The FDA advisory committee's recent vote against Capricor Therapeutics' cell therapy for Duchenne muscular dystrophy looked like a decision about one product. It may ultimately be remembered as something much larger. Beneath the discussion of endpoints, missing data, and statistical analyses lay a larger question that will not disappear after the next advisory committee meeting. Precision medicine has changed the way diseases are understood. Regulatory science now must decide whether its methods are changing quickly enough to keep up.
No one should dismiss the committee's conclusion. FDA reviewers raised legitimate questions about how the trial should be interpreted, and the advisory committee concluded that the evidence fell short of demonstrating efficacy. Fair enough. The more interesting question is whether the tools used to judge today's therapies were designed for yesterday's medicine.
FDA has been here before. The debate has never been whether to lower scientific standards. It has been whether better science demands better tools. HIV forced that conversation. Oncology did too. Rare diseases are forcing it again. Accelerated approval, orphan-drug incentives, adaptive trial designs, Bayesian methods, model-informed drug development, and real-world evidence all grew from the same principle: scientific rigor does not require methodological rigidity. Capricor may represent the next chapter in that story.
Judea Pearl's work on causal inference helped us recognize what we now call denominator collapse. The phrase sounds technical, but the idea is remarkably simple. Precision medicine succeeds by revealing biological differences that once went unseen. Every newly discovered mutation, every validated biomarker, every molecular subtype narrows the group of patients eligible for a therapy. Better biology produces smaller denominators.
That is exactly what biomedical research has been trying to accomplish for decades. The surprise is that success has created a new problem. Every important biological discovery leaves fewer patients available to generate the evidence regulators have traditionally relied upon.
Duchenne muscular dystrophy is only one example. ALS has followed much the same path. What physicians once regarded as a single neurodegenerative disease is increasingly understood as a collection of biologically distinct disorders driven by different genetic mutations, molecular pathways, and rates of progression. Inherited retinal disorders, pediatric epilepsies, and much of modern oncology are moving in the same direction. Precision medicine has not made these diseases more complicated. It has revealed how complicated they always were.
That creates an uncomfortable statistical reality. A therapy may produce substantial benefit in one biologically defined subgroup, little effect in another, and perhaps none in a third. Average those outcomes across a trial enrolling only a few dozen patients and the treatment effect begins to disappear. Biology has not failed. Arithmetic has.
Randomization remains the single most effective protection against bias ever devised. It is indispensable. But randomization is not magic. It depends on numbers. Once trials become very small, chance alone can produce important imbalances in disease severity, progression rates, biomarker profiles, treatment discontinuation, or missing observations. The familiar response is to recommend another, larger study. Increasingly, that recommendation assumes a patient population that no longer exists.
The problem is hardly unique to Capricor. Only a day later another FDA advisory committee wrestled with Replimune's investigational melanoma therapy, RP1. There the challenge was almost the reverse. Reviewers were less concerned about whether patients improved than about what caused the improvement. How much benefit came from RP1 itself? How much reflected concomitant immunotherapy? How much resulted from patient selection or the natural history of the disease? Different therapeutic area. Different study design. The same fundamental question: what caused the observed outcome?
Classical statistics is extraordinarily good at measuring association. Causal inference asks something harder. Why did this happen? What would likely have happened had treatment never occurred? Those counterfactual questions become increasingly important as biology shrinks clinical trials beyond the point where conventional statistics alone can comfortably answer every clinically important question.
Causal inference is not an alternative to randomized trials. It is a way of making them smarter. Causal inference requires investigators to state their scientific assumptions before the trial begins rather than after the results are known. That discipline makes both the science and the disagreements more transparent. Structural causal models then test those assumptions rather than leaving them hidden inside statistical analyses. Bayesian methods can incorporate carefully selected information from natural-history studies, external controls, and earlier investigations without pretending uncertainty has disappeared. The objective is not to manufacture certainty. It is to understand uncertainty more honestly.
None of this weakens FDA's scientific standards. If anything, it raises them. Investigators must explain why they believe a treatment works instead of relying exclusively on whether a p-value crossed an arbitrary threshold. Causal inference also acknowledges something physicians recognize every day: averages treat patients as abstractions. Biology does not.
The practical implications are not revolutionary. Sponsors developing therapies for diseases already experiencing denominator collapse should prospectively submit causal models alongside traditional clinical protocols. Those models should specify anticipated patient subgroups, approaches to missing data, external-control strategies where appropriate, and the biological assumptions underlying key analyses before outcomes are known. FDA review divisions should continue building expertise in structural causal modeling just as they previously built expertise in Bayesian methods, adaptive trial design, pharmacometrics, and real-world evidence. The Agency has done this before. It can do it again.
Whether Capricor's deramiocel ultimately succeeds remains an empirical question. Additional evidence will answer it. The larger issue extends well beyond one company or one product. Duchenne muscular dystrophy, ALS, Huntington disease, inherited retinal disorders, and an expanding list of genetically defined conditions are all moving toward smaller, biologically more homogeneous patient populations. That trajectory is unlikely to reverse.
FDA has spent its history adapting regulatory science to advances in biomedical science without compromising scientific rigor. Precision medicine presents the next test. The Agency does not need lower standards. It needs tools equal to the biology now arriving at its doorstep. The therapies we lose over the next decade may not be those that lack biological activity. They may be the ones whose clinical trials were never designed to reveal where that activity existed. Biology has outrun statistics. Regulatory science shouldn't be far behind.
Peter J. Pitts, a former FDA Associate Commissioner, is President of the Center for Medicine in the Public Interest, Dr. Robert Goldberg is Vice President of Research Programs at the Center for Medicine in the Public Interest.
