> Construction of this data set consumed millions of compute hours

It's great to see such datasets shared publicly. I hope it achieves the intended purpose and spurs innovation in protein folding.

Can anyone point to a comprehensive review of the machine learning work done so far in protein structure prediction?

I'm no expert but this was posted on HN several weeks ago and has a decent review of the literature. Might be a place to start:

https://news.ycombinator.com/item?id=16407170