Evidence grades (GRADE-adapted): A high — multiple well-conducted RCTs or systematic reviews converge. B moderate — single pivotal RCT or consistent observational evidence. C limited — single observational study, mechanistic, or expert consensus. D preclinical / hypothesis-generating.
The biomarker-stratification trade-off
Modern targeted therapies often have substantial benefit in biomarker-positive subgroups and minimal benefit in biomarker-negative subgroups. Trial design must balance:
- Enrichment for the biomarker-positive subgroup — smaller trial size, faster readout, higher likelihood of positive result; but limited evidence about biomarker-negative patients
- All-comer enrollment with biomarker stratification — broader evidence base, larger trial, slower readout; but risk of diluted signal
- Hierarchical analysis — pre-specified hierarchy of biomarker-positive subgroup, then broader subgroups, then ITT, with strict alpha control
Each approach has been used successfully in TNBC, and each has produced failures. The right choice depends on the biological hypothesis, biomarker characterization quality, and the size of the expected treatment effect in biomarker-positive vs negative patients.
Enrichment trial design
Enrichment trials enroll only biomarker-positive patients. Examples in TNBC:
- OlympiAD (olaparib in germline-BRCA-mutated metastatic breast cancer) — enrolled only germline BRCA1/2 carriers. Phase III primary endpoint analysis was therefore in the biomarker-positive population by design. Result: positive PFS endpoint, regulatory approval.
- EMBRACA (talazoparib, similar enrichment) — same biomarker-enriched design. Positive result, approval.
- DESTINY-Breast04 (T-DXd in HER2-low) — enriched for HER2-low (IHC 1+ or 2+/ISH-) breast cancer. Positive result, approval.
Advantages: simpler design, statistical power concentrated on the biomarker-positive subgroup, smaller required trial size. Limitations: no evidence about biomarker-negative patients; the biomarker definition must be operationalized correctly at the trial outset.
Stratification with hierarchical analysis
Stratification trials enroll all-comers but pre-specify hierarchical analyses by biomarker subgroup. Examples in TNBC:
- KEYNOTE-355 — enrolled all-comer metastatic TNBC; pre-specified analyses tested PFS at CPS ≥ 1, CPS ≥ 10, and ITT in hierarchical order. The CPS ≥ 10 subgroup was positive; alpha was allocated to allow formal testing of broader subgroups only if the more restricted one was positive. Result: regulatory approval anchored on the CPS ≥ 10 subgroup analysis.
- IMpassion130 — similar design with hierarchical analysis at SP142 IC ≥ 1% subgroup, then ITT. CPS ≥ 1% subgroup was positive; ITT was not. Accelerated approval was granted on the subgroup analysis.
Advantages: broader evidence base; the biomarker can be evaluated post hoc against alternative definitions. Limitations: larger trial size; risk of biomarker subgroup being underpowered if the hypothesis is wrong.
Basket trials
Basket trials enroll patients with a specific biomarker across multiple cancer histologies (e.g., NTRK fusion patients across all tumor types). The basic question: does this targeted therapy work in this biomarker subset regardless of histology? Examples relevant to TNBC:
- Larotrectinib and entrectinib NTRK basket trials — supported tissue-agnostic FDA approval. TNBC patients with NTRK fusions (rare) are eligible for these approved therapies.
- Pembrolizumab MSI-high basket (KEYNOTE-158) — supported tissue-agnostic approval for MSI-high tumors. TNBC patients with MSI-high disease (rare) are eligible.
- Pembrolizumab TMB-high basket — supported tissue-agnostic approval for TMB ≥ 10 mutations/Mb tumors. ~5–10% of TNBC patients qualify.
In TNBC specifically, basket-trial-supported approvals provide treatment options for the small biomarker-positive subsets that wouldn't have been adequately tested in TNBC-only trials.
Umbrella trials
Umbrella trials enroll patients with a single tumor histology and assign to treatment based on biomarker profile within that histology. The basic question: which biomarker-matched targeted therapy works best for each subset of this disease? Examples in TNBC:
- FUTURE series (FUSCC) — umbrella trial in refractory metastatic TNBC, with patients assigned to biomarker-matched targeted therapy arms based on FUSCC molecular subtype.
- I-SPY 2.2 — can be viewed as an umbrella platform trial in early-stage breast cancer, with multiple investigational arms tested in subtype-matched populations.
Master protocols
Master protocols are over-arching trial frameworks that can include basket, umbrella, or platform designs under a unified administrative structure. The advantage: shared infrastructure, statistical methodology, and regulatory documentation across multiple investigational sub-studies. NCI-MATCH was an early oncology master protocol; multiple commercial sponsors are now developing master protocols for specific tumor types or indications.
The IMpassion130 / IMpassion131 cautionary tale
The atezolizumab program in metastatic TNBC illustrates the risks of biomarker-stratified design:
- IMpassion130 enrolled all-comer metastatic TNBC; pre-specified PD-L1 SP142 IC ≥ 1% subgroup analysis was positive for PFS (HR 0.62) and supportively positive for OS.
- FDA granted accelerated approval based on the subgroup analysis in 2019.
- Confirmatory IMpassion131 (with paclitaxel rather than nab-paclitaxel backbone) failed the same biomarker-positive subgroup endpoint — PFS HR 0.82 (p=0.20) and numerically worse OS.
- Atezolizumab indication for TNBC was voluntarily withdrawn in 2021.
The lessons: biomarker-stratified subgroup analyses can produce false-positive results that fail to replicate; chemotherapy-backbone choices interact with biomarker analyses in ways that pre-specification can miss; confirmatory trial requirements that the FDA imposed worked exactly as intended in protecting patients from a treatment whose benefit was uncertain.
Statistical considerations
Biomarker-stratified design requires careful statistical methodology:
- Pre-specified hierarchical alpha allocation. If multiple subgroup analyses are tested, alpha must be allocated to control overall type-I error.
- Sample size calculations. Trial size must be sufficient for the biomarker-positive subgroup primary analysis, which typically requires more enrollment than an all-comer design seeking a smaller effect size.
- Stratified randomization. Biomarker status should be a stratification factor at randomization to ensure balanced treatment assignment within subgroups.
- Biomarker assay performance. The biomarker measurement must be reproducible and validated before trial enrollment; assay change mid-trial complicates interpretation substantially.
- Multiplicity adjustment for biomarker subgroups. If multiple biomarker thresholds (e.g., CPS 1, 10, 20) are tested, appropriate adjustment is needed.
Evidence table
| Trial | Design | Biomarker | Outcome |
|---|---|---|---|
| OlympiAD | Enrichment | Germline BRCA1/2 | Positive; approved |
| EMBRACA | Enrichment | Germline BRCA1/2 | Positive; approved |
| DESTINY-Breast04 | Enrichment | HER2-low | Positive; approved |
| KEYNOTE-355 | Stratification + hierarchy | PD-L1 CPS | Positive at CPS ≥ 10; approved |
| IMpassion130 | Stratification + hierarchy | PD-L1 SP142 IC | Positive subgroup; later withdrawn |
| IMpassion131 | Same biomarker, different chemo backbone | Same as IMpassion130 | Negative; withdrawal trigger |
| Larotrectinib NTRK basket | Basket | NTRK fusion | Positive; tissue-agnostic approval |
| FUTURE series | Umbrella | FUSCC subtype | Higher response than historical controls |
Open questions and active investigation
- How to integrate emerging biomarkers into trial design. New biomarkers (TMB, ctDNA-MRD status, immune signatures) emerge during ongoing trials; pre-specifying biomarker analyses for trials that take 5–10 years to complete is challenging.
- Multiple-biomarker stratification. If multiple biomarkers (PD-L1, BRCA, HER2-low) each define actionable subgroups, trials may need to stratify on multiple axes. Complexity grows rapidly; sample size requirements scale.
- Adaptive biomarker refinement. Can trials adapt biomarker definitions based on emerging data? Statistical methodology for valid adaptive biomarker refinement is developing.
- Real-world evidence integration. Can post-approval real-world data on biomarker-stratified outcomes complement trial evidence? FDA has begun accepting RWE for some indications.
- Liquid biopsy biomarkers. ctDNA-based biomarkers offer practical advantages (less tissue requirement, possibility of serial monitoring) but have technical performance considerations that affect trial-design choices.
- Single-arm biomarker enrichment. For very rare biomarker-positive populations, single-arm trials may be the only feasible design; methodology for accepting single-arm evidence is being refined.
For specific biomarker-stratified TNBC trials, see the first-line metastatic synthesis (KEYNOTE-355, OlympiAD, EMBRACA), the DESTINY-Breast04 synthesis, and the IMpassion/KEYNOTE umbrella synthesis. For pCR endpoint considerations relevant to enrichment-trial readouts, see the endpoint design synthesis.
References
Each citation links to the original publication via DOI. The same records are searchable in the evidence library by title or DOI.
- Woodcock J, LaVange LM. Master Protocols to Study Multiple Therapies, Multiple Diseases, or Both. N Engl J Med. 2017;377(1):62–70. doi:10.1056/NEJMra1510062. ↩
- Park JJH, Siden E, Zoratti MJ, et al. Systematic review of basket trials, umbrella trials, and platform trials. Trials. 2019;20(1):572. doi:10.1186/s13063-019-3664-1. ↩
Last reviewed: 2026-06-04. Researcher-layer synthesis page. Evidence grades follow the GRADE-adapted rubric defined at the top of this page.