engAbdullahHanif/mutascope-proteingym-v1.3

MutaScope: ProteinGym v1.3 single mutants

Every single amino-acid substitution measured in the 217 deep mutational scanning (DMS) assays of ProteinGym v1.3, cleaned and validated, plus per-residue structure features and Foldseek 3Di tokens for the assays' AlphaFold2 structures. It is the processed data behind MutaScope, a project that predicts and benchmarks the functional effect of protein mutations.

Scores are functional effects measured in laboratory DMS assays (growth, binding, stability, activity, expression). They are not clinical pathogenicity labels and must not be used as such.

At a glance

Assays	217
Proteins (distinct UniProt IDs)	186
Single-mutant variants	696,311
Assays with structure features and 3Di	213
Residues with structure features	77,338
Target sequence length	37 to 3,423 (median 245)
Single mutants per assay	33 to 16,897 (median 1,340)

Function type (coarse_selection_type)	Assays	Proteins	Single mutants
OrganismalFitness	77	64	375,411
Stability	66	66	68,855
Activity	43	40	145,367
Expression	18	18	74,923
Binding	13	12	31,755

The proteins come from humans (96 assays), prokaryotes (50), other eukaryotes (40) and viruses (31).

Source and version

• Release: ProteinGym v1.3, Zenodo 10.5281/zenodo.15293562 (published 2025-04-27), git tag PG_v1.3 = commit 1f8de974dead8ff7501eff087b725d14a965e9f9 of OATML-Markslab/ProteinGym.
• Files used: DMS_substitutions.csv (one row per assay), DMS_ProteinGym_substitutions.zip (the measurements) and ProteinGym_AF2_structures.zip (AlphaFold2 models matched to each assay's target sequence; not AlphaFold DB entries).
• Integrity: each file was checked against the md5 published on Zenodo on first download; its sha256 is committed in MutaScope's data/manifest.yaml and verified on every later run. The same checksums are in this dataset's provenance.json.
• Underlying measurements come from the original publications of each assay (author, year and journal are in the assays table). ProteinGym standardised them; we did not re-derive them.

Contents

Four tables, one Parquet file each. assay_id is ProteinGym's DMS_id and joins every table. Positions are 0-based indices into the assay's target_seq; mutation strings keep ProteinGym's 1-based numbering (A123G is position 122).

variants (default config): one row per (assay, single mutant), 696,311 rows.

Column	Type	Meaning
assay_id	string	ProteinGym assay id
mutant	string	Substitution, 1-based, e.g. H24C
position	int	0-based position in target_seq
wt_aa, mt_aa	string	Wild-type and mutant residue (20 canonical amino acids)
DMS_score	float	Measured effect, as published by ProteinGym. Higher = fitter. Units differ per assay
DMS_score_bin	int	ProteinGym's binary label: 1 = fit, 0 = not fit (see limitations)

assays: one row per assay, 217 rows. A subset of ProteinGym's reference file: UniProt_ID, taxon, source_organism, molecule_name, target_seq, seq_len, coarse_selection_type (function type), selection_assay, selection_type, DMS_number_single_mutants, DMS_binarization_cutoff, DMS_binarization_method (median or manual), first_author, year, jo (journal), pdb_file, pdb_range (1-based, inclusive), plus has_structure (false for the 4 assays without structure features, below).

structure_features: one row per (assay, residue covered by its structure), 77,338 rows.

Column	Meaning
position, residue	0-based position and its residue in target_seq
plddt	AlphaFold2 confidence, 0 to 100
sasa	Solvent-accessible surface area of the residue's heavy atoms, Ų (Shrake-Rupley, 1.4 Šprobe)
rel_sasa	sasa over the residue's theoretical maximum; not clipped, so exposed termini can exceed 1
contact_count	Other residues whose Cβ (Cα for glycine) is within 8 Å of this residue's Cβ, in the whole chain

three_di: one row per (assay, residue), 77,338 rows: position, residue and three_di, the Foldseek 3Di structural state (20 letters), from Foldseek 10-941cd33.

supervised_landscapes (added in the Phase 3 revision): every single substitution (19 × L) of the 198 assays with a structure and at least 300 measured variants, 1,379,989 rows, predicted by MutaScope's per-assay ridge model (model repo engAbdullahHanif/mutascope-ridge-seq-struct, whose card describes the model and its evaluation). These are predictions, not measurements, except dms_score.

Column	Meaning
assay_id, position, wt_aa, mt_aa	The substitution (0-based position)
position_status	measured (all 19 substitutions measured), partially_measured or unmeasured
measured, dms_score	Whether this substitution was measured, and its DMS_score if so
prediction	Predicted DMS_score. Out-of-fold for a measured variant (from a model that never saw it), the final model's for an unmeasured one
prediction_source	out_of_fold or final_model
calibration_scheme	random if the position has (other) measured variants, else contiguous: which fit and which calibration folds the row comes from
lower_80, upper_80, lower_90, upper_90	Constant-width CV+ prediction intervals, DMS_score units
structure_imputed	A structure feature was missing at this residue (outside a partial structure) and set to its mean over measured variants
code_version, config_sha256, manifest_sha256	What produced the row: git commit, resolved-config hash, input-manifest hash

Of the 1,379,989 rows, 663,263 are measured; 585,637 are at positions with no measurement, where the model extrapolates (aggregate Spearman 0.553 on ProteinGym's contiguous split, against 0.721 on positions seen in training). The intervals were checked on held-out folds: 0.908–0.911 mean coverage at 90%, 0.815–0.818 at 80%.

supervised_landscape_assays: one row per ProteinGym assay (217): eligible, reason (why not: no structure, 4 assays; fewer than 300 measured variants, 15), default_model (what MutaScope shows: the ridge, or SaProt 650M / ESM-2 650M zero-shot), has_structure, n_measured, target_length, landscape_rows and the same three version columns.

Also included: reports/ (per-assay cleaning and structure reports, Parquet and Markdown), provenance.json (release pin, input checksums, Foldseek version, row counts) and SHA256SUMS.

How it was built

All steps are commands of the mutascope CLI, chained by a Nextflow pipeline (make pipeline). The rules are recorded as architecture decision records in the repository.

1. Validate (ADR 0004). Every one of the 2,465,767 raw rows is checked three ways: the row matches a schema, the stated wild-type residue is at that position of target_seq, and mutated_sequence equals target_seq with the substitution applied. All rows passed. A mismatch would stop the run with its assay and row; nothing is ever silently fixed or dropped.

2. Clean (ADR 0005). Each removed row is counted under exactly one reason, the first that applies:

   Reason	Rows removed
   Multi-mutant (v1 keeps single substitutions only), in 69 assays	1,769,456
   Missing mutant, mutated_sequence, DMS_score or DMS_score_bin	0
   Exact duplicate (first copy kept)	0
   Same mutant with conflicting values (no copy kept)	0
   Single mutants kept	696,311

   Scores are parsed with correctly rounded decimal conversion, so every stored DMS_score equals the published text exactly. DMS_score is never sign-flipped. Per assay, the kept count equals ProteinGym's DMS_number_single_mutants; the publish step re-checks this.

3. Structures (ADRs 0003, 0006). Each assay's AF2 model is matched to its target_seq over pdb_range and checked residue by residue. 213 assays (185 PDB files) match. Four do not, because of upstream data, and get no structure rows; they stay in variants:

   • CAS9_STRP1_Spencer_2017_positive: the PDB carries a 22-residue C-terminal tag, so pdb_range 1-1390 runs past the 1,368-residue target.
   • P53_HUMAN_Giacomelli_2018_Null_Etoposide, _Null_Nutlin, _WT_Nutlin: the target has R72, the shared PDB the natural P72 variant.

   For the 213 matched assays, every variant falls at a position that has structure features (665,793 of 665,793). Secondary structure is deliberately not included (ADR 0006).

4. Write. Every table is validated against a pandera schema before it is written. Staging the Hub upload twice produces byte-identical files.

Intended use

• Benchmarking and training variant-effect predictors on single substitutions, following ProteinGym's protocols.
• Exploring mutational landscapes alongside structure (pLDDT, burial, contacts, 3Di).
• Structure-aware models such as SaProt, which take the 3Di tokens directly (mask 3Di where plddt < 70, as SaProt does).

Out of scope: clinical interpretation of human variants; multi-mutant or indel effects; any comparison of raw DMS_score values across assays.

Known limitations

• Scores are not comparable across assays. Each assay has its own units and dynamic range (enrichment ratios, ΔΔG, growth rates). Standardise within an assay or use rank metrics. ProteinGym reports Spearman per assay, averaged per protein, then per function type.
• Assay sizes vary 500-fold (33 to 16,897 single mutants), so per-assay metrics are noisy for small assays. 5 assays have fewer than 100 single mutants.
• Coverage is uneven. Targets of up to 1,022 residues have a median of 90% of all 19 × L possible substitutions measured; the 16 longer targets have a median of 4.6%. Those 16 also exceed ESM-2's 1,022-residue context and need sliding windows.
• One study dominates stability. 64 of the 66 stability assays come from a single high-throughput study (Tsuboyama et al., 2023) of small domains of 37 to 72 residues. Stability results mostly reflect that method and protein size range.
• Binary labels inherit ProteinGym's cutoffs, set per assay either at the median score (126 assays) or manually (91). Where the median was taken over multi-mutants as well, the single mutants, which are usually fitter, are skewed towards label 1. Of the 43 median-binarized assays with multi-mutants, 32 have a share of single mutants labelled 1 more than 10 points away from 50% (median share 67%). The 83 without multi-mutants range from 45% to 53%. In F7YBW8_MESOW_Aakre_2015 the cutoff (−0.0017) is the median over 9,192 rows, 9,155 of them multi-mutants, and all 37 single mutants lie above it, so every label is 1. AUC and MCC are undefined for that assay.
• Structures are predictions, not experiments. 23% of residues with structure features have pLDDT below 70, and 18 of the 213 structures have a mean pLDDT below 70 (lowest: BRCA1_HUMAN_Findlay_2018, 41). Treat SASA, contacts and 3Di in those regions with caution. Three structures cover only part of their target (A0A140D2T1_ZIKV_Sourisseau_2019, BRCA2_HUMAN_Erwood_2022_HEK293T, POLG_HCVJF_Qi_2014), though all of their measured variants fall inside the modelled range.
• Four assays have no structure features (above), so structure-based models are evaluated on 213 assays and sequence-only models on 217. Results should say which.
• Assays share proteins. 217 assays cover 186 UniProt IDs (up to 4 assays per protein), so splitting by assay does not separate proteins.
• Single substitutions only. Multi-mutants, insertions and deletions are not included.
• supervised_landscapes are predictions. Use dms_score for measurements. Predictions at unmeasured positions are extrapolations with wide intervals (median 90% width 2.7 assay standard deviations); for A0A140D2T1_ZIKV_Sourisseau_2019 and POLG_HCVJF_Qi_2014 most residues lie outside the structure (structure_imputed). Intervals hold their level on average, not per variant. Never use them as clinical predictions.
• Organism and selection bias. 96 of 217 assays are of human proteins, and the set reflects which proteins and functions DMS studies chose to measure. It is not a random sample of proteins or functions.

Licence and citation

ProteinGym v1.3 is released under the MIT licence (Zenodo record and GitHub repository), and this derived dataset is released under the same MIT licence. The underlying measurements come from the cited original studies; please cite them as well when you use specific assays.

If you use this dataset, cite ProteinGym:

@inproceedings{notin2023proteingym,
  author    = {Notin, Pascal and Kollasch, Aaron and Ritter, Daniel and van Niekerk, Lood and
               Paul, Steffanie and Spinner, Han and Rollins, Nathan and Shaw, Ada and
               Orenbuch, Rose and Weitzman, Ruben and Frazer, Jonathan and Dias, Mafalda and
               Franceschi, Dinko and Gal, Yarin and Marks, Debora},
  title     = {ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {36},
  pages     = {64331--64379},
  year      = {2023}
}

and, for the 3Di tokens, Foldseek (van Kempen et al., Nature Biotechnology 2024).

Loading

from datasets import load_dataset

variants = load_dataset("engAbdullahHanif/mutascope-proteingym-v1.3", "variants", split="train")
structure = load_dataset(
    "engAbdullahHanif/mutascope-proteingym-v1.3", "structure_features", split="train"
)

landscapes = load_dataset(
    "engAbdullahHanif/mutascope-proteingym-v1.3", "supervised_landscapes", split="train"
)

-- DuckDB, straight from the Hub
SELECT v.*, s.plddt, s.rel_sasa
FROM 'hf://datasets/engAbdullahHanif/mutascope-proteingym-v1.3/data/variants.parquet' v
LEFT JOIN 'hf://datasets/engAbdullahHanif/mutascope-proteingym-v1.3/data/structure_features.parquet' s
  USING (assay_id, position);

Reproducing

git clone https://github.com/EngAbdullahHanif/mutascope && cd mutascope
make setup
make pipeline                # download, verify, validate, clean, structure features, 3Di
uv run mutascope data publish  # stage this dataset in build/hf_dataset (no upload)
# with the supervised landscapes (Phase 3: train cv, train calibrate, train final first):
uv run mutascope data publish --landscapes

provenance.json records the inputs and versions; SHA256SUMS lets you compare a rebuilt copy file by file.