Why Preclinical Breakthroughs Often Do Not Translate

A striking result in cells or animals can be real and still fail to become a useful treatment. The experiment may be biased, the effect may be smaller than first reported, another laboratory may not reproduce it, or the model may not represent the human disease closely enough. Dose, timing, pharmacology, manufacturing, and clinical heterogeneity can also break the chain.

Rigor and reproducibility are therefore filters, not slogans, and they help distinguish a durable biological signal from a result that depends on one batch, one analyst, one model, or one favorable decision.

Separate rigor from relevance#

Scientific rigor is the disciplined, unbiased application of design, conduct, analysis, and reporting methods. A rigorous mouse experiment can estimate its intended effect accurately. Translational relevance is different: does that model answer a question that matters for human disease at a plausible dose and route?

Both are necessary. If an experiment lacks randomization or blinded outcome assessment, its estimate may be biased. If the model represents an acute injury while the human condition is chronic and heterogeneous, an unbiased estimate may still have little clinical transport.

Mechanistic plausibility is not enough. A target can alter a pathway in cultured cells but be inaccessible in tissue, redundant in humans, or unsafe to modulate systemically. Translation asks whether the whole causal chain holds.

The experimental unit is a frequent fault line#

The experimental unit is the smallest entity assigned separately to an intervention. It may be one animal, a litter, a cage, a culture dish, or a batch. Measuring ten cells from one dish does not create ten separately randomized biological units.

Treating technical replicates as separate subjects produces pseudoreplication. The nominal sample size increases, standard errors shrink, and a p-value looks more certain than the design permits, so a report should make clear the unit, the allocation mechanism, the number of units per group, and how many repeated measurements sit inside each unit.

Clustered designs need clustered analysis. Litter, cage, plate, operator, and experimental day can all introduce shared variation. Blocking or stratified randomization can balance known sources, while multilevel models can represent the dependence in analysis.

Randomization and blinding target different biases#

Random allocation prevents systematic placement of healthier animals, cleaner samples, or easier cases into one group. The method should be described, not merely labeled “random.” Allocation order, blocking, stratification, and any constraints matter.

Blinding reduces differences in treatment, measurement, and analytic choices after allocation; the person administering an intervention may be impossible to blind, but outcome assessors, image analysts, specimen labels, and data analysts may still be masked. A report should say who knew the group assignments at each stage.

Subjective endpoints are especially vulnerable. Choosing a representative microscopy field after knowing treatment can change the result without conscious misconduct. Automated measurement can help, but an algorithm tuned on labeled groups can carry the same bias.

Sample size should follow the question#

Small experiments can be ethical and informative for feasibility or mechanism. They become misleading when a noisy estimate is treated as definitive. Low power misses real effects more often and, among findings that cross a significance threshold, tends to select exaggerated estimates.

A justification should name the primary outcome, expected effect, and variability. It should name error rates, allocation ratio, and allowance for loss. Pilot data can underestimate variability, so uncertainty in assumptions deserves sensitivity analysis. Precision-based planning may be preferable when estimation, rather than a binary test, is the goal.

More subjects do not cure systematic bias. A large unblinded experiment can estimate the wrong quantity very precisely. Design controls and adequate size solve different problems.

Flexibility creates hidden multiplicity#

Preclinical datasets often permit many endpoints, doses, and time points. They permit normalizations, transformations, exclusions, and subgroups. If only the strongest combination is reported, the stated p-value ignores the search that produced it. It is the winner of a contest nobody described.

Prespecification can identify the primary hypothesis, analysis population, exclusions, and decision rules before outcomes are known. Exploratory work remains valuable, but it should be labeled and confirmed in new data. Keeping a complete record of all experiments, including null and adverse findings, helps prevent a selected subset from defining the literature. Effect sizes and confidence intervals carry more information than a thresholded p-value, and a finding that is statistically compatible with anything from trivial to very large should not be sold as a precise mechanism.

Materials can change the biology#

Cell-line misidentification, microbial contamination, and reagent variability can alter results. So can antibody specificity, passage number, and diet. So can microbiome, housing, temperature, and batch effects. NIH asks applicants to address authentication of key biological or chemical resources where relevant.

Document provenance, lot, and preparation. Document storage, quality checks, and acceptance criteria. Repeat key results across batches and experimental days. A result that appears only with one reagent lot is a clue, not necessarily a general discovery.

Sex, age, strain, disease stage, and comorbidity also matter. Including biological diversity does not mean every small study can estimate all interactions. It means the model should match the question, important modifiers should be planned, and claims should not outrun the sampled conditions.

Replication, reproducibility, and robustness#

Terminology varies across fields. A practical distinction is:

• Repeatability: the same team repeats the work under closely matched conditions.
• Computational reproducibility: another analyst obtains the result from the same data and code.
• Replication: new data and a stated protocol produce a compatible finding.
• Robustness: the conclusion survives reasonable changes in model, method, laboratory, or analysis.

An internal repeat can catch handling errors. A preregistered replication by a separate laboratory tests more of the tacit knowledge and contextual dependence, while a conceptual replication using a different method can strengthen the causal claim if both methods have distinct weaknesses.

Failed replication is informative but not self-interpreting. The original effect may have been biased, the replication may have altered a key condition, or the biology may be context-specific. Protocol exchange, positive controls, measurement verification, and prospective criteria help distinguish these explanations.

Famous reproducibility estimates need caution#

A widely cited 2012 commentary reported confirmation of the main findings in 6 of 53 selected preclinical cancer studies during an industry effort. The observation prompted overdue attention to methods, but it is not a population survey of all biomedical research. The studies and protocols were not fully disclosed, selection was not random, and “confirmation” depended on internal criteria.

The correct lesson is not that a fixed percentage of preclinical science is false. It is that consequential replication attempts can reveal fragility, and that transparent protocols and materials are needed to interpret disagreement.

Reporting audits also show room for improvement. Kousholt and colleagues compared samples of animal studies published in 2009 and 2018. Reporting improved for some items, but key design features remained incompletely described. Missing reporting does not prove a method was absent, yet it stops you from evaluating it.

Reporting guidelines are necessary but insufficient#

ARRIVE 2.0's Essential 10 cover study design, sample size, and inclusion and exclusion. They cover randomization, blinding, and outcome measures. They cover statistical methods, animals, procedures, and results. The Recommended Set adds context and detail.

ARRIVE is a reporting guideline, not a seal of validity. A manuscript can clearly report that it used a biased design. Conversely, an unlabeled method may have been performed but remains impossible to assess. Use the checklist during planning, then evaluate whether each choice fits the scientific question.

Data, code, protocols, and materials should be shared. That holds when ethical, legal, and practical constraints permit. A repository link alone is not enough if the files lack variable definitions, version information, or a workflow you can actually run.

Translation requires a chain of evidence#

Before you move from a preclinical signal to a clinical program, ask whether target modulation is demonstrated at the intended site, whether the dose is pharmacologically achievable, and whether the biomarker connects to the proposed mechanism. Compare models with different strengths. Examine safety margins and off-target effects.

Then state what you still do not know. Animal efficacy may support human testing without predicting the size of clinical benefit. A biomarker can show target activity without being a validated surrogate for how people feel, function, or survive.

Negative findings should change decisions. Stopping an unsuitable program early is a scientific success when the evidence is credible.

References#

1. NIH guidance on rigor and reproducibility
2. NIH initiative on replication and reproducibility, 2026
3. Begley and Ellis, standards for preclinical cancer research
4. ARRIVE guidelines 2.0
5. McGill and Threadgill, rigor, reproducibility, and robustness
6. Kousholt and colleagues, reporting quality in animal research