Artificial intelligence has generated proteins out of thin air that do not exist in nature

The prodigies of artificial intelligence never cease to amaze us. In fact, AI knows some of the underlying principles of biology. But this time it has managed to generate artificial and perfectly functional molecules, or rather enzymes, out of nothing. In laboratory tests by researchers at the University of California at San Francisco and Salesforce Research, some of these molecules functioned exactly like those found in nature, even when their artificially generated amino acid sequences differed significantly from any known natural protein. The study has just been published in Nature Biotechnology.

The new technology, called ProGen, could therefore significantly accelerate the development of new proteins that can be used for a variety of applications, from that of new drug design to degrading plastics. "Artificial models work much better than those inspired by the evolutionary process," said James Fraser, one of the authors of the study. “We now have the ability to fine-tune the generation of these properties to have specific effects. For example, an enzyme that is incredibly thermostable either likes acidic environments or doesn't interact with other proteins."

The steps of the generation

To create the technology, scientists started from an artificial intelligence model created to replicate language and, through machine learning, provided it with the sequences of amino acids from 280 million different proteins, so that he could learn them and be able to identify 56,000 lysozyme sequences. From here, the system was able to generate 100 new artificial sequences, which were then tested on how closely they resembled those of natural proteins. In particular, the researchers selected five which were then compared with an enzyme present in chicken egg white (Lysozyme Hewl), whose activity is very similar to proteins found in tears, saliva and milk. and which have the task of defending against bacteria and fungi.

From subsequent analyses, the researchers observed that two of the artificial enzymes were able to counteract bacteria with an activity comparable to Hewl's lysozyme. The two sequences, the researchers say, were about 90% to 70% identical to any known protein, but in subsequent trials, the team found that the two man-made enzymes showed the same activity even when only 31.4% of their sequence resembled any known natural protein. Not only that: Artificial Intelligence has also been able to learn how enzymes should be modeled, simply by studying the raw data of the sequence. Indeed, measured with X-ray crystallography, the atomic structures of the artificial proteins were exactly as they should be, even if the sequences had never been seen before. "The ability to generate functional proteins from scratch demonstrates that we are entering a new era of protein design," commented first author Ali Madani. "This is a versatile new tool available to protein engineers and we look forward to seeing therapeutic applications."