Skip to content

Dataset Card — SonoForge seed data (v0.1, Phase 0)

Following the spirit of Gebru et al., Datasheets for Datasets (2021).

Motivation

Provide targets and seeds for de novo design of gas-vesicle acoustic-reporter proteins: GvpA (primary shell protein) and GvpC (scaffolding protein that tunes collapse pressure), plus published acoustic-reporter-gene (ARG) variant panels where available.

Composition (planned, Phase 1)

  • Targets: GvpA / GvpC sequences and structures (RCSB PDB, UniProt).
  • Variant panels: ARG directed-evolution variants with measured acoustic properties, where public.
  • Pretraining corpus: a broad protein set for the PLM (transfer learning onto the sparse GV family).

Collection & provenance

Public databases only (RCSB PDB, UniProt) and published literature. Accessions and licenses are recorded in data/manifest.json at download time.

Preprocessing

Length filtering, deduplication by sequence, and featurization to the Candidate schema (src/sonoforge/data/types.py). Simulated oracle labels are marked as such.

Distribution & licensing

Source structures/sequences retain their upstream licenses. No private or personal data. This repository ships download scripts, not bulk redistributed third-party data.

Known limitations

Public GV/ARG measurement data are sparse and heterogeneous across labs and assays — which is precisely why the platform is built around uncertainty-aware, sample-efficient active learning.