🧬 Protein Structure Refinement & Validation — Full Summary

This lecture explains the final stages of X-ray crystallographic structure determination — how we improve, validate, and judge the quality of a protein model before publication.

Think of this stage as:

🧩 You already built a rough protein model → now you polish, test, and verify if it truly matches the experimental data.

🔧 Structure Refinement — Improving the Model

After building an initial model, refinement aims to:

✅ Minimize the difference between:

• Observed structure factors (experimental diffraction data)
• Calculated structure factors (from the model)

This is the central refinement goal.

📏 Using Chemical Knowledge in Refinement

Refinement is not blind fitting — we use known stereochemistry constraints:

• Bond lengths (C–C, C–O etc.)
• Bond angles
• Torsion angles
• Planarity of peptide bonds
• Amino-acid chirality
• van der Waals radii

These help guide the model toward physically realistic conformations.

⛰️ Local vs Global Minimum Problem

Initial models often get stuck in local minima.

To escape this:

1️⃣ Least-squares optimization

• Adjust parameters gradually
• Move toward lower residual error

2️⃣ Simulated annealing 🔥❄️

(Molecular-dynamics style refinement)

• “Heat” atoms → increase mobility
• “Cool” system → settle into better minimum
• Helps escape incorrect conformations

Goal → reach global minimum = best model.

⚙️ Constraints vs Restraints (VERY exam-important)

These control model complexity vs data amount.

🔒 Constraints

Reduce number of parameters.

Example:

• Instead of one B-factor per atom
• Use one B-factor per residue (group B-factor)

Why?

👉 Low-resolution data → fewer reflections 👉 Too many parameters → overfitting

Constraints prevent over-parameterization.

🧷 Restraints

Allow flexibility but within allowed ranges:

• Bond length intervals
• Angle intervals

Model can move — but not unrealistically.

🌫️ B-factor (Atomic Displacement)

Describes atomic mobility / disorder.

• Low B → rigid atoms → sharp diffraction → high resolution
• High B → flexible atoms → blurred diffraction → low resolution

High B-factors cause:

➡ Faster fall-off of scattering ➡ Poor high-resolution density visibility

Especially important for:

• Flexible proteins
• Loop regions
• Ligands with partial occupancy

📊 R-factor — Core Refinement Statistic

Measures mismatch between data and model.

R = rac{sum |F_ - F_|}{sum F_}

• Perfect model → R = 0 (never achieved)
• Good protein model → R < ~20%

Refinement aims to reduce R continuously.

🧪 R-free — Validation Against Overfitting

Super important concept ⭐

Procedure:

• Randomly remove ~5% reflections
• Do NOT use them in refinement
• Calculate R-free using them

Interpretation:

Difference between Rwork and Rfree ≈ 5% is typical.

📐 Ramachandran Plot — Geometry Validation

Plots φ (phi) vs ψ (psi) torsion angles.

Regions:

🔴 Allowed 🟡 Additional allowed 🟨 Generously allowed ⚪ Disallowed

Good model:

• Majority residues in allowed regions
• Very few in disallowed

Exception:

👉 Catalytic residues may appear strained but real — always check electron density.

🔍 Real Space Correlation Coefficient (RSCC)

Measures how well model density matches observed density.

Good value:

RSCC > 0.9

Low RSCC + High B-factor → poorly defined region Typical example: flexible loops or incorrectly modeled ligands.

💊 Ligand Modeling Issues

Ligands often:

• Have higher B-factors
• Lower occupancy
• Weak density

Reasons:

• Not all binding sites occupied
• Conformational disorder
• Incorrect placement by crystallographer

Contour level matters:

• ~1σ = standard map interpretation
• <0.8σ = risky → may see noise instead of real density

📈 Data Collection Statistics (Tables in Papers)

Typical parameters:

🔢 Measured vs Unique Reflections

• More reflections → higher resolution → more model parameters allowed

🔁 Redundancy (Multiplicity)

ext{Redundancy} = rac{ ext{Measured reflections}}{ ext{Unique reflections}}

Higher redundancy → better precision.

🧩 Completeness

How much of reciprocal space was measured.

• Closer to 100% → better dataset
• Must also be high in highest resolution shell

Otherwise resolution claim is unreliable.

📉 Rsym

Agreement between symmetry-related reflections.

• Lower = better
• Higher tolerated in highest shell (weak data)

🔊 Signal-to-Noise (I/σI)

Rule of thumb:

• Good cutoff ≈ 2
• Modern practice accepts values near 1
• CC½ increasingly used instead.

🌍 Wilson B-factor (Overall Dataset Disorder)

Average B-factor for crystal.

• High Wilson B → low resolution
• Membrane proteins often high (~100 Ų)

Again shows disorder limits resolution.

📐 RMSD Bond Length & Angle

Quality indicator of geometry.

Typical targets:

• Bond length RMSD < 0.02 Å
• Angle RMSD < 4°

At low resolution → strong restraints → artificially small RMSD At high resolution → restraints can be loosened.

💧 Modeling Water Molecules

• Visible only at high resolution
• Often absent at low resolution

Structural waters may still appear even at lower resolution.

🧠 Big Conceptual Takeaway

Protein crystallography workflow ends with:

1️⃣ Build model 2️⃣ Refine model (fit data + chemistry) 3️⃣ Validate model (statistics + geometry + density)

Only after passing all checks → structure is considered reliable.

This lecture essentially teaches:

🧬 A protein structure is not just “solved” — it must be statistically and chemically proven correct.