af3_partners

Turn a human gene symbol into a zip of AlphaFold3 (AF3) local-format input JSONs, one per (input-protein isoform × interacting partner).

Install

git clone https://github.com/brianryu87/af3_partners.git
cd af3_partners
pip install .          # or: pip install -e .   (for development)

Usage

af3partners RPS24 --out .
af3partners RPS24 --rna-tsv my_rna_partners.tsv

Without installing, run it from the repo root with:

python -m af3partners RPS24

Output: RPS24.zip containing:

RPS24/
  AF3_inputs/{ribosomal_protein,nonribosomal_protein,rna}/{high,medium,low}/*.json
  manifest.tsv   # per-pair sources, scores, tier
  README.txt     # run summary and caveats

Partner sources (latest release of each)

• Sequences & isoforms: UniProt (reviewed + computational/TrEMBL).
• Protein partners: STRING (combined + per-channel scores) and IntAct (MI score).
• Curated UniProt binary interactions mark uniprot_curated.
• RNA partners: ENCORI (best-effort; bounded), curated RNA from IntAct (RNAcentral URS interactors, sequences resolved via RNAcentral), and optional --rna-tsv (gene, sequence).

Job size limit

Each JSON is one pairwise job. Pairs whose total length (input isoform + partner, counting 1 residue or 1 nucleotide as 1 token) exceeds AF3_MAX_TOKENS (5000, the AlphaFold Server limit) are skipped with a warning.

Confidence tiers

• high: experimental evidence (STRING experiments channel, IntAct MI ≥ 0.45, UniProt-curated, or ENCORI CLIP) or STRING combined ≥ 0.7
• medium: STRING combined in [0.4, 0.7)
• low: below 0.4 / text-mining-only

Tests

python -m unittest discover -s tests -t .
AF3_LIVE=1 python -m unittest tests.test_live_smoke   # opt-in, hits live APIs

Pure Python standard library; no third-party dependencies.