The Predictable Failure: Why Clinical Trials Collapse in Ways We Should Have Seen Coming

Every few months, a phase III trial reads out with a null result and someone, somewhere, writes a post-mortem. The language is usually passive. The animal models did not translate. The biomarker did not hold. The patient population was heterogeneous. What these post-mortems rarely say plainly is that many of the warning signs were there before enrollment began.

This is not a niche complaint. Estimates from multiple pipeline analyses suggest that somewhere between 85 and 90 percent of drugs that enter clinical testing fail to reach approval. Phase II to phase III attrition alone consistently hovers around 50 percent across therapeutic areas. The explanations cluster into the same categories trial after trial: insufficient target validation, biomarker selection based on association rather than mechanism, animal model homogeneity that does not reflect human disease heterogeneity, and sample sizes calculated to detect an effect size more optimistic than preclinical data could honestly support.

The target validation problem has been documented most carefully in oncology, where the pressure to advance assets quickly is intense. A compound moves forward because it hits a target convincingly in cell culture, performs in a mouse model engineered to over-express that target, and generates early safety data clean enough to reassure a review committee. What preclinical packages often do not answer is whether the target is actually causal in the human disease subtype being studied, rather than merely correlated with it. The distinction matters enormously and is expensive to establish, so it gets deferred.

Biomarker selection compounds the problem. Enrichment strategies based on a single companion biomarker assume that the biology is cleaner than it usually is. When the phase II signal looks promising in a biomarker-high subgroup, the temptation is to treat that subgroup as a defined population and power the phase III accordingly. But if the biomarker captures only one node in a redundant pathway, the enriched population is still heterogeneous in the ways that matter, and the phase III collapses in a larger sample exactly as a skeptic reading the phase II data carefully might have predicted.

Effect size inflation is the quieter contributor. Phase II trials are frequently underpowered by design. A nominally positive result in 80 patients produces a point estimate the development team anchors on when powering the phase III. Regression to the mean being what it is, the larger trial finds a smaller effect, often below the threshold for statistical or clinical significance. This is not fraud or negligence in most cases. It is the normal behavior of noisy data in small samples, and it is well understood statistically. The issue is that the incentive structure of drug development does not always reward the caution that understanding would counsel.

None of this means that clinical trials should stop, or that the failures are wasteful in a simple sense. A clean null result in a well-designed trial is real information. It closes a question. The problem is that many trials are designed to test a hypothesis that the underlying preclinical and phase I data never supported at the confidence level the phase III protocol implied.

The post-mortems are valuable precisely because they are consistent. The same failure modes appear in oncology, in neurology, in metabolic disease. The pattern is legible enough that it has generated its own reform literature, from proposals for more rigorous target validation consortia to pre-registered adaptive designs that build in earlier stopping rules. Whether those reforms move faster than the incentives that produced the pattern in the first place is the structural question the field keeps deferring.