Magneton Protein Representation Learning Dataset

## Overview Magneton is a dataset for training and evaluating substructure-aware protein representation learning models. It contains 530,601 SwissProt proteins with associated substructure annotations from DSSP (secondary structure) and InterPro release 103.0 (domain/family/motif annotations). Data is stored as sharded JSONL files with 10,000 proteins per file, designed to be used alongside the Magneton training/evaluation codebase on GitHub. ## Schema - protein_id — string — SwissProt accession identifier - sequence — string — amino acid sequence - dssp_annotations — array/object — per-residue DSSP secondary structure assignments - interpro_annotations — array/object — InterPro 103.0 domain/family/motif spans with IDs - additional fields for substructure metadata as defined by the source ## Sources - HuggingFace: https://huggingface.co/datasets/rcalef/magneton-data — MIT license - Underlying data: UniProt/SwissProt, DSSP, InterPro 103.0 ## Methodology The dataset was constructed by taking SwissProt proteins and annotating them with two complementary substructure sources: DSSP (Dictionary of Secondary Structure of Proteins) for per-residue secondary structure assignments derived from 3D structures, and InterPro release 103.0 for domain, family, and motif annotations integrated from member databases. Proteins are sharded into JSONL files of 10,000 records each for efficient streaming and parallel processing. ## Known gaps & limitations Coverage is limited to SwissProt (curated UniProt subset) — TrEMBL and unreviewed proteins are not included. DSSP annotations require an experimentally determined or predicted structure, so coverage may be partial across the 530k proteins. InterPro 103.0 is a fixed snapshot; newer family/domain definitions are not reflected. Source does not document additional gaps in detail; buyers should validate empirically for their use case. ## Intended use & out-of-scope - IS for: training/evaluating substructure-aware protein language models, protein representation learning, structure-conditioned pretraining, substructure prediction benchmarks. - NOT for: clinical or diagnostic use; not a complete proteome (SwissProt-only); not deduplicated against downstream evaluation splits — leakage risk if reused with overlapping benchmarks. _Federated dataset: 3 parquet shards, 1.19 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: Magneton: Substructure-Aware Protein Representation Learning Dataset 530,601 SwissProt proteins with DSSP secondary structure and InterPro 103.0 substructure annotations, sharded JSONL format. For training/evaluating protein representation learning models.