Association is not causation
An associated position appearing more often in people with a disease does not establish that the variation causes the disease. It can travel alongside the real cause, sit near a gene without affecting it, or reflect some other shared factor in the group being studied. AI can rank a gene highly on the strength of a statistical link while the underlying biology remains unproven.
What the model cannot see
A model works from the data it was given. If the people in the study are not representative, or if the relevant tissue and cell type were never measured, the ranking inherits those gaps. A confident-looking score is not evidence that the model had enough information to be right.
What validation actually tests
Laboratory work asks two separate questions that no ranking can answer. First, does changing this gene or protein actually alter the disease process in cells or in a living organism? Second, is it safe to change — does the protein do something else important elsewhere in the body? A target can be genuinely involved in the disease and still be a poor medicine target because interfering with it causes harm. Clinical testing then asks whether the effect holds in people, which is a further question again.
None of this makes the ranking worthless — it makes it a filter. Testing every gene in the genome against a disease is not feasible, so the value of the list is that it concentrates attention on a small number of candidates that are more likely than average to be worth the cost of an experiment. The ranking narrows the search; the laboratory and the clinic decide what is true.