I agree: Supercomputers can only suggest possible solutions to such problems. They have to be verified with real wet-lab experiments. But just having some of those likely solutions is far better than assembling a stick & ball model of a protein and fiddling around with different folds!
Remember that different folds actually occur in vivo, and we call the really bad ones "prions".
AI can be very good at certain tasks, but it's current capabilities are way, way less sophisticated than most people imagine. For one thing, an AI can only learn about things it has "seen". That's why Tesla's autopilot drove right into a truck parked across the interstate. Even Deep Mind is just a very fast idiot. Not even as smart as a mouse.
