PRIME: Probe-Based Identification of Metal-Binding Sites Using Deep Learning Representations

PRIME^[1] is a hybrid deep learning framework that predicts metal-binding sites in proteins by combining protein language and structure models with a probe-based scanning algorithm. It outperforms existing methods across diverse metal ions and scales to AlphaFold2 and cryo-EM models.

GitHub | Paper | License

Run a Prediction

Upload a structure file (PDB / CIF) or give a PDB / UniProt ID, pick a metal ion, and submit. Each run opens a dedicated, bookmarkable result page.

Metal Ion Type

Upload Structure File (PDB/CIF) Accepts PDB, CIF, or ENT format.

PDB ID / UniProt ID PDB structures are fetched from RCSB^[2]; UniProt accessions use the AlphaFold2^[3] model.

Rotations Random orientations each probe is scored at (averaged). Higher = more robust, but linearly slower. Default 5.

Predictions currently run on a CPU server, so a typical job takes a few minutes. The result page shows live progress, can be bookmarked or shared, and is retained for one week.

How PRIME Works

Architecture of PRIME: PRIME-seq ranks residues, a probe generation algorithm proposes candidate sites, PRIME-probe scores them with a 3D ResNet, followed by post-processing. — Figure 1: Architecture of PRIME. **PRIME-seq** uses a protein language model to rank residues by binding propensity; a **probe generation** step turns the top residues into candidate sites, sharply reducing the search space; **PRIME-probe**, a 3D ResNet pre-trained on PDB structures, scores each probe's binding probability and refines its position; post-processing then clusters the probes into the final predicted sites.

Performance

Comparison of metal-binding-site prediction across methods and metal ions: PRIME versus AllMetal3D, Metal3D and BioMetAll. — Figure 2: Comparison across methods. **(A)** On zinc-binding sites PRIME attains the best F1 (0.892) against AllMetal3D^[4], Metal3D^[5] and BioMetAll^[6], balancing precision and recall with low positional RMSE. **(B)** Across additional metal ions, PRIME (hard-mining and weighting variants) achieves the highest F1 scores for the broadest range of ions.

Supported Metal Ions

Transition metals

Zn²⁺Mn²⁺ Fe³⁺Fe²⁺ Cu²⁺Cu⁺ Co²⁺Ni²⁺ Cd²⁺Hg²⁺

Alkali & alkaline-earth metals

Ca²⁺Mg²⁺ Na⁺K⁺

References

^ Xu, S.; Onoda, A. Probe-Based Identification of Metal-Binding Sites Using Deep Learning Representations. bioRxiv 2025. DOI: 10.1101/2025.10.04.680417
^ Berman, H. M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T. N.; Weissig, H.; Shindyalov, I. N.; Bourne, P. E. The Protein Data Bank. Nucleic Acids Res. 2000, 28 (1), 235–242. DOI: 10.1093/nar/28.1.235
^ Jumper, J.; Evans, R.; Pritzel, A.; Green, T.; Figurnov, M.; Ronneberger, O.; Tunyasuvunakool, K.; Bates, R.; Žídek, A.; Potapenko, A.; et al. Highly Accurate Protein Structure Prediction with AlphaFold. Nature 2021, 596 (7873), 583–589. DOI: 10.1038/s41586-021-03819-2
^ Dürr, S. L.; Rothlisberger, U. AllMetal3D: Joint Prediction of Localization, Identity and Coordination Geometry of Common Metal Ions in Proteins. bioRxiv 2025. DOI: 10.1101/2025.02.05.636627
^ Dürr, S. L.; Levy, A.; Rothlisberger, U. Metal3D: A General Deep Learning Framework for Accurate Metal Ion Location Prediction in Proteins. Nat. Commun. 2023, 14, 2713. DOI: 10.1038/s41467-023-37870-6
^ Sánchez-Aparicio, J.-E.; Tiessler-Sala, L.; Velasco-Carneros, L.; Roldán-Martín, L.; Sciortino, G.; Maréchal, J.-D. BioMetAll: Identifying Metal-Binding Sites in Proteins from Backbone Preorganization. J. Chem. Inf. Model. 2021, 61 (1), 311–323. DOI: 10.1021/acs.jcim.0c00827

Please contact shijie.xu@ees.hokudai.ac.jp for any questions.

Changelogs

2026-05-28Updated web interface with interactive 3D visualization.
2025-10-01First release.