# AlphaFold pLDDT vs. PAE: Which Confidence Metric Should You Trust pLDDT tells you how confident AlphaFold is about each individual residue, while PAE tells you how confident it is about the relative positions of pairs of residues. The rule of thumb is simple. Use pLDDT when your question is about local structure, individual residues, loops, or disordered regions. Use PAE when your question is about how two parts of the model sit relative to each other, such as domain orientations, chain arrangements, or protein interfaces. In the pLDDT vs. PAE comparison, most wrong conclusions in AlphaFold-based projects come from using pLDDT for a question that only PAE can answer. Both numbers appear in every AlphaFold2 and AlphaFold3 prediction, and both are needed to judge a model responsibly. A recent validation study of AlphaFold2 and AlphaFold3 across the AP-1 [transcription factor](/knowledge/molecular-biology/transcription-factor) family and more than 2000 experimentally tested human leucine zippers made this point explicitly. To use AI structure prediction productively, you need to understand what the models can and cannot do, and the confidence metrics are the main tool for doing that [1]. This article explains each metric on its own terms, compares them on the criteria that decide real analysis choices, walks through worked examples, and covers a practical AlphaFold tutorial workflow for pulling these numbers out of your own predictions. ## pLDDT vs. PAE in AlphaFold: The Short Answer ![Decision path asking whether the question is about one residue or about how two parts sit relative to each other, ending at pLDDT or PAE.](/article-images/illustrations/plddt-vs-pae-alphafold-metric-triage.webp) The difference between pLDDT and pAE in AlphaFold comes down to scope. pLDDT is a per-residue score. Every residue in the model gets one number between 0 and 100. PAE is a pairwise score. Every ordered pair of residues gets a number, expressed in angstroms, arranged as a square matrix the size of the total residue count squared. That structural difference drives everything else in the comparison. A compact way to think about the difference between pLDDT and pAE in AlphaFold is this. pLDDT answers "how sure is the model about this residue's local environment?" PAE answers "if I superimpose the true structure on residue j, how far off will residue i be?" A residue can sit in a beautifully predicted helix with pLDDT of 95, yet carry a PAE of 18 angstroms against a distant domain, because the model is unsure where that domain lies relative to the helix. Both statements are true at the same time, and neither metric subsumes the other. The comparison table below is the core of this article. Everything after it explains the rows in detail. | Criterion | pLDDT | PAE | |--|--|--| | What it measures | Expected local accuracy at each residue | Expected positional error of one residue relative to another after superposition | | Format | One number per residue, 0 to 100 | N by N matrix, one angstrom value per residue pair | | Units | Dimensionless score | Angstroms | | Best for judging | Local fold quality, loop reliability, disorder, per-residue filtering | Domain packing, chain arrangements, interface confidence, flexible versus rigid relationships | | Blind spot | Says nothing about where a well-scored residue sits relative to the rest of the model | Says nothing directly about local geometric correctness | | Typical failure it prevents | Trusting an implausible loop or treating a disordered tail as structured | Trusting a beautiful interface that the model itself is unsure about | | Where you see it | B-factor column of PDB and mmCIF files, color scheme in AlphaFold DB | Color matrix in the AlphaFold DB viewer, JSON output files from local runs and the AlphaFold server | ## What pLDDT Measures pLDDT stands for predicted local distance difference test. The lDDT metric itself is a standard structural superposition-free score used in CASP-style assessment. It measures, for a reference structure, the fraction of atom pairs within a fixed inclusion radius whose distances are reproduced within a set of tolerance thresholds. pLDDT is AlphaFold's prediction of the lDDT its own model would achieve at each residue if the true structure were known. Because the prediction is scored against itself in this calibration sense, pLDDT is best read as a self-consistency estimate of local accuracy rather than a guarantee. ### The pLDDT Scale and Color Scheme pLDDT runs from 0 to 100, and the AlphaFold database uses a four-band color scheme that has become the community standard for reading these scores: 1. Dark blue, pLDDT above 90. Very high confidence. Backbone geometry in these stretches is expected to be highly accurate. 2. Light blue, pLDDT between 70 and 90. Confident backbone, though side-chain details deserve more caution. 3. Yellow, pLDDT between 50 and 70. Low confidence. The general fold may be right, but local details can be wrong. 4. Orange, pLDDT below 50. Very low confidence. These regions are frequently intrinsically disordered in reality, meaning the protein has no single fixed structure there for the model to find. The orange band deserves special attention. AlphaFold documentation and the AlphaFold database treat pLDDT below 50 as a strong indicator of intrinsic disorder. Many proteins carry long low-pLDDT termini or linker segments. This is not a failure of the model. It is often the model correctly reporting that the region has no stable structure to predict. Treating low pLDDT as noise rather than as biological signal is one of the most common reading errors among new users. ### Where pLDDT Lives in Your Output Files In a standard AlphaFold2 run, pLDDT is stored in the B-factor column of the output PDB and mmCIF files. Open ranked_0.pdb, look at the B-factor field, and you are looking at per-residue pLDDT values from 0 to 100. The same convention holds for structures deposited in the AlphaFold database. Any molecular viewer that colors by B-factor therefore reproduces the familiar blue-to-orange confidence coloring. This convention matters practically, because it means you can extract, filter, and script against pLDDT with the same tools you already use for crystallographic B-factors. ## What PAE Measures ![Comparison of pLDDT and PAE across what they measure, their format, units, best use and blind spot.](/article-images/illustrations/plddt-vs-pae-alphafold-plddt-pae-scope.webp) PAE stands for predicted aligned error. For each ordered pair of residues i and j, the PAE value is the expected position error of residue i in the model when the predicted and true structures are superimposed using residue j as the alignment anchor. The full result is an N by N matrix. Diagonal cells are near zero by definition. Off-diagonal cells carry the real information. ### Reading the PAE Matrix The standard visual is a heatmap with residues along both axes. Two reading patterns matter most in practice: 1. Dark blocks along the diagonal. Each dark square block corresponds to a region, usually a domain or a chain, whose internal relative geometry the model is confident about. Boundaries between blocks often map to domain boundaries or flexible linkers. 2. Light or dark squares between blocks. If two blocks are darkly connected off the diagonal, the model is confident about their relative arrangement. If the connection between two well-defined blocks is pale, the model is telling you it does not know how those units pack against each other. For multi-chain predictions, the matrix has a block structure by chain. The within-chain blocks describe each chain's internal confidence. The between-chain blocks describe interface confidence. This is the property that makes PAE the metric of choice for judging predicted complexes. A point that trips up careful readers: PAE is a relative error under superposition, not an absolute coordinate error. A large PAE between two domains does not mean either domain is internally wrong. It means their relative placement is uncertain. Conversely, a small PAE between two distant residues means the model is confident about their relative geometry, even if each residue's local chemistry has minor issues that pLDDT would flag. ### Aggregate Scores Derived from PAE AlphaFold reports two aggregate scores that build on this pairwise logic. pTM is a predicted template modeling score covering the whole assembly. ipTM is an interface-focused variant that weights the pairwise errors at chain contacts. For complex predictions, especially on the AlphaFold server, ipTM is the single number most practitioners scan first, because a high pLDDT on both chains means little if the interface itself is unreliable. The full PAE matrix remains more informative than any single aggregate, because it shows where the uncertainty sits. ## When to Use pLDDT The question of when to use pLDDT has a clean answer set. Reach for pLDDT when your analysis unit is a single residue or a contiguous stretch of sequence. **Checking whether a folded region is well modeled.** If you want to know whether the catalytic core of your enzyme is credibly predicted, scan pLDDT across the core. Values in the dark blue band support downstream use of that geometry, for example as a starting point for ligand docking or mutation interpretation. **Identifying disordered regions.** Low pLDDT termini and linkers are the fastest practical screen for intrinsic disorder in a predicted structure. If your protein has a 60-residue tail sitting in the orange band, the correct interpretation is usually "this tail is disordered," not "AlphaFold failed here." That distinction changes how you build constructs, interpret truncations, and present the model. **Filtering residues before downstream computation.** Many pipelines mask or exclude residues below a pLDDT threshold before docking, restraint generation, or comparative modeling. The specific cutoff depends on the application, so choose it based on how much local error your downstream method tolerates rather than copying a universal number. **Comparing local quality between models.** When you sample multiple models from one AlphaFold run, per-residue pLDDT shows you which regions are stable across the ensemble and which fluctuate. Regions that change between ranked models while keeping low pLDDT are usually flexible or disordered in reality. What pLDDT cannot tell you is equally important for deciding when to use pLDDT. It cannot tell you that a residue with a score of 96 is in the right place relative to a distant domain. High pLDDT on every residue of a two-chain prediction is completely compatible with the two chains being wrongly placed relative to each other, or not interacting at all. ## When to Use PAE Reach for PAE whenever your analysis unit is a relationship between two parts of the model. **Judging predicted interfaces.** For any multi-chain prediction, examine the between-chain block of the PAE matrix. Uniformly low values across the interface region support the predicted arrangement. High values mean the model has low confidence in the interface, even if every residue on both chains carries excellent pLDDT. This single check prevents the most expensive error in complex modeling, which is building mechanistic claims on an interface the model itself flagged as uncertain. **Deciding whether two proteins form a complex.** When you predict two proteins as a pair, the PAE pattern is diagnostic. A confident heterodimer shows a coherent dark structure spanning both chains in the off-diagonal blocks. Two proteins that do not interact typically produce a matrix where each chain forms its own confident block while the cross-chain region stays uncertain. Comparing the PAE of the intended pair against a negative control pair is a standard sanity check. **Mapping domain flexibility.** Within a single chain, pale corridors in the PAE matrix between dark diagonal blocks reveal domains whose relative orientation is uncertain. This is genuinely useful information. It tells you the model has not pinned down a hinge, which usually means the hinge is flexible or the domain arrangement depends on a binding partner or ligand not present in the prediction. **Choosing among ranked models.** When AlphaFold returns five ranked structures, they often share a common core but differ in how peripheral domains or chains are placed. Comparing PAE patterns across ranked models shows which arrangements are consistently supported and which are artifacts of a single sampling trajectory. ## Worked Examples Concrete scenarios make the pLDDT vs. PAE decision automatic. The first two are illustrative templates. The third comes from published validation work. ### Scenario 1: The Flexible Activation Loop Suppose you predict a kinase and plan to discuss ligand binding. The per-residue profile shows pLDDT of 92 across the kinase domain and 38 across a 20-residue activation loop. The correct reading is that the catalytic core model is solid for structural reasoning, while the loop is likely disordered or conformationally variable and should not be presented as a single well-defined conformation. PAE adds little here because your question is local. This is a when-to-use-pLDDT case. Mask the loop, annotate it as flexible, and avoid docking a ligand into a loop conformation the model does not actually believe in. ### Scenario 2: The Deceptive Heterodimer Suppose you predict a two-chain complex. Both chains come back with mean pLDDT above 90. A colleague glances at the colored structure, sees two flawless ribbons, and declares the interface validated. Then you open the PAE matrix. Each chain forms a crisp dark diagonal block, and the cross-chain region is uniformly pale. The model is confident about each chain individually and unconfident about their arrangement. The correct conclusion is that the interface prediction is not supported, and the high pLDDT is irrelevant to that conclusion. This is the canonical alphafold plddt vs. pae trap, and it is the reason interface papers report PAE and ipTM rather than per-chain pLDDT. ### Scenario 3: Coiled Coils and Dimer Specificity Leucine zipper coiled-coils are the textbook case where the distinction between the two metrics carries real biological weight. AP-1 transcription factors such as Fos and Jun dimerize through leucine zipper domains, and the identity of the dimer partner determines which DNA sequences get recognized and which gene programs follow. Predicting which zippers pair, and with what relative orientation, is a question about chain relationships, not about individual helix quality. A leucine zipper helix will nearly always show high pLDDT, because an alpha-helical coiled-coil is locally easy for the network to model. Whether the specific Fos-Jun pairing, or any pairing among more than 2000 experimentally tested human leucine zippers, is modeled correctly is a question about inter-chain geometry, which is exactly what PAE reports. A study designed around this problem used the classical Fos-Jun AP-1 dimer together with that large panel of tested human leucine zippers to interrogate how all versions of AlphaFold2 and AlphaFold3 model leucine zippers, with the explicit goal of mapping the methods' capabilities and limits [1]. The design logic is worth copying. When the biological question is dimerization, the validation and the confidence reading both need to live at the level of the interface, which means PAE between chains, while pLDDT remains useful for confirming that each helix is internally sane [1]. ## A Practical AlphaFold Tutorial: Getting the Numbers Knowing which metric to trust is only half the workflow. You also need to locate both numbers in real output. This AlphaFold tutorial section covers the three settings most people work in. ### Reading a Structure on the AlphaFold Database The AlphaFold database, hosted by EMBL-EBI, serves predicted structures for hundreds of millions of protein sequences. For any entry, the viewer offers two tabs that matter here. The structure view applies the standard four-band pLDDT coloring by default, so dark blue, light blue, yellow, and orange map directly onto the confidence bands described earlier. The PAE tab renders the full pairwise error matrix as an interactive heatmap, which you can zoom into to inspect specific domain or chain pairs. For multi-chain entries, the PAE download is available alongside the structure file. Before citing any database structure in a figure, open both tabs. The colored ribbon alone hides exactly the information that most often changes interpretation. ### Using the AlphaFold Server The AlphaFold server, operated by Google DeepMind for AlphaFold 3, accepts a list of interacting molecules, including multiple protein chains, [nucleic acids](/blog/guides/nucleic-acids), and small molecules, and returns ranked predictions. Each prediction comes with per-residue confidence and a downloadable JSON file containing the PAE matrix, the per-atom confidence values, and the aggregate scores including pTM and ipTM. A sensible habit after every server run is to open the PAE visualization first, check the cross-chain blocks if you submitted a complex, and only then look at the ranked structures. For complex submissions, check ipTM next, and treat the per-chain pLDDT as the last and least diagnostic of the three reads. ### Running AlphaFold or ColabFold Locally A local AlphaFold2 or ColabFold run produces a predictable file set. The key items for confidence analysis are: 1. ranked_0.pdb through ranked_4.pdb. The B-factor column of each file holds per-residue pLDDT from 0 to 100. 2. predicted_aligned_error.json. This file contains the full PAE matrix and the maximum PAE value for the model. 3. In ColabFold runs, a scores file per prediction that bundles per-residue pLDDT and the PAE matrix together in one JSON document. A minimal Python check, using only the standard library for parsing, extracts what you need from the PAE file: ```python import json with open("predicted_aligned_error.json") as f: pae_data = json.load(f) matrix = pae_data["predicted_aligned_error"] max_pae = pae_data["max_predicted_aligned_error"] n = len(matrix) # Mean PAE between the first 100 residues (chain A) and last 100 (chain B) cross = [matrix[i][j] for i in range(100) for j in range(n - 100, n)] print(f"max PAE: {max_pae:.2f} angstrom") print(f"mean cross-chain PAE: {sum(cross) / len(cross):.2f} angstrom") ``` That cross-chain mean, compared against the within-chain mean, is the single quickest interface diagnostic you can script. For pLDDT, the same parse applies to the B-factor column of the PDB, or to the scores JSON from ColabFold. Plotting per-residue pLDDT as a line graph alongside the PAE heatmap for every prediction you care about takes minutes and prevents most interpretation errors before they happen. ## Common Mistakes and Limitations The failure modes below account for most bad AlphaFold-based conclusions. Each is avoidable once named. **Reading pLDDT as a certificate of biological correctness.** pLDDT measures expected local structural agreement, not whether the predicted structure is biologically relevant. AlphaFold can produce a locally excellent model of a conformation the protein never adopts, for example a single static snapshot of a state-switching receptor or a confidently folded alternative packing of a repeat protein. Confidence in the geometry is not evidence for the state. **Averaging pLDDT over a chain and calling it model quality.** A mean pLDDT of 85 can describe a protein with a rock-solid core and a fully disordered 80-residue tail, or a uniformly mediocre model. The distribution, not the mean, carries the information. Always look at the per-residue profile. **Using pLDDT to judge interfaces.** As the heterodimer scenario showed, per-residue scores are structurally incapable of answering interface questions. If you take one habit from this article, make it checking the cross-chain PAE block before trusting any multi-chain model. **Reading PAE as absolute coordinate error.** PAE is defined under superposition on the second residue of the pair. Large PAE values between domains mean uncertain relative placement, not broken domains. Small PAE values do not certify local chemical details, which is pLDDT's job. **Treating low pLDDT as model failure.** Below-threshold pLDDT frequently reports true intrinsic disorder. A segment that is disordered in solution has no single correct structure, and a model that hedges there is behaving correctly. Confirm with disorder annotations rather than assuming the predictor failed. **Forgetting that confidence is self-referential.** Both metrics are predictions of error against an unknown truth, calibrated during training. Systematic failure modes, such as shallow multiple sequence alignments for orphan proteins, unusual oligomeric states, membrane environments, and very long disordered regions, can produce miscalibrated confidence. The validation study of Fos-Jun dimers and over 2000 human leucine zippers exists precisely because confidence metrics need empirical stress-testing against experimental data before being trusted for a given protein class [1]. If your protein family has an unusual structural signature, validate a handful of experimentally characterized members before generalizing. **Ignoring ranked-model disagreement.** When ranked_0 and ranked_4 disagree about a domain arrangement, that disagreement is information about multimodality. A single high-confidence model can be one sample from an ambiguous distribution, and the PAE matrix may only partially reflect it. **Skipping the negative control.** For complex predictions, predicting your pair of interest together with a pair known not to interact gives you a baseline PAE pattern for a non-implementation. Without that baseline, "the cross-chain block looks fairly dark" has no reference point. One general limitation applies to all of the above. Confidence metrics rank and qualify predictions, but they do not replace experimental validation for any specific system where the structural claim matters, and individual cases should be interpreted in the context of available experimental data for that protein. ## Frequently Asked Questions ### Can pLDDT be high while the overall structure is still wrong? Yes. pLDDT only certifies local geometry, so a model can carry excellent per-residue scores while placing domains or chains incorrectly relative to each other. Check the PAE matrix for any question involving relative placement. ### What pLDDT cutoff should I use to filter low-confidence residues? There is no universal cutoff. Use the four-band scheme as a guide, exclude orange-band residues for most downstream geometry work, and pick your threshold based on how much local error your downstream method tolerates. ### Where do I find pLDDT and PAE in AlphaFold server output? The structure viewer colors by pLDDT, and the downloadable JSON for each ranked prediction contains the PAE matrix, per-residue confidence values, and the pTM and ipTM aggregate scores. ### What units does PAE use and how do I read the matrix? PAE is in angstroms. Each cell gives the expected position error of one residue when the model is aligned on the other, so dark diagonal blocks mark confident regions and pale off-diagonal blocks mark uncertain relative arrangements. ### Should I trust pLDDT or PAE to decide whether two proteins really interact? Trust PAE, ideally together with ipTM. High pLDDT on both chains says nothing about whether the chains pack as predicted, while the cross-chain PAE block reports interface confidence directly. ### Do AlphaFold database structures include the PAE matrix? Yes. Every AlphaFold database entry exposes a PAE view in the web viewer, and multi-chain predictions provide the PAE data as a downloadable file alongside the structure. ## Sources 1. [Assessing the validity of leucine zipper constructs predicted by AlphaFold.](https://pubmed.ncbi.nlm.nih.gov/41432297/)