Datasets:

stevensu123
/

proteinchameleon-structure-function

---
license: cc-by-4.0
pretty_name: ProteinChameleon Structure–Function Dataset
language:
- en
size_categories:
- 100K<n<1M
task_categories:
- text-generation
tags:
- protein
- biology
- protein-structure
- protein-function
- structure-tokens
- geobpe
- multimodal
configs:
- config_name: alignment
  data_files:
  - split: train
    path: alignment/train.parquet
  - split: validation
    path: alignment/val.parquet
  - split: test
    path: alignment/test.parquet
- config_name: interleaved
  data_files:
  - split: train
    path: interleaved/train.parquet
  - split: validation
    path: interleaved/val.parquet
  - split: test
    path: interleaved/test.parquet
---
# ProteinChameleon Structure–Function Dataset

A leakage-free protein dataset pairing **3D-structure tokens** and amino-acid
sequences with **natural-language function descriptions**. It was built to train and
evaluate [ProteinChameleon](https://github.com/), an early-fusion protein–text model
that ingests discrete structure tokens alongside sequence. Structure is encoded with
**GeoBPE**, a geometric byte-pair tokenizer that turns a protein's 3D backbone into a
sequence of ~2,100 discrete tokens (merged multi-residue structural motifs plus
quantized dihedral/angle bins).

The distinguishing feature of this release is its **split protocol**: train / validation
/ test are separated by *connected components* over both sequence homology and function
text, so the test set measures genuine generalization rather than near-neighbor recall.

## Configurations

| Config | Task | Train | Validation | Test |
|---|---|---:|---:|---:|
| `alignment` | structure + sequence → function text | 243,127 | 13,504 | 13,506 |
| `interleaved` | multimodal protein narratives (domain-segmented) | 63,117 | 4,303 | 4,206 |

### `alignment` fields
| Field | Type | Description |
|---|---|---|
| `accession` | string | UniProt accession |
| `organism` | string | Source organism |
| `sequence` | string | Amino-acid sequence |
| `function_text` | string | Curated natural-language function description |
| `structure_tokens` | list[int] | GeoBPE structure token IDs for the protein's 3D backbone |

### `interleaved` fields
| Field | Type | Description |
|---|---|---|
| `accession` | string | UniProt accession |
| `organism` | string | Source organism |
| `sequence` | string | Amino-acid sequence |
| `narrative` | string | Multimodal narrative interleaving text with domain-level context |
| `n_domains` | int | Number of annotated domains in the narrative |

## Usage

```python
from datasets import load_dataset
# structure -> function
ds = load_dataset("<your-username>/proteinchameleon-structure-function", "alignment")
print(ds["test"][0]["function_text"])
print(ds["test"][0]["structure_tokens"][:16])
# multimodal narratives
di = load_dataset("<your-username>/proteinchameleon-structure-function", "interleaved")
```

The original NumPy artifacts are also included under [`raw/`](./raw) for exact
reproducibility (`np.load(..., allow_pickle=True)`).

## Data Collection & Curation

### Sources
- **Sequences & function annotations** — [UniProtKB / Swiss-Prot](https://www.uniprot.org)
  (manually reviewed entries). `function_text` is derived from curated UniProt function
  annotations; `organism` and `accession` are taken directly from the entry.
- **3D structures** — [AlphaFold Protein Structure Database](https://alphafold.ebi.ac.uk),
  tokenized with **GeoBPE**. Each backbone is quantized into geometric tokens and merged
  into structural-motif tokens, yielding the `structure_tokens` sequence.

### Leakage-free split protocol
Naïve per-accession splits leak badly, because protein databases are highly redundant and
UniProt function text is shared verbatim across orthologs. Measured on the original splits:

- **Homology leakage** — 94% of test proteins had a ≥30%-identity homolog in train/val.
- **Function-text leakage** — 79% of test function descriptions appeared verbatim in train.
- **Cross-task leakage** — 91% of the interleaved-test proteins also appeared in
  alignment-train (the two tasks were split independently but trained jointly).

To remove all three, the splits in this release (`clean2`) are built as follows:

1. **Union of both tasks.** Pool every protein appearing in either the alignment or
   interleaved task into a single set.
2. **Homology graph.** Cluster all sequences with **MMseqs2** at **30% sequence identity**
   and **80% coverage** (`mmseqs easy-cluster --min-seq-id 0.3 -c 0.8`). Add an edge
   between any two proteins in the same cluster.
3. **Function-text graph.** Add an edge between any two proteins that share an *identical*
   function description.
4. **Connected components.** Take connected components of the combined graph (union-find).
   Homologous proteins and proteins sharing boilerplate function text therefore always fall
   in the same component.
5. **Component-level assignment.** Assign whole components to train / validation / test
   (~90 / 5 / 5%), so no protein — and no close homolog or shared description — ever
   crosses the split boundary. The **same assignment is applied to both tasks**, eliminating
   cross-task leakage.

A machine-readable `split_manifest_clean2.json` and the per-protein
`split_assignment_clean2.json` are included under `raw/` for auditing.

### Verification
After construction, the splits were re-checked for all three leakage channels; homology,
function-text, and cross-task overlap between test and train/val are driven to **zero**
under the 30%-identity criterion.

## Limitations & Intended Use
- `structure_tokens` are **GeoBPE codes**, not 3D coordinates; they are only meaningful with
  a model trained on the same GeoBPE vocabulary. They are not interchangeable with other
  structure-token schemes (e.g. Foldseek 3Di, ESM3 structure tokens).
- Function text is curated prose and may contain UniProt-style boilerplate; it is intended
  for conditional generation / understanding, not as a controlled ontology (use the EC/CATH
  ontologies for label-based tasks).
- The dataset reflects the coverage and biases of Swiss-Prot and AlphaFold DB.

## License & Attribution
Released under **CC-BY-4.0**, consistent with the source databases. If you use this dataset,
please attribute the underlying resources:
- The UniProt Consortium, *UniProt: the Universal Protein Knowledgebase* (CC-BY-4.0).
- Varadi et al., *AlphaFold Protein Structure Database* (CC-BY-4.0).

## Citation
```bibtex
@misc{proteinchameleon_dataset,
  title  = {ProteinChameleon Structure--Function Dataset},
  author = {Steven},
  year   = {2026},
  note   = {Leakage-free structure-token / function-text splits derived from
            UniProt/Swiss-Prot and AlphaFold DB via GeoBPE tokenization}
}
```

	---
	license: cc-by-4.0
	pretty_name: ProteinChameleon Structure–Function Dataset
	language:
	- en
	size_categories:
	- 100K<n<1M
	task_categories:
	- text-generation
	tags:
	- protein
	- biology
	- protein-structure
	- protein-function
	- structure-tokens
	- geobpe
	- multimodal
	configs:
	- config_name: alignment
	data_files:
	- split: train
	path: alignment/train.parquet
	- split: validation
	path: alignment/val.parquet
	- split: test
	path: alignment/test.parquet
	- config_name: interleaved
	data_files:
	- split: train
	path: interleaved/train.parquet
	- split: validation
	path: interleaved/val.parquet
	- split: test
	path: interleaved/test.parquet
	---

	# ProteinChameleon Structure–Function Dataset

	A leakage-free protein dataset pairing 3D-structure tokens and amino-acid
	sequences with natural-language function descriptions. It was built to train and
	evaluate [ProteinChameleon](https://github.com/), an early-fusion protein–text model
	that ingests discrete structure tokens alongside sequence. Structure is encoded with
	GeoBPE, a geometric byte-pair tokenizer that turns a protein's 3D backbone into a
	sequence of ~2,100 discrete tokens (merged multi-residue structural motifs plus
	quantized dihedral/angle bins).

	The distinguishing feature of this release is its split protocol: train / validation
	/ test are separated by connected components over both sequence homology and function
	text, so the test set measures genuine generalization rather than near-neighbor recall.

	## Configurations

	\| Config \| Task \| Train \| Validation \| Test \|
	\|---\|---\|---:\|---:\|---:\|
	\| `alignment` \| structure + sequence → function text \| 243,127 \| 13,504 \| 13,506 \|
	\| `interleaved` \| multimodal protein narratives (domain-segmented) \| 63,117 \| 4,303 \| 4,206 \|

	### `alignment` fields
	\| Field \| Type \| Description \|
	\|---\|---\|---\|
	\| `accession` \| string \| UniProt accession \|
	\| `organism` \| string \| Source organism \|
	\| `sequence` \| string \| Amino-acid sequence \|
	\| `function_text` \| string \| Curated natural-language function description \|
	\| `structure_tokens` \| list[int] \| GeoBPE structure token IDs for the protein's 3D backbone \|

	### `interleaved` fields
	\| Field \| Type \| Description \|
	\|---\|---\|---\|
	\| `accession` \| string \| UniProt accession \|
	\| `organism` \| string \| Source organism \|
	\| `sequence` \| string \| Amino-acid sequence \|
	\| `narrative` \| string \| Multimodal narrative interleaving text with domain-level context \|
	\| `n_domains` \| int \| Number of annotated domains in the narrative \|

	## Usage

	```python
	from datasets import load_dataset

	# structure -> function
	ds = load_dataset("<your-username>/proteinchameleon-structure-function", "alignment")
	print(ds["test"][0]["function_text"])
	print(ds["test"][0]["structure_tokens"][:16])

	# multimodal narratives
	di = load_dataset("<your-username>/proteinchameleon-structure-function", "interleaved")
	```

	The original NumPy artifacts are also included under [`raw/`](./raw) for exact
	reproducibility (`np.load(..., allow_pickle=True)`).

	## Data Collection & Curation

	### Sources
	- Sequences & function annotations — [UniProtKB / Swiss-Prot](https://www.uniprot.org)
	(manually reviewed entries). `function_text` is derived from curated UniProt function
	annotations; `organism` and `accession` are taken directly from the entry.
	- 3D structures — [AlphaFold Protein Structure Database](https://alphafold.ebi.ac.uk),
	tokenized with GeoBPE. Each backbone is quantized into geometric tokens and merged
	into structural-motif tokens, yielding the `structure_tokens` sequence.

	### Leakage-free split protocol
	Naïve per-accession splits leak badly, because protein databases are highly redundant and
	UniProt function text is shared verbatim across orthologs. Measured on the original splits:

	- Homology leakage — 94% of test proteins had a ≥30%-identity homolog in train/val.
	- Function-text leakage — 79% of test function descriptions appeared verbatim in train.
	- Cross-task leakage — 91% of the interleaved-test proteins also appeared in
	alignment-train (the two tasks were split independently but trained jointly).

	To remove all three, the splits in this release (`clean2`) are built as follows:

	1. Union of both tasks. Pool every protein appearing in either the alignment or
	interleaved task into a single set.
	2. Homology graph. Cluster all sequences with MMseqs2 at 30% sequence identity
	and 80% coverage (`mmseqs easy-cluster --min-seq-id 0.3 -c 0.8`). Add an edge
	between any two proteins in the same cluster.
	3. Function-text graph. Add an edge between any two proteins that share an identical
	function description.
	4. Connected components. Take connected components of the combined graph (union-find).
	Homologous proteins and proteins sharing boilerplate function text therefore always fall
	in the same component.
	5. Component-level assignment. Assign whole components to train / validation / test
	(~90 / 5 / 5%), so no protein — and no close homolog or shared description — ever
	crosses the split boundary. The same assignment is applied to both tasks, eliminating
	cross-task leakage.

	A machine-readable `split_manifest_clean2.json` and the per-protein
	`split_assignment_clean2.json` are included under `raw/` for auditing.

	### Verification
	After construction, the splits were re-checked for all three leakage channels; homology,
	function-text, and cross-task overlap between test and train/val are driven to zero
	under the 30%-identity criterion.

	## Limitations & Intended Use
	- `structure_tokens` are GeoBPE codes, not 3D coordinates; they are only meaningful with
	a model trained on the same GeoBPE vocabulary. They are not interchangeable with other
	structure-token schemes (e.g. Foldseek 3Di, ESM3 structure tokens).
	- Function text is curated prose and may contain UniProt-style boilerplate; it is intended
	for conditional generation / understanding, not as a controlled ontology (use the EC/CATH
	ontologies for label-based tasks).
	- The dataset reflects the coverage and biases of Swiss-Prot and AlphaFold DB.

	## License & Attribution
	Released under CC-BY-4.0, consistent with the source databases. If you use this dataset,
	please attribute the underlying resources:
	- The UniProt Consortium, UniProt: the Universal Protein Knowledgebase (CC-BY-4.0).
	- Varadi et al., AlphaFold Protein Structure Database (CC-BY-4.0).

	## Citation
	```bibtex
	@misc{proteinchameleon_dataset,
	title = {ProteinChameleon Structure--Function Dataset},
	author = {Steven},
	year = {2026},
	note = {Leakage-free structure-token / function-text splits derived from
	UniProt/Swiss-Prot and AlphaFold DB via GeoBPE tokenization}
	}
	```