# AlphaFold Output Interpretation: What the Results Mean AlphaFold output interpretation is the practice of reading a predicted [protein structure](/blog/guides/protein-structure) together with its confidence scores to decide what the model does and does not establish about that protein. The working rule is that every prediction carries its own measured error estimate, residue by residue and domain by domain, and that estimate must be read alongside the coordinates before any structural claim is made. Predicted structures now arrive before experimental ones. The AlphaFold Protein Structure Database, hosted by EMBL-EBI, holds models for more than 200 million protein sequences, so most researchers meet a prediction long before they collect a crystal structure or a cryo-EM map. The attached scores decide whether you can map a mutation onto an active site, trust the orientation of two domains, or treat a flexible tail as disordered. Read them carelessly and you will build hypotheses on regions the model itself flags as guesses. Read them well and a five-minute download can replace weeks of construct screening. This guide is AlphaFold output interpretation explained from the files up, with the numbers you will actually see and the claims each one supports. ## What Does AlphaFold Output Interpretation Mean? In practice, the task answers two questions about every model. The first is local: where in the chain is the model likely to be geometrically correct. The second is relational: how confident is the model about where one part of the protein sits relative to another. AlphaFold answers the first with a per-residue score called pLDDT and the second with a pairwise error map called PAE. Search phrases such as alphafold output interpretation interpretation, or alphafold output interpretation explained, all reduce to these two questions: which numbers in the result can you trust, and for what claim. A prediction differs from an experimental structure in what it reports about itself. A crystal structure comes with electron density, a resolution value, and validation metrics computed against diffraction data. A cryo-EM model comes with map resolution and half-map correlation curves. An AlphaFold model comes with no new experimental data at all, only coordinates plus the model's own estimate of where those coordinates are likely to be wrong. That self-estimate is the entire subject of interpretation. The table separates the terms that cause most of the confusion. | Term | What it measures | Where you see it | Common misreading | |--|--|--|--| | pLDDT | Local confidence for one residue and its spatial neighborhood, on a 0 to 100 scale | B-factor column of the coordinate file, color bands on the model | Low pLDDT means the model is broken. Usually it means the region is flexible or disordered | | PAE | Expected position error in Ångströms between two residues after alignment | Square matrix shipped as a JSON file with the model | High PAE means the coordinates themselves are wrong. Often only the relative placement of two well-built parts is uncertain | | pTM and ipTM | Predicted similarity of a whole assembly and of its interfaces, for multi-chain predictions | Summary scores from AlphaFold 3 and AlphaFold-Multimer | Reading them as per-residue quality measures | | RMSD | Measured deviation of one structure from a reference structure | Only computable when a reference structure exists | Quoting "the RMSD of a prediction" with no reference structure | | Predicted model vs experimental structure | A hypothesis with self-reported error vs a measurement with resolution and validation metadata | AlphaFold Database entries vs [Protein Data Bank](/blog/guides/protein-data-bank) entries | Citing a prediction as if it were crystallography or cryo-EM | Two families of scores do most of the interpretive work. Per-residue scores grade each position on its own. Pairwise scores grade the relationship between positions. Assembly-level summary scores grade whole complexes. Everything else in a result folder exists to support reading these three. ## The Files and Scores Inside an AlphaFold Result A typical result is a coordinate file, one or more confidence files, and a record of the inputs that shaped the run. Knowing which file holds which number is half of the interpretation task. ### Coordinate files, PDB and mmCIF AlphaFold distributes coordinates in PDB and mmCIF formats. mmCIF is the modern standard and handles the large assemblies and chemical components that legacy PDB format struggles to encode. One detail matters more than the format choice. In AlphaFold outputs, the B-factor column, which crystallographers use for atomic displacement parameters, stores the pLDDT value on a 0 to 100 scale. Any molecular viewer that colors by B-factor therefore colors the model by confidence, and simple viewer commands can isolate weak regions for inspection. Runs of AlphaFold 2 produce five ranked models, conventionally named rank_001 through rank_005, with the ranking based on the model's own summary quality estimates. AlphaFold 3 and the AlphaFold Server follow the same pattern of ranked outputs with confidence summaries. Rank 1 is usually the right starting point, but model selection is a real interpretive step, as the worked example below shows. ### pLDDT, the per-residue confidence score pLDDT stands for predicted local distance difference test. The underlying metric, lDDT, was developed for CASP, the Critical Assessment of protein Structure Prediction, which is the field's blind benchmark for structure prediction. lDDT is reference-free and checks whether atom pairs within a 15 Å neighborhood keep their distances when two structures are compared. pLDDT is the network's prediction of the lDDT its own model would receive, expressed on a 0 to 100 scale. A pLDDT of 95 means the model expects nearly all atom pairs around that residue to hold their distances. A pLDDT of 40 means the model expects most of them to move. The field reads pLDDT in four standard bands, which are also the color bands used in the AlphaFold Database viewers. - Above 90, very high confidence. The backbone is expected to be accurate and side-chain positions are usually correct. Active-site geometry, hydrogen-bond reasoning, and mutation mapping are defensible in this range. - 70 to 90, confident backbone. The chain trace is usually right but side-chain orientations vary. Many domain cores sit here. - 50 to 70, low confidence. Treat coordinates as approximate and avoid mechanistic stories built on these residues. - Below 50, very low confidence. In AlphaFold Database models this band is treated as a marker of likely intrinsic disorder, and viewers color it orange. These bands are conventions from the model developers, not guarantees. Calibration is good on average, which is why the bands are useful, but individual regions can be overconfident, a point that returns in the limitations section. ### PAE, the pairwise error map ![Worked example reading a 12 Å off-diagonal PAE block for a 350-residue protein with domains at residues 1 to 200 and 201 to 350.](/article-images/illustrations/alphafold-output-interpretation-pae-block-gap.webp) The predicted aligned error, PAE, is reported as an N by N matrix for a protein of N residues, usually as a JSON file alongside the coordinates. The value at position (i, j) is the expected position error, in Ångströms, of residue j when the predicted structure is aligned on residue i. Values near zero mean the model is confident about the relative placement of the two residues. Large values mean it is not. Reading the matrix as blocks is the practical skill. A well-determined domain appears as a dark square along the diagonal. Two domains whose relative orientation is known appear as dark squares joined by dark off-diagonal blocks. Two domains connected by a flexible linker appear as dark diagonal squares separated by a pale off-diagonal block. In that last case the model is telling you it built each domain correctly but placed them relative to each other by guesswork. A concrete example makes the scale meaningful. Suppose a 350-residue protein has two domains, residues 1 to 200 and residues 201 to 350. If the off-diagonal block between them shows values near 12 Å, the model is saying the two folded parts could each be correct while their relative orientation is off by roughly that distance, which is wider than an alpha helix. If the same block shows values near 2 Å, the relative arrangement is about as trustworthy as the domains themselves. ### AlphaFold 3 additions AlphaFold 3 extends prediction to complexes containing proteins, [nucleic acids](/blog/guides/nucleic-acids), ions, ligands, and chemical modifications. It adds a pairwise distance error, PDE, which estimates expected errors on atom-pair distances and is useful when a specific contact matters more than global placement. Complex results also carry assembly-level scores, pTM for the whole assembly and ipTM for the interfaces, and the server ranks models by these values. For interface claims, ipTM and the cross-chain PAE blocks carry the interpretive weight, the same way intra-chain PAE does for domains. ### The inputs behind the model Every result also reflects its inputs. The [multiple sequence alignment](/blog/guides/multiple-sequence-alignment), MSA, collects homologous sequences and encodes the evolutionary constraints the network reads. Template structures from the Protein Data Bank, when found and used, anchor parts of the model to known folds. Output logs record MSA depth and coverage, and those numbers explain a great deal of what the confidence scores later say. The next section explains why. ## How AlphaFold Builds a Structure, Step by Step The pipeline runs in six stages, and each stage leaves a fingerprint on the output. 1. The query sequence is submitted, alone or with interaction partners. 2. The system searches sequence databases for homologs and builds the MSA. 3. The system searches the Protein Data Bank for structural templates. 4. The network processes the MSA and a pairwise residue representation through a deep stack of transformer blocks. AlphaFold 2 used 48 Evoformer blocks. AlphaFold 3 separates MSA processing from pair processing and runs 48 Pairformer blocks. 5. A structure module, a diffusion module in AlphaFold 3, renders 3D coordinates. 6. Dedicated confidence heads compute pLDDT, PAE, and, in AlphaFold 3, PDE, plus summary scores for assemblies. Where the fold actually forms inside that stack was an open question until recently. A 2026 causal analysis of AlphaFold 3 probed every Pairformer block and identified the MSA as a structural shortcut to the fold [1]. Removing the MSA largely preserved local secondary structure while disrupting the long-range relationships that define global topology. Restoring the MSA-enriched pair representation at only forty residues recovered most of the lost organization, including at residue pairs that had never been directly modified. The final fold became recoverable by approximately block 9 of 48 for a majority of proteins, roughly twenty-seven blocks before the model's decoder rendered it. Without the MSA, the global fold stayed inaccessible for most proteins throughout the entire pass [1]. The same study showed that the prediction is steered more by which homologs are supplied than by which query is supplied. The effect persisted for a designed query sequence that never evolved, and it collapsed when the sequence was shuffled. An alignment built for a different protein that shares the fold could carry the structural signal [1]. Three consequences follow for interpretation. First, confidence tracks evolutionary information, so proteins with shallow or sparse MSAs, including orphan proteins, rapidly evolving proteins, and designed proteins, tend to show weaker long-range geometry even when local secondary structure looks convincing. Second, homolog curation is an interpretation lever. Excluding a paralog from an alignment can change the predicted state, which is exactly how some studies steer predictions toward a functional state. Third, a pale inter-domain PAE block does not always mean the model failed. Sometimes it means the alignment did not contain the constraints needed to place the domains relative to each other. ## Reading a Result in Practice: A Worked Example Consider a [receptor tyrosine kinase](/knowledge/molecular-biology/receptor-tyrosine-kinase) such as EGFR, a 1,210-residue membrane protein with four extracellular domains, a single transmembrane helix, an intracellular kinase domain, and a long C-terminal tail. Step 1, check the ranking and summary scores. For a database model this is a single entry. For a server run, note which ranked model the service places first and on what basis. Step 2, scan pLDDT along the sequence. The extracellular domains and the kinase domain typically sit in the high bands. The transmembrane helix is usually well determined as a helix, although the membrane environment itself is not modeled. The juxtamembrane segment and the C-terminal tail typically fall into the low bands, because in isolation they sample many positions. Step 3, read the PAE matrix as blocks. Dark squares along the diagonal mark each extracellular domain and the kinase domain. Pale off-diagonal blocks between the extracellular domains tell you their relative orientations are uncertain, which matches what ligand-bound and ligand-free experimental structures show for this family. The kinase domain block is dark, so its internal geometry is trustworthy. Step 4, map the biology you care about. An activating mutation in the kinase domain's activation loop can be interpreted against a well-determined fold. A phosphorylation site in the C-terminal tail cannot be placed relative to the kinase domain from this model, because both the tail's pLDDT and the cross-region PAE say that arrangement is a guess. Step 5, write down the claim the model supports before moving on. Fold-level topology of the kinase domain, yes. A mechanism of tail-mediated regulation from this model alone, no. Model choice is part of interpretation, not an afterthought. A 2026 study of all 437 human catalytic [protein kinase](/knowledge/molecular-biology/protein-kinase) domains illustrates the point. Only 123 of those domains have an active-state structure in the Protein Data Bank, so the authors used AlphaFold 2 with templates resembling substrate-bound structures and shallow sequence alignments of close paralogs and orthologs to push predictions toward the active state, then applied structural criteria for kinase activation derived from 248 crystal structures of 54 kinase-substrate complexes, including the state of the DFG motif in the activation loop and the regulatory salt bridge [2]. To choose among the five models each run produces, they scored the activation loop with intramolecular ipSAE, a confidence metric derived from predicted aligned error, and the highest-scoring models tended to have the lowest RMSD to substrate-bound experimental structures [2]. Interpretation here meant selecting the right model, for the right functional state, using the confidence metric designed for the question. ## What Low Confidence Really Means ### Disorder and molecular recognition features ![Decision path from a low pLDDT stretch to disorder, a MoRF check, or a model failure, ending in a claim you can defend.](/article-images/illustrations/alphafold-output-interpretation-low-plddt-triage.webp) Long stretches below 50 pLDDT usually mean the region is intrinsically disordered, meaning it has no single stable structure in isolation. This is biology, not model failure. Disordered regions are common in signaling proteins, where they act as flexible tethers and binding segments. Low confidence does not tell you which parts of a disordered region matter functionally. Molecular recognition features, MoRFs, are short segments within disordered regions that fold upon binding their partners, and they are the functional hotspots of many tails. Dedicated predictors address this. MoRFchibi 2.0 identifies MoRFs from sequence and outperformed all existing MoRF and general binding-site predictors within disordered regions, including the top-performing models from the CAID assessment rounds 1 through 3 and predictors that use AlphaFold data [3]. The lesson for interpretation is narrow and useful. AlphaFold tells you where the chain has no fixed structure, and specialized tools tell you where within that flexible chain binding is likely to order it. ### One model, one conformation The deeper issue is that AlphaFold returns a single structure, while many proteins adopt several. A 2024 systematic analysis addressed this directly using 2,878 proteins with at least ten distinct experimental structures each, covering 628 distinct folds [4]. AlphaFold 2 predictions returned one specific form at high accuracy, and 99.68 percent of folds, 623 of 628, had an experimental structure within 2.5 Å RMSD of the prediction. Flexibility still showed through. For 27.70 percent of folds, 174 of 628, at least one experimental structure sat more than 2.5 Å from the predicted model, and for 10.82 percent, 68 of 628, at least one sat more than 5 Å away [4]. Those numbers do not mean the model failed in the discordant cases. The protein moves, the prediction captured one state, and the size of the conformational range depends on the fold. Practical habits follow. Before concluding that a prediction contradicts your experiment, pull every experimental structure of your protein from the Protein Data Bank and compare against all of them. When a model and a structure disagree, ask whether the difference looks like a domain rearrangement, which is common, rather than a local error, which is rarer. Treat the single prediction as one sample from a conformational ensemble, and treat fold family as context, because some folds are rigid and some are inherently plastic [4]. ## Where Predicted Structures Change Decisions ### Clinical variant interpretation Clinical sequencing produces variants of uncertain significance, VUSs, whose effect on [protein function](/blog/guides/protein-function) is unknown. Structure converts that list into spatial hypotheses. A 2025 tool built for clinical next-generation sequencing mapped thousands of missense mutations from COSMIC, the Catalogue of Somatic Mutations in Cancer, onto structurally aligned kinase structures across the kinase and juxtamembrane domains of 48 receptor tyrosine kinases, then clustered known activating mutations together with VUSs based on their proximity in three-dimensional structure, and demonstrated in cell-based models that the resource helps identify activating mutations [5]. For interpretation, the point is that a variant's position in a high-confidence structural context, near an active site, on a domain interface, or inside a hydrophobic core, carries mechanistic information that sequence position alone does not, and predicted structures extend that reasoning to the many proteins without experimental structures. ### Binding sites and cross-molecular prediction Predicted structures increasingly feed interaction predictors that add their own calibrated scores. The xBind webserver accepts a single-chain FASTA sequence or a monomer structure in PDB or mmCIF format, runs AlphaFold prediction on the fly for sequence-only inputs, and combines ESM-2 language model embeddings with sequence- and structure-derived features through symmetry-aware graph neural networks to predict residue-level binding sites for protein-protein, protein-DNA, and protein-RNA interactions [6]. Its interface lets users adjust the likelihood threshold used to call binding residues and provides interactive visualizations with a web-based tutorial and results guide [6]. When you use tools like this, interpretation expands. You read pLDDT for the structure, then the tool's calibrated likelihoods for the binding calls, and the threshold you choose changes the residue list you get. ### Machine learning on predicted pockets Drug-discovery pipelines treat AlphaFold models as standard pocket sources. One 2025 drug-target affinity study noted that pocket data typically need to be predicted with software such as AlphaFold, which for datasets of tens of thousands to hundreds of thousands of samples becomes a massive workload, so the authors instead fused pretrained-model embeddings through structure-aware cross-attention to simulate 3D pocket information from sequence [7]. Their feature representation reuses the output of the last layer of a pretrained model as an embedding layer for a new model trained from scratch [7]. The interpretive consequence is that AlphaFold output now flows into downstream models as embeddings and likelihoods, and reading a result can mean reading a vector or a score rather than a ribbon diagram. ## Getting a Prediction: Databases, the AlphaFold Server, and a Short Tutorial ### Check the databases first Most reviewed proteins already have a model. Search UniProt by gene name, and the entry links to the corresponding AlphaFold Database model hosted by EMBL-EBI, viewable in the browser with pLDDT coloring applied automatically. The Protein Data Bank may also hold experimental structures of the full protein or of individual domains, and those take precedence for claims about conformational states. ### Running a new prediction on the AlphaFold Server For new predictions, Google DeepMind's AlphaFold Server runs AlphaFold 3 from a browser with no local setup. You paste sequences and optionally add nucleic acids, ions, ligands, and chemical modifications, and the service returns ranked models with confidence summaries. Access is free for non-commercial use, with daily limits on the number of predictions. The server is the right tool for single biological questions that involve partners, such as whether a protein binds a DNA motif or a small molecule. For batch work, custom MSAs, or full control over templates, laboratories run AlphaFold 2 or 3 locally or through community notebooks such as ColabFold. ### A short AlphaFold tutorial The following six steps turn any result into an interpreted result, and they work for database models and server outputs alike. 1. Search UniProt for your protein and open the linked AlphaFold Database entry, or submit your sequence on the AlphaFold Server and download the ranked outputs. 2. Open the model in a viewer that shows confidence coloring. The AlphaFold Database viewer and Mol* color the pLDDT bands automatically. 3. Inspect the low-confidence regions first. In PyMOL, the pLDDT values sit in the B-factor column, so the command "spectrum b" recolors the model by confidence, and "select lowconf, b < 50" isolates the weakest residues for measurement. In UCSF ChimeraX, "color byattribute bfactor palette alphafold" applies the standard four-band palette. 4. Open the PAE matrix that ships with the model and read it as blocks. Dark diagonal squares are well-determined domains. Pale off-diagonal blocks mark relative placements you should not interpret. 5. Map what you know onto the confident regions, including active sites, known domains, and your mutations of interest, and check each mapped feature against its local pLDDT band. 6. Write one sentence stating the claim the model supports, for example, "the kinase domain fold and activation-loop geometry are reliable, the C-terminal tail placement is not." Attach that sentence to the file in your notes. ### Choosing between the server and local runs The choice affects interpretation. Server runs fix the MSA pipeline for you, which keeps results reproducible but removes control over homolog selection. Local runs let you curate the alignment, and because homolog choice measurably steers the predicted fold [1], a curated alignment is itself an interpretive act. If a result looks wrong, one of the first diagnostics is to inspect the MSA the run used. ## Common Mistakes and Limitations The following mistakes account for most misinterpretations in coursework, manuscripts, and reviews. 1. Reading pLDDT as guaranteed accuracy. The score is the model's self-estimate, well calibrated on average but capable of overconfidence on unusual folds, designed proteins, and shallow alignments. Very high pLDDT on a novel designed protein deserves more skepticism than the same score on a kinase domain. 2. Ignoring the PAE matrix. Two domains can each be beautifully built while their relative orientation is a guess. Any claim about how domains contact each other needs a dark cross-domain PAE block or experimental support. 3. Calling every low-confidence segment disordered. Low pLDDT correlates with disorder but is not a disorder measurement. Use dedicated disorder and MoRF predictors when a functional claim depends on it. 4. Treating the single model as the only conformation. Flexible proteins sample multiple states, and the prediction represents one of them. Compare against all available experimental structures before declaring a conflict. 5. Presenting a prediction as an experimental structure. Label models as predictions, cite the database entry or the run, and report confidence values alongside any figure. Journals and depositories treat the two categories differently. 6. Reading interfaces from monomer models. A monomer prediction says nothing about which face binds a partner. Interface claims need multi-chain predictions with strong cross-chain ipTM and PAE, or experimental data. 7. Over-reading side chains at moderate confidence. Between 50 and 70 pLDDT, side-chain orientations are not reliable enough to support hydrogen-bond or salt-bridge stories. 8. Forgetting the alignment shaped the answer. The MSA carries the structural signal, homolog selection steers the trajectory, and a poor alignment can silently reshape global topology while local secondary structure still looks convincing. The standing limitation is that a prediction is a hypothesis. For claims that carry consequences, such as a mutation mechanism, a drug design decision, or a clinical interpretation, experimental validation remains the standard, and individual projects benefit from review by a structural biologist before conclusions are locked in. ## Quick Review - pLDDT is per-residue local confidence. Read the four bands, above 90, 70 to 90, 50 to 70, and below 50, and treat the lowest band as a disorder flag in database models. - PAE is the pairwise error map. Dark blocks mean trusted relative placement, pale blocks mean guesswork, and cross-domain claims need dark off-diagonal blocks. - The MSA is the structural shortcut to the fold, and homolog choice steers the prediction more than the query sequence does. - A prediction is one conformation. Flexible proteins sample many states, and fold family predicts how much they move. - Low confidence usually means disorder or flexibility, not a broken model, and dedicated tools identify binding-ready segments within disordered regions. - Model selection is part of interpretation. Rank, functional state, and the right confidence metric all matter. - Always label a predicted model as a prediction and report its confidence values with any figure. ## Frequently Asked Questions ### What does a pLDDT score of 60 mean? It means the model has low confidence in that residue, and the local geometry should not support structural claims. Between 50 and 70, treat the coordinates as approximate and avoid side-chain reasoning. Below 50, the region is likely disordered rather than mispredicted. ### What is the difference between pLDDT and PAE? pLDDT grades each residue on its own neighborhood, while PAE grades how confidently one residue is placed relative to another after alignment. High pLDDT with high PAE across a boundary means two well-built parts whose relative orientation is uncertain, which is a different problem from a badly built residue. ### Can I publish or deposit work based on an AlphaFold model? Yes, if the model is labeled as a prediction, the database entry or run is cited, and confidence values are reported with the figures. For claims that carry mechanistic weight, support the prediction experimentally. Journals and structure depositories treat predicted and experimental structures as separate categories. ### Why does my protein's N-terminal tail have low confidence? Most often the tail is intrinsically disordered or flexible in isolation, so no single stable position exists for the model to find. Confirm with a disorder predictor before calling it a model failure, and check whether the tail contains a molecular recognition feature that folds upon binding. ### Does high confidence mean the protein does not move? No. Confidence measures coordinate certainty for one predicted state, not rigidity. A protein can have a high-confidence single model while sampling multiple conformations experimentally, and the size of that conformational range depends on the fold. ### How do I get a model for my protein of interest? Check the AlphaFold Database through the UniProt entry first, since most reviewed proteins already have one. For new predictions involving partners, ligands, or nucleic acids, use the AlphaFold Server. For batch work or custom alignments, run AlphaFold locally or through community notebooks. ## Sources 1. [A Shortcut to Structure in AlphaFold 3.](https://pubmed.ncbi.nlm.nih.gov/42926731/) 2. [Defining active conformations from substrate-bound structures enables active-state AlphaFold2 modeling of all 437 human catalytic protein kinase domains.](https://pubmed.ncbi.nlm.nih.gov/42257422/) 3. [Predicting molecular recognition features in protein sequences with MoRFchibi 2.0.](https://pubmed.ncbi.nlm.nih.gov/42288727/) 4. [Systematic analysis of the relationship between fold-dependent flexibility and artificial intelligence protein structure prediction.](https://pubmed.ncbi.nlm.nih.gov/39591473/) 5. [A structure-based tool to interpret the significance of kinase mutations in clinical next generation sequencing in cancer.](https://pubmed.ncbi.nlm.nih.gov/40831936/) 6. [xBind: an integrated webserver for large language model-enabled cross-molecular protein binding site prediction.](https://pubmed.ncbi.nlm.nih.gov/42083872/) 7. [Interpretable drug-target affinity prediction based on pre-trained models' output as embeddings and based on structure-aware cross-attention for feature fusion.](https://pubmed.ncbi.nlm.nih.gov/40279085/)