DeepMind AI cracks 50-year-old problem of protein folding | DeepMind | The Guardian

This is an incredible milestone.

To learn how proteins fold, researchers at DeepMind trained their algorithm on a public database containing about 170,000 protein sequences and their shapes. Running on the equivalent of 100 to 200 graphics processing units – by modern standards, a modest amount of computing power – the training took a few weeks.

Okay, my main takeaway on that is that this wasn't much data, and didn’t need much computing power. For AI purposes that isn't a big footprint. So, cheap and easy to get your hands on.

AlphaFold not only outperformed other computer programs but reached an accuracy comparable to the laborious and time-consuming lab-based methods. When ranked across all proteins analysed, AlphaFold had a median score of 92.5 out of 100, with 90 being the equivalent to experimental methods. For the hardest proteins, the median score fell, but only marginally to 87.

Cheap hardware without a ton of data and it outperformed the other methods. Amazing.