A knowledge graph proposes a target
Researchers built a knowledge graph from published literature and databases, asked which proteins were linked to the disease but under-explored, and got RIPK1 back as a candidate. Laboratory experiments then tested whether blocking RIPK1 changed disease biology. The AI's role was to assemble scattered evidence into a ranked suggestion; the experiments decided whether the suggestion held.
What the case demonstrates
The value is breadth and ordering, not proof. A model can weigh thousands of weak connections at once and surface a candidate that is not the field's current favorite. But the ranking is a hypothesis: the target only becomes credible when an experiment confirms that changing it changes the disease. Published cases also over-represent successes, so the visible record of AI-informed targets is more favorable than the true hit rate.
Why this decision is hard to undo
Once a target is chosen, years of chemistry and testing are built on top of it. If the target turns out to be wrong, that work is not recoverable — which is why a ranked list is treated as a starting point for experiments, not as a decision.