Why Did DeepMind Solve Protein Folding?

They didn’t solve protein folding, in that they did not generate an explanation as to why common structural motifs are common, etc. people have been trying to explain things at that level for decades.

What it did was provide a perhaps better, perhaps more general structure generator than existing tools such as protein homology. Both, I think, rely on structural similarity as a means of guesstimating the structure of a new protein. No clue how any of these approaches would work with truly novel structures that are outside the training set.

A first principles approach would probably do better as the energetics of the structural motifs would be modeled properly.