Dataset Card for PFArena
Timeline
2026-11-16: PFArena_v1 update. Retained more representative assays, fields, and protein metadata.
2026-09-03: PFArena_v0 launch.
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 170 unique assays, 248 assay–task instances, and 466,069 target candidate rows, PFArena defines four task interfaces that correspond to common decisions in protein-engineering workflows:
| Task | Directory | Instances | Task purpose |
|---|---|---|---|
| T1 | T1_single_mutant_generation |
101 | Propose promising mutations without target-specific mutation measurements |
| T2 | T2_measurement_free_multi_mutant_ranking |
65 | Prioritize a supplied pool of multi-mutant candidates |
| T3 | T3_anchor_informed_multi_mutant_ranking |
60 | Rank successors of an experimentally measured mutant |
| T4 | T4_single_mutant_informed_multi_mutant_ranking |
22 | Prioritize combinations using measured effects of their component mutations |
Every assay follows 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, as defined by its experimental objective.
Score processing, normalization, and direction are checked at the assay level against source data and experimental context. The 41 ProteinGym assays retain their accepted upstream scales without additional z-scoring. The other 129 assays use verified z-scores calculated from documented source populations, with population standard deviation (ddof=0). Normalization parameters are not fitted separately to each task's candidate pool.
Supported Tasks
The four tasks map to two model capabilities:
- T1 — protein-mutation-generation: produce a ranked list of exactly 40 distinct, legal 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, including every supplied candidate exactly once. 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 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 (rootassay.csvand per-taskassay.csv).norm_data_columns.md— candidate-table columns (norm_data/files).
A single norm_data row (T1) looks like:
candidate_group_id,mutant,mutated_sequence,DMS_score
CDKN2A_elife_95347_supp4_supp6_T1,M1A,AEPAAGSSMEPS...,0.5788823699525986
A single norm_data row (T2) looks like:
candidate_group_id,mutant,mutated_sequence,DMS_score
COMBINGYM_eqFP611_blue_fluorescence_T2,D20N+L63M+T127P+A172C+A174L+N207K+K231R,MSELIKENMHMK...,0.9008113352529036
Data Fields
Root assay.csv — one row per assay (170 rows). Shared assay metadata:
| Column | Description |
|---|---|
assay_id |
Stable identifier for the assay. |
source_dataset |
Curation source (one of the seven sources below). |
primary_task_class |
Primary benchmark class (activity_function, binding, stability, abundance_expression). |
uniprot_id |
UniProt accession (blank when no reliable accession was available). |
fitness_type |
Second-level biological objective category. |
readout_subclass |
Fine-grained assay readout description, including relevant experimental context. |
wildtype_sequence |
Supplied reference sequence used for mutation notation; it may be a natural sequence, source construct, fragment, or inferred reference. |
sequence_length |
Number of amino-acid residues in the supplied reference, excluding chain separators. |
preferred_DMS_score_direction |
Preferred score direction; higher for all assays. |
agent_execution_policy |
Model-input policy; context_and_direction_required for all assays. |
agent_assay_context_json |
Six-key experimental context describing the endpoint, objective, score direction, reference, critical context, and measurement method. |
label_schema_version |
Classification-schema identifier. |
context_schema_version |
Experimental-context schema identifier. |
n_rows |
Representative assay-index count; not the normalization-population size or a sum across tasks. |
a3m_relative_path |
Relative path to the assay-linked A3M alignment; multiple chain paths are separated by ;. |
paper_doi |
DOI of the associated source publication. |
The three classification fields form a hierarchy:
primary_task_classdistinguishes four broad classes.fitness_typedistinguishes seven categories:folding_thermal_stability,abundance_expression,binding_interaction,biochemical_function,cellular_regulatory_function,cellular_fitness_phenotype, andfluorescence_output.readout_subclassspecifies the experimental readout, preserving relevant distinctions such as interaction partners, cellular backgrounds, measurement methods, and treatment conditions.
Per-task assay.csv adds task_name, DMS_filename, task_data_path, and the task-specific count/context columns:
- T2 adds
n_random_candidates. - T3 adds
anchor_mutantandanchor_DMS_score. - T4 adds
n_context_singleandsingle_context_data_path.
In a per-task index, n_rows is the actual number of rows in the corresponding target table. All data paths are relative to the dataset root, and task_data_path = task_name / DMS_filename.
norm_data/ core columns (all target and context 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:
- T3 target tables contain the successor candidate set; the anchor constraint and measured
anchor score are stored in the task
assay.csvmetadata, not per row. - T4 uses two files per assay: a multi-mutant candidate table with hidden target scores, and a single-mutant context table with visible measured scores.
The dataset contains 248 target tables with 466,069 candidate rows, together with 22 T4 context tables containing 27,363 measured single-mutant rows. An assay or variant may occur in more than one task setting.
Data Source
The 170 assays are drawn from seven sources:
| Source | Assays |
|---|---|
| ProteinGym | 41 |
| MaveDB | 34 |
| Human_Domainome | 35 |
| MegaScale | 34 |
| FLAb | 18 |
| CombinGym | 3 |
target_dms |
5 |
Biological objectives span activity/function (54 assays), abundance/expression (54),
stability (36), and binding (26). Sequence lengths range from 60 to 1,368 residues,
including multi-chain antibody constructs (heavy/light chains joined by :).
For each assay, an MSA is provided under msa_mmseqs_uniref100/ (188 .a3m files generated with
MMseqs2 against UniRef100; multi-chain assays have one alignment per chain).
Assays outside the benchmark's biosafety and task-scope criteria are excluded.
Considerations for Using the Data
- Ground truth must not be leaked to inference. The benchmark is intended to measure a model's predictions. T1 receives no target mutation measurements; T2 receives candidate identities without target scores. T3 additionally receives its designated anchor measurement, and T4 receives the designated measured single-mutant context. Other target scores must remain hidden from the inference backend. Associated papers may also reveal target measurements and are not default measurement-free inputs.
- Keep the downloaded data unchanged. Candidate tables, reference sequences, and MSAs are frozen; modifying them will invalidate evaluation.
- Score direction.
DMS_scoreis already oriented so that larger is better per assay. Do not invert scores again, re-normalize individual candidate pools, or apply a logarithm to z-scores. Zero is not automatically wild-type performance. uniprot_idmay be blank where no reliable accession could be assigned.
Citation Information
If you find PFArena useful, please kindly cite our paper:
@article{ouyang2026pfarena,
title={PFArena: Benchmarking Language Models for Protein Modification},
author={Ouyang, Yawen and Zhang, Xinbo and Ma, Ziyuan and Wu, Yixin and Liao, Wenbin and Zhang, Feiran and Li, Wenjie and Wang, Lihao and Wang, Hao and Zheng, Xiaoqing and others},
journal={arXiv preprint arXiv:2609.28921},
year={2026}
}
- Downloads last month
- 174