README.md

metadata

license: other
task_categories:
  - tabular-classification
language:
  - en
pretty_name: cytolexmuta ProteinGym DMS substitutions zero-shot scores

cytolexmuta

cytolexmuta is a zero-shot map-fusion method for ProteinGym DMS substitutions.

It treats each component model as a complementary protein energy map: sequence-language plausibility, ProSST structure-aware likelihood, evolutionary coupling, and inverse-folding structural compatibility. These maps are projected into a shared assay-local coordinate system, and the released score follows a fixed convex path through that aligned map family.

The exact deterministic scoring rule is implemented in code. Normalization is computed independently within each DMS assay from prediction values only. The score files do not contain DMS_score, DMS_score_bin, labels, targets, or benchmark metrics.

Coverage

• Benchmark: ProteinGym DMS zero-shot substitutions
• Assays: 217
• Variants: 2,465,767
• Per-assay score column: cytolexmuta
• Score direction: higher is better

Frozen performance

• Average Spearman: 0.5347273342803031
• Official rounded score: 0.535
• Activity: 0.524200
• Binding: 0.469375
• Expression: 0.536833
• OrganismalFitness: 0.490910
• Stability: 0.652318

Files

• cytolexmuta_scores_full217.tar.gz: 217 per-assay CSV files, each with columns mutant,cytolexmuta.
• full217_cytolexmuta_scores.csv.gz: combined full217 prediction table with row_index,DMS_id,mutant,cytolexmuta.
• summary.json: package manifest and validation summary.
• METHOD_CARD.md, pr_body.md, issue_body.md: method and ProteinGym submission drafts.

License and upstream note

This dataset is provided for benchmark review and reproducibility. Upstream model/code/data licenses remain with their respective authors. GEMME/JET2 runtime artifacts are not redistributed here.