It's the ultimate prediction problem. And short term, computational bio methods are driving R&D costs up, not down. But arguably, techniques like DeepFold constitute current best hope
Garbage in / garbage out.
You can’t get away from reliable and predictable data. At that part is shockingly lacking. We have much better, sensitive and high-though put methods these days, but we are still relying on mice or cell culture data. We should be using ML to find what’s dissonance and missing links in the data, not to predict what will work.
I'm putting my money on CryoEM solving all the relevant structural problems first. Plus, I like exotic electron optics and UHV much more than ML - at least you get some expensive paperweights when a turbopump grenades...