The result worth dwelling on is zinc. Metal3D was built for that one metal, and the 14-ion model outscored it anyway [9]. The authors read that as evidence that training on many ion types at once lifts performance overall [9], which is a claim about the data rather than the code: the coordination geometry around a magnesium apparently teaches the network something useful about a calcium. It also suggests the training set was dense. More than 35,000 mapped ions across more than 12,000 complexes [4] works out to roughly 2.9 ions per structure [19], so the model learned in crowded environments rather than from one tidy site at a time.
Then the headline ratio. Two- to threefold better accuracy than existing predictors, AlphaFold 3 among them [2], arrives in the material without a named metric, and for a site-detection task "accuracy" can mean precision within a distance cutoff, recall of known sites, or a ranked-detection score. Those move independently. The per-ion breakdown is the more useful object: better scores for calcium, potassium, magnesium, phosphate and sulfate, with AlphaFold 3 keeping a slight edge on carbonate and sodium [7], five ions to two [18].
Popov is careful about the framing, and the care is warranted. He says the two systems answer different questions, and that the finding is about a compact task-specific model annotating an existing structure very accurately and very fast [10]. AlphaFold 3 places ions in the course of building a whole structure from sequence [17]. BiteNetI needs a structure to start from. In a working pipeline those are sequential steps.
Which brings up the condition I would footnote. Precision rose significantly when surrounding water molecules were included, since waters often help hold an ion in place [11]. Ordered waters are a property of well-resolved experimental structures. The paper reports accuracy holding across imaging methods including cryo-electron microscopy [12], but the supplied material says nothing about solvent-free predicted models, which is exactly the input you would reach for if you wanted to annotate a proteome nobody has crystallised. That condition travels with every number above.
The speed buys something real regardless. Determining these sites experimentally means high-resolution X-ray crystallography or spectroscopy [13], and the computational alternatives have generally been single-ion, sequence-based, or unable to scale to many proteins [14]. Seconds per structure across 14 ion types [1][6] turns the question of whether a given point mutation weakens or restores a site [15] from a project into a query, and it makes pocket triage for charged chemical groups a cheap first pass in early drug design [20]. The reason to care about ion sites at all is that they are not readable from sequence: a bound ion stabilises one particular conformation, and which one it stabilises is structural information [22].
