The Substudies You Never Hear About: Why Most ComboMATCH Protocols Die Before They Reach a Patient

ComboMATCH is described as having opened with multiple treatment trials and more in development. The design paper lists the scientific rationale, the biomarker-matching infrastructure, and the cooperative group coordination. Press releases count the substudies that activated. The ECOG-ACRIN treatment trials page lists the ones currently open or closed to accrual.

What none of these sources show is the graveyard.

For every ComboMATCH substudy that reaches a patient, others were proposed, developed, reviewed, and killed at various stages of the pipeline. Some died at concept. Some made it to full protocol drafts with complete statistical sections, eligibility criteria, and drug information, then never activated. The attrition is invisible to the public because dead substudies don't get ClinicalTrials.gov entries, don't appear on treatment trial listings, and aren't discussed in design papers.

I worked on substudies at SWOG that spanned this lifecycle. One completed accrual. Others didn't make it. The reasons they didn't are worth understanding, because they reveal something about platform trial design that the success stories don't capture.

The pipeline a substudy has to survive

Before a ComboMATCH substudy can treat a single patient, it must pass through a multi-stage review process involving at least nine distinct entities. The concept is first vetted by the Agents and Genes Working Group, which evaluates scientific rationale, and the Sub-Protocol Design Development Working Group, which reviews the statistical design. Biomarker feasibility is assessed by the Molecular Biomarker and Specimen Management Committee. The Precision Medicine Analysis and Coordination Committee evaluates whether the matching algorithm can identify and assign patients. The Steering Committee grants LOI approval.

Only then does internal protocol development begin. The cooperative group writes the full protocol. More committee reviews follow. The pharmaceutical partner must formally commit to drug supply. CTEP reviews the protocol for scientific and regulatory adequacy, which is its own multi-round negotiation. Central IRB review. CTEP monitors. Final activation.

This is not a fast process. Concept to activation takes a year or more under favorable conditions. And at every stage, a substudy can be killed, shelved, or sent back for redesign.

The pipeline exists for good reasons. Platform trials that activate bad substudies waste screening slots, consume drug supply, and expose patients to combinations that shouldn't have made it past review. The committees are the immune system. But the immune system has consequences for the substudies trying to get through it.

Three ways to die

I'll describe three substudies that illustrate distinct failure modes. One completed accrual. One was terminated before it treated a patient. One never made it past the concept stage. The details are drawn from protocol documents and public records.

The one that made it. A single-arm substudy targeting a rare molecular alteration in non-breast solid tumors. Thirty evaluable patients, two-stage design, objective response rate as the primary endpoint. The concept was approved with specific operating characteristics: a null response rate of 20%, an alternative of 40%, 82% power, one-sided alpha of 6%. Positive if 10 or more of 30 patients responded.

Then the protocol was submitted, and the statistical section didn't match the concept. The null had shifted to 15%, the alternative to 35%, the power to 90%, the alpha to 10%. The positivity threshold had dropped to 8 out of 30, which is 26.6%. Nobody made a deliberate decision to change the design. The parameters drifted during protocol writing, the way they do when a team is iterating on a document over months.

CTEP caught it. They also noticed that the existing single-agent response rate for a related drug in the same molecular population was 29%, which was above the 26.6% positivity threshold. A substudy designed to detect combination activity would have declared success at a rate the single agent had already exceeded. The team corrected back to the approved concept parameters. Then CTEP recalculated the power and got 87%. The team recalculated and got 82%. The final protocol matched the concept: 20%/40%, 82% power, 6% alpha, 10 out of 30.

Separately, CTEP challenged the scientific rationale. Two Phase III trials of a related drug combined with chemotherapy in a related population had been negative. Why should this combination work? The investigators' response was precise: those trials selected patients on broader pathway alterations, not the specific activating mutations this substudy required. The protocol also included a kill switch: if subgroup data from one of the negative trials showed no activity in patients with the specific mutation, the substudy would be reconsidered.

The substudy activated in 2023 and has since closed to accrual. Even this "success" took roughly two years from concept to first patient. The CTEP review document runs to dozens of pages. What looks in retrospect like a clean path was anything but.

The one that died in development. A three-cohort substudy proposed for a well-known oncogenic driver mutation. Two randomized cohorts (one restricted to a single tumor type, one open to all others) and one single-arm cohort for patients who had already been treated with the same drug class. Total sample size: 150 patients. The design was ambitious: three separate statistical analyses, different endpoints across cohorts, a crossover mechanism that allowed patients from other ComboMATCH substudies to enroll in the single-arm cohort.

The first protocol draft was completed in mid-2021, with ComboMATCH's opening still more than a year away. The document had placeholder fields throughout: section numbers marked with X, no statistician named yet, multiple TBD entries. It was clearly early-stage.

By the time ComboMATCH opened for enrollment in late 2022, the therapeutic landscape for this mutation had transformed. A competing drug in the same class had received FDA approval. The original drug chosen for the substudy also gained accelerated approval. The clinical question the substudy was designed to answer hadn't become irrelevant, but the competitive environment for accrual had fundamentally changed. A different substudy, led by a different cooperative group, using a different drug combination for the same mutation, took its place in the active portfolio.

The terminated substudy left no public trace. No ClinicalTrials.gov entry. No NCT number. No mention on any treatment trial listing, even as "closed." If you weren't involved in its development, you wouldn't know it existed.

The one that never left concept. A randomized substudy targeting the intersection of two pathway alterations in two rare histologies. Cohort 1 required a specific receptor mutation or fusion plus a concurrent mutation in a second pathway in one cancer type. Cohort 2 required the same receptor alteration in gynecologic cancers, with at least half the patients also carrying the second mutation. Two cohorts, 70 patients each, 154 total accrual.

The science was elegant. Preclinical data supported the combination. The resistance mechanism was well-characterized. The concept form included a full statistical design: 85% power, one-sided alpha 0.10, hazard ratio 0.55, Wieand-type futility monitoring, same accrual kill switches as every other substudy.

The problem was arithmetic. The concept required finding 154 patients whose tumors harbored a specific receptor alteration and, in at least one cohort, a concurrent second mutation, in specific rare histologies, among patients who had progressed on prior therapy and were eligible for all the standard ComboMATCH criteria. Each of those conditions narrows the funnel. The intersection of all of them produces a population that may simply not exist at the rate needed to power a randomized trial through cooperative group accrual.

The concept projected 2-3 patients per cohort per month. For 77 patients per cohort (accounting for ineligibility), that's 26-39 months of accrual per cohort. In a platform where the accrual kill switch fires at 35% of projected enrollment after one year, a substudy that misses its accrual target by even a modest margin gets flagged for termination at 12 months. The margin for error was essentially zero.

The concept was submitted in 2021. It never advanced to protocol. No protocol document was drafted. No statistician was assigned. Its designation was eventually reused for a different substudy led by a different group targeting a different mutation.

What the dead substudies teach

These three outcomes represent three distinct failure modes, and none of them reflect bad science or bad design.

The substudy that died in development was killed by timeline. The pipeline from concept to activation takes long enough that the drug development landscape can shift underneath you while you're traversing it. An approval, a competitor's Phase III readout, a change in standard of care, any of these can make your substudy's clinical question less compelling or your accrual projections less realistic. The substudy's design was sound. The question it asked was important. The world moved.

The substudy that died at concept was killed by prevalence arithmetic. Dual-biomarker eligibility in rare histologies is scientifically justified when resistance mechanisms involve pathway crosstalk. But the number of patients who meet all the criteria, at a cooperative group site, with progressive disease, willing to enroll, may not support the sample size the design requires. The concept form's own accrual projection, 2-3 patients per month, was likely optimistic for a dual-biomarker population in uncommon cancers.

The substudy that made it threaded a narrow needle: a single biomarker in a molecularly defined population, a single-arm design requiring only 33 patients, and an accrual timeline that stayed ahead of the kill switches. Even it nearly had its scientific rationale rejected on the basis of negative trials in a related but distinct population.

The attrition is the feature

It's tempting to read substudy attrition as waste. All that protocol development, all those committee reviews, all those hours of statistical design work, for substudies that never treated a patient.

But the attrition is a feature of good platform design, not a failure. The pipeline is supposed to kill substudies that can't accrue, that target populations too rare to study in a cooperative group setting, or whose clinical questions have been overtaken by the pace of drug development. Every slot in the platform that goes to a substudy that can't complete enrollment is a slot that doesn't go to one that can. The committees are performing triage.

What gets lost in the triage is the cost. Not financial cost, though that's real, but the intellectual and human cost. Months of protocol development by investigators who believed in the science. Statistical designs that were carefully constructed, reviewed, and negotiated, then filed away when the substudy was killed. Clinical questions that were worth asking but couldn't be answered in this particular vehicle.

The public sees ComboMATCH as a set of active treatment trials. The people who work on it see a much larger landscape: the active trials, the ones in development, and the ones that didn't make it. The last category is the largest. Understanding why substudies die, and accepting that most of them should, is part of understanding what platform trials actually are.

They are not machines for running studies. They are machines for deciding which studies to run.

Many evidentiary problems appear during trial design, not after the analysis. I work with teams to review trial designs and run simulation studies to evaluate operating characteristics before protocols are finalized.

For consulting inquiries: maggie@zetyra.com
For more essays on statistical design and regulatory evidence, subscribe to the Evidence in the Wild newsletter.