swhitfield/proteingym-dms_substitutions

ProteinGym dms_substitutions

ProteinGym v1.3 DMS substitution assays, normalized for use by the evaluation workflows in this repository. Each row represents one scored variant from one assay.

Provenance

• Reference metadata: DMS_substitutions.csv
• Assay archive: DMS_ProteinGym_substitutions.zip
• Project: OATML-Markslab/ProteinGym
• Assays in this track: 217
• Variants in the combined artifact: 2465767
• Card generated (UTC): 2026-09-28

Data Files

• data/<dms_id>.parquet: one normalized parquet file per assay.
• data/combined.parquet: all normalized assays concatenated.
• data/combined_<split>.parquet: canonical seeded and pooled random/stratified split views.
• reference.parquet: assay metadata including UniProt IDs, selection type, and wild-type sequences.
• stats.json at the dataset root: row counts, columns, and SeqKit sequence-length diagnostics.

Intended use

Variant-effect prediction and assay-level functional assessment. Scores support ranking or classification within an assay.

Schema

Every assay file contains id, dms_id, and mutated_sequence. Normalized score columns present: dms_score, dms_score_bin. Optional columns present: mutant.

Processing

• combine_assays: true.
• create_split_subsets: true.

Splits

When enabled, each assay receives deterministic train, validation, and test labels using seed 1957723 with fractions 0.8, 0.1, and the remaining rows. The generated pooled *_random and *_stratified views are alternative evaluation partitions and may overlap in rows; they are not additional independent data. MMseqs clustering is not used because variants within an assay share one wild-type sequence. Assays with fewer than three variants are retained without split labels and are excluded from pooled split views.

Loading

These are ordinary parquet files rather than a Hugging Face DatasetDict with standard train/validation/test shards. Load an assay or a named combined view explicitly:

import polars as pl

assay = pl.read_parquet("data/<dms_id>.parquet")
reference = pl.read_parquet("reference.parquet")

Limitations

ProteinGym scores and labels are assay-specific and should not be interpreted as directly comparable measurements across assays. The synthetic split labels are provided for repository evaluation and do not replace ProteinGym's official fold columns when those are present.