How to Evaluate Protein Structure Quality: DOPE, QMEAN, MolProbity and Ramachandran Plots Explained
Whether you’ve just built a homology model or downloaded an AlphaFold prediction, the next step before using any protein structure is the same: assess its quality. This guide explains every major quality metric, what it actually measures, and what values reviewers consider acceptable for publication.
Why quality assessment is not optional
A protein structure can look perfectly reasonable in PyMOL and still contain errors that invalidate downstream analysis. Steric clashes between atoms that never appear in a real protein. Backbone dihedral angles that no known protein adopts. Bond lengths and angles that violate basic chemistry. These problems don’t announce themselves visually — they hide in the coordinates and silently corrupt docking poses, MD simulations, and structure-function interpretations built on top of them.
Quality assessment exists to catch these problems before they become published errors. The metrics described in this article assess different aspects of structural quality — some measure energetic plausibility, others measure geometric correctness, others compare the model to statistics derived from thousands of experimentally validated structures. A structure that passes all of them can be used with confidence; one that fails should be refined or discarded before any downstream work begins.
DOPE Score
What it measures: DOPE is a statistical potential derived from a database of known protein structures. It calculates the probability that each atom in the model is in a chemically and physically reasonable environment, given what is known about how proteins are built. Regions with unfavorable atomic contacts, unlikely side chain conformations, or improbable backbone geometries produce poor (less negative) DOPE scores.
Why it’s useful: Unlike global quality scores, DOPE can be calculated per residue — plotting DOPE score along the sequence immediately highlights which loops and regions are problematic. Peaks in the DOPE profile (less negative values) mark regions with poor local geometry that may need loop refinement or closer scrutiny.
How to interpret it: DOPE scores are always negative — more negative is better. They are not normalized, so absolute values are not meaningful across different proteins of different sizes. Use DOPE to rank multiple models of the same protein (the model with the most negative DOPE is best) and to identify problematic regions via the per-residue profile.
How to calculate: DOPE is built into MODELLER and calculated automatically when you run AutoModel with assess_methods = (assess.DOPE). It can also be calculated on any PDB file using the MODELLER Python API: atmsel = model.assess_dope().
QMEAN Z-score
What it measures: QMEAN evaluates overall model quality by combining multiple structural features — local geometry, long-range interactions, secondary structure agreement — and comparing the result to a reference set of experimentally determined crystal structures of similar size. The Z-score expresses how many standard deviations the model is from the mean quality of real structures.
Why it’s useful: Unlike DOPE, QMEAN Z-scores are normalized by protein size, making them comparable across different proteins. A score near 0 means the model quality is consistent with real crystal structures. This makes QMEAN the primary global quality metric used by Swiss-Model and other web servers for assessing predicted models.
How to calculate: QMEAN Z-scores are generated automatically by Swiss-Model for every model it produces. For models built outside Swiss-Model, submit the PDB file to the SWISS-MODEL Structure Assessment server (swissmodel.expasy.org/assess) or to ProSA (prosa.services.came.sbg.ac.at), which calculates an equivalent Z-score and provides a visualization showing where your model falls relative to experimental PDB structures.
MolProbity and clashscore
What it measures: MolProbity is a comprehensive stereochemical validation suite that checks bond lengths, bond angles, Ramachandran statistics, rotamer quality, and steric clashes simultaneously. It is the tool that structural biology journals — including Nature, Science, PNAS, and the Journal of Molecular Biology — require for validation of deposited structures, making it the de facto standard for any published structure or model.
Clashscore is MolProbity’s most widely used single metric: it counts the number of serious steric clashes per 1,000 atoms. A clash occurs when two atoms that are not covalently bonded overlap by more than 0.4 Å. Real proteins don’t have such clashes — finding them in a model indicates an error in atom placement.
MolProbity score combines clashscore, Ramachandran outliers, and rotamer outliers into a single number that corresponds to the resolution of an X-ray structure with equivalent quality issues. A MolProbity score of 1.5 means the model has the stereochemical quality of a 1.5 Å crystal structure — excellent. A score of 3.5 means it has the quality of a 3.5 Å structure — poor.
How to access: MolProbity is available as a free web server at molprobity.biochem.duke.edu. Upload your PDB file, and within a few minutes you receive a complete validation report including per-residue scores, the Ramachandran plot, clash list, and a summary score. No registration required.
Ramachandran plot
What it measures: The Ramachandran plot displays the backbone dihedral angles (φ and ψ) for every residue in a protein. Because of steric constraints between backbone atoms, only certain combinations of φ and ψ are physically possible for non-glycine residues. Residues in “favored” regions adopt the angles seen in well-determined experimental structures. Residues in “outlier” regions adopt angles that are rare or physically strained — typically indicating a modeling error at that position.
Glycine residues are excluded from standard Ramachandran analysis because they lack a Cβ atom and can adopt a much wider range of angles. Proline residues are analyzed separately because the ring constrains their geometry. MolProbity handles these exceptions automatically.
How to run these checks
A complete quality assessment workflow for any protein model takes about 15 minutes and uses three free web servers:
- MolProbity (molprobity.biochem.duke.edu) — upload your PDB, get full stereochemical validation including clashscore, Ramachandran statistics, rotamer outliers, and a per-residue report identifying every problem. This is the mandatory check for publication.
- ProSA (prosa.services.came.sbg.ac.at) — upload your PDB, get the Z-score and a per-residue energy plot. Peaks in the plot identify the same problematic regions as DOPE, with a global score normalized for comparison across structures.
- Swiss-Model Structure Assessment (swissmodel.expasy.org/assess) — additional QMEAN scoring with visualization. Best for homology models built outside Swiss-Model.
Publication-ready checklist
-
Ramachandran favored > 95%, outliers < 0.5%Core requirement for any published model. Outlier residues should be identified and discussed if not zero.MolProbity
-
Clashscore below 20 (ideally below 10)Counts severe steric clashes per 1,000 atoms. High clashscore indicates unresolved geometric conflicts.MolProbity
-
MolProbity score below 2.0Combined metric. Corresponds to equivalent quality of a crystal structure at that resolution. Below 2.0 is acceptable; below 1.5 is excellent.MolProbity
-
QMEAN Z-score between 0 and −2Global quality normalized to protein size. Values below −4 indicate serious model quality problems requiring investigation.ProSA / Swiss-Model
-
Rotamer outliers below 2%Residues with unusual side chain conformations. High rotamer outlier percentages indicate poorly modeled side chains.MolProbity
-
For homology models: best DOPE score selectedWhen generating multiple models with MODELLER, confirm the selected model has the lowest (most negative) DOPE score.MODELLER
-
Quality scores reported in methods sectionReport MolProbity clashscore, Ramachandran statistics, and QMEAN Z-score. Reviewers increasingly check for these explicitly.All tools
Quality assessment in four checks
Run every protein structure through MolProbity before using it for docking, MD simulation, or publication. DOPE identifies problem regions and ranks multiple models of the same protein. QMEAN Z-score gives a normalized global quality estimate comparable across proteins — near 0 is good, below −4 is a problem. MolProbity clashscore and Ramachandran statistics are the publication-standard stereochemical checks that journals require. A structure that passes all four can be used with confidence; one that fails should be refined or its limitations acknowledged explicitly in the methods section.