Protein Structure Cleaning
Protein structure cleaning converts raw PDB/mmCIF records into model-ready structure inputs. It should be treated as part of the data protocol, not as invisible preprocessing.
A cleaning pipeline can be written as:
where includes chain selection, residue mapping, alternate-location handling, missing-residue masks, heteroatom policy, and coordinate normalization.
Core Decisions
- Chain selection: which biological chain or assembly is used?
- AltLoc handling: which alternate conformation is kept?
- Missing residues: are gaps masked, modeled, or removed?
- Non-standard residues: are they mapped to canonical residues or marked unknown?
- Heteroatoms: are waters, ions, cofactors, glycans, and ligands retained?
- Hydrogen atoms: are hydrogens added, stripped, or ignored?
- Coordinate units and frame: are coordinates in Å and in the intended frame?
Structure Identity
A cleaned structure should keep a stable identity key:
This matters for feature caching, split checks, and reproducibility.
Interaction and Pocket Tasks
For structure-based tasks, cleaning choices can change the binding site:
- stripping a cofactor can remove a real interaction partner
- keeping a crystallographic ligand can leak the binding site
- removing waters can remove water-mediated contacts
- changing protonation can change hydrogen-bond patterns
- selecting a ligand-defined pocket can leak test information if not available at deployment
These choices should be documented in public notes at the protocol level without exposing private target details.
Checks
- Is chain selection explicit?
- Are missing residues and chain breaks represented?
- Are waters, ions, cofactors, and ligands handled with a stated policy?
- Are non-standard residues mapped consistently?
- Is residue indexing preserved after cleaning?
- Is pocket definition independent of unavailable test-time information?
- Is the cleaned structure hash tied to downstream features?