We’ve had some interesting new developments in AI-based protein structure prediction, and yeah, I realize that I could have recycled that opening sentence several times over the last couple of years. Most folks who read this site will have followed at least the broad outlines of the AlphaFold and RoseTTAFold story, where machine learning has come through with an extraordinary number of predicted structures for proteins (and even protein complexes). And while I reiterate that this is not going to cause a sudden leap forward in drug discovery per se, I should also reiterate that it’s a tremendous accomplishment that immediately began speeding up work in structural biology and whose usefulness will only increase.

What we’ve seen in the last month or so are reports of another attack on protein structure from protein sequence, this time from language models. And here’s where the needs of biology and the life sciences piggyback once again on the work that’s been done through the really heavy-duty funding in machine learning and AI. Two of the areas that have had very large, very serious investment put into them (huge corporations, governments, national security and defense, etc.) are image recognition and language processing. Look at it from an intelligence-gathering perspective, and it’s easy to see why. There is an absolute flood of imaging data pouring down every minute of the day, coming out of everything from high-resolution multiwavelength spy satellites to inadvertently revealing backgrounds in casual photos on Instagram. Similarly, the amount of written and spoken language communication that’s out there to be read and heard is almost beyond measuring, and some of it (a very small fraction) is of great interest and importance.