⇧ [VIDÉO] You might also like this partner content (after ad)
Machine learning is a branch of artificial intelligence that uses algorithms to solve problems or create new molecules — like proteins. Last year, DeepMind used its AI to predict the shapes of almost any protein, to better understand their function and hopefully develop more effective treatments, or even vaccines. It is for this purpose that researchers have developed an artificial intelligence system capable of creating proteins, based on two different approaches.
The biochemical functions of proteins are generally provided by a small number of residues which constitute a functional site (for example, an active site of an enzyme or a binding site to a molecule). Therefore, creating proteins involves identifying the amino acids that produce the desired activity and the geometric conformations of the functional sites. In the past, protein design has found sequences that fold into a desired conformation, but it remains a challenge to obtain functional proteins.
Artificial intelligence expands the realm of possibilities, using multiple neural networks trained and trained on multiple protein data — a public database of hundreds of thousands of protein structures. ” In this work, we show that machine learning can be used to design proteins with a wide variety of functions. David Baker, lead author of the study and professor of biochemistry at UW Medicine, said in a statement. ” The proteins we find in nature are amazing molecules, but proteins designed this way can do so much more “.
Two different approaches: “hallucination” and “inpainting”
The researchers describe two machine learning approaches to scaffold proteins with novel functions, without the need to specify the molecule’s fold or secondary structure. The first, called ‘hallucination’, optimizes the amino acid sequences of proteins so that their final structures contain the desired functional site. It should be understood that amino acids are like the letters of a text and code for functional proteins.
In fact, the research team compares this first approach to the way one might write a book: You start with a random assortment of words – total gibberish. Then you impose a requirement, for example that the first paragraph be a dark and stormy night. The computer then changes the words one by one and asks itself: ‘Does it give more meaning to my story?‘. If so, it continues editing until the story is complete “. In the case that interests us, from a random chain of amino acids, the software mutates the sequence until a final version coding for the desired function is generated.
The second approach, called “inpainting”, works in the opposite direction. It starts from the functional site of the protein, then adds additional sequences to create a viable protein scaffold. Neural networks fill in the “missing pieces” of a protein’s structure in just seconds.
Laboratory tests have shown that many proteins generated by these approaches work as expected, for example by binding to the cancer receptor PD-1. In addition, these methods could be useful for the design of vaccines, often complicated by the molecular shape to be obtained. The researchers were thus able to create new proteins including the fragment of the pathogen necessary for the vaccine against the respiratory syncytial virus. The software was free to create any structure around this fragment, resulting in several potential vaccines with diverse molecular shapes.
#produces #proteins #vaccines #drugs