via todd richmond
In a previous life, I was a “protein engineer.” We joked that we performed stupid protein tricks, one of which was creating a light switch on green fluorescent protein. In the world of biochemistry, there is a mantra – “structure determines function.” For a biochemist that looks at proteins (particularly enzymes, which are proteins that catalyze chemical reactions), that means if we can understand the structure, then we can get insight into function, then we can manipulate it (protein engineering) or stop it (drug design).
At some point in the 90’s, isolating and sequencing DNA because relatively easy and cheap. Since DNA holds the code for proteins, by sequencing DNA we understood the so-called primary structure of proteins – the linear sequence of amino acids. But proteins work by folding up into 3-dimensional structures, and going from sequence information to spatial information wasn’t well understood. We could predict some local structures (secondary structures – alpha helices, beta sheets), but the overall fold typically required further experimental techniques on purified proteins. Often getting the purified proteins was difficult, and even if you had it, getting it to crystalize (for X-ray crystallography) or behave in solution (for NMR spectroscopy) was more art than science.
But today we have AI – specifically Deep Mind’s AlphaFold algorithm. Deep Mind started training algorithms to solve 3D protein structures. And the algorithm got pretty good at it, and Deep Mind started a database of probable structures for proteins that did not have an experimental structure determined. Today they announced that they now have 200 million (with an m) structures available in their database.
The implications for health and medicine are massive. We are in the midst of a number of medical crises where new generation antibiotics and anti virals will need to be discovered or created. Having the structure of probable target proteins makes that process much less, “spray and pray,” and much more rational design. When we did our protein engineering back in the 90’s, we used solved structures and molecular modeling to help figure out where we needed to make changes. But for proteins where we didn’t have a structure, making progress was a mix of art and science, and you either needed to be very clever to infer structure/function relationships, or get lucky. Or in my case, we generated every possible mutation at certain points and looked at the results (site saturation mutagenesis). It was slow and often frustrating work.
Certainly the work of finding new agents to help humans survive and thrive in an increasing chaotic environment will remain challenging. But these AI-driven structures will at least unbind the one hand tied behind the back and open new avenues of inquiry. It is a great time to be a protein engineer…