From a laboratory model to people.

Start by naming the model and the claim

“The compound worked” is too vague. Ask what was tested, in which cells or animal species, how it was given, at what exposure, against what comparison and what was measured. A molecule can change a pathway in cells grown in a laboratory without improving an outcome that matters to patients. An animal response does not necessarily reproduce human disease or how a treatment would be used.

The FDA describes nonclinical work as including cell or tissue tests, computer models and animal studies. These help researchers assess whether a study in people is likely to be safe; they do not replace evidence from people.12

What cell and tissue models can show

Cell models can isolate a receptor, biological pathway, toxicity signal or interaction under controlled conditions. Cells taken from humans add relevant biology, but a culture cannot reproduce an entire person, including behaviour, treatment-taking and interactions among organs.

More complex models can include several cell types and fluid flow. FDA researchers, for example, have studied a liver model with several liver-associated cell types to examine toxicity, metabolism and drug distribution. It models selected interactions, not a complete human circulation or immune system. Researchers still need to show that a model gives reproducible results and fits the question.

Results also depend on the cells, their history and conditions, exposure time, test method and concentration. A high concentration that changes a pathway does not establish a useful or safe dose in people. Researchers need to connect the exposure and selectivity to human pharmacology.110

What animal studies add—and what they cannot settle

Animal studies can help researchers examine exposure throughout the body, organ effects, behaviour and toxicity over time. Good laboratory practice sets expectations for personnel, facilities, study plans, records, reporting and quality assurance in regulated nonclinical studies.

An animal model is a selected approximation, not a complete copy of a human disease. Differences in receptors, metabolism, immune responses, anatomy, lifespan, background disease and housing can change both benefits and harms. A model may capture one feature of a condition while missing others that matter to patients.12

Reproducibility is part of translation

Strong preclinical work explains why a model was chosen, sets outcomes in advance, uses suitable controls and reports both positive and negative findings. Small studies can be sensitive to one missing or unusual observation. Selective reporting can make a result seem more reliable than it is. Reviews have identified model relevance, standard methods, validation and reporting negative results as recurring gaps.

ARRIVE 2.0 is a reporting guideline for animal research. It asks authors to describe the experimental unit and sample size, randomisation, blinding, outcomes and statistical methods. A complete report helps readers assess bias, but a checklist alone cannot guarantee a sound design or valid result.48

Nonclinical methods are changing

Nonclinical evidence can come from animal studies, cell and tissue systems, computer models and other methods. Each method needs to be reliable and suited to the question. A model is not automatically predictive just because it uses human cells.

In September 2026, the FDA issued a US direct final rule updating terminology in specified regulations to recognise nonclinical methods more broadly. As issued, it is due to take effect on 4 February 2027, subject to the direct-final-rule comment process. FDA says it does not require a particular method or change evidence standards. It does not mean animal studies are never needed or that every alternative is validated for every decision.9213

Safety and the first human dose are separate questions

Before a first-in-human study, nonclinical programmes examine toxicity, pharmacology and pharmacokinetics to identify potential target organs, dose relationships, exposure relationships, reversibility where relevant and monitoring parameters. ICH M3(R2) describes development as stepwise: early human studies generally begin with relatively low exposure and expand only as safety information supports the next stage.

The starting dose is a risk-management decision informed by the full nonclinical profile, not a direct conversion of an effective animal dose. The appropriate species, a no-observed-adverse-effect level or another justified exposure-based approach (such as a minimum anticipated biological effect level), pharmacodynamic effects, pharmacokinetics, molecule-specific risks and study design all matter. A safety margin does not prove that an intervention will work, and a reassuring early study cannot exclude rare or delayed harms.312

Safety pharmacology is a separate evidence layer

General toxicology is not the only safety question before human studies. ICH S7A describes safety pharmacology as testing whether a substance affects vital functions. Before human exposure, a core set of tests usually covers the heart and blood vessels, central nervous system and breathing. Further product-specific studies may be needed when the substance’s action or an initial signal warrants them.

These studies can identify hazards and help plan monitoring. They do not prove a product is safe in people. Interpretation still depends on exposure, relevance of the species and test, study quality and the clinical evidence that follows.11

What changes when people enter the evidence

Studies in people bring human variation, informed consent, planned monitoring and outcomes that may matter directly to patients. Phase 1 studies may focus on tolerability, exposure and how a product acts in the body. Later studies test effects in people with the condition, using an appropriate comparison. The first study in people is a new experiment, not automatic confirmation of the laboratory result.

Researchers must define who is eligible, how exposure will be monitored, which outcomes and harms will be collected, and when the study will stop or change. A signal in a small early study can justify the next question; it cannot establish comparative benefit, how long benefit lasts or a complete safety profile.56

Follow exposure, outcome and population together

Ask whether the concentration or tissue level tested can be reached safely in people, whether the active drug or metabolite is comparable and whether timing matches the proposed use. Total blood concentration is not necessarily the amount free to reach a target. Giving the same milligrams per kilogram to different species does not ensure the same exposure.

Then ask whether the outcome matters directly to patients or is only a substitute, whether follow-up is long enough and whether participants resemble the people who may use the treatment. Age, other conditions and medicines, genetics, organ function, treatment-taking and social context can all affect results. A finding in a selected model or trial group may not apply more widely.76

A five-question translation check

1. What model was used, and what part of human biology does it represent? 2. How was the product given, at what exposure, against what comparison and with what outcome? 3. Was the model validated, was the result repeated, and were negative or uncertain findings reported? 4. What nonclinical evidence supports human safety, and what remains untested? 5. Has the finding been tested in people with an outcome that matters to patients, an appropriate comparison and enough follow-up?

If the last answer is no, “preclinical evidence” or “a biological hypothesis” is usually more accurate than “a proven treatment.”