SuperMetal

Metal ions are among the most common protein cofactors, and where an ion sits determines catalysis, folding, and interaction specificity. Structure prediction has largely solved the backbone, but ion placement remains awkward: geometry-based methods scan for coordinating residue constellations, voxel-based learned methods discretize the space around candidate residues, and both inherit assumptions that limit accuracy or speed.

SuperMetal treats ion placement as a generative problem rather than a classification one. A score-based diffusion model learns the spatial distribution of zinc ions conditioned on a protein structure, so sampling produces candidate positions anywhere in the structure rather than scores attached to a predefined grid or residue list. A confidence model then evaluates each sampled ion and discards the poor ones, and surviving positions are clustered into final predictions. The practical consequence is that the model never needs to be told how many ions a protein binds — the number falls out of the clustering.

The framework was built at Vanderbilt University's Data Science Institute with collaborators at Heriot-Watt University and Leipzig University, posted as a preprint in March 2025, and published in the Journal of Cheminformatics that July.

#Key Features

• Ion count is inferred, not supplied: Predictions emerge from clustering high-confidence samples, so a structure with one zinc and a structure with six are handled by the same call.
• Sub-angstrom localization: Predicted zinc positions land within 0.52 ± 0.55 Å of experimentally determined positions, and spatial precision tightens further as the confidence cutoff is raised.
• Runtime that barely tracks protein size: A hierarchical graph representation, coarse-grained at distance and all-atom nearby, keeps inference under 10 seconds for proteins of about 2,000 residues.
• Tunable confidence threshold: The filtering cutoff trades coverage against precision explicitly, letting a user pick an operating point rather than accept a fixed one.
• Reusable as a probe framework: The same sample-filter-cluster construction applies to other small entities placed in a protein interior, which the authors have extended to water sites and to protein-ligand and protein-protein complexes.

#Technical Details

The score model learns the gradient of the conditional density of ion positions given the protein, using a variance-exploding forward SDE and an SE(3)-equivariant network over a heterogeneous graph of residues, atoms, and candidate ions; it was trained for 400 epochs. At inference, 100 candidate ions are initialized randomly across the structure and driven to favorable positions by reverse diffusion. The confidence model is a second SE(3)-equivariant convolutional network, adapted from the same architecture and trained with cross-entropy against a 5 Å mean-absolute-deviation label, that emits an invariant score per sampled ion. Survivors are grouped with DBSCAN using ε = 5 Å and a minimum of two points per cluster, and each cluster's centroid becomes one predicted ion. The released implementation is built on the DiffDock codebase and featurizes residues with ESM-2 650M embeddings, which are cached on first run.

Training data comes from ZincBind, a non-redundant collection of 19,154 physiologically plausible zinc sites across 19,103 PDB files, from which 10,253 structures were extracted after removing exogenous ligands and excluding proteins above 3,000 residues; 1,000 structures were held out for validation and a 350-structure test set was assembled to include the original Metal3D test set. On that test set SuperMetal reaches 94% precision at 90% coverage. The precision-coverage curve dominates Metal3D across its range: where Metal3D holds 100% precision at roughly 30% coverage, SuperMetal reaches roughly 70%, and at 88% coverage Metal3D falls to about 84% precision against SuperMetal's 95%. Runtime diverges sharply with size — near 2,000 residues Metal3D needs about 500 seconds, roughly 60 times longer. In a case study on zinc-binding structures, AlphaFold 3 achieved perfect precision and coverage only when given the correct ion count; supplying one ion instead of two dropped coverage to 50%, and supplying six produced spurious sites.

#Applications

The direct use is annotating metalloproteins: locating catalytic and structural zinc sites in newly solved or predicted structures, a prerequisite for mechanistic modeling, molecular dynamics setup, and docking against metal-dependent enzymes. In protein engineering, reliable placement supports the design and validation of engineered metal sites; in drug discovery it helps identify zinc-dependent pockets targeted by chelating inhibitors. Runtime nearly flat in protein length also makes proteome-scale annotation tractable.

#Impact

SuperMetal shows that diffusion generative modeling, already established for ligand pose prediction, transfers cleanly to ion placement and beats voxel-based predictors on both accuracy and speed. The released model's scope is deliberately narrow: trained on ZincBind, it predicts zinc, and other metals require new training runs on scarcer data. It also models the protein alone, ignoring RNA, small molecules, and structural water that in reality participate in coordination shells. The work was peer reviewed and published in the Journal of Cheminformatics in July 2025 under CC BY 4.0, but the GitHub repository ships no LICENSE file, leaving reuse terms for the software undefined even though weights and preprocessed training data are publicly downloadable.