Synthetic Control Arms at FDA: Agreed in Principle, Untested at Review

On 15 October 2020 Medicenna Therapeutics announced that, after an end-of-phase-2 meeting, it had been "advised to proceed with an innovative open-label hybrid control design" for a phase 3 trial of MDNA55 in recurrent glioblastoma. Matched external controls would make up two-thirds of the control arm. Medidata says it did the analysis behind the design, and the case became the flagship for the Medidata Synthetic Control Arm. The vendor's case study is headlined as a precedent-setting FDA approval.

Six years on, there is no trial for FDA to review. In its results for the quarter to 30 June 2026, published on 14 August 2026, Medicenna calls the drug, now named bizaxofusp, its "Phase 3-ready" asset and says it is still seeking a partner. The most-cited synthetic control arm precedent is an agreement about a protocol that has not produced a single comparison. The randomised side of the subject, covariate adjustment and EMA's PROCOVA qualification, is covered in our piece on digital twins and external controls.

What FDA actually agreed to in the Medicenna case

In Medicenna's account, patients would be randomised 3:1 between MDNA55 and standard of care, and a matched external arm would be added to the randomised controls. Overall survival, the primary endpoint, would compare everyone on MDNA55 with the pooled controls. Medidata's case study adds that the external patients would come from past clinical trials, and that the hybrid would need roughly 100 fewer prospective controls. Both are the vendor's account. No protocol has been published.

Read the wording. Medicenna says it was advised to proceed. The text of Medidata's case study says FDA "agreed to consider the use of a hybrid SCA". Its headline says approval to design one. The first two describe advice, which is what an end-of-phase-2 meeting produces: FDA's views, recorded in minutes. The third describes something that did not happen.

The statute creates something firmer for trial design in one situation. Under section 505(b)(5)(C) of the Federal Food, Drug, and Cosmetic Act, a written agreement on trial design and size is fixed once testing begins. FDA can change it only if its review division identifies a substantial scientific issue. Medicenna's release does not mention one. Even a written agreement covers the design, not the result. An agreed protocol can still produce data that do not persuade.

There is a second gap. FDA's draft guidance on externally controlled trials says it does not cover using external data "to supplement a control arm" in a randomised trial. That is precisely what a hybrid does. The most-quoted precedent sits outside the scope of the guidance most often quoted beside it.

Hybrid designs were not new at FDA. A 2021 review of FDA approvals found at least three products between 2000 and 2019 where external data were added to a concurrent randomised control: velaglucerase alfa, corticotropin and a scorpion antivenom. What Medicenna claimed as new was the size of the external share in an oncology registration trial.

Has FDA ever accepted a historical-trial control arm?

Yes, rarely, and mostly in rare diseases. The same review found 45 approvals from 2000 to 2019 in which FDA accepted external control data in its benefit-risk assessment. Of these, 80 per cent were in rare diseases and 87 per cent used an objective endpoint. Retrospective natural history data, including chart reviews, were the commonest source at 44 per cent. Data from a previous clinical study, the source a synthetic control arm uses, accounted for 11 per cent. Most of the authors worked for BioMarin, a rare-disease sponsor with an interest in wider use of external controls, and the review covered selected therapeutic areas, so treat the shares as indicative.

The cleanest precedent for pooled historical-trial controls is not in oncology. Lamictal XR's indication for conversion to monotherapy in partial seizures was approved in April 2011. The sponsor compared 223 patients on one of two doses with the control arms of eight earlier trials of other antiepileptic drugs. It judged a placebo arm unethical. The review records that FDA presented an evaluation based on criteria Stuart Pocock proposed in 1976, covering the timeframe of seizure assessments, baseline medicines and regional differences. The advisory committee voted 14 to 0 that the approach was acceptable in that specific circumstance.

The oncology example FDA itself cites used patient records from study and treatment sites rather than a pooled set of trials. FDA's real-world evidence framework of December 2018 describes Blincyto's initial accelerated approval. Its single-arm remission rates were compared with historical data on 694 comparable patients, drawn from more than 2,000 records in the EU and US. FDA then required a randomised trial to verify the clinical benefit.

What binds and what is draft: position current at 12 October 2026

In the United States, two provisions bind on this question. Section 505(d) of the Act requires substantial evidence from adequate and well-controlled investigations. And 21 CFR 314.126(b)(2)(v) lists the historical control as one of the recognised kinds of control. The same paragraph says such designs are "usually reserved for special circumstances". Its examples are diseases with high and predictable mortality, and drugs whose effect is self-evident.

Everything else is guidance, and most of it is draft.

Comments on the externally controlled trials draft closed on 2 May 2023, according to its Federal Register notice, and no final version has followed. FDA's revised draft on substantial evidence of effectiveness, published on 24 June 2026, still cites it as a draft.

The June 2026 revision names six conditions under which an externally controlled trial can support effectiveness. The natural history is well defined. The external population is very similar to the treated group. Concomitant treatments are very similar. The endpoint is objective and is defined and measured the same way in both groups. The observation period starts at a comparable point. And the pre-specified estimate of effect is so large that bias is unlikely to explain all of it. The same draft adds that natural history data offered as confirmatory evidence "should be distinct from any data used as a control". One dataset should not do both jobs.

The 2018 framework is worth one sentence. It says "Typically, the external control arm uses data from past traditional clinical trials". Real-world data were then the newcomer, and the final guidances that followed, on registries, data standards, and health records and claims, govern those sources.

ICH E10, the older text, limits external controls to cases where the "effect of treatment is dramatic and the usual course of the disease highly predictable". In Europe the reflection paper on external controls is still at concept stage. For real-world control data in EU trials, ICH E6(R3) Annex 2 applies from 15 January 2027, as set out in our analysis of the E6(R3) clauses.

Where bias gets into a historical control arm

In 1982 Sacks, Chalmers and Smith compared 56 historically controlled trials with 50 randomised trials of the same six therapies. Of the historically controlled trials, 79 per cent found the new therapy better. Of the randomised trials, 20 per cent did. The difference came mostly from the control groups, which did worse in the historical trials. Adjusting for prognostic factors did not change the picture much. ICH E10 makes the same point: untreated historical controls tend to do worse than apparently similar randomised controls, possibly through selection bias.

FDA's 2023 draft splits the problem into parts. Four matter most here.

Selection

The draft calls choosing a control from a completed trial whose outcomes are known a particular concern. It asks sponsors to finalise the protocol, including the external control, before the trial starts. It also asks them to document every data source they looked at, to justify each one they excluded, and to keep audit trails of access and analyses. Matching then thins the sample: in the GEN-1 analysis below, 15 of 41 candidates survived.

Time

Standards of care change. So do supportive care, diagnostic criteria and the rules for judging response or progression. Older trials may not record biomarkers that now define the population. The draft says such differences are hard to fix with statistics alone. An arm drawn from trials run a decade ago compares your drug with a different era of medicine.

Endpoints and start dates

The endpoint must be defined and assessed the same way in both arms, on the same schedule. Where outcomes depend on judgement, the draft suggests blinded re-adjudication of the external data. The start of follow-up matters as much, and the draft gives an example of immortal time bias. Suppose both arms start the clock when prior therapy fails, but the treated arm only includes patients who went on to receive the drug. Those patients had to survive long enough to get it, while controls count however soon they died. Early deaths land only in the control arm, and the drug looks better than it is.

Missing data

The draft asks sponsors to measure missingness in the external arm before committing to the design, and again in the analysis. Old trials often lack variables nobody thought to collect. Real-world sources add patients who drop out of the record when they change provider.

Historical-trial controls versus real-world data controls

The draft accepts that old-trial data may have advantages, because collection followed a protocol. But it adds a condition: comparability on eligibility, treatment, patterns of care, concomitant medicines and how outcomes were assessed.

Three replication studies show what each source can do in a known setting. A 2022 study by Medidata statisticians with FDA and Friends of Cancer Research co-authors built an external arm from lung cancer trials started between 2004 and 2013. Once baseline characteristics were balanced, its overall survival was very similar to the target randomised control, provided most important prognostic variables were known and measured. On the real-world side, a Genentech study using oncology health records reproduced hazard ratios from lung cancer trials closely, with one exception, and a correlation of 0.86 between log hazard ratios. In glioblastoma, Medicenna's own disease, a Dana-Farber analysis found that combining prior trials with real-world data cut the inflated false-positive rate of single-arm trials by up to 30 per cent.

None of these tests a new drug with an unknown effect. Reproducing a known answer is necessary, not sufficient.

What the case studies show, and what they do not

The Medicenna case in our library shows that FDA was willing, in 2020, to consider a large external share in a confirmatory oncology trial. It cannot show acceptance, because no data exist.

The best real test of a Medidata external arm is less famous. For Imunon's ovarian cancer drug GEN-1, Medidata matched 15 of 18 phase 1b patients to 15 of 41 patients from earlier trials. The estimated hazard ratio for progression was 0.53, with a 95 per cent confidence interval from 0.16 to 1.73, and it helped size the randomised phase 2. That trial, OVATION 2, enrolled 112 patients. Imunon's topline results of 30 July 2024 gave a progression hazard ratio of 0.79 in the intent-to-treat population, from an open-label trial not powered for statistical significance. The external arm pointed the right way but was more optimistic. That is what a 15-patient comparator is good for: a go decision and a sample size.

Medidata-sourced past-trial comparators were also used for obe-cel in adult leukaemia. The paper, published in July 2026, reports overall remission of 67.3 per cent against 51.4 per cent. But FDA had already approved the drug in November 2024 on single-arm data: 27 of 65 patients reached complete remission within three months. The comparison is context, and its authors include Medidata and Autolus staff.

The Cytel case repeats the Medicenna pattern. FDA would consider a phase 3 design showing superiority over inadequately treated historical controls. The 33 per cent enrolment cut and $2.4 million saving are Cytel's figures, for an unnamed sponsor, with no published outcome.

Unlearn's work needs sorting into two piles. The AbbVie poster assesses variance reduction and sample-size savings in an Alzheimer's phase 2 trial. That is covariate adjustment, the PROCOVA side, not an external control. The VectorY study and a March 2026 partnership with SOLA Biosciences use digital twins from a model trained on historical patient-level ALS data as external comparators in single-arm phase 1/2 trials. VectorY's use is labelled exploratory, and neither claims regulatory acceptance. If a model-generated comparator were ever offered to FDA as evidence, the draft AI credibility guidance would become relevant, as our piece on that framework explains.

Phesi's digital twin for chronic graft-versus-host disease drew a standard-of-care arm of 438 patients from existing trial data. Its six-month prednisone response of 52.7 per cent agreed with the literature. That is not the same as matching a randomised arm, and two of its three authors work for Phesi.

What this means in practice

If you are weighing a synthetic or historical control arm, work in this order.

1. Decide the job first, months before the protocol. An external arm that informs a go decision or a sample size never has to survive a review. One offered as the basis for approval has to meet the six conditions in FDA's June 2026 draft, above all an effect too large to be bias. If your expected effect is modest, randomise.
2. Run feasibility on the candidate data before writing anything. Apply your eligibility criteria and count the matchable patients. Check that the prognostic factors, endpoint definitions, assessment timing and a usable start date all exist. Log every source you open and why you dropped it.
3. Secure data rights early. FDA expects patient-level data for both arms and access to source records for the external arm. For old trials, check that consent and data-sharing terms allow regulatory reuse.
4. Pre-specify, then submit before enrolment. Fix the external control, the matching method, the analysis and the sensitivity analyses before anyone sees external outcomes. FDA's draft asks for the protocol and analysis plan before the trial enrols.
5. Meet FDA early, with four things. Bring the case for the design, the data sources and why they fit, the planned analyses, and how the data will reach FDA. If you need certainty, seek a written agreement on design, which covers the design, not the result.
6. In a hybrid, keep enough randomised controls to check the external ones. Pre-specify what happens if the two disagree.
7. Plan for confirmation. Blincyto's route is the realistic one: an external comparison for accelerated approval, then a randomised trial to confirm benefit.

The sentence to keep in the room: six years after FDA agreed to consider the best-known synthetic control arm, no regulator has judged its result, because the trial had still not started in August 2026.