bindome-query

Query the Human Bindome for a list of protein targets. It pulls the metrics for every designed binder, writes a per-target summary, and draws a figure of how those designs score.

The Human Bindome (Wenckstern et al., bioRxiv 2026.07.30.741542) is a proteome-scale atlas of about 306,000 in-silico BindCraft binder candidates covering roughly 8,300 human proteins (40.9% of the proteome). Each target comes with one record per designed binder, and each record carries AlphaFold-style confidence metrics plus links to the predicted complex.

Heads up: these are computational design candidates, not experimentally validated binders. There is no measured affinity in the Bindome. The metrics (i_pTM, i_pAE, pLDDT, and so on) are design-time confidence only, so use them to triage, not to conclude that a target "has a binder".

Install

pip install -r requirements.txt

Usage

python bindome_query.py --targets targets.example.json --outdir out

Add --top-structures N to also download the top N binder CIFs per target (ranked by i_pTM) into out/structures/.

Targets file

A JSON list of {name, uniprot}. A bare list of accession strings works too.

{
  "targets": [
    { "name": "PD-L1",  "uniprot": "Q9NZQ7" },
    { "name": "VEGF-A", "uniprot": "P15692" },
    { "name": "IL-2",   "uniprot": "P60568" }
  ]
}

Name resolution

If you leave out uniprot, the name is looked up against UniProt (reviewed, human). The lookup is tolerant of case, spacing, hyphens, and gene vs protein names, so CLDN6, Claudin-6, claudin 6, PD-L1, pdl1, and HER2 all resolve correctly. It matches on the exact normalized name, so it never silently picks the wrong protein. (Naive fuzzy search maps claudin 6 to Claudin-9 and her-2 to HERC3; this tool does not.)

For a real typo it will not guess. It prints the closest "did you mean" candidates and skips that target:

[resolve] 'claudn 6': no exact match; did you mean: CLDN6 (P56747), CLDN16 (Q9Y5I7), ...
          -> skipping (add an explicit "uniprot" to use one)

UniProt has no fuzzy search, and for design work picking the wrong target is worse than a clean miss. If a target matters, put its accession in the file. That is the key the Bindome API uses anyway.

Outputs

File	Contents
out/bindome_designs.csv	one row per designed binder: target, the UniProt region hit (uniprot_start/end), parsed design fields (domain, binder_length, seed, mpnn_variant, af_model), all seven metrics, and model_url/pae_url
out/bindome_summary.csv	one row per target: n_binders, n_regions, and the best and median of every metric
out/bindome_metrics.png	two metric-vs-metric gate scatters (i_pTM vs i_pAE, pLDDT vs ipSAE) colored by target, with the passing corner shaded and dashed cutoff lines, plus a target key and a per-target pass-rate table

Metrics

All values are normalized to 0 to 1 as returned by the API. Higher is better for i_pTM, pTM, pLDDT, i_pLDDT, and ipSAE; lower is better for i_pAE and pAE. The uniprot_start and uniprot_end columns tell you which region of the target each binder was designed against, which is handy for checking whether a candidate overlaps an epitope you care about.

Cutoffs used in the figure (normalized units): i_pTM ≥ 0.85, i_pAE ≤ 0.19, pLDDT ≥ 0.90, ipSAE ≥ 0.90. About i_pAE: it is BindCraft's normalized interface PAE, defined in ColabDesign as the interface PAE divided by 31 (see get_pae_loss in colabdesign/af/loss.py). So i_pAE ≤ 0.19 is about 6 Å, and 0.6 would be roughly 19 Å. Note that the Bindome already applies BindCraft's inclusion filter of iPAE < 0.35 (about 11 Å, median 0.23 in the paper), so every design in the atlas has passed that; 0.19 is a stricter triage bar. Edit THRESHOLDS in bindome_query.py to change any of them. These are in-silico BindCraft design metrics, not measured affinity.

API notes

The tool uses one unauthenticated endpoint:

GET https://bindome.epfl.ch/uniprot/{UNIPROT_AC}/metrics

There is no batch or list endpoint, so the tool just loops over your targets. If you want everything in bulk (all sequences, structures, and ML splits), the authors publish the full dataset on Hugging Face: wjulius/HumanBindome.

Tests

pip install -r requirements-dev.txt
pytest -q

The tests run fully offline, with no API calls. They cover design-name parsing, the name-resolver safety logic, the metric tables (including the no-double-count and lower-is-better handling), and a figure smoke test. CI runs them on Python 3.10 to 3.12 (.github/workflows/ci.yml).

License and attribution

This tool (the code in this repo) is MIT licensed.

The data it fetches is not covered by that license. The Human Bindome (the sequences, structures, and metrics served by the API) is released under CC-BY-4.0 by the authors at EPFL. You are free to use it, including commercially, but you must give credit. Note that CC-BY covers copyright and database rights only; it says nothing about patents, and the sequences are publicly disclosed, so treat them as an open starting library rather than as patentable matter.

If you use the data (or this tool) in your work, please cite the paper:

Wenckstern, J., Diaz-Rovira, A. M., Kuhn, J., et al. The Human Bindome: A Proteome-scale Atlas of Designed Binder Candidates. bioRxiv 2026.07.30.741542 (2026). https://www.biorxiv.org/content/10.64898/2026.07.30.741542v1

Dataset: wjulius/HumanBindome (CC-BY-4.0). The Bindome is built on BindCraft, which is worth citing too if the designs are central to your work.