Developing SCL2205: a protein sequence-based spatial modelling dataset for the protein language model frontier.

motivationDeep learning (DL) has substantially advanced protein subcellular localization (SCL) prediction, yet its potential remains constrained by suboptimal input preparation and limited high-quality reference data. Furthermore, existing state-of-the-art (SoTA) predictors suffer from performance metric inflation due to unmitigated training-to-testing data leakage during homology augmentation. We address these challenges by introducing SCL2205, a leak-minimized benchmark dataset and pipeline curated specifically to support trustworthy, scalable, and reproducible DL-based SCL modelling.

resultsSCL2205 was constructed from the universal protein knowledgebase (UniProtKB) using rigorous preprocessing, manual label mapping, and stringent partitioning. When evaluated on independent test sets, SCL2205 yielded up to a 10.8 percentage point improvement in macro area under the precision-recall curve (PR-AUC) over SoTA baselines (mean Δ 95% CI = 0.07-0.12), with maximum benefits observed when paired with modern protein language models (PLMs). Crucially, we quantify for the first time a systemic 5.2%±0.32 data leakage rate in conventional homology augmentation workflows-even when restricting sequence similarity searches to just 10% of the training set. AVAILABILITY AND IMPLEMENTATION: The dataset is openly available on Dryad under a CC0 1.0 Universal licence (https://doi.org/10.5061/dryad.2ngf1vj1t). The dataset interface is available as an installable Python package, p-scldata (v2026.2.0), under the MIT licence on the Python Package Index (PyPI). Full code and data repositories are hosted on GitHub (https://github.com/ousodaniel/scldata) and archived on Zenodo (https://doi.org/10.5281/zenodo.21796423).