PRIME 2

Metal-binding sites from a structure

API PRIME Source

Where the ion sits, in 3D

Give PRIME 2 a structure and it returns the coordinates of every metal site it finds, for 14 ions, in about a quarter of a second per protein per ion. Submit one structure, or five hundred in an archive.

PRIME 2 is a faster implementation, not a new model. It runs the published PRIME networks unchanged — same architecture, same weights, same training split, no retraining and no distillation — through a rewritten inference path. On the held-out benchmark the predictions come out slightly better rather than cheaper: recall 0.878 against the released 0.877, precision 0.924 against 0.832.

N N N O Zn 2.0 Å
A zinc site as the network sees it: four donor atoms inside 2.6 Å. PRIME 2 scores 30 candidate positions per chain at 3 fixed orientations and keeps the clusters that survive.
or

PDB entries come from RCSB, UniProt accessions from the AlphaFold database.

Sample structure if you want to check the format first.

Ions to scan
Transition metals
Alkali and alkaline earth

Circle size is the ion's radius. Each ion is a separate network and a separate scoring pass, so the run time grows with the number selected.

Keeping 30 is the cheapest setting that stays as accurate as the published method.

Three reach the published recall. A fourth and fifth add nothing.

No account, no installation. A job is private to whoever holds its link, and results are kept for one week.

What it costs

Both rows were measured on the machine serving this page, over 64 chains drawn at random from the BioLiP zinc test split, and scored against 441 deposited Zn²⁺ ions in 267 chains.

Configuration Proteins/s per ion Recall @1.5 Å Precision Repeats itself
PRIME as released, every probe at 5 random rotations 0.0330.8770.832 0.94 of its own sites
PRIME 2, top 30 probes at 3 fixed rotations 4.350.8780.924 exactly

The 130-fold difference splits in two, and neither half is paid for in accuracy. Run at the released configuration, this engine is 0.37 proteins per second, so 11 times comes from the implementation and 12 from the operating point. The released code draws its rotations from a generator seeded once per process and reproduces 0.94 of its own sites between runs; PRIME 2 uses a fixed set, so resubmitting a structure returns the same answer.

The figures above are zinc. Of the other 13 ions, only calcium and magnesium have test splits large enough to quote a recall for, and the rest run the same code on their own published checkpoints.

Reading the result

One structure opens a viewer with the ions placed in it. A batch opens a table and three files. Both carry the same fields per site.

element
The ion this site was predicted for. Fe³⁺ and Fe²⁺, and Cu²⁺ and Cu⁺, are separate networks, so both may report a site in the same pocket.
x, y, z
The ion position in the input structure's own frame. predicted.pdb is these coordinates as HETATM records, ready for a viewer.
confidence
The network's probability for the best probe in the cluster, from 0 to 1. Sites above 0.5 land within 1.5 Å of a deposited ion 92% of the time on the benchmark.
chain
The chain whose probes produced the site.
n_probes
How many candidate positions clustered into it. Several probes agreeing is a pocket, not a single lucky guess.

A prediction is a position, not an occupancy: it does not say the ion is present in your sample. Apo structures are the known blind spot, since an unformed site has nothing for a structure-based method to see, and AlphaFold models of metalloproteins are often apo-like. If an expected site is missing, try the holo conformation before concluding anything about the protein.

From a script

The page is a thin layer over a REST API, and a job submitted from a script is the same job. The API page documents every endpoint and ships a Python client that runs as it stands.

curl -s -F 'protein_id=12CA' -F 'metal_type=ZN' \
     https://onodalab.ees.hokudai.ac.jp/prime2/api/jobs

Licence and citation

Free for everyone, academic or commercial, with no registration. The code is MIT licensed and this server's output may be reused under CC BY-SA 4.0. If it helps a publication, cite the method paper:

Xu, S.; Onoda, A. Probe-based identification of metal-binding sites using deep learning representations. Nature Communications 17, 8146 (2026).

Questions and problems: shijie.xu@ees.hokudai.ac.jp.