Except a blackbox that simply spits out a resultant folded protein doesn't actually improve our understanding of anything. Are we just going to fuzz the blackbox with different protein combos hoping to find something useful? In that case, aren't we just more likely to find some stupid error in the ML predictor?