Datasets:

CollagenHelixLabs
/

cdsm_benchmarking_data

---
pretty_name: CDSM Collagen Structure Benchmark
license: other
license_name: mixed-see-licensing-section
tags:
  - biology
  - protein-structure
  - collagen
  - structure-prediction
  - benchmark
size_categories:
  - n<1K
configs:
  - config_name: manifest
    data_files:
      - split: train
        path: experimental/manifest.parquet
  - config_name: scores_summary
    data_files:
      - split: train
        path: scores/scores_summary.parquet
  - config_name: scores_per_residue
    data_files:
      - split: train
        path: scores/scores_per_residue.parquet
  - config_name: cdsm_stage_table
    data_files:
      - split: train
        path: scores/cdsm_stage_table.parquet
  - config_name: method_summary_table
    data_files:
      - split: train
        path: scores/method_summary_table.parquet
---
# CDSM Collagen Structure Benchmark — Data

Structures and scores for a benchmark comparing a deterministic collagen
triple-helix builder (CDSM) against four co-folding models — Boltz-2, Chai-1,
Protenix and AlphaFold3, the last in both with-MSA (`af3_msa`) and no-MSA
(`af3_nomsa`) conditions — on 80 experimentally resolved collagen triple
helices from the RCSB PDB.

Code: https://github.com/bm-howard/cdsm_benchmarking
## Layout
| Prefix | Contents | Size |
|---|---|---|
| `experimental/` | 80 filtered experimental triple helices (`.cif`) + `manifest.csv/.parquet` | 8.3 MB |
| `cdsm/<stage>/` | Deterministically built structures, one directory per pipeline stage | 16 MB |
| `cdsm/<stage>/trajectories/` | MD trajectories (`.dcd` + starting `.pdb`) for the relaxed and annealed stages | 95 MB |
| `predictions/<model>/` | `<PDB>_<model>.cif` for `boltz`, `chai`, `protenix`, `af3_msa`, `af3_nomsa` | 19 MB |
| `scores/` | Scoring tables, as both `.csv` and `.parquet` | 4.4 MB |
| `compute_cost/` | Runtime/cost benchmark records, seed-sweep scores, and the seed-sweep structures | 15 MB |

CDSM stages, in pipeline order: `coreonly`, `fullseq`, `fullseq_reregistered`,
`fullseq_reregistered_relaxed`, `fullseq_reregistered_annealed`.

The prefixes are deliberate: reproducing the scoring needs `experimental/`,
`cdsm/` and `predictions/` — about 40 MB — rather than the full 142 MB. The
trajectories are the bulk of the dataset and are almost never needed.

## Loading

The benchmark code resolves these paths for you:

```python
from data_locations import experimental_cif_dir, predictions_dir, cdsm_dir
exp = experimental_cif_dir()                      # downloads on first use
af3 = predictions_dir("af3_msa")
mdt = cdsm_dir("fullseq_reregistered_relaxed", trajectories=True)
```

Directly, without the repo:

```python
from huggingface_hub import snapshot_download
path = snapshot_download(
    repo_id="CollagenHelixLabs/cdsm_benchmarking_data",
    repo_type="dataset",
    allow_patterns=["predictions/**", "experimental/**"],   # skip the trajectories
)
```

The four tabular configs load as datasets:

```python
from datasets import load_dataset
scores = load_dataset("CollagenHelixLabs/cdsm_benchmarking_data", "scores_summary")
```

## Schemas

**`manifest`** — one row per PDB entry (80 rows):
`pdb_id`, `kind` (`homotrimer` | `heterotrimer`), `n_distinct_chains`,
`gly_start`, `frame_offset`, `has_hyp`, `len_a/b/c`,
`chain_a/b/c_sequence`. Sequences are one-letter codes with **`O` =
hydroxyproline (HYP)**.

**`scores_summary`** — one row per (structure, variant), 700 rows:
`pdb_id`, `variant`, `tm_score`, `global_rmsd_allatom`,
`global_rmsd_backbone`, `global_lddt_allatom`, `global_lddt_backbone`,
`coverage`.
**`scores_per_residue`** — long format, 55,684 rows:
`pdb_id`, `variant`, `chain`, `resnum`, `rmsd_allatom`, `rmsd_backbone`,
`lddt_allatom`, `lddt_backbone`.
**`cdsm_stage_table`** — mean/median per CDSM pipeline stage and metric,
written by `figures/make_figures.py`.

**`method_summary_table`** — mean/median per method and metric, formatted for
the manuscript. **A static snapshot, not a derived table**: no script in the
repository regenerates it, and it was built over a shared-target subset that
differs from the current one, so its values will not match `scores_summary`
exactly. Treat `scores_summary` as authoritative.

**`compute_cost/`** — the compute-cost benchmark (CSV only, no parquet):
- `*_records*.csv` — raw per-call records (wall seconds, peak VRAM, cold/warm
  phase, GPU, seed) for each method. `*_records_seeds_shuffled*.csv` are the
  randomised 5-seed sweeps on the 10 length-stratified targets; bare
  `*_records.csv` are the single-seed timing runs. The co-folder sweeps cover
  the 10-target subset (both L40S and L4 tiers); the CDSM records cover all 75
  buildable targets (deterministic, single-threaded CPU, no seed).
- `seed_sensitivity_scores.csv` — every seed-sweep structure scored against
  its experimental reference with the same machinery as `scores_summary`
  (one row per model × tier × seed × target).
- `seed_cifs/<model>[_L4]/seed_<n>/` — the structures themselves
  (`<PDB>_<model>.cif`). `seed_cifs/af3/` holds locally generated AlphaFold 3
  output (see Licensing). Each model's sweep is complete at 50 calls per tier
  (10 targets × 5 seeds), except protenix which has an L40S sweep only.
- `bench_manifest.json` — the target/seed fingerprint every sweep runner
  asserts against; `cdsm_machine.json` — the CPU the CDSM records were
  timed on.
RMSD is computed in US-align's global-fit frame; lDDT is superposition-free.
TM-score is Cα, reference-normalised (`USalign -mm 1`).
## Coverage notes
- All five prediction variants cover all 80 targets. The CDSM stages cover 75: five
  entries (`1EI8`, `6M80`, `5K86`, `7LXQ`, `7LXP`) have internal Gly-X-Y
  register interruptions that divide by zero in the builder's propensity
  step, so they have no deterministic build at any stage.
- The annealed stage covers 19 structures, not 75 — it is a targeted
  comparison against the relaxed stage, not a full sweep.
## Licensing
**This dataset is mixed-licence. The prefixes are not interchangeable.**

| Prefix | Source | Terms |
|---|---|---|
| `experimental/` | RCSB PDB | CC0 — public domain |
| `cdsm/` | This work | See repository licence |
| `scores/` | This work | See repository licence |
| `compute_cost/` (except `seed_cifs/af3/`) | This work | See repository licence |
| `predictions/af3_msa/`, `predictions/af3_nomsa/` | **AlphaFold Server (Google DeepMind)** | **Output Terms of Use — non-commercial only** |
| `compute_cost/seed_cifs/af3/` | **Local AlphaFold 3 run (Google DeepMind model)** | **Output Terms of Use — non-commercial only** |
| `predictions/boltz/`, `chai/`, `protenix/` | Generated locally with the respective open models | Each model's own licence |

`predictions/af3_msa/` and `predictions/af3_nomsa/` are AlphaFold Server Output.
Google DeepMind's Output Terms
of Use apply to it and to anything substantially derived from it, including a
non-commercial restriction. Anyone redistributing or building on that prefix is
bound by those terms; see https://alphafoldserver.com/terms. The scores in
`scores/` include `variant == "af3_msa"` and `"af3_nomsa"` rows derived from
that Output. `compute_cost/seed_cifs/af3/` is Output of a local AlphaFold 3
run; the same terms apply, and the `model == "af3"` rows in
`compute_cost/seed_sensitivity_scores.csv` are derived from it.

## Revision history

Structures are replaced in place rather than versioned, so changes that alter
scores are recorded here. Superseded files remain in the dataset's commit
history.

- **2026-09-01** — Added `compute_cost/`: the raw compute-cost benchmark
  records and the 5-seed sweep scores (`seed_sensitivity_scores.csv`,
  10 length-stratified targets × 4 models × 5 seeds, L40S and L4 tiers), with
  the seed-sweep structures under `seed_cifs/`.

- **2026-08-14** — `predictions/boltz/` regenerated from a local Boltz run; all
  80 structures replaced. Median TM-score moved 0.913 → 0.914 and backbone RMSD
  1.138 → 1.085 Å. `scores/` was recomputed against the new structures.
- **2026-08-14** — AlphaFold3 prefixes renamed to name their MSA condition
  explicitly: `af3` → `af3_msa`, `AF3_no_MSA` → `af3_nomsa`. Scores were
  relabelled, not recomputed; the `af3_msa` values are unchanged from the
  earlier `af3` rows.

## Citation

<!-- TODO: replace on publication -->

```bibtex
@unpublished{cdsm_collagen_benchmark,
  title  = {Deterministic and co-folding model predictions of collagen triple helices},
  author = {Howard, Bruno},
  note   = {Manuscript in preparation},
  year   = {2026}
}
```

Please also cite the RCSB PDB, US-align, and whichever prediction models you use.

	---
	pretty_name: CDSM Collagen Structure Benchmark
	license: other
	license_name: mixed-see-licensing-section
	tags:
	- biology
	- protein-structure
	- collagen
	- structure-prediction
	- benchmark
	size_categories:
	- n<1K
	configs:
	- config_name: manifest
	data_files:
	- split: train
	path: experimental/manifest.parquet
	- config_name: scores_summary
	data_files:
	- split: train
	path: scores/scores_summary.parquet
	- config_name: scores_per_residue
	data_files:
	- split: train
	path: scores/scores_per_residue.parquet
	- config_name: cdsm_stage_table
	data_files:
	- split: train
	path: scores/cdsm_stage_table.parquet
	- config_name: method_summary_table
	data_files:
	- split: train
	path: scores/method_summary_table.parquet
	---

	# CDSM Collagen Structure Benchmark — Data

	Structures and scores for a benchmark comparing a deterministic collagen
	triple-helix builder (CDSM) against four co-folding models — Boltz-2, Chai-1,
	Protenix and AlphaFold3, the last in both with-MSA (`af3_msa`) and no-MSA
	(`af3_nomsa`) conditions — on 80 experimentally resolved collagen triple
	helices from the RCSB PDB.

	Code: https://github.com/bm-howard/cdsm_benchmarking

	## Layout

	\| Prefix \| Contents \| Size \|
	\|---\|---\|---\|
	\| `experimental/` \| 80 filtered experimental triple helices (`.cif`) + `manifest.csv/.parquet` \| 8.3 MB \|
	\| `cdsm/<stage>/` \| Deterministically built structures, one directory per pipeline stage \| 16 MB \|
	\| `cdsm/<stage>/trajectories/` \| MD trajectories (`.dcd` + starting `.pdb`) for the relaxed and annealed stages \| 95 MB \|
	\| `predictions/<model>/` \| `<PDB>_<model>.cif` for `boltz`, `chai`, `protenix`, `af3_msa`, `af3_nomsa` \| 19 MB \|
	\| `scores/` \| Scoring tables, as both `.csv` and `.parquet` \| 4.4 MB \|
	\| `compute_cost/` \| Runtime/cost benchmark records, seed-sweep scores, and the seed-sweep structures \| 15 MB \|

	CDSM stages, in pipeline order: `coreonly`, `fullseq`, `fullseq_reregistered`,
	`fullseq_reregistered_relaxed`, `fullseq_reregistered_annealed`.

	The prefixes are deliberate: reproducing the scoring needs `experimental/`,
	`cdsm/` and `predictions/` — about 40 MB — rather than the full 142 MB. The
	trajectories are the bulk of the dataset and are almost never needed.

	## Loading

	The benchmark code resolves these paths for you:

	```python
	from data_locations import experimental_cif_dir, predictions_dir, cdsm_dir

	exp = experimental_cif_dir() # downloads on first use
	af3 = predictions_dir("af3_msa")
	mdt = cdsm_dir("fullseq_reregistered_relaxed", trajectories=True)
	```

	Directly, without the repo:

	```python
	from huggingface_hub import snapshot_download

	path = snapshot_download(
	repo_id="CollagenHelixLabs/cdsm_benchmarking_data",
	repo_type="dataset",
	allow_patterns=["predictions/", "experimental/"], # skip the trajectories
	)
	```

	The four tabular configs load as datasets:

	```python
	from datasets import load_dataset

	scores = load_dataset("CollagenHelixLabs/cdsm_benchmarking_data", "scores_summary")
	```

	## Schemas

	`manifest` — one row per PDB entry (80 rows):
	`pdb_id`, `kind` (`homotrimer` \| `heterotrimer`), `n_distinct_chains`,
	`gly_start`, `frame_offset`, `has_hyp`, `len_a/b/c`,
	`chain_a/b/c_sequence`. Sequences are one-letter codes with **`O` =
	hydroxyproline (HYP)**.

	`scores_summary` — one row per (structure, variant), 700 rows:
	`pdb_id`, `variant`, `tm_score`, `global_rmsd_allatom`,
	`global_rmsd_backbone`, `global_lddt_allatom`, `global_lddt_backbone`,
	`coverage`.

	`scores_per_residue` — long format, 55,684 rows:
	`pdb_id`, `variant`, `chain`, `resnum`, `rmsd_allatom`, `rmsd_backbone`,
	`lddt_allatom`, `lddt_backbone`.

	`cdsm_stage_table` — mean/median per CDSM pipeline stage and metric,
	written by `figures/make_figures.py`.

	`method_summary_table` — mean/median per method and metric, formatted for
	the manuscript. A static snapshot, not a derived table: no script in the
	repository regenerates it, and it was built over a shared-target subset that
	differs from the current one, so its values will not match `scores_summary`
	exactly. Treat `scores_summary` as authoritative.

	`compute_cost/` — the compute-cost benchmark (CSV only, no parquet):

	- `_records.csv` — raw per-call records (wall seconds, peak VRAM, cold/warm
	phase, GPU, seed) for each method. `_records_seeds_shuffled.csv` are the
	randomised 5-seed sweeps on the 10 length-stratified targets; bare
	`*_records.csv` are the single-seed timing runs. The co-folder sweeps cover
	the 10-target subset (both L40S and L4 tiers); the CDSM records cover all 75
	buildable targets (deterministic, single-threaded CPU, no seed).
	- `seed_sensitivity_scores.csv` — every seed-sweep structure scored against
	its experimental reference with the same machinery as `scores_summary`
	(one row per model × tier × seed × target).
	- `seed_cifs/<model>[_L4]/seed_<n>/` — the structures themselves
	(`<PDB>_<model>.cif`). `seed_cifs/af3/` holds locally generated AlphaFold 3
	output (see Licensing). Each model's sweep is complete at 50 calls per tier
	(10 targets × 5 seeds), except protenix which has an L40S sweep only.
	- `bench_manifest.json` — the target/seed fingerprint every sweep runner
	asserts against; `cdsm_machine.json` — the CPU the CDSM records were
	timed on.

	RMSD is computed in US-align's global-fit frame; lDDT is superposition-free.
	TM-score is Cα, reference-normalised (`USalign -mm 1`).

	## Coverage notes

	- All five prediction variants cover all 80 targets. The CDSM stages cover 75: five
	entries (`1EI8`, `6M80`, `5K86`, `7LXQ`, `7LXP`) have internal Gly-X-Y
	register interruptions that divide by zero in the builder's propensity
	step, so they have no deterministic build at any stage.
	- The annealed stage covers 19 structures, not 75 — it is a targeted
	comparison against the relaxed stage, not a full sweep.

	## Licensing

	This dataset is mixed-licence. The prefixes are not interchangeable.

	\| Prefix \| Source \| Terms \|
	\|---\|---\|---\|
	\| `experimental/` \| RCSB PDB \| CC0 — public domain \|
	\| `cdsm/` \| This work \| See repository licence \|
	\| `scores/` \| This work \| See repository licence \|
	\| `compute_cost/` (except `seed_cifs/af3/`) \| This work \| See repository licence \|
	\| `predictions/af3_msa/`, `predictions/af3_nomsa/` \| AlphaFold Server (Google DeepMind) \| Output Terms of Use — non-commercial only \|
	\| `compute_cost/seed_cifs/af3/` \| Local AlphaFold 3 run (Google DeepMind model) \| Output Terms of Use — non-commercial only \|
	\| `predictions/boltz/`, `chai/`, `protenix/` \| Generated locally with the respective open models \| Each model's own licence \|

	`predictions/af3_msa/` and `predictions/af3_nomsa/` are AlphaFold Server Output.
	Google DeepMind's Output Terms
	of Use apply to it and to anything substantially derived from it, including a
	non-commercial restriction. Anyone redistributing or building on that prefix is
	bound by those terms; see https://alphafoldserver.com/terms. The scores in
	`scores/` include `variant == "af3_msa"` and `"af3_nomsa"` rows derived from
	that Output. `compute_cost/seed_cifs/af3/` is Output of a local AlphaFold 3
	run; the same terms apply, and the `model == "af3"` rows in
	`compute_cost/seed_sensitivity_scores.csv` are derived from it.

	## Revision history

	Structures are replaced in place rather than versioned, so changes that alter
	scores are recorded here. Superseded files remain in the dataset's commit
	history.

	- 2026-09-01 — Added `compute_cost/`: the raw compute-cost benchmark
	records and the 5-seed sweep scores (`seed_sensitivity_scores.csv`,
	10 length-stratified targets × 4 models × 5 seeds, L40S and L4 tiers), with
	the seed-sweep structures under `seed_cifs/`.

	- 2026-08-14 — `predictions/boltz/` regenerated from a local Boltz run; all
	80 structures replaced. Median TM-score moved 0.913 → 0.914 and backbone RMSD
	1.138 → 1.085 Å. `scores/` was recomputed against the new structures.
	- 2026-08-14 — AlphaFold3 prefixes renamed to name their MSA condition
	explicitly: `af3` → `af3_msa`, `AF3_no_MSA` → `af3_nomsa`. Scores were
	relabelled, not recomputed; the `af3_msa` values are unchanged from the
	earlier `af3` rows.

	## Citation

	<!-- TODO: replace on publication -->

	```bibtex
	@unpublished{cdsm_collagen_benchmark,
	title = {Deterministic and co-folding model predictions of collagen triple helices},
	author = {Howard, Bruno},
	note = {Manuscript in preparation},
	year = {2026}
	}
	```

	Please also cite the RCSB PDB, US-align, and whichever prediction models you use.