Approval requires controlled human evidence that the drug produces a meaningful benefit relative to a real alternative, and the developer carries the burden of producing it.
Three things the evidence has to do
The evidence has to come from people, not models. It has to show a benefit that matters to patients, not just a change in a measurement. And it has to be strong enough that more than one well-run study points the same way.
A marker is not an outcome
Suppose an AI-designed molecule reliably lowers a protein that is elevated in a disease. That is a real result, and it may be a good sign. But regulators ask a different question: do patients live longer, feel better, or function more normally? A drug can move the marker and still fail to change the disease, because the marker may be a bystander rather than a cause. The gap between 'changed a measurement' and 'helped a patient' is where many candidates are lost.
The direction of proof matters. Regulators do not have to demonstrate harm to reject a drug; the developer has to demonstrate benefit to get it approved. This asymmetry is deliberate — it protects patients from treatments whose risks are unknown.