What Are Platform Trials and How Do They Work?

A platform trial is a clinical trial designed to test multiple treatments under one permanent structure, rather than running a separate trial for each drug. Treatments can enter or leave the trial as data accumulate, all compared against a single shared control group that runs continuously. The design gained worldwide attention during the COVID-19 pandemic, when trials like RECOVERY produced life-saving answers in weeks instead of years, but platform trials were already being used in cancer and other diseases well before 2020. What makes them genuinely different from a standard trial is not just efficiency but the underlying logic: the trial itself is the infrastructure, and the individual treatments are temporary guests.

How a Platform Trial Differs From a Standard Trial

In a conventional clinical trial, researchers pick one experimental treatment, design a protocol around it, recruit patients, compare the treatment to a control, and publish the results. If they want to test a second treatment, they start over: new protocol, new ethics approval, new recruitment. Each trial is a standalone project with its own timeline and its own control group.

A platform trial flips that model. The trial is a standing structure with a single protocol, a shared control arm, and room for multiple experimental arms running at the same time. The multi-arm element means several treatments can be tested simultaneously against the same control group, while interim analyses allow the trial to assess accumulating data and ensure only treatments showing promise continue recruiting.

This matters because the slowest, most expensive part of any clinical trial is the infrastructure: setting up sites, getting regulatory approvals, recruiting patients, and training staff. A platform trial pays that cost once and then reuses the infrastructure for every new treatment that enters. When one experimental arm finishes or gets dropped, a new treatment can slot into the existing framework without starting from scratch.

The Mechanics of Adding and Dropping Treatments

The adaptive part of platform trials is what gives them their speed. At scheduled interim analyses, a data monitoring committee reviews how each experimental arm is performing relative to the control. Arms that look ineffective get dropped, freeing up patients and resources for more promising candidates. Arms that meet pre-specified success criteria can “graduate” to confirmatory testing or regulatory submission. And new arms can enter the trial at any point, inheriting the existing infrastructure and control group.

Some platform trials go further and use a technique called response-adaptive randomization. Instead of assigning patients to treatment arms with equal probability, the randomization shifts over time so that more patients are assigned to arms that appear to be working better. The I-SPY 2 breast cancer trial, for instance, uses clinical biomarkers to classify breast cancer into subtypes and then uses adaptive randomization to steer patients toward the arms most likely to benefit them based on emerging data.¹ A glioblastoma platform trial called INSIGhT similarly incorporated Bayesian adaptive randomization based on biomarker-specific progression-free survival data.²

This sounds elegant, but response-adaptive randomization is not without risk. REMAP-CAP, a large perpetual platform trial for pneumonia that pivoted rapidly to test COVID-19 therapies in 2020, found that its response-adaptive randomization amplified random noise in some treatment domains, exposing patients to interventions that were later shown to be ineffective.³ The lesson: adaptive randomization can accelerate answers, but if the early signal is just noise, the algorithm chases the wrong treatment for a while before the data catch up.

The Shared Control Group and Why It Gets Complicated

One of the biggest practical advantages of a platform trial is that all experimental arms share a single control group. In a world of separate trials, each one needs its own control arm, meaning many patients receive standard care without contributing to comparisons across treatments. Sharing the control group means fewer total patients are needed, and direct comparisons between treatments become possible because they were tested under the same conditions.

But sharing the control group creates a statistical wrinkle that researchers have spent considerable effort solving. Because new treatment arms can enter the platform at any time, some of the control-group patients were enrolled before a given experimental arm even existed. These are called non-concurrent controls, as opposed to concurrent controls who were randomized at the same time as the experimental arm they are being compared against.⁴

Should those earlier control patients be included in the analysis for a later-entering treatment? If nothing about the patient population or standard of care has changed over time, including them gives you a bigger control group and more statistical power. But if something has shifted, like improvements in supportive care, seasonal variation in disease severity, or a change in which patients choose to enroll, then lumping older control data in can introduce bias. Simulation studies have confirmed that naively pooling all control data increases power when no time trends exist, but under drift, it can inflate or deflate error rates and produce biased estimates.⁵

Researchers have developed several approaches to handle this. Methods that partially borrow non-concurrent data, either through fixed weighting or by letting the degree of similarity between old and new data determine how much borrowing happens, perform better than naive pooling when time trends exist. But none of them can fully guarantee the kind of error control you get from using only concurrent controls. The current consensus is that for confirmatory analyses meant to support regulatory approval, concurrent controls should remain the primary comparator.⁵ Model-based time trend adjustments, such as treating each period of the trial as a separate block, can also help control error rates, though these work best when the time trend is consistent across all arms.⁶

How Much Time and Money Platform Trials Save

The efficiency gains are not hypothetical. An economic evaluation published in JAMA Network Open compared the cost and time requirements of a platform trial against running the same treatments as a series of separate conventional trials. When three treatments were tested sequentially in separate trials, cumulative setup costs rose by roughly 390% compared with the platform approach. Total costs went up by about 58%, and cumulative trial duration increased by about 310%.⁷

Even for just two treatments tested sequentially, cumulative setup costs still more than tripled relative to the platform trial, total costs increased by about 17%, and the cumulative trial timeline stretched by roughly 170%.⁷ The savings come from not having to duplicate infrastructure, not re-recruiting a control group, and not re-negotiating site contracts for every new treatment. The initial setup cost of a platform trial is higher than a single conventional trial, which is worth noting. The payoff materializes once the second and third treatments enter the platform.

The HEALEY ALS Platform Trial illustrates what this looks like in practice. The trial was able to enroll patients concurrently into four distinct treatment regimens and, once the infrastructure was running, accelerated the start-up time for adding a new regimen compared with what a standalone trial would have required.⁸ For a disease like ALS, where patient populations are small and time matters enormously, that acceleration is not an abstract benefit.

Where Platform Trials Are Being Used

Oncology has been the most active testing ground. The I-SPY 2 trial, running since 2010, tests neoadjuvant therapies for high-risk breast cancer. It has cycled through 17 agents or combinations, with 7 treatments graduating from the platform, meaning they showed at least an 85% probability of succeeding in a larger confirmatory trial. Two of those graduated drugs received accelerated FDA approval, and one earned breakthrough therapy designation.¹ The trial’s biomarker-driven approach has also been influential: by classifying patients into molecular subtypes and matching them to treatments, I-SPY 2 demonstrated that platform trials and precision medicine are natural partners.

That pairing is now spreading. A Canadian prostate cancer platform called PC-BETS screens patients with metastatic castration-resistant prostate cancer using circulating tumor DNA to identify genomic biomarkers, then enrolls them into treatment arms matched to their molecular profiles.⁹ This kind of biomarker-first enrollment, where you screen for the biology and then assign the treatment, is something platform trials are uniquely suited to because the infrastructure to add new biomarker-matched arms already exists.¹⁰

Outside oncology, infectious disease has been the other major proving ground. RECOVERY, launched in the UK in March 2020, became the trial that established dexamethasone as a COVID-19 treatment, producing results in months rather than the years a conventional trial would have taken. REMAP-CAP ran across dozens of countries and tested treatments in domains like antivirals, immune modulators, and anticoagulants simultaneously. Both demonstrated that platform trials are especially valuable in emergencies, when you need answers fast and cannot afford to test one drug at a time.¹¹

The Statistical Error Problem That Keeps Researchers Busy

When a platform trial tests many treatments over its lifetime, a subtle statistical challenge emerges. Each time you compare an experimental arm to the control and ask “is this treatment better?”, you are running a hypothesis test. The more tests you run, the more likely you are to get a false positive just by chance. In a conventional trial testing one treatment, you set a threshold (usually a 5% false-positive rate) and move on. In a platform trial that might test dozens of treatments over its lifetime, the total false-positive risk across the entire platform can creep much higher if you do not account for it.

One approach is to apply a strict correction, like the Bonferroni method, which divides the error budget equally among all tests. That keeps the overall false-positive rate low but makes it harder to detect real effects, since the bar for each individual comparison gets very high. Researchers have developed online error-rate control methods specifically for the platform trial setting, which test hypotheses sequentially over time and adjust the threshold dynamically based on past decisions. Simulations show these methods achieve a substantially lower overall false-positive rate than uncorrected testing while still gaining meaningful power compared with a rigid Bonferroni correction.¹²

Whether a platform trial even needs to control the family-wise error rate across all its arms, or whether each arm can be treated as an independent test, is itself a matter of regulatory debate. The answer often depends on whether the trial is generating exploratory signals or making confirmatory claims for drug approval.

Informed Consent Is Harder Than It Looks

Explaining a platform trial to a potential participant is genuinely difficult. In a standard trial, you can tell someone: you will either get drug X or a placebo. In a platform trial, the treatment you are randomized to might not even exist yet when you first hear about the trial. The probability of being assigned to a given arm might shift as data accumulate. And if your arm gets dropped mid-trial, your experience changes in ways that were not fully predictable at enrollment.

Research on participant understanding confirms this is not a trivial concern. When participants were shown a consent form that included response-adaptive randomization, enrollment increased by about 13% compared to a standard consent form, suggesting the design appeals to people. But those same participants were significantly less likely to correctly identify how treatment allocation would actually work, even though they reported similar levels of understanding as the control group.¹³ In other words, people like the idea of adaptive randomization but do not always grasp what it means in practice.

A review of consent approaches in adaptive trials found that participants tended to prioritize information about treatment risks over details of the study design itself.¹⁴ That makes practical sense, since most people joining a trial care more about what might happen to them than about the statistical architecture. But it does mean that consent processes need to be carefully designed so that participants understand the most important features, like the possibility that their arm could be stopped early, without drowning them in methodological detail they will not retain.

Regulatory Frameworks Are Still Catching Up

Platform trials do not fit neatly into the regulatory structures that were built around conventional one-drug, one-trial designs. In a standard trial, a single sponsor submits a single protocol for a single drug. In a platform trial, multiple sponsors (often competing pharmaceutical companies) may each have a treatment arm within the same trial, governed by a single master protocol. Questions about data ownership, intellectual property, and who is responsible for safety reporting become more complex.

The EU-PEARL project tackled these issues by reviewing existing guidelines in both the European Union and the United States, consulting with the European Medicines Agency and the FDA, and engaging ethics committees and health technology assessment bodies. The result was a master protocol template designed to address the governance gaps that platform trials expose.¹⁵ But this is still an evolving area. Regulators are broadly supportive of the concept but cautious about specifics, particularly around how non-concurrent controls are handled and whether the statistical methods used to control error rates are adequate for confirmatory claims.

Running a Platform Trial Across Countries With Different Resources

Expanding a platform trial internationally introduces challenges that go beyond language and logistics. When REMAP-CAP and other pandemic-era platform trials added sites in low- and middle-income countries, they encountered a fundamental tension. On one hand, a platform trial needs a universal protocol so that data from all sites are comparable and can be pooled for analysis. On the other hand, the treatments that are realistically available in a given country may differ substantially from those in wealthier settings, and the research priorities of local communities may not perfectly align with those of the trial’s central leadership.¹⁶

One advantage of the platform design in this context is its modularity. Treatment arms can be added or removed based on what is feasible and relevant at specific sites without disrupting the rest of the trial. Local researchers can potentially run independent analyses on their own site’s data to identify findings relevant to local practice, even while contributing to the pooled global analysis. But making this work requires genuine partnership, not just data extraction. The lessons from pandemic-era implementation suggest that building local research capacity, respecting local priorities, and ensuring that the benefits of the trial flow back to the communities that participated are essential for the model to be sustainable and ethical.¹⁶

Platform Trials and the Shift Toward Biomarker-Driven Medicine

The growth of platform trials is converging with a broader shift in how diseases, particularly cancers, are classified and treated. Instead of defining a disease purely by where it occurs in the body, precision medicine increasingly defines it by its molecular features. A lung cancer driven by a specific gene mutation may have more in common with a breast cancer harboring the same mutation than with another lung cancer driven by a different pathway.

Platform trials are well suited to this reality because they can accommodate biomarker-based enrollment as a core feature rather than an afterthought. A patient enters the platform, undergoes molecular screening, and is assigned to whichever treatment arm matches their tumor’s biology. As new biomarkers are identified and new targeted therapies developed, corresponding arms can be added to the platform without redesigning the entire trial.¹⁰ The PC-BETS prostate cancer trial, which uses circulating tumor DNA screening to match patients to genomically targeted arms, is one example of this approach in practice.⁹

This is different from basket trials or umbrella trials, which are sometimes confused with platform trials. A basket trial tests one drug across multiple cancer types that share a biomarker. An umbrella trial tests multiple drugs within a single cancer type, matching treatments to biomarker subgroups. A platform trial can incorporate elements of either design, but its distinguishing feature is permanence: it does not end when the initial set of treatments finishes, and new treatments can keep entering indefinitely.¹⁷ The trial outlives the drugs it tests. That permanence is what makes the infrastructure investment worthwhile and what makes the statistical challenges around non-concurrent controls and error rate control an ongoing concern rather than a one-time problem to solve.