Artificial intelligence algorithms solve problems in structural biology

2022-05-13 0 By

A new artificial intelligence algorithm can pick out the correct three-dimensional structure of an RNA molecule from the wrong one.Computer predictions of RNA folding are often important and difficult — because so few structures are known.Determining the three-dimensional structure of biomolecules is one of the most challenging problems in modern biology and medical discovery.Companies and research institutions often spend millions of dollars trying to determine the structure of a molecule, and even such efforts often end in failure.Using clever, new machine learning techniques, Stanford doctoral students Stephan Eismann and Raphael Townshend, under the guidance of Ron Dror, an associate professor of computer science, have developed a way to overcome this problem by using computers to predict precise structures.Of particular concern is the success of their method despite learning only a few known structures, making it suitable for molecules whose structures are difficult to determine experimentally.Their studies, published in the Aug. 27, 2021, and December 2020, respectively, in The journal Science, detail RNA molecules and multiprotein complexes.The Science paper was written in collaboration with Rhiju Das, an associate professor of biochemistry at Stanford.”Structural biology is the study of molecular shape, and it has a credo: structure determines function.”Townshend said.Algorithms designed by researchers can predict the exact structure of the molecule, allowing scientists to explain how different molecules work.Applications of this approach range from basic biology research to practice in drug design.The various levels of protein structure.The structure of a protein is spatially complex because of its folding.”Proteins are molecular machines that do all sorts of things.In order to perform these functions, proteins tend to bind together, “Eismann said.”If you know a pair of proteins that are associated with a disease and you know how they bind at a three-dimensional level, then you can try to influence their interaction with a very targeted, targeted drug.”Eismann and Townshend co-authored a paper published in Science with Andrew Watkins, a Stanford postdoctoral scholar in Das’s lab,And co-wrote the paper published in Proteins with Nathaniel Thomas, a former Stanford doctoral student.Instead of specifying exactly what makes structure prediction more accurate, researchers have left the algorithms to discover these molecular features themselves.They did so because they found that the traditional techniques that provided this knowledge predisposed the algorithm to certain features and prevented it from finding other information features.”The problem with manually filtering features in an algorithm is that it becomes biased — biased toward what the filter thinks is important.You may miss out on information you need to complete your study.”Eismann said.”In the absence of explicit instructions, the network learned to look for basic concepts that are critical to the formation of molecular structure.””What’s exciting is that algorithms have clearly picked up the important things we know about them, and they’ve found properties we’ve never heard of before,” Townshend said.Having succeeded with proteins, the researchers then applied the algorithm to another important class of biomolecules, RNA.They tested their algorithm in a series of “RNA puzzles” from a long-standing competition in their field.In each case, the tool outperformed all the other puzzle participants, and the algorithm itself was not designed specifically for RNA structure.The study has so far had success with protein complexes and RNA molecules, and the researchers are excited to see how it could be used elsewhere.”Most of the recent breakthroughs in machine learning have been trained by large amounts of data.Our approach was successful with only a small amount of training data — which tells us that relevant approaches can help us solve unsolved problems in areas where data is scarce.””Said Dror, senior author of the Proteins paper and co-senior author of the Science paper.New molecules developed using artificial intelligence could be used in drug research and development.In structural biology in particular, the team says they have only scratched the surface of the scientific progress to be made.”Once you’ve mastered this basic technique, you’ve taken your understanding to another level, and then you can start asking a bunch of questions.””With this kind of information, for example, you can start designing new molecules and drugs, and that’s an area that everyone is really excited about,” Townshend said.From: China Digital Science and Technology Museum