ELEN_af2_loops_28_1M

The ELEN-Loop dataset comprises approximately 1.5 million protein loop pockets extracted from high-quality, diverse protein crystal structures filtered from the Protein Data Bank (PDB) as of September 2023. Each loop pocket includes the loop residues themselves (2 to 10 residues in length) as well as their immediate structural context, enabling detailed modeling and quality assessment of protein loops within their local environment. The dataset contains: Per-residue structural quality metrics comparing AlphaFold2 predicted models with ground truth crystal structures (including lDDT, CAD-score, and normalized RMSD). Loop pockets curated to maximize sequence diversity and structural quality using stringent filtering criteria (maximum 90% sequence identity, resolution ≤ 2.0 Å, and R-value ≤ 0.25). Data split into training (~1.2 million samples), validation (~200,000 samples), and test (~100,000 samples) subsets for robust benchmarking and model development. Provided in standard .pdb format and LMDB database files for efficient loading and use in machine learning workflows. This dataset supports development and evaluation of protein loop quality assessment models such as ELEN (Equivariant Loop Evaluation Network) and facilitates benchmarking in structural bioinformatics and computational protein design.