Use SuperMetal Online
Commercially Available SuperMetal No-Code Web Server
SuperMetal: Fast, Precise Zinc Ion Location Prediction
SuperMetal is a generative AI framework that predicts where zinc ions bind in a protein structure. It pairs a score-based diffusion model with equivariant graph neural networks. Instead of scoring a grid of positions, it treats metal binding as a generative modeling problem: it learns a distribution over metal positions given the protein structure, then samples from it.
Metric | SuperMetal result |
|---|---|
Precision | 94% |
Coverage | 90% |
Localization accuracy | 0.52 ± 0.55 Å from experimental zinc positions |
Speed | Under 10 seconds for proteins of about 2,000 residues |
SuperMetal also does not need to be told how many metal ions to look for.
Why Metal-Binding Site Prediction Matters
About one-third of Protein Data Bank structures contain metal ions.
Zinc binds to roughly 10% of all human proteins, and it is essential to the activity of more than 300 enzymes across all six enzyme classes.
Zinc binding also stabilizes the folded conformations of protein domains.
Determining zinc sites experimentally is costly and time-consuming, which makes accurate computational prediction valuable.
Where Other Approaches Fall Short
The paper describes four families of existing methods and their limits:
Template-based methods are limited by the templates available and may miss novel binding sites.
Sequence-based methods can't give atomic-level detail on protein–metal interactions.
Physics-based methods are hard to apply, because molecular dynamics struggles to find force fields that reproduce transition-metal coordination distances, and QM/MM is too expensive for typical design tasks.
Structure-based 3D CNNs such as Metal3D need fine voxel grids, whose cost grows cubically with resolution, and they are sensitive to input orientation, so they need rotational data augmentation.
SuperMetal works in continuous 3D space with an SE(3)-equivariant model. It needs no voxelization and no rotational augmentation.
How SuperMetal Works
The pipeline has three stages.
Graph preprocessing. The protein structure becomes a heterogeneous geometric graph covering protein residues, protein atoms and metal ions, with diffusion time included.
Diffusion sampling. SuperMetal places 100 candidate metal positions at random across the protein. A learned score function then moves each one toward its most favorable position through reverse diffusion. Random starting points help sample sparse or unusual coordination environments.
Confidence filtering and clustering. An SE(3)-equivariant confidence model scores each candidate. Low-confidence positions are discarded, and DBSCAN clustering merges the rest. Each cluster yields one metal ion, placed at the cluster's average position.
Multi-Scale Efficiency
SuperMetal uses a hierarchical interaction scheme. Metal ions far from residues interact only through coarse-grained terms, and nearby ions use the full atomic structure of the residues. This avoids building very large graphs, which is how it stays fast on large proteins.
Training Data
The model was trained on the ZincBind database, a non-redundant collection of 19,154 zinc-binding sites across 19,103 PDB files. From it, 10,253 PDB files were used, with exogenous ligands removed and only zinc retained. Structures above 3,000 residues were excluded. The test set contains 350 structures, none with binding sites similar to those in the training sets of SuperMetal or Metal3D.
How SuperMetal Compares
Versus Metal3D
Coverage at high precision: When Metal3D reaches 100% precision, its coverage is about 30%. SuperMetal reaches about 70% coverage at the same precision.
At high coverage: At 88% coverage, Metal3D's precision is about 84%, against about 95% for SuperMetal.
Spatial accuracy: SuperMetal's mean absolute deviation improves from 0.61 ± 0.66 Å to 0.44 ± 0.58 Å as the confidence cutoff rises. Metal3D's median deviation worsens at stricter cutoffs, from 0.36 Å to 0.87 Å.
Speed: For a protein near 2,000 residues, Metal3D takes about 500 seconds. That is around 60 times longer than SuperMetal.
Versus AlphaFold 3
A case study compared the tools on two proteins: 5IN2 (a Cu/Zn superoxide dismutase) and 6BTP (bone morphogenetic protein 1).
AlphaFold 3 needs the number of zinc ions to be specified in advance. On 5IN2, giving it the correct number (two) produced 100% precision and coverage. Giving it one ion cut coverage to 50%, and giving it six produced multiple incorrect predictions.
On 6BTP, AlphaFold 3 reached only 50% precision and 50% coverage even with the correct ion count. SuperMetal reached 100% on both.
The authors also note that AlphaFold 3's source code is not public, and its server accepts sequences rather than existing PDB structures.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with SuperMetal on Tamarind Bio
The authors point to several uses for accurate metal-ion placement:
Metalloprotein engineering: Locate zinc sites before designing or modifying a metal-dependent protein.
Structural studies and binding site prediction: Add precisely placed metal ions to structures that lack them.
Multi-body docking: Include metal cofactors when docking complexes.
Understanding metal-dependent biology: Reliable positions support work on protein stability and enzyme catalysis.
Screening unknown sites: Because the ion count is not an input, you can run it on proteins whose metal content is unknown.
How to Use SuperMetal on Tamarind Bio
Access the platform. Log in to tamarind.bio.
Prepare your protein structure. SuperMetal takes a 3D structure as input. If you only have a sequence, generate a structure first with a structure-prediction tool on Tamarind. For this you should confirm the tool and structure quality.
Check protein size. The training data included proteins up to about 3,000 residues. Larger structures are supported if memory allows. For very large proteins or limited resources, split the protein into sections, predict each piece, and merge the results.
Run SuperMetal. Use the open-source code. The model samples 100 candidate positions per structure.
Review confidence scores. Each candidate gets a confidence score. Higher thresholds trade coverage for precision, and the paper reports results across a range of cutoffs.
Inspect the clustered ions. Each DBSCAN cluster gives one predicted ion. A prediction counts as correct in the paper if it lies within 5 Å of an experimental site.
Cross-check if needed. Compare against a cofolding prediction on Tamarind, as the paper's case study did with AlphaFold 3, keeping in mind that those models need the ion count supplied.
Source: Lin, Su, Liu, et al., SuperMetal: a generative AI framework for rapid and precise metal ion location prediction in proteins, Journal of Cheminformatics 17, 107 (2025). DOI: 10.1186/s13321-025-01038-9