Metal sites no longer need high-resolution crystals: prune-then-optimize predicts location and ion type on cryo-EM and AlphaFold models

A protein structure can now yield both the position of a metal-binding site and the identity of the bound metal, rather than only flagging regions that might bind. Most earlier structure-based methods locate sites without specifying the metal type, and even the best-reported approaches reach only moderate localization accuracy on high-resolution crystal structures, with worse results on medium-resolution data. On a held-out test set separated at 30% sequence identity, PrOMet reached 80.0% recall and 65.0% precision at a 2 Å cutoff, where Metal3D reached 75.8% and 37.2%.

The method first narrows the whole protein surface to a small set of candidate points, then scores those candidates at atomic level to fix the site and the metal type. Scoring the full surface directly produces many false positives, so a high-recall, strongly contracting filter is needed first. The pruning stage retained 98.8% metal-site coverage while keeping only 2.76% of the initial grid points, cutting the false-to-true grid ratio from 1224:1 to 37:1 on the independent test set.

On inputs carrying structural error, such as cryo-EM structures and AlphaFold3 models, site and type predictions degrade only modestly. Structure-based methods are generally sensitive to structural noise, and their performance outside high-resolution crystal structures drops noticeably. On 200 AF3 models with global backbone RMSD at most 5 Å and binding-pocket side-chain RMSD at most 3 Å, PrOMet showed only a small drop relative to crystal structures, while Metal3D dropped more sharply.

The report covers metal-binding site and type prediction from protein structure, evaluated on X-ray crystal structures, cryo-EM structures and AlphaFold3-generated models, with training data drawn from X-ray and cryo-EM protein-metal complexes in BioLiP. The method accepts inputs with bound ligands or nucleic acids and produces predictions at protein-ligand and protein-nucleic acid interfaces. The authors propose as next steps building large-scale standardized benchmarks that better cover minority metal species and non-protein coordination environments, and integrating multi-level ensembles or dynamics such as equilibrium ensembles and MD-sampled structures.

Minority metal classes such as Mn, Co and K are underrepresented, giving relatively lower performance and less reliable uncertainty estimates for those types. Confusion analysis shows systematic swaps between chemically similar ions, with Zn over-represented and many Mn or Fe sites predicted as Zn. The method takes static structures as input and does not directly capture side-chain reorientation, conformational change or long-range allostery associated with metal binding. Poor AF3-model predictions occurred mainly when coordinating side chains were modeled poorly (RMSD above 2 Å) and the coordination number was low (fewer than 3).