Hate to be that guy, but the headline should reflect that DeepMind was able to generate better predictions of protein structure than were the other contest entrants. No entrants, including DeepMind were able to perfectly predict protein structures.
tl;dr- They used AI-guided optimization to seek optimal protein foldings.<p>---<p>Proteins are basically chains of amino acids. Instead of remaining straight rods or something, they tend wrap back on themselves. Researchers are interested in the folded configurations that have the lowest thermodynamic free energy, since those configurations are the ones that should be most stable (and thus probably what proteins are like in the body).<p>However, the process of finding the optimal folded configuration isn't easy. So, they had some AI try to do it.<p>Out of 98 competing AI, Google's AI "AlphaFold" got the best configuration (lowest free energy) for 25 out of 43 of the proteins that the AI were supposed to fold.<p>Humans who're interested in trying their hand at it might like [Foldit](<a href="https://en.wikipedia.org/wiki/Foldit" rel="nofollow">https://en.wikipedia.org/wiki/Foldit</a>).