Experimental structures: X-ray, NMR and cryo-EM

Every predicted model is ultimately judged against experiment, so an advanced practitioner must read experimental structures critically rather than trusting them wholesale. Three methods dominate the Protein Data Bank: X-ray crystallography, solution NMR, and cryo-electron microscopy. Each reports different quality signals, and knowing which numbers to trust tells you how much weight a given coordinate actually deserves.

X-ray crystallography

X-ray crystallography fits atoms into an electron-density map reconstructed from crystal diffraction, and resolution in angstroms sets the detail, with sub-2.0 angstrom maps resolving side chains while 3.0 angstrom and beyond blurs them. The R-factor measures agreement between the model and the observed amplitudes, and R-free computes the same on a held-out roughly 5 percent of reflections never used in refinement, so a large R-free minus R-work gap signals overfitting. Per-atom B-factors, the isotropic atomic displacement parameters in square angstroms, capture positional uncertainty from thermal motion and static disorder, which makes them a useful within-structure confidence proxy.

NMR and cryo-EM

Solution NMR infers geometry from experimental restraints, namely interproton distances from NOEs, backbone dihedral angles from J-couplings, and bond-vector orientations from RDCs, rather than from a density map, so it has no single resolution value and entries are deposited as an ensemble, typically 10 to 20 models, whose spread reflects both genuine flexibility and restraint ambiguity. Cryo-EM images many single particles and averages them into a 3D density, and the resolution revolution around 2013, driven by direct electron detectors and better software, pushed it from roughly 8 angstrom blobs to near-atomic maps. Global cryo-EM resolution is reported from the Fourier shell correlation at the 0.143 gold-standard cutoff, but resolution varies locally, so a well-ordered core and a fuzzy periphery can share one headline number.

PDB and file formats

All three methods feed the wwPDB archive, mirrored by RCSB, PDBe, and PDBj. The legacy PDB format is a fixed 80-column text layout with hard limits, notably 99,999 atoms and single-character chain IDs, that large complexes outgrew, so mmCIF has been the canonical default format since 2014, using extensible category.item key-value pairs. Despite hundreds of thousands of entries, the PDB still covers only a fraction of known sequence space, and membrane proteins, large assemblies, and disordered regions remain underrepresented, which is exactly the coverage gap that structure prediction sets out to fill.

# Fetch the same entry as legacy PDB and as modern mmCIF
curl -sO https://files.rcsb.org/download/6LU7.pdb
curl -sO https://files.rcsb.org/download/6LU7.cif

# Method, resolution and refinement R-factors from the PDB header
grep -E '^EXPDTA|^REMARK   2 RESOLUTION|R VALUE +\(WORKING SET\)|FREE R VALUE +:' 6LU7.pdb

EXPDTA    X-RAY DIFFRACTION
REMARK   2 RESOLUTION.    2.16 ANGSTROMS.
REMARK   3   R VALUE            (WORKING SET) : 0.202
REMARK   3   FREE R VALUE                     : 0.235

# Columns 61-66 of each ATOM record hold the isotropic B-factor in square angstroms
# (the same value lives in _atom_site.B_iso_or_equiv in the mmCIF file)
awk '/^ATOM/ {b=substr($0,61,6)+0; sum+=b; n++; if(b>max)max=b} END {printf "atoms=%d  mean_B=%.2f  max_B=%.2f\n", n, sum/n, max}' 6LU7.pdb

atoms=2387  mean_B=42.68  max_B=89.99