Hauser7733/ProteinGym-MCQ-v0.1

ProteinGym-MCQ: LLM Reasoning over Protein Variant Effects

Multiple-choice reformulation of the ProteinGym v1 benchmark for evaluating large language models' ability to reason about the functional effects of protein mutations.

Overview

Sub-task	# assays / proteins	# MCQ items	Format
DMS Substitution MCQ	217 assays	696,311	3-way
DMS Indel MCQ	66 assays	287,207	3-way
Clinical Substitution MCQ	2,525 proteins	62,727	binary
Clinical Indel MCQ	858 proteins	2,039	binary
Total		1,048,284

Schema

Each row is a single MCQ item with the following flat fields:

Field	Type	Description
dms_id	str	ProteinGym DMS_id (empty for Clinical)
protein_id	str	RefSeq/UniProt ID (empty for DMS)
molecule_name	str	Human-readable protein name
fitness_type	str	Activity / Stability / Binding / Expression / OrganismalFitness
selection_assay	str	Original assay description
source_organism	str	Source organism
target_seq	str	Wild-type protein sequence
mutated_seq	str	Variant sequence
mutant	str	Mutation notation (L123G for sub, N/A for indel)
mutations_json	str (JSON)	Parsed mutation structure
options_json	str (JSON)	Shuffled MCQ options with labels
is_indel	bool	True for insertion/deletion items
dms_score	float	Continuous fitness score (Clinical: 0.0)
answer	str	Correct letter (A/B/C)
label	str	Gold label (increases / decreases / neutral / Pathogenic / Benign)
task_type	str	T1_variant_level or clinical_binary

Construction

Data is built from ProteinGym v1 on HuggingFace (OATML-Markslab/ProteinGym_v1) joined with per-assay metadata from the ProteinGym GitHub reference CSVs.

DMS three-way partition

Variants are labeled as increases / decreases / neutral using each assay's own DMS_binarization_cutoff (provided by ProteinGym authors) plus a margin of 0.3 × std(DMS_score):

• DMS_score > cutoff + margin → increases
• DMS_score < cutoff - margin → decreases
• otherwise → neutral

Single-point substitutions only (multi-site variants filtered out).

Clinical binary classification

Variants with annotation ∈ {Pathogenic, Benign} are kept as-is (no re-labeling).

Reproducibility

The build script (scripts/build_mcq_dataset.py) and configs (config/) are part of this repository. Same config + same seed (random_seed: 42) produces identical output.

python scripts/build_mcq_dataset.py \
    --config config/full_config.yaml \
    --output ./data_raw \
    --ref-dir ./reference_files

Reference CSVs are auto-downloaded on first run.

Data split provenance

ProteinGym maintains two parallel evaluation tracks: a zero-shot track (models score all variants without fine-tuning) and a supervised track (models are trained on per-assay 5-fold CV splits). This dataset is built from the zero-shot track only — all variants from OATML-Markslab/ProteinGym_v1 are included as test data. The supervised CV folds (cv_folds_*.zip on the ProteinGym website) are not used and do not apply here, because MCQ evaluation targets zero-shot LLMs that never train on DMS fitness labels.

Filtering applied to the upstream data:

• DMS substitutions: multi-site variants (mutant notation containing :) are excluded; 696,311 / 2,465,767 single-point variants retained.
• DMS indels: all 287,207 variants retained.
• Clinical substitutions: all 62,727 Pathogenic/Benign annotations retained.
• Clinical indels: 839 rows with annotation "0" (neither Pathogenic nor Benign) excluded; 2,039 / 2,878 retained.

Citation

If you use this MCQ reformulation, please cite the original ProteinGym:

@inproceedings{notin2023proteingym,
  title={ProteinGym: Large-Scale Benchmarks for Protein Design and Fitness Prediction},
  author={Notin, Pascal and others},
  booktitle={NeurIPS Datasets and Benchmarks},
  year={2023}
}