# What the MolProbity score and clashscore mean > The MolProbity score combines clashscore, rotamer outliers and Ramachandran favored residues into one number on the scale of crystallographic resolution. Learn what counts as good and how AlphaFold models score. The MolProbity score is one number that combines a structure's clashscore, its share of rotamer outliers and its share of residues outside the favored Ramachandran regions. It is scaled so that the value matches the crystallographic resolution at which those numbers would be average: a 2.0 Å crystal structure with a MolProbity score of 1.5 is better than a typical 2.0 Å structure, and lower is always better. The clashscore is the number of serious all-atom overlaps, 0.4 Å or more, per 1,000 atoms once hydrogens have been added. You can calculate both, with the residue-level outliers behind them, by uploading a PDB or mmCIF file to [MolProbity](/app/molprobity) on ProteinIQ. This guide covers each number in a MolProbity report, the formula that turns three of them into the MolProbity score, how to judge a result against structures at similar resolution, and why AlphaFold models often score better than the crystal structures they are compared with. A worked example on three models of sperm whale myoglobin shows the arithmetic. ## What MolProbity checks MolProbity validates the atomic model itself. It does not look at the diffraction data or the density map, so it tells you whether atoms sit in physically plausible places, not whether they fit the experiment. Before any contact is scored, the program Reduce adds hydrogens, optimizes rotatable OH, SH and NH3 groups, and flips Asn, Gln and His side chains whose terminal atoms were fitted 180° backwards. The program Probe then measures contacts between every pair of nonbonded atoms, hydrogens included. The torsion checks compare each residue with reference distributions from the Top8000 set: 7,957 protein chains filtered at 70% sequence identity, solved by X-ray crystallography at better than 2.0 Å, with a MolProbity score better than 2.0 and residues with any main-chain B-factor above 30 removed. Table 1 lists the numbers a report contains and the goal values MolProbity prints beside them. Table 1. The main numbers in a MolProbity report. Goal values as printed in the MolProbity summary table; clashscore and MolProbity score are judged by percentile rather than a fixed goal. Sources: Chen et al. (2010), Williams et al. (2018), MolProbity source code (2026). Metric What it counts Goal Clashscore Serious all-atom overlaps of 0.4 Å or more per 1,000 atoms Percentile against similar resolution Ramachandran favored Residues inside the contour that holds 98% of reference data Above 98% Ramachandran outliers Residues outside the allowed contour Below 0.05% Rotamer outliers Side chains in conformations seen in under 0.3% of reference data Below 0.3% Cβ deviations Residues whose Cβ atom is 0.25 Å or more from its ideal position 0 CaBLAM outliers Residues with backbone geometry in the rarest 1% of reference data Below 1% CaBLAM Cα geometry outliers Residues with a Cα trace in the rarest 0.5% of reference data Below 0.5% MolProbity score Log-weighted combination of clashscore, rotamer outliers and Ramachandran favored Below the structure's resolution; percentile against similar resolution ## Clashscore Probe works by rolling a probe sphere 0.25 Å in radius over the van der Waals surfaces and measuring how far nonbonded atoms overlap. When two atoms that are not a hydrogen-bond donor and acceptor overlap by 0.4 Å or more, the contact counts as a serious clash. An overlap that large cannot occur in the real molecule, so at least one of the two atoms has been placed incorrectly. The clashscore is the number of these serious clashes per 1,000 atoms. Hydrogens are part of the count, and most crystal structures are deposited without them, which is why Reduce runs first. Where the hydrogens sit also matters. Until 2018 MolProbity used one set of X-H bond lengths for every method, and even the best reference structures bottomed out at about 5 clashes per 1,000 atoms. It now places hydrogens at electron-cloud centers for X-ray models (0.97 Å for a C-H bond) and at nuclear positions for other methods (1.09 Å), and Top8000 clashscores now approach zero. The Richardson lab still describes the goal as few clashes, as in the best reference data, rather than zero. wwPDB validation reports always place hydrogens at nuclear positions so that X-ray and NMR entries can be compared, which is one reason a clashscore from the PDB report can differ from one you calculate yourself. Clashscore is a count of atom pairs, not of bad residues. One side chain fitted into the wrong pocket can produce several clashes with its neighbours, so a clashscore of 10 does not mean that 1% of the structure is wrong. The residue list in the report is the place to see where they concentrate. ## Ramachandran favored, allowed and outliers Each residue's backbone φ and ψ angles are compared with a reference distribution for its residue type. MolProbity uses six such plots: general, Ile and Val, Gly, pre-proline, trans-proline and cis-proline. The favored contour encloses 98% of the high-quality reference data. In the general case the allowed contour encloses 99.95%, so only one in 2,000 well-determined residues falls outside it, and a residue there is an outlier. Outliers become more common as resolution gets worse. Each one is either a modeling error or a rare strained conformation that the data clearly support, so it is worth checking the density before accepting it. To see where every residue falls rather than a summary percentage, plot φ and ψ with the [Ramachandran plot](/app/ramachandran-plot) tool. CCTBX reports also include a Rama-Z score, which judges the shape of the whole φ,ψ distribution rather than individual residues; an absolute value below 2 is good, 2 to 3 is suspicious and above 3 is poor. ## Rotamer outliers Side-chain χ angles are compared with Top8000 rotamer distributions built from 983,574 residues. Since 2016 MolProbity sorts each side chain into three bands: favored conformations hold 98% of the reference data, allowed ones occur at 0.3% to 2%, and outliers occur in fewer than 0.3% of good side chains. Older software used a single 1% cutoff for poor rotamers, so rotamer percentages from different MolProbity versions are not directly comparable. A common cause of a rotamer outlier is a branched side chain (Thr, Val, Ile, Leu or Arg) fitted backwards into ambiguous density. ## Cβ deviations The position of the Cβ atom is fixed by the backbone N, Cα and C atoms. MolProbity calculates where the Cβ should be from that backbone and measures how far the modeled Cβ is from it. Residues with a deviation of 0.25 Å or more are flagged, because deviations that large are very often linked to local misfitting of the backbone or the side chain. The report gives a count of flagged residues, and the goal is zero. ## CaBLAM Ramachandran analysis needs accurate backbone positions, which become uncertain at the 2.5 to 4 Å resolutions common in cryo-EM. CaBLAM works from the Cα virtual dihedrals across a five-residue stretch, which are better determined, and uses them to judge the much less certain carbonyl directions. A CaBLAM outlier is in the rarest 1% of the reference data, a disfavored residue is in the rarest 5%, and a separate Cα geometry check flags the rarest 0.5% of Cα traces. CaBLAM is more reliable than Ramachandran analysis between 2.5 and 4 Å, and the MolProbity web service runs it by default for structures at 2.5 Å or worse. It is expected to find some outliers even in good structures, with false positives in loops. ## How the MolProbity score is calculated The MolProbity score was added so users could have a single number for the central protein quality statistics. It is a log-weighted combination of clashscore, the percentage of residues outside the Ramachandran favored regions and the percentage of rotamer outliers, scaled to the crystallographic resolution at which those values would be expected. The coefficients come from a log-linear fit to resolution on a filtered set of PDB structures. As implemented in CCTBX, the library behind MolProbity, it is: $$ \begin{aligned} S ={}& 0.426\ln(1 + c) \\ &+ 0.33\ln\bigl(1 + \max(0,\, r - 1)\bigr) \\ &+ 0.25\ln\bigl(1 + \max(0,\, 98 - f)\bigr) \\ &+ 0.5 \end{aligned} $$ Here S is the MolProbity score, c is the clashscore, r is the percentage of rotamer outliers and f is the percentage of Ramachandran favored residues. The rotamer term ignores the first 1% of outliers and the Ramachandran term ignores the first 2% of residues outside favored (hence 98 − f), so a structure with no clashes that meets those two levels scores the floor of 0.5. MolProbity's own source code records more precise coefficients (0.42574, 0.32996 and 0.24979); they change the scores in this guide by less than 0.01. Clashscore carries the largest weight. Ramachandran outliers enter only through the favored percentage, and Cβ deviations, CaBLAM, covalent geometry and ligands do not enter at all, so a good MolProbity score can sit beside other problems in the same report. ## What counts as a good score A MolProbity score lower than the structure's resolution means the model is better than the average structure at that resolution. The fit is distorted at very high and very low resolution, and there the percentile is the better guide. The MolProbity web service ranks clashscore and MolProbity score against PDB structures within 0.25 Å of the file's resolution. Its summary colors percentiles below 33 as poor and 66 or above as good; the comparison sets were chosen in 2004 for clashscore and 2006 for the MolProbity score. wwPDB validation reports use their own percentiles. The rank is the percentage of entries that are equal to or poorer than the structure, so 100 is best, and the similar-resolution comparison uses the smallest resolution bin that contains at least 1,000 entries. Because the two systems use different comparison sets and hydrogen conventions, compare structures scored by the same pipeline. NMR models and predicted structures have no crystallographic resolution, so the resolution reading of the MolProbity score does not apply to them. Judge them by the goal values in Table 1 and by the residue-level outliers. ## Worked example: three myoglobin models We validated three models of sperm whale myoglobin, a protein with crystal structures from several eras of crystallography and a high-confidence AlphaFold prediction. 1MBN is the 2.0 Å crystal structure deposited in 1973, and 1A6M is the 1.0 Å oxy-myoglobin structure deposited in 1998. AF-P02185-F1 is the AlphaFold DB model, with a mean pLDDT of 97.5. Each file was run through the CCTBX 2026.4 MolProbity validation with Reduce, Probe and the Top8000 reference data at the same versions that [MolProbity](/app/molprobity) on ProteinIQ uses. The [script](/data/guides/molprobity-score/run_molprobity.py), [results table](/data/guides/molprobity-score/molprobity_results.csv) and full reports for [1MBN](/data/guides/molprobity-score/1MBN_molprobity.txt), [1A6M](/data/guides/molprobity-score/1A6M_molprobity.txt) and [AF-P02185-F1](/data/guides/molprobity-score/AF-P02185-F1_molprobity.txt) are public, so the numbers in Table 2 can be reproduced from the downloaded coordinates. Table 2. MolProbity results for three sperm whale myoglobin models. CCTBX 2026.4 MolProbity validation of the deposited coordinates, including heme and any waters and ions, with electron-cloud hydrogens added by Reduce. Sources: RCSB PDB (2026), AlphaFold DB (2026), ProteinIQ analysis (2026). Metric 1MBN 1A6M AF-P02185-F1 Method and resolution X-ray, 2.0 Å X-ray, 1.0 Å AlphaFold prediction Clashscore 39.68 7.91 0.00 Ramachandran favored 76.16% 98.66% 98.68% Ramachandran allowed 20.53% 1.34% 1.32% Ramachandran outliers 3.31% 0.00% 0.00% Rotamer outliers 19.20% 4.03% 0.00% Cβ deviations 99 3 0 CaBLAM outliers 0.00% 0.00% 0.67% CaBLAM Cα geometry outliers 0.67% 0.00% 0.00% Rama-Z (whole model) −6.21 1.57 0.79 MolProbity score 3.84 1.89 0.50 Table 3 rebuilds each MolProbity score from its three inputs with the formula above. For 1MBN, the clashscore term is 0.426 × ln(1 + 39.68), the rotamer term is 0.33 × ln(1 + 19.20 − 1), and the Ramachandran term is 0.25 × ln(1 + 98 − 76.16). Table 3. MolProbity score term by term. Each term rounded to two decimals; totals match the CCTBX output in Table 2. Source: ProteinIQ analysis (2026) using the CCTBX formula. Term 1MBN 1A6M AF-P02185-F1 Clashscore term 1.58 0.93 0.00 Rotamer term 0.98 0.46 0.00 Ramachandran term 0.78 0.00 0.00 Constant 0.50 0.50 0.50 MolProbity score 3.84 1.89 0.50 1MBN was built from a 2.0 Å map, but its MolProbity score of 3.84 is what an average structure at about 3.8 Å would score. Every component is poor: a quarter of residues sit outside the favored Ramachandran regions, one side chain in five is a rotamer outlier and 99 residues have Cβ deviations. Its wwPDB validation report puts its clashscore at the 0.3rd percentile of 11,152 entries at 2.0 Å. That report lists a clashscore of 56.08 rather than 39.68 because it comes from a different MolProbity build with hydrogens at nuclear positions, which is why scores from different pipelines should not be mixed. 1A6M shows that high resolution does not guarantee a clean model. Its backbone is excellent, with no Ramachandran outliers and nothing for the Ramachandran term to count, but a clashscore of 7.91 and 4.03% rotamer outliers give a MolProbity score of 1.89, roughly what an average 1.9 Å structure would score. Its wwPDB report places its clashscore at the 11.7th percentile of 1,411 entries between 0.96 and 1.04 Å. The structure was deposited in 1998, four years before MolProbity's all-atom contact analysis became available, and clashscores across the PDB have improved since then. The AlphaFold model hits the floor of 0.5: no clashes, no rotamer outliers and Ramachandran favored above 98%. It has no heme and no waters, and the score says nothing about whether the heme pocket or any other region matches a real state of myoglobin. The reasons are in the next section. ## How predicted models score AlphaFold 2 ends each prediction with an Amber relaxation step. Its authors report that relaxation does not improve accuracy but removes distracting stereochemical violations. Relaxed predictions therefore tend to have near-zero clashscores and MolProbity scores at or below those of experimental structures. In a comparison of 102 AlphaFold predictions with the deposited crystal structures of the same proteins, the predictions averaged 98% Ramachandran favored against 97% for the deposited models, and 0.2% rotamer outliers against 1.5%, even though the authors concluded that predictions are hypotheses that do not replace experiment. Across 918 paired segments from medium-resolution cryo-EM entries, the mean MolProbity score was 0.96 for AlphaFold DB models and 1.98 for the PDB models. For a predicted model, then, MolProbity is [a check for broken geometry](/guides/structure-prediction-confidence), not an accuracy score. A relaxed loop the model has little confidence in can still have clean geometry, so read it alongside [pLDDT](/guides/plddt) and the predicted aligned error. MolProbity is most useful on models that did not go through a relaxation step, such as designs, docked complexes, threaded mutants and edited structures, because nothing in those pipelines guarantees physical contacts. ## Limits of the MolProbity score MolProbity judges the model without the data. A model can have ideal geometry and still be placed in the wrong density or be the wrong conformation, and resolution remains the most informative single estimate of a crystal structure's accuracy. Scores can also be improved without improving the model. When Ramachandran and rotamer restraints are used during refinement at 2.5 to 3 Å, and especially at 3.5 to 4 Å, outliers can be pushed just across the nearest contour boundary, which makes the model look better than it is. CaBLAM, clashscore and the residue-level outlier lists are harder to satisfy this way, so look at them as well as the summary. The MolProbity score is defined for proteins and was calibrated on crystal structures. Nucleic acids have their own checks for sugar puckers and backbone conformers, and ligands are not part of the score. A single global number also hides local problems: if your conclusions depend on an active site or an interface, read the outliers for those residues rather than the summary. ## MolProbity tools on ProteinIQ - [MolProbity](/app/molprobity) calculates clashscore, Ramachandran and rotamer statistics, Cβ deviations, CaBLAM, Rama-Z and the MolProbity score from a PDB or mmCIF file or a PDB ID, with a residue-level outlier table and a Coot script. - [Ramachandran plot](/app/ramachandran-plot) shows where every residue falls in φ,ψ space. - [AlphaFold database download](/app/alphafold-database-download) fetches a predicted model by UniProt accession so you can validate it beside an experimental structure. ## Frequently asked questions ### Is a lower MolProbity score better? Yes. The score is on the scale of crystallographic resolution, so lower means better, and the minimum is 0.5. For a crystal structure, compare it with the resolution: a score below the resolution means the model is better than the average structure at that resolution. ### What is a good clashscore? There is no fixed cutoff, because clashscore depends on resolution and on how hydrogens are placed. With current MolProbity parameters the best reference structures approach zero. The percentile against structures at similar resolution is the fairer comparison. ### Why does my AlphaFold model have a MolProbity score of 0.5? 0.5 is the floor of the formula. Relaxed AlphaFold models usually have no serious clashes, almost no rotamer outliers and more than 98% Ramachandran favored residues, so every penalty term is zero. It means the geometry is clean, not that the prediction is correct. ### Why do MolProbity and the PDB validation report give different numbers? The wwPDB pipeline uses its own MolProbity build, places hydrogens at nuclear positions and ranks against its own comparison sets. A local CCTBX or web MolProbity run uses electron-cloud hydrogens for X-ray models by default. Both are valid, but compare structures scored the same way.