If you combine the two it sounds like a solution with some remaining risk, but if the disease is bad enough that risk may be worth it.

No there are not magic AI tools where you can wave your hands and design proteins or evaluate the properties of arbitrary compounds.

The way these models were is more or less by association. There is no foundational understanding of underlying pathways or likely decomposition products. The AI models can be used as a first stage filter: generate a bunch of putative products computationally, filter them through predicted toxicity profile, and then start real testing (though you are likely leaving out solid leads from false positives).

AI can be a human assisting tool, but it is in no way sufficient for leap-frogging into human testing.