Questions: Structure Validation and Model Quality

Question 1

A crystallographic structure reports an R-factor of 0.20 and an R-free of 0.35. What does the large gap between R and R-free indicate?

• A. The structure is excellent — a low R-factor proves the model is correct
• B. The large R-R-free gap (0.15) suggests overfitting: the model has been adjusted to match the data used in refinement (producing low R) but does not predict the reserved test set well (high R-free). This typically indicates problems like over-refined B-factors, incorrect ligand placement, or an incorrect space group
• C. R-free is always much higher than R — this is normal and expected
• D. The gap indicates the crystal had unusually high mosaicity

Answer: (B)

R-free (Brunger, 1992) is computed on a small subset of reflections (typically 5-10%) excluded from refinement — the model never sees this data during fitting. It serves as a cross-validation metric: if the model captures genuine structural features, R-free should track R closely (typical gap: 0.02-0.05 for well-refined structures). A gap of 0.15 indicates the model has been tuned to match the refinement data in ways that do not generalize — adding too many parameters (high B-factors, alternative conformations, or waters) without genuine signal. This overfitting warning is one of the most important contributions of R-free to structural biology, as R alone can always be lowered by adding parameters regardless of whether they represent real features.

Question 2

A protein structure has 97% of its residues in the favored regions of the Ramachandran plot. This automatically means the structure is correct.

• T. True
• F. False

Answer: False

Good Ramachandran statistics are necessary but not sufficient for a correct structure. The Ramachandran plot checks whether backbone dihedral angles (phi/psi) fall in sterically allowed regions — a knowledge-based validation against known protein geometry. A structure with 97% favored Ramachandran is geometrically reasonable, but it could still have incorrect chain tracing (the backbone follows a plausible path through the wrong part of the density), wrong sequence register (the right fold threaded through the density with a shift in residue numbering), incorrectly placed ligands, or unresolved regions built as a best guess. Conversely, genuine Ramachandran outliers do exist in correct structures — functional sites like cis-prolines and strained active-site residues. Validation requires combining Ramachandran analysis with data-fit metrics (R-free, real-space correlation), clash analysis, rotamer quality, and visual inspection of the electron density.

Question 3

What is the role of real-space correlation coefficient (RSCC) in structure validation, and why is it more informative than global R-factor for evaluating specific regions of a structure?

Model answer: The global R-factor averages the model-data agreement over the entire structure, so a well-ordered core can mask problems in poorly resolved regions (loops, termini, ligands). The real-space correlation coefficient (RSCC) evaluates the agreement between the model and the electron density map at the level of individual residues or ligands. It computes the correlation between the observed electron density and the density calculated from the model atoms in a local region around each residue. An RSCC near 1.0 indicates excellent fit; values below 0.7-0.8 flag residues where the model may not match the density — suggesting incorrect placement, conformational disorder, or insufficient resolution in that region. RSCC is particularly important for validating ligand binding poses in drug design, where an incorrect ligand placement in a globally good structure could lead to entirely wrong mechanistic conclusions.

The wwPDB validation pipeline reports per-residue RSCC values and highlights residues with poor density fit. Ligand validation is especially critical: a 2019 study found that roughly 10% of ligand-bound structures in the PDB have significant problems with the modeled ligand — incorrect stereochemistry, impossible contacts, or density that does not support the deposited binding mode. Tools like EDSTATS and Privateer provide specialized ligand and carbohydrate validation.

Question 4

Why was the introduction of R-free considered one of the most important methodological advances in protein crystallography?

Model answer: Before R-free, the R-factor was the primary measure of model quality, but R can always be reduced by adding parameters to the model — more atoms, alternative conformations, high B-factors, water molecules — regardless of whether these additions represent real structural features. There was no independent check on whether the model genuinely explained the diffraction data versus merely fitting noise. R-free (Brunger, 1992) introduced cross-validation to crystallography by withholding a random subset of reflections from refinement and computing the agreement only against this unseen test set. This immediately exposed overfitting: models with artificially low R but genuinely poor structure showed high R-free. The impact was transformative — it changed how refinement was practiced (favoring restraints and conservative parameterization), how structures were evaluated by journals and the PDB, and how the community detected errors in published structures.

R-free fundamentally changed the culture of crystallographic refinement from R-factor minimization to genuine model improvement. Before R-free, some practitioners would add waters and alternative conformations aggressively to reduce R, producing models that appeared good by R but contained many spurious features. R-free penalizes overfitting, aligning the statistical incentive with the scientific goal of an accurate model.