Why can a promising laboratory result fail?

A model deliberately leaves things out

An in-vitro experiment might show that a molecule binds a receptor, slows an enzyme, or protects cultured cells from a stress. That is evidence about a mechanism under chosen conditions. It does not establish that the target drives illness in a person, that the same concentration can be reached safely, or that changing the pathway improves how people feel, function, or survive.

An animal model adds an intact organism, but it remains a model. Species differ in metabolism, immune responses, anatomy, disease course and behaviour. An induced injury or engineered mutation may reproduce one feature of a human disease while omitting others. The useful question is which aspect it represents, which it leaves out, and whether the study was designed and reported well enough to support the next experiment.18

The molecule has to get there

A substance can affect a target in a laboratory yet fail to reach it in a person at a useful and tolerable exposure. Absorption, distribution, metabolism and excretion change the journey. A short pulse, continuous exposure and treatment after established disease are different interventions even when the molecule is identical.

ICH M3(R2) describes nonclinical pharmacology, toxicology and pharmacokinetic studies used to support human trials and marketing authorisation. It is a framework for decisions, not a guarantee that an animal exposure predicts a therapeutic human exposure.28

The outcome has to matter

A biomarker can help researchers investigate a mechanism. Whether changing it reliably predicts a meaningful clinical outcome is a separate question. Later trials need a defined population, an appropriate comparator, outcomes that matter to patients, and follow-up long enough to observe benefit and harm. Ask what was measured, whether that measurement was the intended target of the study, and what interpretation it supports.34

A benefit can come with unacceptable trade-offs

Development evaluates adverse effects as well as the desired response. Sometimes the exposure needed for an effect cannot be used safely; sometimes the overall clinical result is less useful than hoped. A convincing mechanism does not prevent either outcome. A candidate may fail because it is unsafe, because the effect is smaller than expected, because the endpoint misses the proposed benefit, or because an apparent effect exists only in a subgroup the trial was not designed to identify.51

A real example: CEP-1347 in Parkinson disease

Cell and animal work supported the idea that inhibiting mixed-lineage kinases could protect vulnerable neurons. The PRECEPT trial randomly assigned 806 people with early Parkinson disease who did not yet need dopaminergic therapy to one of three CEP-1347 dose groups or matching placebo; participants were assessed without knowing assignment. Its clinical question was whether treatment delayed the need for dopaminergic therapy. After an average of 21.4 months, a planned interim analysis found it would be futile to continue the experimental treatment, and the trial found no evidence of the intended clinical benefit.

That result is evidence about the tested intervention, regimens, population and endpoint. It does not show that the pathway was irrelevant in every context or that the animal experiments were worthless. The trial alone cannot establish whether an imperfect model, insufficient target engagement, intervention at the wrong stage, or another explanation accounted for the result.7

Controls can make the claim narrower and stronger

Small preclinical studies can produce unstable estimates. Selective reporting, unblinded assessment, flexible outcome choices, inadequate randomisation, and failure to reproduce a result independently can all inflate an apparent effect without requiring misconduct. Vehicle controls, positive controls, blinded assessment, randomisation, prespecified outcomes, adequate sample size and independent replication help separate a robust signal from a fragile one. A result that survives these tests is stronger, but it remains evidence in that model.9

Follow the chain of questions

A translation claim is easier to assess when each link is visible. Did the human exposure engage the intended target? Was treatment started at a disease stage when that mechanism could plausibly matter? Was the clinical outcome prespecified and important to patients? Did the study have a credible comparator, adequate follow-up and a plan for missing data? Were harms and tolerability measured alongside benefit? These are questions for evaluating the evidence, not a universal sequence that every development programme must follow. A missing answer limits the conclusion; it does not automatically explain why a particular study failed.38

Stopping is also information

Development is not a progress bar that inevitably fills. A programme may stop, change direction or require a different study. A negative trial can narrow a claim without answering every biological question; an early positive signal can be overturned by a larger or better-controlled study. Reporting those outcomes helps the reader understand the scientific question rather than simply follow a success narrative.6

Read “promising” as an invitation to ask more

Promising in what model? At what exposure? Compared with what? With which harms measured? What result would be needed next? Does the evidence support a mechanism, a safety decision, a clinical benefit, or only a reason to study the question further? These questions do not dismiss the research. They locate it.

Target engagement is a bridge, not a verdict

A translational programme can ask whether a candidate reached and changed its intended target in people. That bridge may use a biomarker, an imaging signal, a concentration measurement or a physiological response. Each answers a narrower question than ‘did patients get better?’ A target can be engaged without producing a useful clinical effect, and a negative biomarker result can reflect the dose, timing, tissue access or assay rather than disprove the biology. Keep the chain visible: exposure, target engagement, biological response and patient-important outcome are related measurements, not interchangeable proof.134