Determining the three-dimensional structure of biomolecules is one of the biggest challenges in modern biology. Traditionally, pharmaceutical companies and research institutions often spend millions of dollars to determine the structure of a molecule, but even with such a large investment, they often fail.
AlphaFold on the cover of Nature magazine
In the past two years, the application of artificial intelligence (AI) technology in the prediction of the structure of biological macromolecules has attracted more and more attention. Among them, Google’s DeepMind’s new-generation AlphaFold system has made breakthroughs in accurately predicting the 3D structure of proteins based on amino acid sequences, which shocked the industry. Its accuracy can be compared with 3D analyzed by experimental techniques such as cryo-electron microscopy, nuclear magnetic resonance, or X-ray crystallography. The structure is comparable.
Proteins are vital to life, and understanding their three-dimensional structure is the key to understanding their functions. However, to date, only 17% of the human proteome has been structured through experimental techniques. In a paper published in Nature on July 22 [1], DeepMind and collaborators released the AlphaFold Protein Structure Database (AlphaFold Protein Structure Database) predicted by the new generation of AlphaFold. The database contains about 350,000 protein structures predicted by the AlphaFold system, covering humans and 20 commonly used model organisms in biological research. Among them, in terms of human proteome, AI has a strong influence on 98.5% (20 296) of human proteins. The structure makes a prediction.
RoseTTAFold system on the cover of Science magazine
In addition to AlphaFold, another system called RoseTTAFold developed by the research team at the University of Washington has also made important progress in predicting protein structure. Also in a paper published in Science in July, the RoseTTAFold system broke through an important limitation of AlphaFold: it can be used not only to predict the structure of a single protein, but also to predict the conformation of protein complexes [2]. These achievements are considered to be major breakthroughs in the field of life sciences and may have inestimable significance for promoting the research and development of new drugs.
AI reveals RNA structure on the cover of Science Magazine | AI algorithm selects the three-dimensional shape of RNA molecule from a large number of wrong and wrong shapes. The calculation and prediction of RNA folding structure is very difficult because there are too few known structures. The success of machine learning has opened the door to understanding and designing various molecules including drugs.
While AI technology has solved major challenges in the field of protein structure analysis for decades, scientists from Stanford University have also applied AI technology to RNA structure prediction. Their research results were newly published in Science on August 27 and appeared on the cover of the current issue [3].
RNA molecules are usually folded into complex three-dimensional shapes, which are critical to their function, but it is therefore difficult to determine the structure of RNA through experimental means. Since there are few known RNA structures, it is also very challenging to use algorithms to predict the structure of such biological macromolecules.
In this new study, Raphael JL Townshend et al. introduced a machine learning algorithm that can significantly improve the prediction of RNA structure. Although only 18 known RNA structures are used for training, this type of machine learning method (ARES network) has been able to identify accurate structural models. Compared with previous methods, it has always shown the highest level in the blind RNA structure prediction challenge.
Scientists believe that this machine learning method is expected to accelerate the deciphering of the RNA molecular structure, thereby helping to find drugs for the treatment of diseases that are currently incurable, such as the development of RNA targeted therapies.
In addition, the scientists emphasized that most recent advances in deep learning require large amounts of data for training. The algorithm in this research overcomes a major limitation of standard deep neural networks by effectively learning from a small amount of data. Therefore, the related methods are expected to solve many unsolved problems in data-scarce fields.
Reference materials:
[1] Kathryn Tunyasuvunakool et al. Highly accurate protein structure prediction for the human proteome. Nature (2021).
[2] Minkyung Baek et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science(2021).
[3] Raphael JL Townshend et al. Geometric deep learning of RNA structure. Science(2021).
Read more here: Source link