Why FOCUS4 Used a Stratified Platform
Metastatic colorectal cancer is not a single molecularly uniform disease, and patients’ disease courses can differ after initial treatment. FOCUS4 was designed to connect molecular information with randomized treatment comparisons at a clinically meaningful point: after patients had received initial therapy and their disease had not progressed. Its architecture therefore had two linked features. Molecular results helped determine which treatment comparison a patient could enter, while randomization within an available comparison was used to assess treatment.
This timing matters. FOCUS4 did not randomize every screened patient at diagnosis to one of several platform treatments. Patients first received initial therapy; only those who met the relevant post-induction criteria could be considered for randomization. The trial’s conclusions consequently concern this eligible group, not all people with newly diagnosed metastatic colorectal cancer.
The FOCUS4 Patient Pathway
At a high level, FOCUS4 screened patients with metastatic colorectal cancer, obtained tumour molecular information, and treated eligible patients with initial therapy. Patients whose disease had not progressed after the planned initial-treatment period could then be assessed for a randomized maintenance or other post-induction comparison. Molecular findings and the trial modules open at that time affected which comparison was available.
The molecular assessment included markers used to define clinically relevant groups, such as RAS, BRAF, PIK3CA, and TP53 alterations, as well as mismatch-repair status. The precise route was governed by the protocol and the module’s eligibility rules; a marker result alone did not guarantee entry. A patient also had to satisfy clinical eligibility requirements, and the relevant comparison had to be recruiting.
FOCUS4 included molecularly directed modules as well as a non-stratified comparison, FOCUS4-N, for patients who did not enter a molecularly selected module under the platform’s rules. Examples of molecularly directed comparisons included FOCUS4-C for PIK3CA-mutant disease and FOCUS4-D for RAS/TP53-mutant disease. These examples illustrate the routing principle; they do not imply that every module was open simultaneously or that every patient in a marker group was eligible for every relevant comparison.
Identify patients with metastatic colorectal cancer and obtain the molecular information needed for classification. An unsuccessful or incomplete assay can prevent molecular routing, even if the patient is otherwise clinically suitable.
Patients receive initial treatment. At the planned post-induction point, assess disease status and clinical eligibility. Progression before this point generally means the patient does not enter a post-induction randomization.
Use molecular classification together with clinical criteria and module availability to identify the comparison or comparisons for which the patient qualifies.
Assign eligible patients randomly to the treatment options specified for that module. The randomized contrast, rather than a comparison between molecular groups, is the basis for a causal treatment estimate.
Worked Example: Following a Hypothetical Screening Cohort
Suppose, purely for illustration, 240 patients are screened. Molecular results are usable for 204, and 132 of those 204 have disease that has not progressed after initial therapy and meet the clinical criteria for post-induction assessment. The following classification is hypothetical and is used only to show how the flow can be summarized.
The proportion with usable profiling is \(204/240=0.850\), or 85.0%. Among patients with usable profiling, the proportion who reach the eligible post-induction group is \(132/204=0.647\), or 64.7%. The latter denominator is not all screened patients: it is the group with usable profiling.
Suppose the 132 eligible patients consist of 30 assigned to an available PIK3CA-related comparison, 24 to an available RAS/TP53-related comparison, 18 to another available molecularly directed comparison, and 60 to the non-stratified route. The counts sum to \(30+24+18+60=132\), so this schematic accounts for the full eligible group.
These are not reported FOCUS4 accrual figures. They show why platform flow should be reported with explicit denominators: screening, successful profiling, post-induction eligibility, and entry into each treatment comparison are distinct stages.
What Is Randomized—and What Is Not
The key inferential distinction is between classification and allocation. A molecular result classifies a patient for routing. Clinical status after initial therapy determines whether the patient can proceed. Randomization then assigns treatment among patients eligible for a particular comparison. Patients in different molecular groups are not randomized to their molecular status, so differences in their outcomes cannot by themselves establish that one molecular group benefits more than another.
Let \(E=1\) indicate that a patient meets the post-induction eligibility criteria, \(M=m\) indicate membership in molecular group \(m\), and \(A\) indicate randomized treatment within an eligible module. Let \(Y(1)\) and \(Y(0)\) denote the outcomes that the patient would have under the two randomized options. If randomization is implemented within the relevant eligible group, then, in expectation, treatment assignment is independent of these potential outcomes conditional on eligibility and group:
Thus, the randomized comparison estimates a treatment effect for the eligible patients in that comparison. For a simple mean outcome, the justification can be written as follows. By randomization, the mean potential outcome under treatment is the same in the observed treatment arm as in the full eligible group; the same holds for control. Therefore:
This is a conditional treatment effect: it applies to eligible patients in molecular group \(m\), under the module’s criteria and trial context. It does not automatically apply to patients whose disease progressed during initial treatment, patients with missing molecular results, or all patients with metastatic colorectal cancer. Nor does the equation make a comparison between two molecular groups causal.
Worked Example: Interpreting Randomization Within a Molecular Group
Imagine a hypothetical module in which 44 eligible patients with a specified molecular profile are randomized equally: 22 to an experimental option and 22 to its control. If the planned primary outcome is progression-free survival, the randomized contrast concerns outcomes for this eligible molecular group at the post-induction point.
Equal allocation gives \(22/44=0.500\), or 50.0%, in each arm. This balance is a design feature, not evidence that the treatments have equal effects. The treatment effect must be estimated from the outcome data using the prespecified analysis.
If a separate molecular group has a different observed outcome, that difference alone does not show that the treatment works differently between groups. A treatment-by-group interaction analysis, with appropriate design and uncertainty assessment, would be needed to support such a claim. Randomization within each group protects the treatment comparison there; it does not randomize patients between groups.
How to Read a Module Result in Context
A treatment estimate should be read against the exact comparison and target population that produced it. The population is shaped by screening, assay success, initial-treatment course, post-induction disease status, clinical eligibility, molecular routing, and module availability. The comparator and endpoint are defined for that module; they should not be assumed to be identical across every platform comparison.
For example, suppose a hypothetical module reports a progression-free survival hazard ratio of 0.72 with a 95% confidence interval from 0.50 to 1.04. The point estimate indicates an estimated 28% lower hazard in the experimental arm over follow-up, under the hazard-ratio interpretation. Because the interval includes 1, it is compatible with no hazard difference as well as with effects in either direction covered by the interval. It does not establish a 28% increase in survival time, and it is not a confidence interval for the effect in every screened patient.
Platform operation can also change over time. Modules may open or close as evidence and recruitment develop. Consequently, the set of comparisons available to a patient can depend on when and where screening occurs. Such changes are part of the platform context and should be considered when describing accrual or comparing module populations. They do not invalidate a randomized comparison automatically, but the analysis and interpretation must respect the actual randomization and eligibility rules.
Worked Example: Defining the Population Behind a Hypothetical Estimate
Suppose a hypothetical module reports a hazard ratio of 0.72 (95% confidence interval 0.50 to 1.04) for its experimental option versus control. Assume the endpoint is progression-free survival and the comparison enrolled patients who had not progressed after initial therapy and met the module’s molecular and clinical criteria.
The point estimate is below 1, so the estimated hazard of progression or death is lower in the experimental arm under the model used. The interval includes 1, so these data do not rule out no hazard difference at the conventional two-sided 5% level. The estimate is for patients eligible for that module, not for everyone screened or for patients who progressed during initial therapy.
The appropriate summary is therefore conditional: in this hypothetical eligible post-induction group, the estimated hazard ratio favored the experimental option, but the confidence interval includes no difference. The result does not establish a benefit in other molecular groups or in the broader diagnosis-time population.
Comparing the Architecture With Alternatives
| Approach | How patients enter comparisons | Key distinction from FOCUS4’s architecture |
|---|---|---|
| Conventional single-population randomized trial | Eligible patients enter one treatment comparison, commonly without molecular routing. | May not organize several molecularly defined comparisons within a platform. |
| FOCUS4-style stratified platform | Patients are screened and assessed after initial therapy; molecular and clinical criteria route eligible patients to available comparisons. | Connects molecular classification and a post-induction randomization point to multiple treatment modules. |
| Observational comparison across molecular groups | Patients receive or are classified by characteristics without randomized treatment allocation across the groups being compared. | Differences between groups can reflect prognosis or other factors; they are not a randomized treatment effect. |
The platform can make efficient use of screening and support several related randomized questions. However, the existence of a common platform does not mean that all modules share a control group, use identical endpoints, or form one interchangeable treatment comparison. Those details must be checked in each module’s protocol and analysis.
Diagnostics, Limitations, and Common Pitfalls
A sound interpretation begins by checking the flow. Ask how many patients were screened, how many had usable molecular results, how many remained eligible after initial therapy, and how many entered each module. Large losses at any stage affect generalizability and may reveal practical constraints, such as assay failure, progression before randomization, or a module not being open. These counts describe selection; they do not, on their own, show that the randomized comparison is biased.
Also check that the treatment comparison is defined by randomization within the eligible module. A comparison of outcomes between a molecular subgroup that entered one module and a different subgroup that entered another is not a substitute. Differences may reflect biology, prognosis, eligibility, calendar time, or treatment options as well as treatment effects.
- Wrong reasoning: “The patient has the target mutation, so the patient was randomized to the molecular treatment.” Correction: The result supports routing or eligibility; randomization occurs only after the patient meets the module’s criteria and is assigned among its treatment options.
- Wrong reasoning: “The module result applies to every patient with metastatic colorectal cancer.” Correction: It applies most directly to the randomized population that met the post-induction and module-specific criteria.
- Wrong reasoning: “One module’s result can be compared directly with another module’s result to rank treatments.” Correction: Cross-module differences may involve different patients, times, comparators, and designs. A valid comparative claim requires an appropriate randomized contrast or a justified comparative analysis.
- Wrong reasoning: “A shared platform guarantees one shared control and one overall type I error rate.” Correction: Control arrangements and multiplicity procedures must be established from the relevant protocol; they cannot be inferred from the platform label.
The design’s principal limitation for interpretation is its conditional target population. Patients who do not reach post-induction randomization are not represented by that randomized comparison. In addition, molecular classification depends on adequate testing, and changing module availability can shape which patients are enrolled in which comparisons. These features do not negate the value of the platform, but they make transparent reporting of eligibility, assay success, timing, and module-specific methods essential.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Why does a FOCUS4-style post-induction randomized comparison not directly estimate a treatment effect for all patients screened at diagnosis?
- In a hypothetical cohort, 180 patients have usable molecular results and 117 meet post-induction clinical criteria. What proportion of those with usable results meet the criteria?
- Why is a difference in outcomes between two molecular groups not, by itself, evidence that a treatment works differently in those groups?
- A hypothetical module reports a hazard ratio of 0.80 with a 95% confidence interval of 0.62 to 1.03. State a careful interpretation of the estimate and interval.
Solutions. (1) Randomization occurs only after initial therapy and eligibility assessment. Patients who progress, lack usable profiling, or otherwise fail the criteria are outside the randomized module population, so the estimate is conditional on reaching that point. (2) The required denominator is the 180 patients with usable results: \(117/180=0.650\), or 65.0%. (3) Molecular group is not randomized. Group differences may reflect prognostic and other characteristics; evidence of different treatment effects requires an appropriate treatment-by-group comparison. (4) The point estimate suggests a 20% lower hazard in the experimental arm under the hazard-ratio interpretation. The interval includes 1, so it is compatible with no hazard difference at the conventional two-sided 5% level; the result is not proof of benefit and applies to the module’s eligible randomized population.