Shaharyar Lakhani published on June 24, 2019:
Artificial intelligence is a powerful tool that is impacting almost every facet of our existence, including and especially science and technology. An artificial neural network, a computer system modeled on the human neural system, allows a computer to figure out what constitutes a particular subject on its own. Take a cat for instance; telling a computer that a cat is defined as an animal with whiskers, a tail, and paws is one way for the computer to identify a cat, but this way can lead to many errors since there are so many variations of cats. However, if a computer were fed multiple images of different cats, it would itself be able to come up with what constitutes a cat and continuously update its perception with every image.
The team of Raghav Shroff, Danny Diaz, and Austin Cole in the Ellington lab are applying this same concept not to cats, but to amino acids in the context of the proteins they comprise. Their project aims to use convolutional neural networks to recognize the ‘amino acid-ness’ of individual amino acids in a protein structure. This is a very different way of looking at protein structures, and draws inspiration from Wen Torng and Russ Altman (Stanford). Instead of calculating the physical or energetic properties of how each amino acid fits into a protein, they used neural nets to essentially ask protein structures on a position by position basis, “What amino acid would fit best here?” Just like the cats example, the team feeds in images and first lets the neural network learn what amino acids are already present at different spots in a protein. Most of the time, the answer matches the amino acid that’s already there, giving them confidence that their network is working correctly. However, in some cases the answer is different … and this gives the team the insight that maybe there is beneficial mutation waiting to happen! They call the project “JMBLYA” because like the dish, they mix various components together to make a final product, in this case the mutated amino acids to make a protein. Various tests of JMBLYA have now proven it to be a time machine of sorts, where it can predict the future evolution of a protein to be more stable, and allows us to intervene in the present to make the beneficial mutation.
The biotechnology implications of this insight are enormous, and the team has already used this method to alter the structure and function of three different proteins: the antibiotic resistance protein beta-lactamase, a blue fluorescent protein (shown below), and the enzyme phosphomannose isomerase. They are working towards ever more difficult protein targets of industrial and biomedical relevance, and have formed a company, AI Protein Solutions, to work with others on how to speed up evolution. Their entrepreneurial spirit has already led to negotiations on the use of their (now very blue) blue fluorescent protein, which they call “Blue Bonnet,” a nod to their Texas origins. Into the future, as computational power grows, we are confronted with the odd prospect of machines knowing more about our evolutionary futures than we do, and one of the reach goals of the team is to begin to predict the evolution of the human proteome (including mutations that abet or resist cancer) for millions of years to come.