How to Validate a Predicted Protein Structure Before Using It in Research

AlphaFold embeds its own uncertainty estimate in every prediction. These confidence metrics are the fastest and most direct information available about which regions can be trusted — and they’re available for free alongside every structure from the AlphaFold Database and ColabFold.

pLDDT — per-residue confidence

Load the structure in PyMOL and color immediately by pLDDT, stored in the B-factor column. Every residue in your region of interest should be checked before any analysis proceeds.

# Color by pLDDT: blue = high confidence, red = low
spectrum b, blue_cyan_yellow_orange_red, minimum=50, maximum=100

# Record mean pLDDT for your site of interest
select site, resi 175+248+249+273+282
iterate site and name CA, print(resi, b)

Residues below pLDDT 70 are unreliable and should be excluded from structural conclusions. Document the mean pLDDT of your region of interest — this number goes in your methods section.

PAE — inter-domain confidence

For any protein with more than one domain, open the PAE plot. Download the JSON file from the AFDB or ColabFold output, or view the PAE directly on the AFDB protein page. Look at the off-diagonal blocks: dark blue means confident relative domain positioning; light or white blocks mean the relative orientation of those domains is uncertain, regardless of how well each domain scores individually.

pLDDT tells you whether AlphaFold was confident. MolProbity tells you whether the resulting structure is chemically and geometrically sensible. These are different questions — and both answers are required before publishing any work based on a predicted structure.

Submit your PDB to MolProbity (molprobity.biochem.duke.edu — free, no registration required). The report covers four key metrics:

• Clashscore — severe steric clashes per 1,000 atoms. Real proteins don’t have clashes. Target below 10; investigate above 20.
• Ramachandran statistics — backbone dihedral angle distribution. Target: favored above 95%, outliers below 0.5%.
• Rotamer outliers — residues with unlikely side chain conformations. Target below 2%.
• MolProbity score — composite metric corresponding to the resolution of a crystal structure of equivalent quality. Below 2.0 is acceptable; below 1.5 is excellent.

When outliers appear in or near your region of interest, investigate before proceeding. For outlier residues in the binding site: apply energy minimization (AlphaFold structures) or remodel with MODELLER’s LoopModel (homology models). Document which residues had outlier geometry and how they were handled.

Comparing a predicted structure to any available experimental data is the most powerful validation you can do — it tests whether the prediction captures real biology rather than a plausible but incorrect conformation. Most proteins with enough importance to study computationally have some experimental data somewhere in the literature. The question is which data type is available for your target.

• Mutagenesis
  Mutational mapping

  If loss-of-function mutations cluster in the predicted binding site or active site, the prediction places the right functional residues in the right positions. Inconsistencies between published mutational data and the predicted structure are a red flag that warrants investigation.
• B-factors
  B-factor and pLDDT concordance

  If a crystal structure of the same protein or a close homolog exists, compare its B-factors to the pLDDT map. Regions with high B-factors (flexible in the crystal) should correspond to low-pLDDT regions in the prediction. Consistent correlation validates that AlphaFold is correctly identifying the flexible and rigid regions.
• HDX-MS
  Hydrogen-deuterium exchange

  HDX-MS reports which protein regions are flexible and solvent-exposed versus rigid and protected. Rigid HDX-protected regions should correspond to high-pLDDT, low-RMSF regions in the prediction. Multiple consistent HDX-prediction correlations across the protein validate the overall fold.
• SAXS
  Small-angle X-ray scattering

  SAXS provides the protein’s overall shape in solution — radius of gyration, maximum dimension, low-resolution envelope. Comparing the predicted structure to a SAXS envelope using CRYSOL or DAMMIF validates overall architecture without requiring atomic-resolution experimental data.
• Homolog crystal
  Structural superposition to homolog

  Superimpose the predicted structure onto a close homolog crystal structure in PyMOL. Calculate RMSD for the core secondary structure elements. Below 2 Å is consistent with a reliable prediction; larger deviations in functional regions are worth investigating before drawing conclusions.

Even a geometrically clean, high-pLDDT structure may not represent a conformation stable in solution. Running a short MD simulation (50–100 ns) provides a physical test that neither pLDDT nor MolProbity can give: does this conformation persist under physiological conditions in a real force field?

What to look for

The backbone RMSD of the high-pLDDT core should plateau within the first 10–20 ns and remain stable throughout. A steadily rising RMSD signals that the predicted conformation is not in a stable energy minimum — the structure is drifting toward a different state. Rapid unfolding of predicted secondary structure elements is a particularly concerning signal: helices and sheets in high-pLDDT regions should be stable in MD. Instability suggests either a force field artifact or that AlphaFold placed those elements in an unrealistic structural context.

What stable MD tells you

A structure that simulates stably for 100 ns with consistent secondary structure, stable binding site geometry, and converged RMSF profiles is significantly more trustworthy as a research foundation. MD simulation acts as a physical filter — selecting for conformations energetically accessible under physiological conditions, not just geometrically plausible ones.