Prediction of biological molecular structures accuracy doubled! Google AI drug development model AlphaFold upgraded significantly

Author: Li Dan

Source: AI Hard

Google's artificial intelligence (AI) model in the field of biomedicine has been significantly upgraded, making a greater contribution to accelerating drug development.

On Wednesday, May 8th, Eastern Time, Google announced that its AI research lab, Google DeepMind, and its sister company dedicated to enhancing drug discovery through AI, Isomorphic Labs, jointly developed a new AI model - AlphaFold 3. It is described as a "revolutionary model" that can predict the structure of biological molecules such as proteins, DNA, RNA, and how they interact with each other.

In a paper published in the journal Nature on Wednesday, Google officially introduced the third generation of this model, AlphaFold 3, stating:

"It can predict the structure and interactions of all biological molecules with unprecedented accuracy."

"Compared to existing prediction methods for interactions between proteins and other molecule types, we found at least a 50% improvement, and for some important interaction categories, our prediction accuracy has doubled."

DeepMind stated that many drugs are small molecules called ligands, which change the way they interact in human health and disease states by binding to receptors. By accurately predicting the structure and interactions of proteins and other molecules through AlphaFold 3, Google hopes it will change human understanding of the biological world and drug discovery.

DeepMind demonstrated that AlphaFold 3 can generate 3D structures of proteins, DNA, RNA, and smaller molecules, while also revealing how they combine together. The model can also model chemical changes that control cellular health functions, pointing out that once these chemical changes are disrupted, humans may become ill.

DeepMind has also launched an online free platform called AlphaFold Server, where scientists from around the world can use it for non-commercial research. Regardless of their technical expertise, with just a few clicks of the mouse, they can leverage the predictions of AlphaFold 3 and test hypothetical scenarios.

Alphabet and Google CEO Sundar Pichai stated that over 1.8 million researchers have used DeepMind's AlphaFold protein prediction in research on vaccine development, cancer treatment, and more. AlphaFold 3 represents the latest breakthrough, with unprecedented accuracy in predicting the structures and interactions of all biological molecules. We are sharing these capabilities through the AlphaFold Server to facilitate further scientific discoveries.

DeepMind's CEO and scientist Demis Hassabis described AlphaFold 3 as a significant milestone for them. "Biology is a dynamic system, and you need to understand how physiological properties emerge from interactions between different molecules in cells. You can think of AlphaFold 3 as a big step in that direction for us."

Proteins play crucial roles in cells, and their functions are closely related to their three-dimensional structures. Over 60 years ago, scientists determined that the amino acid sequence encodes the protein structure, but due to conformational diversity, it is almost impossible to try all possible structural arrangements. To overcome this challenge, scientists have used various methods, including using protein database fragments to predict local structures, but there are limitations. Conventional experimental methods to understand molecular interactions not only take time, possibly requiring several years, but are also extremely costly.

In 2018, scientists began applying machine learning methods that could spontaneously discover data patterns in protein structure prediction. DeepMind's team introduced the first generation of AlphaFold that same year and won first place in the 13th Critical Assessment of Structure Prediction (CASP) competition with a machine learning system, surpassing the runner-up by nearly 50% in accuracy. This model demonstrates the potential to accelerate biological research by understanding interactions through computation with sufficient accuracy In the 2020 CASP, AlphaFold2 performed remarkably well, with predictions for nearly 100 protein targets almost matching experimental results. This technology has had a huge impact in biomedicine and other fields, helping to understand the nuclear pore complex, redesign proteins for drug delivery and gene therapy, and has applications in pharmaceuticals and environmental protection. For creating AlphaFold and using AI systems to solve the challenge of protein structure prediction, last September, DeepMind's Demis Hassabis and another scientist John Jumper were awarded the Lasker Award for Basic Medical Research in the field of biomedicine.

After Google DeepMind officially released AlphaFold 3 on Wednesday, many netizens congratulated DeepMind on its new achievements.

One highly praised comment from a netizen stated that DeepMind is the lifeline of Google and humanity, and that DeepMind's co-founder and Chief AGI Scientist Shane Legg and his team have done a great job. I believe that Artificial General Intelligence (AGI) will consider DeepMind as its home and Shane Legg as its creator. Another comment mentioned that this is fantastic, and commercializing through Isomorphic is a wise move