DeepMind, owned by Alphabet, may be best known for building AI that beats the world’s best Go players, but the company announced another, perhaps more important, breakthrough on the morning of December 1, 2012. It has been shown that it is possible to guess how certain proteins will fold with a surprising degree of accuracy. In some cases, these results are considered to be “competitive” with actual experimental data.
Although CASP has been in operation for 26 years, the scientific community has been able to make huge leaps in computing power and machine learning in the past few years to meet the challenge. In DeepMind’s case, this involved training AlphaFold 2’s predictive model on approximately 170,000 known protein structures, as well as a large number of protein sequences for which 3D structures had not yet been determined.
The team acknowledges that the test data is very similar to that used in 2018, when the original AlphaFold system scored top marks in CASP 13. (At the time, the organizers applauded the unprecedented advancement in DeepMind’s ability to predict computational methods for protein architectures.)
On top of that, DeepMind relies on Google’s approximately 128 cloud-based TPUv3 cores, which ultimately enabled AlphaFold 2 to accurately determine the structure of proteins in just a few days (if not quickly) – in some cases, predictions can be generated within hours, the media noted.
All of this sounds impressive – and it is – but there is still a lot of work to be done. Overall, the AlphaFold results represent a significant improvement in accuracy over the past few years, and as noted, some of DeepMind’s predictions are accurate enough to be comparable to experimental results at the atomic level.
However, the company notes, “For the most difficult protein targets, i.e., those in the most challenging free modeling category, AlphaFold’s average score is 87.0 GDT. “-slightly below the 90 GDT metric hurdle that CASP co-founder Moult used to compare results to real data. In other words, DeepMind hasn’t quite solved the protein problem yet, but it’s closer than many people think.
Recent Comments