processed_data/README.md

processed_data

The small tables the notebooks read, committed so the figures redraw from a fresh clone. Every one is the output of a script in this repo run against inputs far too large to commit — the deposited mmCIFs (~2 GB), the prediction store (tens of GB), a count-ECFP4 fingerprint matrix over half a million PDB molecules. Committing the results is what makes the notebook runnable without fetching any of that.

run_pipeline.sh writes the curation half of this directory. The analysis half needs inputs that are not obtainable by a script — see docs/reproducing.md.

What run_pipeline.sh writes, and what is here now

file	written by	grain
peptide_manifest.json	characterize_peptides.py	one record per peptide chain considered
edge_cases.csv	characterize_peptides.py, via edge_cases.py	one row per flagged case per chain
cyclization_classes.csv	cyclization_classes.py	one row per buildable chain: the collapsed class, and whether the entry list's annotation agrees

The two checks' outputs are not here. They describe one config tree, so they live beside it: <CONFIG_DIR>/config_fidelity_<MODEL>.tsv, check_summary.csv and check_summary.md, together with the edge_case_decisions.yaml that produced them. A check result separated from the configs it judged is not interpretable.

At the time of writing that is 291 buildable peptide chains out of 945 peptide chain records across 347 passing entries. It was 315 until 2026-09-14, when characterize_peptides.py stopped inheriting an entry's is_cyclic flag onto every peptide chain of that entry and started asking each chain whether any of its own bonds closes a ring: 26 chains left and 2 joined. n is not a constant anywhere: it changes with the entry list, the notebook prints it, and a change shows up in edge_cases.csv.

The analysis tables

All four are present and all were regenerated at n = 291 on 2026-09-14. notebooks/pretzel_ground_truth.ipynb runs end to end from this directory and writes ten figures.

directory	written by	regenerating it needs	how long
novelty/	analysis/novelty/novelty_search.py	the 477 MB fingerprint matrices, see fingerprints/README.md	~10 s
peptide_profile/	analysis/dataset_profile.py	the manifest only	~5 s
pocket/	analysis/pockets/*	the mmCIF mirror and a foldseek binary	~15 min
bond_geometry/<MODEL>/	analysis/bond_geometry.py	the mmCIFs, the CCD, and that model's configs and predictions	~20 s per model

pocket/ does not need pxmeter installed, whatever earlier drafts of this file and docs/reproducing.md said. analysis/pockets/pocketlib.py transcribes PXMeter's pocket definition — alpha carbons within 10 A, a fixed radius PXMeter does not make configurable — and cites it in a comment. Nothing imports the package. The dependencies are gemmi, numpy, rdkit and a foldseek binary on the path.

The 15 minutes for pocket/ is almost all two steps: the all-against-all foldseek search in similarity.py (9 min, of which the SuCOS pass over 10,332 scoring pairs is most) and the solvent-accessible-surface pass in pocket_chemistry.py (5 min).

bond_geometry/ holds Protenix only. It is the one model re-run against the audited config set; AF3, Boltz2 and HighFold3 were last run on v8, whose inputs differ for 84 of the chains they cover, and they are left out until they are re-run rather than put on the same forest plot. AF2 is out of scope — 13 chains is too few to read as a distribution. The notebook derives its model list from the directories here, so re-adding one is exactly "drop its table in"; bond_geometry/README.md has the command and the provenance.

fingerprints/

Ignored except for its README, because the RCSB-wide matrix is 477 MB and over GitHub's 100 MB per-file limit. novelty/ holds the results of the search those matrices feed, which is what the notebook reads — so the figures redraw without them and only re-running the search itself needs them.

Why outputs/ is not committed beside this

outputs/ssot/ would hold the same numbers as this directory, regenerated from it in under a minute. Two copies of one set of counts is two places able to disagree about n. This directory is the committed input; outputs/ is the product, and it is ignored.