archiitecture/dynamic-proteins

Protein Conformational Dynamics Embeddings

This dataset contains 8,598 protein sequences with operational dynamic or static labels, provenance, homology-aware frozen train/validation/test assignments, and residue-level embeddings used in the associated study. The class balance is 4,289 dynamic and 4,309 static proteins.

Files

• manifest.csv and manifest.parquet contain the same protein-level records and exact shard/member locations.
• embeddings/esmfold/*.tar contains 8,598 ESMFold folding-trunk embeddings with width 1,024.
• embeddings/bioemu/*.tar contains 8,572 BioEMU AlphaFold2 Evoformer single embeddings with width 384.
• embeddings/bioemu-no-msa/*.tar contains 8,572 query-only BioEMU embeddings with width 384.
• checksums.sha256 records SHA-256 checksums for every published file.

Each tar archive is flat and contains files named by sequence SHA-256. The manifest identifies the archive and member for every available representation. BioEMU is unavailable for 26 proteins, which is recorded explicitly rather than imputed. Randomly initialized trunk embeddings are not part of this release. ESM2 3B embeddings are not included in this initial upload because their local artifacts were unavailable for verification.

Splitting and reuse

Use the provided split and homology_group_id fields for the frozen homology-aware evaluation. Do not make random row-level splits when comparing with the reported results. Labels are operational dataset annotations based on the evidence and screening procedures described by their source provenance, not universal experimental proof of rigidity or motion.

Licensing

The project-authored manifest and annotations are released under CC BY 4.0. Underlying sequences, identifiers, derived embeddings, and source-linked metadata retain the terms and attribution requirements of their originating resources and model licenses. Users are responsible for complying with those upstream terms.

Repository: archiitecture/dynamic-proteins.