Questions: Atomic Model Refinement

Question 1

Why does crystallographic refinement require geometric restraints, and what happens if they are removed?

• A. Restraints are purely cosmetic and can be safely removed at any resolution
• B. At typical protein crystallographic resolution (1.5-3.0 A), the number of independent observations (reflections) is comparable to or fewer than the number of model parameters (x, y, z, B per atom). Without geometric restraints that encode known chemistry (ideal bond lengths, angles, dihedral angles from small-molecule structures), the optimization is underdetermined and the model overfits the data — atomic positions drift to unreasonable values that happen to reduce R-factor but violate basic chemistry
• C. Restraints are only needed for hydrogen atoms
• D. Removing restraints improves the model because the data are perfectly informative

Answer: (B)

The data-to-parameter ratio is the fundamental issue. At 2.0 A resolution, a protein structure has roughly 1 observation per parameter (the ratio improves at higher resolution). Without restraints, the optimization has too much freedom — it can reduce R by moving atoms to positions that fit the noise in the data, producing unrealistic bond lengths and angles. Geometric restraints (derived from high-resolution small-molecule crystal structures where bond lengths and angles are determined to 0.001 A precision) act as additional 'observations' that regularize the optimization, preventing overfitting. At very high resolution (< 1.0 A, where the data-to-parameter ratio exceeds ~5), restraints can be loosened because the data are sufficient to determine atomic positions independently — this is why small-molecule structures are refined with minimal or no restraints.

Question 2

Real-space refinement in COOT and reciprocal-space refinement in REFMAC/PHENIX optimize the same objective function.

• T. True
• F. False

Answer: False

They optimize different but complementary objective functions. Reciprocal-space refinement (REFMAC, phenix.refine) minimizes a target function in reciprocal space — the difference between observed and calculated structure factor amplitudes (|F_obs| - |F_calc|), typically using maximum likelihood targets that weight observations by their estimated uncertainty. It adjusts all atomic parameters simultaneously through gradient-based optimization. Real-space refinement (COOT, phenix.real_space_refine) optimizes the fit of the model to the electron density map in real space — maximizing the correlation between model density and observed density, typically for local regions (a few residues at a time) during manual rebuilding. The crystallographer uses COOT to correct errors (flipped peptides, wrong rotamers, misplaced loops) that reciprocal-space refinement cannot fix because they require large-scale coordinate changes that cross energy barriers in the optimization landscape. The two approaches are complementary: reciprocal-space refinement optimizes globally, real-space rebuilding fixes local errors.

Question 3

Describe the typical iterative workflow of model refinement and explain why multiple rounds are usually required.

Model answer: Refinement proceeds in cycles: (1) reciprocal-space refinement (REFMAC or PHENIX) optimizes coordinates, B-factors, and occupancies against the diffraction data, producing an updated model and improved electron density maps; (2) the crystallographer examines the updated maps in COOT, comparing the model to the density, correcting errors (wrong side chain rotamers, peptide flips, missing or incorrectly placed residues, water molecules, ligands), and rebuilding disordered regions; (3) the manually corrected model goes back into reciprocal-space refinement. Multiple rounds are needed because each round of reciprocal-space refinement produces better phases (and therefore better maps), revealing errors and features not visible in previous maps. Early maps may be too noisy to build side chains confidently; after a few rounds of refinement, the maps improve enough to resolve ambiguities. The process converges when R-free stops decreasing, the maps no longer show features requiring correction, and the model geometry is satisfactory.

A typical refinement might involve 5-20 rounds, depending on the initial model quality and resolution. Automated pipelines (AutoBuild in PHENIX, Buccaneer in CCP4) can handle many rebuilding tasks computationally, but challenging regions (crystal contacts, active sites, ligand binding poses, disordered loops) usually require expert manual intervention in COOT. The quality of the final model depends critically on the crystallographer's skill in map interpretation.

Question 4

What is the difference between an Fo-Fc difference map and a 2Fo-Fc map, and why are both needed during refinement?

Model answer: The 2Fo-Fc map (coefficients 2|F_obs| - |F_calc|, phases from the model) shows the electron density that corresponds to the current model — it reveals the overall shape of the protein and confirms where the model matches the data. The Fo-Fc difference map (coefficients |F_obs| - |F_calc|, phases from the model) shows what the model gets wrong — positive peaks indicate density present in the data but not in the model (missing atoms, unmodeled ligands, water molecules), and negative peaks indicate model features not supported by the data (incorrectly placed atoms, wrong conformations). Both are needed because the 2Fo-Fc map guides overall building (confirming correct placement) while the Fo-Fc map diagnoses errors and reveals missing features. A crystallographer examines both simultaneously in COOT: the 2Fo-Fc map at 1.0 sigma contour to see the overall density, and the Fo-Fc map at +/- 3.0 sigma to spot errors and additions.

The Fo-Fc map is particularly important for identifying ligands and water molecules: a strong positive Fo-Fc peak (> 3 sigma) in the active site after refining the protein alone indicates a bound molecule. The shape of this peak, combined with the biochemical context, guides ligand identification and placement. Many errors in published structures stem from insufficient attention to Fo-Fc maps during refinement.