related to the study drug if it was clearly stated by the study authors. Expressions such as “AEs at least possibly related to study therapy” or “AEs suspected to be drug-related” or “AEs attributed to treatment” or “AEs possibly or probably related to study drug” were also acceptable. If an AE was not clearly described as treatment related, we excluded it from the risk analysis.

We assessed the risk of bias (RoB) for each basket study arm with the results published in the article/s. We did not perform RoB assessments for other data sources (abstracts or summary of results in the ClinicalTrials.gov registry) as they usually provide limited information to conduct comprehensive RoB. We adapted the Cochrane RoB tools for non-randomized trials as all included studies were non-randomized [24]. Two researchers (KK, KS, LZ, AW, TK, or UBK) independently performed RoB directly after data extraction. Whenever any discrepancies occurred, we discussed and resolved them.

We calculated rates of objective response, rates of grade 5 (death), and grade 3–4 drug-related AEs by dividing the number of patients who experienced the outcome(s) by the total number of patients evaluated for response or toxicity in this arm. We estimated pooled rates by a meta-analysis for proportion. We performed a meta-analysis when results for a particular outcome were available from at least two studies. We used random-effects modeling and the restricted maximum likelihood estimator to account for heterogeneity between studies. We calculated the I2 statistic to provide a measure of the proportion of overall variation that is due to heterogeneity between studies. Differences in response rates and grades 5 and 3–4 drug-related AEs rates between the categories of phase, type of therapy, number of drugs evaluated, cancer types, and funding were assessed using the Q test for heterogeneity in meta-regression. We present results as rates with a 95% confidence interval (CI) in each category and a P value from the Q test for heterogeneity in meta-regression. Unweighted median with 95% CI was calculated for PFS and OS by bootstrap methods using the “boot” package in the R software [25] and compared between categories using the Mann–Whitney test or Kruskal–Wallis test. We performed a meta-analysis using the metafor package (R Version 4.3.2). P < 0.05 was considered statistically significant. All tests were two-sided.