Nature uses a variety of metal ions as cofactors for specific tasks. The three main classes of metals in biological systems are alkali ions (e.g. sodium and potassium), alkaline earth ions (e.g. magnesium and calcium) as well as transition metal ions (e.g. zinc or iron and many others). To date no selective predictor to localize each of these metals in a given protein structure exists. Current methods require either preselecting a specific ion (e.g MIB2 (Lu et al., 2022), ESMBind (Dai et al., 2024) or AlphaFold3 (Abramson et al., 2024) or a location (MIC Shub et al. (2024)). In this work, we describe an extension of the recently introduced Metal3D framework (Dürr et al., 2023) that was originally trained on zinc sites to predict the location of all biologically relevant classes of metal ions as well as to classify their coordination geometry. Our model is the first of its kind and even outperforms Metal3D for the prediction of the location of Zn2+ and generalizes well to the other metals in terms of location prediction as well as identity classification. Comparing the model to several other available tools such as MetalSiteHunter, AlphaFold3, MIC, MIB2 and MetalHawk highlights important shortcomings in these tools with respect to data bias, prediction of negative sites and selectivity. However, concerning coordination geometry prediction, similar to other work, we find that our method cannot accurately make correct classifications beyond the most common classes in natural proteins i.e. tetrahedral and octahedral arrangements. AllMetal3D is available as ChimeraX extension, standalone web app as well as python package https://github.com/lcbc-epfl/allmetal3d

AllMetal3D: joint prediction of localization, identity and coordination geometry of common metal ions in proteins