PDB (Protein Data Bank)

The PDB exclusively archives experimentally determined structures derived from three primary methods: X-ray crystallography (providing high-resolution electron density maps), NMR spectroscopy (capturing dynamic ensembles in solution), and 3D cryo-electron microscopy (resolving large macromolecular complexes in near-native states). Each entry is assigned a unique 4-character PDB ID and includes the atomic coordinates, structure factors, and extensive metadata describing sample preparation, data collection, and refinement statistics. This experimental grounding makes the PDB the gold-standard benchmark for training and validating AI models like AlphaFold2 and RoseTTAFold.

The PDB is managed by the Worldwide Protein Data Bank (wwPDB) consortium, comprising founding members RCSB PDB (USA), PDBe (Europe), PDBj (Japan), and BMRB (USA). This federated governance model ensures:

• Unified data standards and format compliance across all deposition sites
• Continuous data curation and remediation of legacy entries
• Global mirroring for high-availability access
• Community-defined validation via wwPDB Validation Task Forces The consortium enforces strict deposition policies requiring coordinate release upon publication, ensuring the archive remains comprehensive and timely.

The legacy PDB file format (column-fixed, 80-character records) has been officially deprecated in favor of PDBx/mmCIF (macromolecular Crystallographic Information File). This extensible, dictionary-driven framework:

• Supports large complexes exceeding the legacy format's 99,999-atom limit
• Captures multi-scale metadata including assembly information, symmetry operators, and biological context
• Enables machine-readable validation reports with structured quality metrics
• Provides ontological mappings to external resources like UniProt and Pfam All modern structure prediction tools consume mmCIF natively for training pipelines.

Every PDB entry undergoes rigorous automated validation upon deposition, generating a standardized report with key geometric and fit-to-data metrics:

• Ramachandran plot statistics identifying backbone dihedral angle outliers
• Clashscore quantifying steric overlaps per 1,000 atoms
• RSRZ (Real-Space R-value Z-score) measuring per-residue fit to electron density
• Rotamer normality assessing side-chain conformation quality These metrics directly inform the pLDDT and PAE confidence scores used in AI-predicted structures, providing a quantitative bridge between experimental and computational quality assessment.

The PDB provides comprehensive RESTful APIs and bulk data services essential for training large-scale machine learning models:

• Search API with complex query operators for sequence, structure, and metadata filtering
• GraphQL endpoint enabling precise field selection and reduced payload sizes
• Weekly snapshot downloads of the entire archive in mmCIF, PDB, and PDBML/XML formats
• Customized cluster datasets at 30%, 40%, 50%, 70%, 90%, and 95% sequence identity thresholds These pre-computed sequence clusters are critical for creating non-redundant training/validation/test splits that prevent data leakage in structure prediction benchmarks.

The asymmetric unit deposited in the PDB often represents only a fraction of the functional biological molecule. The PDB provides computed biological assemblies generated by applying crystallographic and non-crystallographic symmetry operations to reconstruct the oligomeric state observed in vivo. This is critical for:

• Training multi-chain structure prediction models on correct quaternary contacts
• Evaluating protein-protein docking algorithms against physiological interfaces
• Understanding allosteric regulation and binding cooperativity AI models like AlphaFold-Multimer rely on these biologically relevant assemblies to learn inter-chain interaction patterns.

The macromolecular Crystallographic Information File is the current PDB archive standard, replacing the legacy PDB format. It uses a tag-value structure organized into categories and supports large, multi-chain complexes (>62 chains, >99,999 atoms) that the legacy column-based format could not handle. Key features include:

• Hierarchical data model with explicit relationships between entities
• Unlimited atom records for massive assemblies like ribosomes and viral capsids
• Rich metadata capturing experimental details, refinement statistics, and biological assembly information
• Programmatic parsing via libraries like GEMMI and BioJava All modern structure prediction models, including AlphaFold2, output coordinates in mmCIF format.

The RCSB PDB Search API provides programmatic access to the entire PDB archive through a powerful query language supporting attribute-based, sequence-based, and structure-based searches. Capabilities include:

• GraphQL and REST endpoints for flexible data retrieval
• SMILES and InChI substructure searches for ligand-bound structures
• Sequence similarity queries using BLAST or MMseqs2 against all deposited chains
• Batch downloads of custom datasets filtered by resolution, organism, or experimental method
• Real-time access to newly released entries through the weekly update cycle This API is the primary data pipeline for training large-scale protein structure prediction models.

Every PDB entry undergoes rigorous automated validation by the worldwide Protein Data Bank consortium, producing standardized reports that assess structural quality using multiple orthogonal metrics:

• Clashscore: Number of serious steric overlaps per 1,000 atoms, with lower scores indicating better geometry
• Ramachandran outliers: Percentage of residues in disallowed phi/psi backbone conformations
• Rotamer outliers: Side-chain conformations deviating from preferred statistical distributions
• RSRZ (Real-Space R-value Z-score): Per-residue fit of the model to experimental electron density
• Bond length and angle RMSDs: Deviation from ideal covalent geometry targets These validation metrics are critical for filtering high-quality training data for machine learning models.

A critical distinction in PDB data interpretation: the asymmetric unit is the smallest portion of a crystal structure that can generate the complete unit cell through symmetry operations, while the biological assembly is the functionally relevant oligomeric state. Key considerations:

• Crystal packing artifacts: Interfaces present in the asymmetric unit may not exist in solution
• PISA (Proteins, Interfaces, Structures and Assemblies): Algorithm that predicts the likely quaternary structure
• Author-assigned vs. software-predicted assemblies: Both are stored in the PDB
• AlphaFold-Multimer and similar tools predict biological assemblies directly from sequence Training structure prediction models requires careful curation of biological assembly annotations to avoid learning non-physiological contacts.

The Chemical Component Dictionary (CCD) is the definitive reference database defining all small molecule ligands, cofactors, inhibitors, and modified residues found in PDB entries. Each three-letter residue code maps to:

• Idealized covalent geometry including bond lengths, angles, and torsion parameters
• SMILES, InChI, and IUPAC nomenclature for cross-referencing with cheminformatics databases
• Atom typing and connectivity used by refinement software like Phenix and Refmac
• Dictionary versioning to track corrections and updates to ligand definitions Accurate ligand geometry from the CCD is essential for training molecular docking and drug-target interaction prediction models.

PDB-REDO is a project that systematically re-refines all PDB X-ray structures using modern crystallographic software and consistent protocols, producing geometry-optimized models with improved fit to experimental data. Benefits for machine learning include:

• Reduced noise from inconsistent historical refinement practices across decades of depositions
• Improved side-chain rotamer assignments and hydrogen bonding networks
• Consistent treatment of alternate conformations and disordered regions
• Updated ligand geometries using the latest CCD definitions
• Paired original and re-refined structures enabling quality improvement quantification Many state-of-the-art protein structure prediction models use PDB-REDO as their primary training set to learn from the highest-quality structural data available.