proteins don't fold to their energy minimum, they typically "collapse" to an intermediate state and then sample many different states (kinetically), rather than adopting the absolute thermodynamic minimum. In many proteins, the structure snapshot doesn't even correspond to a functionally active protein. The "protein folding problem is NP-hard is an entirely inaccurate view of the biophysics of folding", and if solved, would not address any useful question about folding.
What DeepMind solved was a far simpler problem: reproduce the structures that get produced by some experimental method. That's a well posed problem but ultimately a less interesting one, even if it's immediately "useful".
