If we knew how to design small-molecule drugs to attach to binding pockets on any given RNA molecule to interrupt or modulate its functions, it could open up a whole new realm of medical treatments. The problem is, if all you know about an RNA molecule is its nucleotide sequence, it’s very hard to predict where those binding pockets might be and what kind of drug might fit into them. As a PhD student at Stanford, Raphael Townshend designed a deep learning model to tackle that problem. Called ARES, the model started with a proposed structure for an RNA molecule with a known nucleotide sequence, and predicted whether that structure would turn out to be correct compared to real-world data. It turned out to be stunningly accurate—and unlike the algorithms behind generative AI models like ChatGPT or DALL-E, it built up its skills based on a tiny data set consisting of just 18 examples of known RNA structures. Now Atomic AI is building on Townshend's original model to create an engine for discovering new small-molecule drugs that could potentially interrupt any disease where RNA is a player.