README.md

---
annotations_creators:
- expert-generated
- machine-generated
language:
- en
language_creators:
- found
license:
- other
multilinguality:
- monolingual
pretty_name: PFArena
size_categories:
- 100K<n<1M
source_datasets:
- original
tags:
- protein
- mutation
- protein-engineering
- deep-mutational-scanning
- fitness
- stability
- binding
- benchmark
task_categories:
- text-generation
- text-ranking
---

# Dataset Card for PFArena

## Dataset Description

- **GitHub:** https://github.com/AMix-Bio/PFArena
- **Dataset:** https://huggingface.co/datasets/AMix-Bio/PFArena
- **Leaderboard:** https://huggingface.co/spaces/AMix-Bio/PFArena-Leaderboard

### Dataset Summary

**PFArena** is an assay-grounded **protein-mutation benchmark** that systematically evaluates protein language models (PLMs), large language models (LLMs), and LLM-based agents under four protein modification settings. Comprising 202 unique assays, 293 assay--task instances, and 607,269 target candidate rows, PFArena defines four task interfaces that correspond to common decisions in protein-engineering workflows:

| Task | Directory | Task purpose |
| --- | --- | --- |
| T1 | `T1_single_mutant_generation` | Propose promising mutations without target-specific mutation measurements |
| T2 | `T2_measurement_free_multi_mutant_ranking` | Prioritize a supplied pool of multi-mutant candidates |
| T3 | `T3_anchor_informed_multi_mutant_ranking` | Rank successors of an experimentally measured mutant |
| T4 | `T4_single_mutant_informed_multi_mutant_ranking` | Prioritize combinations using measured effects of their component mutations |

Every assay is normalized to a shared schema. The canonical ground-truth label is `DMS_score`,
an assay-resolved score in which **larger values always correspond to the better-performing
direction** for that assay (higher fitness, higher stability, higher binding, etc.).

### Supported Tasks

The four tasks map to two model capabilities:

- **T1** — *protein-mutation-generation*: produce a ranked list of single-point substitutions
  predicted to be most beneficial. Evaluated with **NMS@40** (normalized max score) and
  **Recall@40**.
- **T2–T4** — *protein-mutation-ranking*: rank a fixed candidate set of multi-point mutants.
  Evaluated with global ranking agreement (**Spearman**, **NDCG**) and top-five quality/recovery
  (**NMS@5**, **Recall@5**).

## Dataset Structure

### Data Instances

The dataset is a directory of CSV tables and FASTA/A3M alignment files. It has no train/validation/test split — the data are benchmark inputs and ground-truth labels, not a training corpus.

Two documentation files describe the schema precisely:

- `assay_columns.md` — assay-level metadata columns (root `assay.csv` and per-task `assay.csv`).
- `norm_data_columns.md` — candidate-table columns (`norm_data/` files).

A single `norm_data` row (T1) looks like:

```csv
candidate_group_id,candidate_label,mutant,mutated_sequence,DMS_score
CDKN2A_elife_95347_supp4_supp6_T1,single_mutant,M1A,AEPAAGSSMEPS...,-0.5788823699525986
```

A single `norm_data` row (T2) looks like:

```csv
candidate_group_id,mutant,mutated_sequence,DMS_score
COMBINGYM_CreiLOV_fluorescence_log_mean_T2,G3E+R5D+T7H+...,MAELDHHFVVADA...,0.8368092913493479
```

### Data Fields

**Root `assay.csv`** — one row per assay (202 rows). Shared assay metadata:

| Column | Description |
| --- | --- |
| `assay_id` | Stable identifier for the assay–task record. |
| `source_dataset` | Curation source (one of the seven sources below). |
| `primary_task_class` | Primary benchmark class (`activity_function`, `binding`, `stability`). |
| `uniprot_id` | UniProt accession (blank when no reliable accession was available). |
| `fitness_type` | Broad biological objective. |
| `readout_subclass` | Fine-grained assay readout subclass. |
| `assay_modality` | Experimental measurement modality. |
| `fitness_subtype` | Specific phenotype/readout taxonomy. |
| `wildtype_sequence` | Wild-type protein sequence. |
| `sequence_length` | Length of the wild-type sequence. |
| `single_site_mutation_coverage` | Fraction of wild-type positions with measured single substitutions (empty only for the FLAb T2 library that has no single-mutant measurements). |
| `biophysical_directness` | `low` (former scores 1–2), `medium` (3), or `high` (4–5). |
| `n_rows` | Number of retained normalized measurement rows. |
| `n_positions` | Number of sequence positions represented among retained variants. |
| `a3m_relative_path` | Relative path to the assay-linked A3M alignment. |
| `a3m_status_detail` | Status and coverage detail for the linked A3M alignment. |

**Per-task `assay.csv`** adds the task-specific count/context columns. For example:

- **T3** adds `anchor_mutant`, `anchor_DMS_score`, `anchor_mutation_count`,
  `anchor_effect_provided`, `anchor_is_multi_mutant`, `n_successors_available`,
  `n_successors_selected`.
- **T4** adds `n_context_single`, `n_combo_candidates`, `n_combo_available`,
  `single_context_data_path`.

**`norm_data/` core columns** (all candidate tables):

| Column | Description |
| --- | --- |
| `candidate_group_id` | Candidate-group identifier, one per assay per task (e.g. `{assay_id}_T2`). |
| `mutant` | Canonical mutation notation, e.g. `A12V` or `A12V+G35L`. |
| `mutated_sequence` | Protein sequence carrying the listed mutations. |
| `DMS_score` | Canonical assay score used for evaluation; larger = better (assay-resolved direction). |

Task-specific variations:

- **T1** adds `candidate_label` (the single-mutant candidate-set label).
- **T3** files are named `anchor_successor.csv` and contain the successor candidate set; the
  anchor constraint is stored in the task `assay.csv` metadata, not per row.
- **T4** uses two files per assay: `single_context_combo.csv` (hidden multi-mutant candidates to
  rank) and `single_mutant_context.csv` (the visible measured single-mutant context).


## Data Source

The 202 assays are drawn from seven sources:

| Source | Assays |
| --- | --- |
| ProteinGym | 53 |
| MaveDB | 44 |
| Human_Domainome | 36 |
| MegaScale | 34 |
| FLAb | 23 |
| CombinGym | 7 |
| `target_dms` | 5 |

Biological objectives span stability (81 assays), binding (37), enzymatic activity (37),
organismal/cellular fitness (31), fluorescence (13), electrophysiology (1), polyreactivity (1),
and abundance/expression (1). Sequence lengths range from roughly 62 to over 1,360 residues,
including multi-chain antibody constructs (heavy/light chains joined by `:`).

For each assay, an MSA is provided under `msa_mmseqs_uniref100/` (256 `.a3m` files generated with
MMseqs2 against UniRef100; multi-chain assays have one alignment per chain).


## Considerations for Using the Data

- **Ground truth must not be leaked to inference.** The benchmark is intended to measure a model's
  predictions; do not expose `DMS_score` (or other ground-truth columns) to an inference backend.
- **Keep the downloaded data unchanged.** Candidate tables, wild-type sequences, and MSAs are
  frozen; modifying them will invalidate evaluation.
- **Score direction.** `DMS_score` is already oriented so that larger is better per assay; do not
  re-normalize or re-orient without consulting the per-assay metadata.
- **`uniprot_id` may be blank** where no reliable accession could be assigned.

## Citation Information

If you use PFArena, cite the dataset and the original source datasets. Refer to the
`AMix-Bio/PFArena` repository for the canonical citation.