The zhang / i-tasser
New AI model allows evolutions in creating new proteins – and goes far beyond those in nature.
US scientists have used evolutionarryscale model 3 (ESM3) to synthesize a new protein called esmGFP (fluorescent green protein), which shares only 58% of its material with its nearest natural relative, says a.
Fluorescent proteins give some oceanic organisms their brightness, and their use as markers are enormous in medicine by biotechnology.
Now, “more than three billion years of evolution have produced an image of the coded biology in the space of natural proteins,” write the authors of this month in the Science.
“Here we show that large -scale -trained language models in evolutionary data can generate functional proteins that are far from known proteins“They add.
ESM3 was trained with 3.15 billion sequences236 million structures and 539 million protein notes.
In fact, the base that makes this AI is similar to that of our well -known chatgpt: just as it can create a new text based on prior information to which it has access, also the new technology is able to recognize based on its training as proteins that will or will not work.
“Proteins can be seen as existing in an organized space where Each protein is neighboring all the others that are at a mutational event away ”, comment the authors. “The structure of evolution appears as a network within this space, linking all the proteins by the paths that evolution can take between them.”
“We choose the functionality of fluorescence because it is difficult to obtain, easy to measure and one of the most beautiful mechanisms of nature,” scientists also justify.
It is still necessary to submit this method to tests, but scientists expect to start soon produce proteins to various areas of science.
“Protein language models do not work explicitly within the physical restrictions of evolution, but they can implicitly build a model Of the multiplicity of potential paths that evolution could have followed, ”concludes the investigators.
