README.md

---
license: other
license_name: mixed-upstream
language:
  - en
tags:
  - biology
  - proteins
  - protein-language-model
  - contrastive-learning
  - embeddings
  - sentence-transformers
  - pfam
  - alphafold
  - string-db
  - proteingym
pretty_name: ProtSent Training Data
size_categories:
  - 100M<n<1B
task_categories:
  - feature-extraction
  - sentence-similarity
configs:
  - config_name: pfam
    data_files: pfam_sorted.parquet
  - config_name: afdb
    data_files: afdb_sorted.parquet
  - config_name: stringdb
    data_files: stringdb_train.parquet
  - config_name: dms
    data_files: dms_cosent.parquet
---

# ProtSent Training Data

Training data for **ProtSent (Protein Sentence Transformers)** — a contrastive fine-tuning
framework that adapts protein language models (ESM-2, ESM-C) into general-purpose sequence
embeddings, so that nearest-neighbor distance reflects functional, evolutionary, and structural
similarity between proteins.

- **Paper:** *ProtSent: Protein Sentence Transformers* (Ofer, Perets, Linial, Rappoport), arXiv:2605.06830
- **Code:** https://github.com/oriel9p/ProtSent
- **Models:** https://huggingface.co/collections/oriel9p/protsent

Four of the paper's five training sources are included as parquet files. The synthetic **Pfam
hard-negatives** set is excluded: the paper's ablation (Tables 4, 7) shows removing it improves
aggregate performance (20/23 tasks improved, +7.9% mean delta, vs. 16/23 at +6.7% with it included).

| Source | Signal | Positive-pair criterion | Loss |
|---|---|---|---|
| **Pfam** | evolutionary / functional homology | same Pfam family | MNRL / GIST |
| **AFDB** | 3D structural similarity | same Foldseek structural cluster | MNRL / GIST |
| **STRING** | physical / functional interaction | pair-native (interacting proteins) | MNRL / GIST |
| **DMS** | continuous mutational fitness | continuous score | CoSENT |

## Files and schema

Two conventions are used. **Grouped** files store one sequence per row plus a group label; positive
pairs are formed at training time from sequences sharing a group (capped per group via
`--max_pairs_per_cluster`). **Pair-native** files store explicit pairs, one per row.

### `pfam_sorted.parquet` — Pfam families (grouped, 2.4 GB)
| column | type | notes |
|---|---|---|
| `sequence` | string | amino-acid sequence (Pfam domain) |
| `family_id` | string | Pfam family (PF accession, version stripped) |
| `clan_id` | string | Pfam clan; orphan families inherit `clan_id := family_id` |
| `group_id` | string | positive-pair label; equals `family_id` |

Sorted `clan_id → family_id`; singleton families dropped. 28.5M domains after MMseqs2
linclust@70% deduplication. Do not shuffle — training relies on the sort order for windowed slicing.

### `afdb_sorted.parquet` — AlphaFold DB Foldseek clusters (grouped, 13 GB)
| column | type | notes |
|---|---|---|
| `sequence` | string | amino-acid sequence (pLDDT > 70, non-fragment) |
| `cluster_id` | string | Foldseek structural-cluster representative |
| `afdb50_cluster_id` | string | AFDB50 (50%-identity) representative |
| `group_id` | string | positive-pair label; equals `cluster_id` |

Sorted by `cluster_id`; single-member clusters dropped. The positive label is the Foldseek
structural cluster (`cluster_id`), coarser than AFDB50 sequence identity.

### `stringdb_train.parquet` — STRING-DB v12 PPI (pair-native, 34 GB)
| column | type | notes |
|---|---|---|
| `seq1` | string | interacting protein A |
| `seq2` | string | interacting protein B |

`combined_score ≥ 400`, Bernett-decontaminated and cluster-deduplicated (see Provenance).
Endpoint length filter `[10, 2048]`.

### `dms_cosent.parquet` — ProteinGym DMS / clinical (pair-native, scored, 106 MB)
| column | type | notes |
|---|---|---|
| `sentence_0` | string | sequence A |
| `sentence_1` | string | sequence B |
| `score` | float | fitness-similarity target for CoSENT (0–1) |

2.18M pairs over 3,576 wild-type targets and 2.09M unique mutants.

## Usage

### Load with `datasets`
```python
from datasets import load_dataset

pfam     = load_dataset("GrimSqueaker/protsent-data", "pfam",     split="train")
afdb     = load_dataset("GrimSqueaker/protsent-data", "afdb",     split="train")
stringdb = load_dataset("GrimSqueaker/protsent-data", "stringdb", split="train")
dms      = load_dataset("GrimSqueaker/protsent-data", "dms",      split="train")
```

### Download raw parquets (for the ProtSent training code)
```python
from huggingface_hub import snapshot_download
snapshot_download("GrimSqueaker/protsent-data", repo_type="dataset", local_dir="data")
```
Files land flat in `data/`. Then train:
```bash
accelerate launch --num_processes 8 --mixed_precision bf16 \
  protein_pipeline.py train \
  --model Synthyra/ESMplusplus_large \
  --files data/pfam_sorted.parquet data/afdb_sorted.parquet data/stringdb_train.parquet \
  --dms_file data/dms_cosent.parquet \
  --loss_mode multi --multi_primary_loss gist \
  --max_seq_length 1024 --max_pairs_per_cluster 100
```
The training code detects file type by schema: `seq1`/`seq2` → PPI pairs; `sentence_0`/`sentence_1`/`score`
→ DMS/CoSENT; otherwise grouped (`sequence` + `group_id`), with positive pairs formed per group.

## Provenance and filtering

- **Pfam** — `Pfam-A.fasta.gz` + `Pfam-A.clans.tsv.gz` (EBI current release). MMseqs2 `easy-linclust`
  at 70% identity (`--cov-mode 1 -c 0.8`), family→clan join, singleton families dropped, sorted by
  `(clan_id, family_id)`.
- **AFDB** — sequences from `willdaspit/afdb_clustered_seqs` (pLDDT > 70, non-fragment) inner-joined to
  the Steinegger-Lab AFDB Foldseek v6 S-cluster mapping (`cluFlag ∈ {1,2}`) on the AFDB50
  representative; positive label = Foldseek cluster.
- **STRING** — STRING v12.0 physical links (`combined_score ≥ 400`). Bernett decontamination: any
  STRING protein clustering (MMseqs2 linclust, 50% id / 80% cov) with a `Synthyra/bernett_gold_ppi`
  test sequence is removed. Survivors two-stage clustered (65%/85% then 50%/75%); pairs canonicalized
  and deduplicated per cluster-pair; endpoint length filter `[10, 2048]`.
- **DMS** — `OATML-Markslab/ProteinGym_v1` (DMS + clinical, substitutions + indels). Per-assay
  z-scored, clipped to [−3, 3], rescaled to [0, 1]; clinical Pathogenic→0 / Benign→1. GB1 and GFP
  assays dropped (benchmark overlap); supervised test folds removed.

Thresholds, identity cutoffs, and seeds are fixed in `data_prep.py`.

## Leakage controls

- DMS and STRING are decontaminated against their downstream benchmarks (ProteinGym excluded from
  evaluation; Bernett test set removed from STRING).
- Pfam training uses family-membership labels, disjoint from the fold-level remote-homology eval split.
- AFDB uses Foldseek-cluster co-membership rather than SCOPe labels. AFDB sequences are not filtered
  against SCOPe test domains, so partial sequence overlap with SCOPe-40 retrieval is possible — noted
  in the paper as a limitation.

## Reproduce

```bash
# requires MMseqs2 on PATH for pfam + stringdb
python data_prep.py --dataset pfam
python data_prep.py --dataset afdb --limit_gb 0
python data_prep.py --dataset stringdb --max_seq_len 2048 --min_seq_len 10
python data_prep.py --dataset dms
```

## Citation

```bibtex
@article{ofer2026protsent,
  title         = {ProtSent: Protein Sentence Transformers},
  author        = {Ofer, Dan and Perets, Oriel and Linial, Michal and Rappoport, Nadav},
  journal       = {arXiv preprint arXiv:2605.06830},
  year          = {2026},
  eprint        = {2605.06830},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2605.06830}
}
```

## License

Pfam (EBI), AlphaFold DB (CC-BY-4.0), STRING (CC-BY-4.0), and ProteinGym each carry their own
license.