Humanity Obtains Structures of All Proteins on Earth [Reading Science]

Google DeepMind Develops AI AlphaFold2
Released 200 Million 3D Database at End of Last Month
Revolutionary Foundation for Biotechnology Research Including Drug Development
Korean Dr. Baek Min-kyung Gains Attention for Developing More Advanced RoseTTAFold

[Asia Economy Reporter Kim Bong-su] DeepMind, a Google subsidiary well known for AlphaGo, the AI that defeated 9-dan Lee Sedol, has once again made a breakthrough. In November 2020, they developed an AI called AlphaFold2 that predicts the structure of proteins. On the 28th of last month (local time), they finally released a 3D database (DB) containing the structures of nearly 200 million protein forms existing on Earth. This includes protein structures found in humans as well as almost all animals, plants, bacteria, and other organisms on the planet. Protein structure analysis has long been a major challenge for scientists in medicine, biotechnology, chemistry, and all fields dealing with proteins. While it has been possible to directly analyze protein structures using experimental results such as X-ray crystallography or cryo-electron microscopy (Cryo-EM), analyzing a single protein structure could take from several days to years.

AlphaFold2 can complete this task in minutes to hours. It first learns from accumulated protein structure data and amino acid sequences, then predicts the structure from new amino acid sequences based on that knowledge. Before AlphaFold2 was developed, the number of protein structures solved by human effort was only about 170,000, which accounted for just 0.085% of all protein structures. AlphaFold2 independently revealed the remaining 99% of protein structures in just over a year. It is praised for leveraging the pattern prediction capabilities of AI to establish a foundation that opens a new chapter in the history of human scientific and technological development.

By quickly and easily solving the previously difficult and time-consuming task of determining protein structures, it has become possible to conduct biotechnology research more rapidly and accurately. This includes identifying disease causes, developing treatments and vaccines, creating enzymes that degrade plastics, treating rare genetic diseases, and improving seeds of plants and animals.

‘RosettaFold’ is another AI for protein structure prediction. It boasts even better performance than AlphaFold2. Developed by a team at the University of Washington’s Protein Design Institute, including former postdoctoral researcher Min-Kyung Baek, who will join Seoul National University’s Department of Life Sciences as an assistant professor this September, RosettaFold was created based on AlphaFold2’s code. It provides more detailed and efficient 3D prediction models of organism protein structures and was selected as the best innovative research achievement of the year by the international journal Science in December last year.

Former researcher Baek said, "After hearing about AlphaFold2 developed by DeepMind and waiting for the DB release, we decided to try developing one ourselves and started the research." She added, "We plan to continue upgrading the research domestically and aim to study interactions with molecules other than proteins in the future."