MPLID

MPLID curates a large-scale, experimentally validated dataset of membrane protein-lipid contact residues derived from PDB crystal and cryo-EM structures for machine learning applications.

Key Features:

• 100% Experimental Labels: All labels derive solely from experimentally resolved lipids in PDB structures with no computational database dependencies.
• Residue-Level Annotations: The dataset provides 80,685 residue-level lipid contact annotations across 4,712 membrane proteins (v2.0.0: 80,439 contact residues across 4,704 proteins).
• Binary and Continuous Labels: Pre-computed binary contact labels and continuous distance values are included using a 4.0 Angstrom all-atom heavy-atom distance cutoff.
• Sequence-Identity-Based Clusters: Proteins are grouped into sequence clusters (813 clusters at 30% identity in v2.0.0) to enable cluster-aware data splitting.
• Ready-Made Splits: Train-validation-test splits are provided to prevent data leakage, with v2.0.0 splits comprising 2,578/1,051/1,075 proteins respectively.
• Broad Lipid Coverage: The dataset recognizes 117 distinct lipid codes.
• Full Reproducibility: The dataset is fully reproducible from public PDB data.

Scientific Applications:

• Machine Learning Model Development: Ready-made train-validation-test splits enable development and benchmarking of predictive models for membrane protein-lipid contacts.
• Membrane Protein Structural Analysis: Large-scale annotation of lipid contact residues across all membrane protein classes supports systematic structural biology studies.

Methodology:

Lipid contact residues were identified from experimentally determined structures in the Protein Data Bank using a 4.0 Angstrom all-atom heavy-atom distance cutoff, with proteins clustered at 30% sequence identity to generate non-redundant train-validation-test splits.