Next-Gen AI for Science (Category Science)

Interesting. report John Jumper, Nobel laureate in chemistry, at AI Startup School Y Combinator. Interestingly, John received the Nobel Prize along with Demis Hassabis, who is much better known for playing the role of CEO of Google DeepMind. (I told you about him once. speech about AI reshuffle at Google). But I didn't know much about John, and this report gave me a little more information. John's scientific career began with physics, but after disillusioning with his academic career, he moved to computational biology, working on applying machine learning to the study of proteins.

Key ideas are presented below. 1. Revolution in Predicting the Structure of Proteins The main achievement of John’s team was the development of AlphaFold, an artificial intelligence system that decided to create a new system. 50The problem of predicting the three-dimensional structure of proteins by their amino acid sequence. At the CASP14 competition in 2020 AlphaFold 2 demonstrated an accuracy comparable to experimental methods, achieving a GDT-TS score 90 for most proteins. By the way, I've already talked about the complexity of predicting proteins in survey From Orgasm to Immortality. Drag designer notes" 2. Importance of Research in Machine Learning John emphasized the critical role of scientific research along with data and computing power. As he noted in the report, the team was able to show that AlphaFold 2trained in 1The percentage of data available was as accurate as the previous version of AlphaFold. 1, trained on a full set. This demonstrates that new ideas and research can be 100 More valuable than just increasing the amount of data. 3. Components of success According to John, three key components led to the breakthrough:

• Data: 200,000 known protein structures from an open PDB database
• Computing resources: 128 TPU v3 cores within two weeks for final model
• Research: Small team (About two main researchers)Developing new architectural solutions 4. Openness and accessibility A critical decision was the discovery of AlphaFold’s source code and the creation of a database predicting protein structures. K 2022 The AlphaFold database contains more 200 millions of predictions of protein structures, covering virtually all known proteins. This openness has led to massive acceptance of the tool by the scientific community. 5. Impact and application AlphaFold is already in use. 2 millions of researchers in 190 countries. John gave an example of a special issue of the journal Science on the nuclear pore complex, where three out of four papers actively used AlphaFold, and the DeepMind team was not directly involved in these studies. Users have found ways to use AlphaFold that the team did not foresee. For example, researcher Yoshitaki Moriwaki showed two days after the code was released how AlphaFold could be used to predict interactions between proteins, which became the best method for predicting protein interactions. John also spoke about the work of MIT’s lab, which used AlphaFold to reengineer a “molecular syringe,” a protein system capable of delivering targeted proteins to specific cells. The researchers were able to modify the protein to target drug delivery to mouse brain cells.

If we sum up, we can see that such works accelerate scientific progress in the whole field. AlphaFold demonstrates how AI can work as an amplifier for experimental researchers, filling gaps in scientific knowledge. The most intriguing question, according to Jumper, remains the degree of commonality of the application of AI in science. Will it be a few narrow areas with transformational impacts or very broad systems? He expects that, eventually, broad systems will be created that will revolutionize science as a whole.

#AI #Science #Architecture #ML #Software #Engineering #Startup