For as long as cells have needed ions to function, scientists have needed to know exactly where those ions attach to proteins — a question that has demanded years of painstaking crystallography and hand-drawn maps. A research team has now trained a deep learning model called BiteNetI on more than ten thousand protein-ion structures, teaching it to recognize the three-dimensional signatures of fourteen biologically critical ions at once, with an accuracy two to three times greater than any existing tool. The work, published in Nature, suggests that a fundamental bottleneck in structural biology