PRIME

Where a protein binds a metal ion, and what the site's coordination chemistry is. Fourteen ions, from a structure, in seconds. Free, no registration.

API Paper GitHub License PRIME v1
Structure sample.cif
Drop a structure here, or browse
PDB, mmCIF or ENT, gzipped or not.
or
What the limits are
A PDB ID is fetched from RCSB and a UniProt accession uses the AlphaFold model, which is a single apo-like conformation — a site that only forms in the holo structure can be missed, and the result says when the input was a model. One structure or id per submission, capped at 64 MB, 250,000 atoms per structure and at 500 chain–ion units of work, one unit being one chain scanned for one ion — that is 500 chains at a single ion, or 35 chains at all fourteen, in one structure. For work beyond that, run the engine yourself.
Metal ions
Selected (0):
Show the periodic table
Alkali metal Alkaline earth Transition metal Nonmetal
Advanced options
Orientations each probe is scored at. A fixed set, not random draws, so the same structure always returns the same sites. Three reach the published recall; four and five add nothing.
Seed for generating the fixed rotation set. The same seed reproduces the same rotations. Default: 0.
Candidate probes ranked by coordination statistics. Set to 0 to keep all probes.
Predicted ion positions closer than this are merged into one site. 3.0 Å is the published default. For ion channels, where sites sit close together along the selectivity filter, use a smaller value such as 1.0 Å.
PDB IDs only. Prediction runs one chain at a time, so a site completed by a neighbouring chain is seen with half its shell missing. Such sites are 4% of the deposited ions on the 267-chain zinc test set and are recovered at 0.78 against 0.88 for the rest. The cost is duplication: a site on a repeated chain is reported once per copy.

One structure at one ion is typically under a second, all fourteen a few seconds. Each run opens its own bookmarkable result page, which is retained for one week. Submitted structures are private and are not visible to anyone else.

License. Free for everyone, academic or commercial, with no registration, login or e-mail address. The code is MIT. Results from this server are released under CC BY 4.0; a structure you upload stays yours.

Privacy and cookies. This server sets no cookies and runs no analytics or tracking scripts. Your recent-jobs list is held by your own browser in local storage and is never sent to us, so clearing your browser data removes it. A result page is reachable only through its own URL, which is neither listed nor indexed, and uploaded structures and results are deleted one week after the job. The only third party the page contacts is Google Fonts, which serves the web fonts and sets no cookies.

Method. Xu S. and Onoda A., Nat Commun 17, 8146 (2026). The networks are the published ones — same architecture, same weights, same training split, no retraining and no distillation; what is new is the implementation and the service around it. Questions to shijie.xu@ees.hokudai.ac.jp.