BILBY-PRO / construction /README.md
EdeBruyn's picture
Initial upload for camera-ready version of NeurIPS paper.
23a728e verified
|
Raw History Blame Contribute Delete
5.21 kB
# Benchmark construction
This directory contains the `motifbench_prosite` pipeline developed to construct
the original BILBY-PRO benchmark, together with subsequent release changes.
The construction code was developed by Emile de Bruyn within PROFOUND. It and
the release scripts are licensed under [Apache-2.0](../LICENSE-APACHE-2.0).
Molecular dynamics and dynamics-aware evaluation are handled separately by
[BILBY-MD](https://github.com/PROFOUND-HFMI/BILBY-MD)
The main stages are `prosite_patterns.py` (PROSITE parsing and span filters),
`pdb_io.py` and `motif_extraction.py` (PDB retrieval and motif instances),
`structure_analysis.py` (RMSD and Foldseek conservation),
`export_benchmarks.py` and `unified_export.py` (selection and export), and
`normalisation.py` (renumbering and segment remapping). The CLI also
extracts PROSITE documentation into its output; those documentation files
must **not** be published as part of this dataset.
## Reproducibility and historical inputs
The benchmark used PROSITE 2021_04 (29 September 2021). Its `3D` records
contain PDB cross-references, parsed as `pdb_structs` and used to find motif
structures; those records are absent in the current 2026_03 data file. The
older release was chosen deliberately for this field.
The local `full_search.yaml` references PROSITE 2021_04 and a historical
MotifBench reference table. Its candidate export matches the four corresponding
historical submission CSV tables byte-for-byte and the shared metadata fields
semantically. The public release includes the corrections described in the
[changelog](../CHANGELOG.md).
Nevertheless the run log, exact source state, historical reference table,
Foldseek version, PDB revisions and locked environment are not established.
The code reads PDBs from live servers and cached paths; a fresh run
can therefore differ. No representative subset or full reconstruction has been
shown to reproduce the published release. The published data remains the
authoritative benchmark.
The portable `configs/reference.yaml` pins 2021_04 and records the
full-search configuration values, with portable paths, no reused cache and no
skipped failed downloads. This does not establish that those values were used
in the historical run. In particular,
the original configuration has `multi_state_ratio: 0.75` and
`multi_state_min_n: 3`, but `PipelineConfig` does **not** load
those two fields: the export code instead defaults to 0.9 and 10. Do not
silently "correct" this scientific ambiguity when reproducing the release.
The pipeline's 10 MB structure limit, external reference table,
download failures, live PDB state and clustering can also affect selection.
Small release safety changes make failed/skipped downloads, missing Foldseek
statistics, empty Foldseek clusters and an empty conservation selection fail
explicitly rather than producing a misleading result. The pipeline still
contains historical skip/error behavior in other stages; inspect logs and
compare all outputs. These guards do not modify the published artifacts.
The export now reduces overly large internal contig minima when necessary to
fit the requested scaffold length, preserving the motif and preferring gaps
of at least 10 residues. It also emits canonical `_cN` case identifiers in
cluster membership rows. These changes apply to future exports; no full
reconstruction has been verified.
## Inputs and run
From the repository root, with Python 3.10+ and a working Foldseek executable
on `PATH`:
```bash
python3 -m venv .venv
. .venv/bin/activate
pip install -e 'construction[dev]'
mkdir -p inputs
curl -fL https://ftp.expasy.org/databases/prosite/old_releases/prosite2021_04.tar.bz2 \
-o inputs/prosite2021_04.tar.bz2
echo '32db1518ae124f42f659b4a1151e4510b3aae6cb6a87f379508783342a2c99d7 inputs/prosite2021_04.tar.bz2' | sha256sum -c -
tar -xjf inputs/prosite2021_04.tar.bz2 -C inputs prosite.dat prosite.doc
```
Place the **historical** MotifBench reference `test_cases.csv` at
`inputs/motifbench_test_cases.csv`. Its precise contents were not recovered;
substituting another version is exploratory and may change selections. After
obtaining the missing provenance, use:
```bash
python scripts/reconstruct.py --config construction/configs/reference.yaml
```
The wrapper verifies the exact PROSITE files by SHA-256, requires the reference
table and Foldseek, and refuses an existing output directory. Paths in the
configuration are relative to the repository root. This is a potentially
large CPU/network job; it downloads many structures and may stop on
unavailable or oversized structures. Do not use it as an unattended release
rebuild. Output and input resources are ignored by Git.
Dependencies in `pyproject.toml` reflect imports rather than historical
versions. Once a verified environment is found, record platform, Python,
Foldseek and `pip freeze` output; use that record or a platform-specific
lockfile for later installations. A lock generated today is not evidence of
the historical environment. Run unit tests with `python -m pytest construction/tests`;
some historical tests write `motif_conservation_stats.csv` in the current
directory, so run them from a disposable directory if needed.