Datasets:

HFMI-PROFOUND
/

BILBY-PRO

---
pretty_name: "BILBY-PRO: PROSITE-derived functional motif scaffolding benchmark"
license: cc-by-4.0
task_categories:
  - other
tags:
  - biology
  - protein-design
  - protein-structure
  - motif-scaffolding
  - benchmark
  - prosite
size_categories:
  - n<1K
---
# BILBY-PRO

BILBY-PRO is a benchmark of structurally conserved functional motifs derived
from PROSITE-linked experimental protein structures. It provides fixed
motif-scaffolding objectives for comparing protein design methods. It is the
benchmark construction and **static** dataset. The related
[BILBY-MD](https://github.com/PROFOUND-HFMI/BILBY-MD) molecular dynamics and
dynamics-aware evaluation pipeline is maintained separately.

It contains **220 cases** spanning **174 PROSITE patterns**. Construction
software was developed by Emile de Bruyn within PROFOUND, a pilot project of
the Helmholtz Foundation Model Initiative (HFMI) of the Helmholtz Association.
Associated publication: *Static Recovery Is Not Dynamic Stability:
Dynamics-Aware Benchmarking of Protein Motif Scaffolding*, by Emile de Bruyn,
Moritz Schäffler, Jannik Schneider, Sebastian R. Schmidt, Chanwoong Hwang,
Aleena Siji, Utkarsh Upadhyay, Stefan Bauer, Alexander Schug, Stefan Kesselheim,
Artur Yakimovich and Alina Bazarova. See [CITATION.cff](CITATION.cff).
The first three authors contributed equally.

## Quick start

Once published, download with
`hf download HFMI-PROFOUND/BILBY-PRO --repo-type dataset --local-dir BILBY-PRO`
or clone the dataset repository with Git LFS. From this checkout:

```bash
python3 scripts/validate_release.py
```

The validator checks case and pattern counts, table relationships, residue
coordinates, contig feasibility and checksums, including regression against
the historical submission. See [croissant.json](croissant.json) for
machine-readable table fields, links between tables and structure collections,
and for the Responsible-AI and provenance metadata (Croissant 1.1 and
`http://mlcommons.org/croissant/RAI/1.0`, `rai:*` and `prov:*` properties).
It is regenerated by `python3 scripts/build_croissant.py` and passes
`mlcroissant validate --jsonld croissant.json` without errors. For example,
after `pip install mlcroissant`, read a local case with:

```python
import mlcroissant as mlc
dataset = mlc.Dataset("croissant.json", mapping={"test_cases.csv": "test_cases.csv"})
print(next(iter(dataset.records(record_set="test_cases"))))
```

## Release integrity and citation

`provenance/release-sha256.txt` pins all 445 released data files with one
SHA-256 per line. The published release is
[commit `b0312e0433191ec425c67e2f9a620c2f614551c2`](https://huggingface.co/datasets/HFMI-PROFOUND/BILBY-PRO/tree/b0312e0433191ec425c67e2f9a620c2f614551c2),
uploaded on 2026-11-02; `main` may receive later revisions, so cite the commit
or the dataset DOI rather than the branch. The dataset DOI
`10.57967/hf/11435` (DataCite via Hugging Face, state findable, registered
2026-11-04) is a `Dataset` record whose `contentUrl` is exactly that commit
tree, so the DOI resolves to this same snapshot rather than to a moving branch:
<https://api.datacite.org/dois/10.57967/hf/11435>. The v1.0.0 data files and
their digests are unchanged by that registration. This release supersedes the
earlier anonymous dataset, which remains the historical artifact behind the
paper.

`provenance/huggingface-manifest.json` records every hosted
`resolve/main/<path>` URL with its exact content type, size and LFS marker: 450
entries, namely the two structure directories, the five CSV tables, the hosted
`README.md`, `LICENSE` and `.gitattributes`. The 440 PDB files are LFS objects;
the CSV and JSON files are plain Git objects, so Git LFS filter rules cannot be
used as an integrity check for them. `provenance/published-sha256.txt` retains
the historical digests.

This metadata revision changes no hosted data file. `croissant.json`,
`CITATION.cff`, `CHANGELOG.md`, `construction/`, `provenance/` and `scripts/`
are part of this repository release and are not yet in the hosted tree recorded
by the manifest; refresh that manifest after they are uploaded.

## Dataset contents

| File | Contents |
|---|---|
| `test_cases.csv` | Case IDs, normalized motif residues, redesignable residues, scaffold length and group |
| `motif_cases_metadata_croissant.csv` | PROSITE accessions and patterns, structural annotations, original and normalized residue labels, and PDB paths |
| `contig_specifications.csv` | Normalized and original contigs and target lengths |
| `contigs.csv` | Original-numbering contigs |
| `cluster_membership.csv` | Cluster membership, including non-representative members |
| `motif_pdbs/` | One normalized motif PDB per case; one chain per motif segment |
| `original_pdbs/` | Corresponding source structures |

Cluster `case_id` values use the same `_cN` identifiers as the case metadata.
Member rows can name a different PROSITE entry within the representative's
cluster; join to cases using `benchmark_pdb_id`.

Group 4 denotes multi-state motifs, not a particular segment count; groups
1–3 follow the published segment-based grouping. Redesignable positions in the
published motif PDBs retain their original residue identities: use the
MotifBench adapter below to encode them as `UNK` for evaluation.

This corrected release adjusts one infeasible contig and standardizes cluster
case IDs. Historical paper results were not recomputed; see
[CHANGELOG.md](CHANGELOG.md) for the exact changes.

## Construction and provenance

The construction code parses PROSITE sequence patterns and associated PDB
identifiers, extracts structural motif instances, analyzes structural
conservation using RMSD and Foldseek clustering, selects representatives, and
exports motif structures, case tables and normalized contigs. See
[construction/README.md](construction/README.md) for the audited snapshot,
inputs, entry point and limitations.

The benchmark used **PROSITE 2021_04, dated 29 September 2021**. Its `3D`
field supplied the PDB cross-references used to select structures; that field
is absent from the current PROSITE data file. The wrapper verifies input
hashes. The historical code revision, PDB snapshot and software environment
are not fully established, so a fresh reconstruction has not been validated.
PROSITE is maintained by the [SIB Swiss Institute of Bioinformatics](https://prosite.expasy.org/).
No PROSITE database archive, complete database entry or PROSITE documentation
file is distributed here. **The published metadata does contain verbatim
PROSITE sequence-pattern strings**, alongside accessions and links; these
fields have not been removed. The PROSITE 2021_04 input identifies its license
as CC BY-NC-ND 4.0; see [PROSITE's licensing conditions](https://prosite.expasy.org/prosite_license.html).
That license and PDB terms apply to upstream material; the dataset's CC BY
notice does not grant rights to the pattern strings or PDB structures. See
[provenance/README.md](provenance/README.md).
## MotifBench quick start
The following creates a separate BILBY-PRO working copy of the tested
[MotifBench](https://github.com/blt2114/MotifBench) revision. It does not
overwrite the original checkout or the published PDBs. Use absolute paths
without spaces: upstream shell scripts do not consistently quote paths.
```bash
export BILBY="$PWD"
git clone https://github.com/blt2114/MotifBench.git ../MotifBench
git -C ../MotifBench checkout d5a714c6a376c701e29dfa66d42a1b13c5e339e0
python3 scripts/prepare_motifbench.py ../MotifBench ../MotifBench-BILBY
export WORK="$(cd ../MotifBench-BILBY && pwd)"
# Upstream evaluation environment; this can take substantial time and space.
conda env create -f "$WORK/motif_bench.yml"
conda activate motif_bench
pip install -e "$WORK/Scaffold-Lab"
```
The helper prefixes IDs with their **one-based CSV row number** (e.g.
`01_1CK7_PS00023_c1`), writes a separate `test_cases.csv`, adapts PDBs by
marking redesignable residues `UNK` and retaining only backbone atoms there,
and writes compatible normalized contigs. All other motif coordinates are
unchanged. The source checkout must be clean and at the pinned commit.
Generate scaffolds with a separately installed MotifBench-compatible method.
For example, after installing [RFdiffusion](https://github.com/RosettaCommons/RFdiffusion)
and its model weights, a possible invocation for the first case is:
```bash
export CASE=01_1CK7_PS00023_c1
export SCAFFOLDS="$PWD/scaffolds"
mkdir -p "$SCAFFOLDS/$CASE"
python /path/to/RFdiffusion/scripts/run_inference.py \
  "inference.input_pdb=$BILBY/motif_pdbs/1CK7_PS00023_c1.pdb" \
  "inference.output_prefix=$SCAFFOLDS/$CASE/$CASE" \
  'contigmap.contigs=[0-185/A1-42/0-185]' \
  'contigmap.length=225-225' inference.num_designs=100
```
This generation example is **not tested here**; consult the generator's
documentation and honor the redesignable-residue constraints during sequence
design. Contigs are ranges, not literal placements. Each case directory must
contain `{case}_{sample_num}.pdb` files with 1-indexed residues and a
`scaffold_info.csv` with `sample_num,motif_placements`, e.g. `0,50/A/133`
for a 225-residue scaffold containing the 42-residue segment A. Use each
design's **actual** placement, not the example. For multiple segments use
their actual chain order, e.g. `12/B/15/A/29`. Upstream's
`example/write_scaffold_info.py` is usable only if the generator's B-factors
actually distinguish motif from scaffold; inspect its output.

To evaluate, prepare the Foldseek PDB database as described in the upstream
README, then create a configuration (these paths must exist):

```bash
export RESULTS="$PWD/evaluation"
export FOLDSEEK_DB=/path/to/foldseek/database/pdb
mkdir -p "$RESULTS"
cat > "$WORK/config-bilby.txt" <<EOF
scaffold_base_dir=$SCAFFOLDS
benchmark_dir=$WORK
foldseek_db_path=$FOLDSEEK_DB
base_output_dir=$RESULTS
python_path=$(command -v python)
EOF
cd "$WORK"
./scripts/evaluate_bbs.sh "$CASE" config-bilby.txt
# After evaluation of all intended cases:
./scripts/summarize_results.sh config-bilby.txt
python "$BILBY/scripts/summarize_motifbench.py" "$WORK" "$RESULTS"
```

The last command preserves the upstream per-problem summary, orders cases
before applying upstream bootstrap scoring, and emits ordered per-case scores
and an overall summary with a fixed random seed. It requires 100 evaluated
designs per included case; a subset is explicitly labeled partial. Upstream's
overall group metrics average group means rather than weighting by case count.
The interface and summary scripts were smoke-tested locally; scaffold
generation, refolding, Foldseek database search and full GPU evaluation were
**not** run. See [provenance/README.md](provenance/README.md) for validation.

## Attribution and citation

Please cite the dataset release itself as described in
[CITATION.cff](CITATION.cff): *BILBY-PRO: PROSITE-derived functional motif
scaffolding benchmark*, DOI [10.57967/hf/11435](https://doi.org/10.57967/hf/11435),
v1.0.0, commit `b0312e0433191ec425c67e2f9a620c2f614551c2`, and the associated
publication *Static Recovery Is Not Dynamic Stability: Dynamics-Aware
Benchmarking of Protein Motif Scaffolding*. The machine-readable preferred
citation is the `citeAs` field of [croissant.json](croissant.json).

Benchmark case definitions and added annotations are [CC BY 4.0](LICENSE).
Construction code and release scripts are
[Apache-2.0](LICENSE-APACHE-2.0), copyright 2026 Emile de Bruyn. PDB structures
and PROSITE material retain their upstream terms. See [CITATION.cff](CITATION.cff)
for the publication's full author list.

Please also cite:

* Sigrist et al., “New and continuing developments at PROSITE,” *Nucleic Acids
  Research* (2013), [doi:10.1093/nar/gks1067](https://doi.org/10.1093/nar/gks1067).
* MotifBench, [arXiv:2502.12479](https://arxiv.org/abs/2502.12479).
* The source PDB structures and their corresponding experimental publications.

	---
	pretty_name: "BILBY-PRO: PROSITE-derived functional motif scaffolding benchmark"
	license: cc-by-4.0
	task_categories:
	- other
	tags:
	- biology
	- protein-design
	- protein-structure
	- motif-scaffolding
	- benchmark
	- prosite
	size_categories:
	- n<1K
	---

	# BILBY-PRO

	BILBY-PRO is a benchmark of structurally conserved functional motifs derived
	from PROSITE-linked experimental protein structures. It provides fixed
	motif-scaffolding objectives for comparing protein design methods. It is the
	benchmark construction and static dataset. The related
	[BILBY-MD](https://github.com/PROFOUND-HFMI/BILBY-MD) molecular dynamics and
	dynamics-aware evaluation pipeline is maintained separately.

	It contains 220 cases spanning 174 PROSITE patterns. Construction
	software was developed by Emile de Bruyn within PROFOUND, a pilot project of
	the Helmholtz Foundation Model Initiative (HFMI) of the Helmholtz Association.
	Associated publication: *Static Recovery Is Not Dynamic Stability:
	Dynamics-Aware Benchmarking of Protein Motif Scaffolding*, by Emile de Bruyn,
	Moritz Schäffler, Jannik Schneider, Sebastian R. Schmidt, Chanwoong Hwang,
	Aleena Siji, Utkarsh Upadhyay, Stefan Bauer, Alexander Schug, Stefan Kesselheim,
	Artur Yakimovich and Alina Bazarova. See [CITATION.cff](CITATION.cff).
	The first three authors contributed equally.

	## Quick start

	Once published, download with
	`hf download HFMI-PROFOUND/BILBY-PRO --repo-type dataset --local-dir BILBY-PRO`
	or clone the dataset repository with Git LFS. From this checkout:

	```bash
	python3 scripts/validate_release.py
	```

	The validator checks case and pattern counts, table relationships, residue
	coordinates, contig feasibility and checksums, including regression against
	the historical submission. See [croissant.json](croissant.json) for
	machine-readable table fields, links between tables and structure collections,
	and for the Responsible-AI and provenance metadata (Croissant 1.1 and
	`http://mlcommons.org/croissant/RAI/1.0`, `rai:` and `prov:` properties).
	It is regenerated by `python3 scripts/build_croissant.py` and passes
	`mlcroissant validate --jsonld croissant.json` without errors. For example,
	after `pip install mlcroissant`, read a local case with:

	```python
	import mlcroissant as mlc

	dataset = mlc.Dataset("croissant.json", mapping={"test_cases.csv": "test_cases.csv"})
	print(next(iter(dataset.records(record_set="test_cases"))))
	```

	## Release integrity and citation

	`provenance/release-sha256.txt` pins all 445 released data files with one
	SHA-256 per line. The published release is
	[commit `b0312e0433191ec425c67e2f9a620c2f614551c2`](https://huggingface.co/datasets/HFMI-PROFOUND/BILBY-PRO/tree/b0312e0433191ec425c67e2f9a620c2f614551c2),
	uploaded on 2026-11-02; `main` may receive later revisions, so cite the commit
	or the dataset DOI rather than the branch. The dataset DOI
	`10.57967/hf/11435` (DataCite via Hugging Face, state findable, registered
	2026-11-04) is a `Dataset` record whose `contentUrl` is exactly that commit
	tree, so the DOI resolves to this same snapshot rather than to a moving branch:
	<https://api.datacite.org/dois/10.57967/hf/11435>. The v1.0.0 data files and
	their digests are unchanged by that registration. This release supersedes the
	earlier anonymous dataset, which remains the historical artifact behind the
	paper.

	`provenance/huggingface-manifest.json` records every hosted
	`resolve/main/<path>` URL with its exact content type, size and LFS marker: 450
	entries, namely the two structure directories, the five CSV tables, the hosted
	`README.md`, `LICENSE` and `.gitattributes`. The 440 PDB files are LFS objects;
	the CSV and JSON files are plain Git objects, so Git LFS filter rules cannot be
	used as an integrity check for them. `provenance/published-sha256.txt` retains
	the historical digests.

	This metadata revision changes no hosted data file. `croissant.json`,
	`CITATION.cff`, `CHANGELOG.md`, `construction/`, `provenance/` and `scripts/`
	are part of this repository release and are not yet in the hosted tree recorded
	by the manifest; refresh that manifest after they are uploaded.

	## Dataset contents

	\| File \| Contents \|
	\|---\|---\|
	\| `test_cases.csv` \| Case IDs, normalized motif residues, redesignable residues, scaffold length and group \|
	\| `motif_cases_metadata_croissant.csv` \| PROSITE accessions and patterns, structural annotations, original and normalized residue labels, and PDB paths \|
	\| `contig_specifications.csv` \| Normalized and original contigs and target lengths \|
	\| `contigs.csv` \| Original-numbering contigs \|
	\| `cluster_membership.csv` \| Cluster membership, including non-representative members \|
	\| `motif_pdbs/` \| One normalized motif PDB per case; one chain per motif segment \|
	\| `original_pdbs/` \| Corresponding source structures \|

	Cluster `case_id` values use the same `_cN` identifiers as the case metadata.
	Member rows can name a different PROSITE entry within the representative's
	cluster; join to cases using `benchmark_pdb_id`.

	Group 4 denotes multi-state motifs, not a particular segment count; groups
	1–3 follow the published segment-based grouping. Redesignable positions in the
	published motif PDBs retain their original residue identities: use the
	MotifBench adapter below to encode them as `UNK` for evaluation.

	This corrected release adjusts one infeasible contig and standardizes cluster
	case IDs. Historical paper results were not recomputed; see
	[CHANGELOG.md](CHANGELOG.md) for the exact changes.

	## Construction and provenance

	The construction code parses PROSITE sequence patterns and associated PDB
	identifiers, extracts structural motif instances, analyzes structural
	conservation using RMSD and Foldseek clustering, selects representatives, and
	exports motif structures, case tables and normalized contigs. See
	[construction/README.md](construction/README.md) for the audited snapshot,
	inputs, entry point and limitations.

	The benchmark used PROSITE 2021_04, dated 29 September 2021. Its `3D`
	field supplied the PDB cross-references used to select structures; that field
	is absent from the current PROSITE data file. The wrapper verifies input
	hashes. The historical code revision, PDB snapshot and software environment
	are not fully established, so a fresh reconstruction has not been validated.

	PROSITE is maintained by the [SIB Swiss Institute of Bioinformatics](https://prosite.expasy.org/).
	No PROSITE database archive, complete database entry or PROSITE documentation
	file is distributed here. **The published metadata does contain verbatim
	PROSITE sequence-pattern strings**, alongside accessions and links; these
	fields have not been removed. The PROSITE 2021_04 input identifies its license
	as CC BY-NC-ND 4.0; see [PROSITE's licensing conditions](https://prosite.expasy.org/prosite_license.html).
	That license and PDB terms apply to upstream material; the dataset's CC BY
	notice does not grant rights to the pattern strings or PDB structures. See
	[provenance/README.md](provenance/README.md).

	## MotifBench quick start

	The following creates a separate BILBY-PRO working copy of the tested
	[MotifBench](https://github.com/blt2114/MotifBench) revision. It does not
	overwrite the original checkout or the published PDBs. Use absolute paths
	without spaces: upstream shell scripts do not consistently quote paths.

	```bash
	export BILBY="$PWD"
	git clone https://github.com/blt2114/MotifBench.git ../MotifBench
	git -C ../MotifBench checkout d5a714c6a376c701e29dfa66d42a1b13c5e339e0
	python3 scripts/prepare_motifbench.py ../MotifBench ../MotifBench-BILBY
	export WORK="$(cd ../MotifBench-BILBY && pwd)"

	# Upstream evaluation environment; this can take substantial time and space.
	conda env create -f "$WORK/motif_bench.yml"
	conda activate motif_bench
	pip install -e "$WORK/Scaffold-Lab"
	```

	The helper prefixes IDs with their one-based CSV row number (e.g.
	`01_1CK7_PS00023_c1`), writes a separate `test_cases.csv`, adapts PDBs by
	marking redesignable residues `UNK` and retaining only backbone atoms there,
	and writes compatible normalized contigs. All other motif coordinates are
	unchanged. The source checkout must be clean and at the pinned commit.

	Generate scaffolds with a separately installed MotifBench-compatible method.
	For example, after installing [RFdiffusion](https://github.com/RosettaCommons/RFdiffusion)
	and its model weights, a possible invocation for the first case is:

	```bash
	export CASE=01_1CK7_PS00023_c1
	export SCAFFOLDS="$PWD/scaffolds"
	mkdir -p "$SCAFFOLDS/$CASE"
	python /path/to/RFdiffusion/scripts/run_inference.py \
	"inference.input_pdb=$BILBY/motif_pdbs/1CK7_PS00023_c1.pdb" \
	"inference.output_prefix=$SCAFFOLDS/$CASE/$CASE" \
	'contigmap.contigs=[0-185/A1-42/0-185]' \
	'contigmap.length=225-225' inference.num_designs=100
	```

	This generation example is not tested here; consult the generator's
	documentation and honor the redesignable-residue constraints during sequence
	design. Contigs are ranges, not literal placements. Each case directory must
	contain `{case}_{sample_num}.pdb` files with 1-indexed residues and a
	`scaffold_info.csv` with `sample_num,motif_placements`, e.g. `0,50/A/133`
	for a 225-residue scaffold containing the 42-residue segment A. Use each
	design's actual placement, not the example. For multiple segments use
	their actual chain order, e.g. `12/B/15/A/29`. Upstream's
	`example/write_scaffold_info.py` is usable only if the generator's B-factors
	actually distinguish motif from scaffold; inspect its output.

	To evaluate, prepare the Foldseek PDB database as described in the upstream
	README, then create a configuration (these paths must exist):

	```bash
	export RESULTS="$PWD/evaluation"
	export FOLDSEEK_DB=/path/to/foldseek/database/pdb
	mkdir -p "$RESULTS"
	cat > "$WORK/config-bilby.txt" <<EOF
	scaffold_base_dir=$SCAFFOLDS
	benchmark_dir=$WORK
	foldseek_db_path=$FOLDSEEK_DB
	base_output_dir=$RESULTS
	python_path=$(command -v python)
	EOF
	cd "$WORK"
	./scripts/evaluate_bbs.sh "$CASE" config-bilby.txt
	# After evaluation of all intended cases:
	./scripts/summarize_results.sh config-bilby.txt
	python "$BILBY/scripts/summarize_motifbench.py" "$WORK" "$RESULTS"
	```

	The last command preserves the upstream per-problem summary, orders cases
	before applying upstream bootstrap scoring, and emits ordered per-case scores
	and an overall summary with a fixed random seed. It requires 100 evaluated
	designs per included case; a subset is explicitly labeled partial. Upstream's
	overall group metrics average group means rather than weighting by case count.
	The interface and summary scripts were smoke-tested locally; scaffold
	generation, refolding, Foldseek database search and full GPU evaluation were
	not run. See [provenance/README.md](provenance/README.md) for validation.

	## Attribution and citation

	Please cite the dataset release itself as described in
	[CITATION.cff](CITATION.cff): *BILBY-PRO: PROSITE-derived functional motif
	scaffolding benchmark*, DOI [10.57967/hf/11435](https://doi.org/10.57967/hf/11435),
	v1.0.0, commit `b0312e0433191ec425c67e2f9a620c2f614551c2`, and the associated
	publication *Static Recovery Is Not Dynamic Stability: Dynamics-Aware
	Benchmarking of Protein Motif Scaffolding*. The machine-readable preferred
	citation is the `citeAs` field of [croissant.json](croissant.json).

	Benchmark case definitions and added annotations are [CC BY 4.0](LICENSE).
	Construction code and release scripts are
	[Apache-2.0](LICENSE-APACHE-2.0), copyright 2026 Emile de Bruyn. PDB structures
	and PROSITE material retain their upstream terms. See [CITATION.cff](CITATION.cff)
	for the publication's full author list.

	Please also cite:

	* Sigrist et al., “New and continuing developments at PROSITE,” *Nucleic Acids
	Research* (2013), [doi:10.1093/nar/gks1067](https://doi.org/10.1093/nar/gks1067).
	* MotifBench, [arXiv:2502.12479](https://arxiv.org/abs/2502.12479).
	* The source PDB structures and their corresponding experimental publications.