Artificial Intelligence has learned to create enzymes that don't exist in nature, starting from scratch and using the same types of algorithms used to master natural language.
The success achieved by the collaboration of researchers from the University of California in San Francisco with the IT company ProGen and published in the journal Nature Biotechnology concretely paves the way for the design of new drugs and completely new molecules.
After having decoded the three-dimensional structure of practically all the proteins existing in nature, with AlphaFold by Google and then with Meta, the AI algorithms center another important objective, that is to design new and concretely useful molecules, surpassing what nature has done in billions of years of evolution.
The work of American researchers aimed to produce new molecules capable of replicating some of the functions of an enzyme present in the albumen of chicken eggs, known as lysozyme (Hewl), used in the food industry and which also has analogs in human tears, saliva or milk with the ability to protect against bacteria and fungi.
To develop something similar, the researchers used an AI born to replicate human language and provided it with data relating to the amino acid sequences (the 'syllables' that make up proteins) of 280 million proteins but identifying in particular 56 thousand enzyme sequences of the lysozyme family.
Using that data, the AI came up with a sort of new grammar and semantics of its own and thus reworked a first batch of 100 new enzymes.
These were analyzed by human researchers who selected the 'best' 5, of which 2 demonstrated the ability to fight bacteria similar to that of Hewl's lysozyme.
The most surprising thing is that the sequence of these new enzymes has very large differences compared to the natural molecule (a similarity between 90 and 70%), a molecule that never existed before.
