Two different jobs, one shared logic
Target identification is a ranking problem: which protein is most plausibly involved in the disease and reachable by a drug. Molecule generation is a proposing problem: which chemical structures might act on that protein. Property prediction and screening sit alongside both, estimating whether a proposed molecule is likely to work and be safe enough to make. In every case AI narrows or generates options; the decision to test still belongs to the laboratory.
A concrete shape of the work
Suppose a research group is studying a disease and has a list of several thousand proteins that might be involved. Testing each one experimentally would take years. An AI system can rank that list using existing gene-activity and structural data and hand back a shortlist of a few dozen. Separately, given one chosen protein, a generative system can propose hundreds of chemical structures predicted to fit it. Neither output is a medicine. Both are starting points that a laboratory then has to test.
What this does not mean
A high-ranked target or a well-scored molecule is a hypothesis, not a result. The ranking reflects patterns in the data the system was given, and the data reflects what has already been studied. Whether the hypothesis holds in a living body is a separate question, and it is not answered by the model.