It takes mere seconds for a quiet computer in Bremen to do what once required days in a lab with expensive X-ray machines. Researchers there have built a new AI called BiteNetI that can scan any protein and instantly find where 14 different kinds of ions—charged particles like calcium, sodium, and potassium—like to dock. And it's two to three times more accurate than most tools already on the market, including Google DeepMind's famous AlphaFold 3.
Proteins are the tiny molecular workhorses inside our cells, but they rarely work alone. They need ions to hold their shape, spark chemical reactions, and pass signals between cells. When those ion connections go wrong, they can set off serious problems like neurological disorders, heart disease, or metabolic illnesses. So knowing exactly where an ion binds matters a lot—it helps scientists understand disease and design new medicines.
The trouble is, finding these sites the old-fashioned way means using high-resolution X-ray crystallography or spectroscopy, which is slow and costly. Older computer programs could usually only handle one ion type at a time, or they worked on the protein's sequence rather than its actual 3D shape. BiteNetI solves all of that.
Built by Igor Kozlovskii and Petr Popov at Constructor University and Constructor Labs, the model borrows the image-recognition technology found in modern smartphone cameras. It turns a protein into a 3D grid and scans its complex shape across 11 different atom types, like watching a short animated movie. It learned by studying a digital library of more than 12,000 protein complexes containing over 35,000 precisely mapped bound ions.
Because it works on many ion types at once, BiteNetI can predict the exact location and coordinating residues for all 14 ion types in a few seconds—about ten times faster than running separate specialist models one by one, without losing accuracy. On two of the field's main benchmarks, MIonSite and IonCom, it beat its rivals for most ion types. It even outperformed Metal3D, a model built specially for zinc, suggesting that training on many ions at once makes the whole system stronger.
Curiously, the AI got even sharper when water molecules were included, since water often helps "glue" ions into place. And its accuracy held up even on proteins imaged with cryo-electron microscopy, a different technique.
Popov says the team's next dream echoes the shift in generative AI: "Every binding site in a protein is a potential point of intervention for a drug. Today we predict them class by class—small molecules, peptides, ions—each with its own model. The next step forward would be one general model that recognizes all of them." That single all-in-one tool could give researchers one place to answer questions they currently tackle with many.