Protein language models are statisticians, not oracles. They are good at what evolution has rehearsed many times and weaker where it has left few traces — and they lack physical common sense: cut a chunk out of a protein, and the model still folds the fragment as if the rest were there. Yet what matters in biology is often the switch between states: “I’m off until something comes along and tells me to turn on.” These active, minor states rarely make it into the crystal structures AlphaFold was trained on.

To build Dyna-1, Hannah and her team read the absence of signals in archived NMR data as the presence of motion. This turned about a hundred carefully curated data sets into more than ten thousand usable ones — at the price of some noise, which, as Hannah says, “is actually not a bad thing for machine learning.” Dyna-1 predicts where in a protein biologically relevant motion is likely to occur, and researchers who focus on the spots it flags are already more successful at improving proteins. The next goal: a model “that can tell me the actual 3D structure of the other states I care about.”