The End of Human Clinical Trials?

A fair question from the floor

At a conference recently, someone asked me a question I have not stopped thinking about. If animals are on the way out, and AI keeps improving at the rate it is, when will we be able to phase out in vitro and wet lab work altogether?

It is a reasonable question. If we really are on the cusp of recursively self-improving AI that can perfectly predict real biology, what value does the bench add? But the question led me somewhere less comfortable. If practical biology can be fully displaced by computation, why would we stop at the wet lab?

What does that really mean?

If a model can predict what a liver spheroid will do, it can in principle (and with enough compute) predict what a liver will do. Following that logic, could we predict every organ, then whole systems, then a complete human?

Surely an appropriately powerful AI could take a patient’s history, genome and other omics data and tell us whether a specific drug will work for that person, at what dose, and with what risk. At that point the trial is no longer a population average applied to an individual. It is a prediction made for one patient. This addresses one of the core challenges existing clinical trials pose, “will these results translate to safe and effective medicine for the entire human population?”

That is the real promise of AI replacing practical experimentation. It’s not just that we can end slow and expensive trials. It is the arrival of genuinely personalised medicine that can go from idea, to candidate, to treatment in a matter of weeks instead of years.

The 3Rs already apply to humans

We usually talk about NAMs as tools to reduce, refine and replace testing in animals. Psychologically, for most people, that means non-human animals. But good NAMs should also reduce, refine and replace parts of human clinical trials. This is already happening in regulatory practice.

• Refine: smarter first-in-human dosing. Conventional MABEL calculations often land far below the pharmacologically relevant range, so early cohorts receive doses with no chance of benefit. Mechanistic PBPK and QSP models can justify a higher, still-safe starting dose. In one published T-cell engager case, a modified MABEL approach gave a starting dose more than four times the traditional estimate, cutting subtherapeutic cohorts (Allucent). Fewer cohorts means fewer participants and a shorter Phase 1.
• Replace: clinical DDI studies. ICH M12 (2024) accepts that PBPK or mechanistic static models can, in some situations, translate in vitro results to the clinic without a clinical DDI study (Simulations Plus). In August 2025 the EMA qualified the Simcyp Simulator for CYP-mediated DDI prediction (Certara). Every waived study is a cohort of healthy volunteers who did not need to be dosed.
• Reduce: smaller control arms. In 2022 the EMA qualified PROCOVA, which uses AI-generated digital twins as a prognostic covariate to cut sample size in Phase 2 and 3 trials (Unlearn). A retrospective simulation of the donanemab TRAILBLAZER-ALZ 2 trial estimated up to 24% fewer control participants for the same conclusions (Alzheimer’s & Dementia, 2025).

None of these fully eliminates the human trial. Each one makes it smaller, shorter or better targeted. That is the shape of what comes next.

How the evidentiary requirements might shift

In this possible future where AI can replace practical science, human clinical trials will still likely be the last piece to be fully eliminated. My expectation is that animals are eliminated first, human trial participant numbers draw down next, in vitro work goes second to last, and human clinical trials end last, if they end at all.

There are three considerations in this chart. In vitro work rises before it falls, because it absorbs much of what animals used to do. In silico work grows alongside it, then pulls away in the computation-first phase as it takes over what the bench used to do. And human trial participation falls in two stages: a long drawdown driven by models that refine and reduce, then a much later drop that only arrives if computation can be trusted to replace the trial outright. This process may take 30 years or it may take 5. Time will tell.

Why the bench might outlast the animal, and the human outlast the bench

The order follows cost, time and ethical weight. The steps that are most expensive, slowest and hardest to justify have the most pressure to disappear; the step that is cheapest, fastest and least ethically challenging has the least pressure.

• Animals go first because they are slow, expensive and because they answer human questions indirectly. The regulatory machinery to replace them is already being built.
• Trial size draws down next because every participant carries cost and risk, and the tools above already shrink cohorts without asking anyone to trust a model with the whole decision. This has the advantage of decreasing the cost and increasing the speed of the clinical trials that proceed.
• In vitro goes second to last because it is the least expensive, least time-intensive and least ethically challenging step we have. It is also the ground truth that computational models are trained and verified against. Removing it means trusting models that can no longer be checked against human biology outside the body. That is a high bar, and there is little cost pressure to clear it early.
• Complete human trials go last, if ever because they are the only experiment that measures the thing we actually care about. A model that replaces them has to be qualified for a context of use as broad as “any drug, any patient”, and no Weight of Evidence package today comes close.

The personalised medicine version is not a trial at all. It requires an extremely accurate prediction for one person, and the regulatory question becomes how we validate a model whose output is never tested on a population before it is used.

So, the end of human clinical trials?

Not soon, and not all at once. But the question is worth asking seriously, because the same tools that are retiring the animal model are already shrinking the human trial. Every PBPK model that justifies a higher starting dose, every waived DDI study, every digital twin in a control arm is a small step toward a future where we test fewer people and predict more.

NAMs-first is viable now. NAMs-only is the preclinical future. Computation-first medicine, where the trial confirms rather than discovers, may be the clinical future after that.

The bottleneck at every step is the same one we already know: validation, qualification and trust. Context of use does not get easier as the stakes rise. It gets harder.

Sources

• Allucent: Modified MABEL approach for first-in-human dose selection
• Simulations Plus: PBPK acceptability under ICH M12
• Certara: PBPK for drug-drug interactions (EMA Simcyp qualification, Aug 2025)
• Unlearn: EMA qualification opinion for PROCOVA (Sep 2022)
• Digital twins in a simulated TRAILBLAZER-ALZ 2 trial (Alzheimer’s & Dementia, 2025)