AlphaFold and the state of AI in biotech

AlphaFold was all over the news this week. What it’s done is hugely important, but also nowhere near enough. In case you’ve forgotten your high-school biology, proteins are complex molecules made up of amino acids. They can be extremely complicated molecules, but one of the more interesting things about them is that their function can largely, if not entirely, be determined by their shape. Therefore, understanding what shape a given protein takes plays a key role in two parts of biology: understanding what naturally occurring proteins do (e.g., hemoglobin in the blood), and constructing new proteins to do what we want (e.g., “consume” industrial waste). We’ve been slowly and painstakingly discovering what shape proteins take using a variety of experimental methods since Max Perutz first did it in 1959 with x-ray crystallography. A dozen years later, it was postulated that a protein’s shape could be wholly deduced from its amino acid sequence — which has led to a 50-year attempt to get computers to predict the shape of proteins. So far, they’ve been terrible. This is not inherently surprising, as it would take longer than the age of the known universe to enumerate all of the potential configurations of a typical protein. Our computers just aren’t that fast yet. However, the folks over at DeepMind — the AI startup acquired by Google in 2014 — are getting better and better at it. In one of my very first newsletters, I wrote about how they crushed their competitors in the 2018 CASP (a biennial competition between computers at solving this problem). At the time, doing eight times better than the next closest competitor, they were at about 68% accuracy in their predictions. This year, they upped that to over 90% accuracy. This is tremendously important research, and a huge advancement, but there’s still a long way to go. AlphaFold is crushing it, but even at these accuracy levels, they’re only getting the fully-correct final shape about 67% of the time — which means we still are using experimental verification to even know which ones they’ve gotten right. That said, at a minimum, AlphaFold is already good enough to provide guidance to researchers, and tremendously speed up the search for the shapes of proteins. That alone is a major breakthrough. More details should become available when their final, peer-reviewed paper becomes available.

Leave a Reply

Your email address will not be published. Required fields are marked *