ELEN_loop_dataset

ELEN-Loop Dataset v2.0 The ELEN-Loop dataset is a large-scale structural biology dataset for machine-learning-based evaluation of protein loop conformations. It contains approximately 1.5 million protein loop pockets extracted from high-quality and sequence-diverse protein crystal structures from the Protein Data Bank (PDB), filtered as of September 2023. Each sample represents a local protein loop environment. It contains a loop segment of 2 to 10 residues together with its immediate three-dimensional structural context, enabling residue-level assessment of loop geometry within the surrounding protein environment. The dataset was developed for training and benchmarking models such as ELEN, the Equivariant Loop Evaluation Network, which predicts local protein structure quality at per-residue resolution. The dataset includes AlphaFold2-based loop models and corresponding reference crystal structures, together with per-residue structural quality labels. These labels include lDDT, CAD-score, and normalized RMSD values, allowing detailed evaluation of local structural accuracy from complementary geometric perspectives. Protein structures were curated using stringent quality and redundancy filters to support robust model development and benchmarking. The underlying crystal structures were filtered for a maximum sequence identity of 90%, resolution of 2.0 Å or better, and R-value of 0.25 or lower. The resulting dataset is split into training, validation, and test subsets, comprising approximately 1.2 million, 200,000, and 100,000 samples, respectively. The release provides both structure files and machine-learning-ready database files. Loop pocket structures are provided in standard PDB format, while LMDB files enable efficient loading in large-scale machine learning workflows. The dataset also includes precomputed per-residue physicochemical features used by ELEN. SaProt sequence embeddings are not included in the public archive because of their size, but they can be made available upon request. This dataset is intended to support research in protein structure quality assessment, loop modelling, structural bioinformatics, geometric deep learning, protein design, and benchmarking of computational methods for local protein structure evaluation. This archive contains: Loop pocket structure files in PDB format. LMDB databases for efficient machine learning workflows. Per-residue structural quality labels comparing AlphaFold2-predicted models against crystal structures, including lDDT, CAD-score, and normalized RMSD. Precomputed physicochemical residue features. Predefined training, validation, and test splits. SaProt sequence embeddings are not included in this archive due to file size constraints, but can be provided upon reasonable request.