No single dataset says 'this is the target'. What researchers have instead is many partial views of the same disease, each noisy on its own. AI's contribution at this stage is to combine those views into one ranked list of candidate targets, so that limited laboratory time goes to the most promising few.
The inputs are heterogeneous. Tumor sequencing shows which genes are mutated, amplified, or deleted in patients. Gene expression measurements show which proteins the tumor is actually producing, which matters because a mutated gene that is never expressed is a poor target. Patient outcome data — how long people with a given tumor profile survived, and on which treatments — links molecular features to real consequences. Tissue studies show where a protein is present in the body, which speaks to side-effect risk. Databases of known biological relationships describe which proteins act together in the same pathway, so that a candidate with no drug yet can still be judged by its neighbors.
A model trained on these sources learns to score a candidate protein by how strongly its pattern of evidence resembles proteins that have already been successfully drugged, or how consistently it separates tumors that respond from those that do not. The output is a ranking with a confidence attached, not a verdict. Two things follow from that. First, the ranking is only as good as the data feeding it — a target that no cohort has ever been sequenced for will not surface, however important it is. Second, a high rank means 'worth testing', not 'correct'. The model has found a statistical resemblance to past successes; whether the biology actually behaves that way is still an open question that only an experiment can settle.