# Interpreting AlphaFold Confidence Scores for Non-Model Organisms AlphaFold confidence scores are numerical estimates that ship with every predicted structure and describe how closely each part of the model is expected to match the true structure. For non-model organisms, meaning species without extensive experimental structural data, these scores are the main evidence for deciding which parts of a prediction to trust, which parts are probably flexible or disordered, and which parts to set aside. Genome projects produce protein sequences much faster than laboratories solve structures, so predicted models now carry a large share of structural biology outside the classic laboratory species. AlphaFold predictions have been used to survey the photosystem II D1 protein across a broad range of photosynthetic organisms and to infer the structures of odorant receptors in two longhorn beetle genomes, both settings where experimental structures are scarce [1][2]. Generating a model is the easy part. Interpreting it is where the work lies, because accuracy varies along the chain and the confidence scores are the map of that variation. A high score on a conserved core licenses mechanistic reasoning, while a low score on a loop or terminus often signals real flexibility rather than a failed prediction. Misreading the scores in either direction produces overconfident claims about active sites, mutations, or evolutionary history, which is why score interpretation has become a core skill for anyone working beyond the standard laboratory species. ## What AlphaFold Confidence Scores Mean for Non-Model Organisms A model organism is a species with deep experimental resources, such as *Escherichia coli*, mouse, fruit fly, or thale cress. A non-model organism is everything else, which in practice means most of the tree of life. Every AlphaFold prediction reports coordinates together with confidence annotations, and three quantities matter most for single-chain work: pLDDT, PAE, and pTM. Each is explained below with the reading rules that matter most when the species of interest has few or no experimental structures. ### pLDDT: Per-Residue Local Confidence pLDDT stands for predicted local distance difference test. The underlying metric, lDDT, measures how well distances between atom pairs in a model agree with a reference structure within a local neighborhood, typically within 15 Å, and it does so without superimposing the two structures. AlphaFold predicts the lDDT value at the Cα position of every residue, where Cα is the alpha carbon at the center of each amino acid, scales it from 0 to 100, and reports the result as pLDDT. Read pLDDT as a per-residue grade for local backbone geometry. The standard color legend used in structure viewers divides the scale into four bands: - **Above 90 (dark blue), very high.** Backbone and side-chain detail are usually reliable enough for mechanism-level discussion. - **70 to 90 (light blue), confident.** The fold and secondary structure are generally trustworthy, while fine details deserve caution. - **60 to 70 (yellow), low.** Secondary structure elements may be correctly placed while exact coordinates are not. - **Below 60 (orange), very low.** Treat the region as a hypothesis rather than a result. Regions with pLDDT below 50 are frequently unstructured in isolation, which means a very low score often reflects intrinsically disordered regions, segments that lack a stable structure on their own, rather than model failure. Two practical facts follow. First, in AlphaFold output files the pLDDT value is stored in the B-factor column, the file field that normally holds experimental uncertainty measures, so any viewer that colors by B-factor displays confidence automatically. Second, pLDDT describes the local environment of each residue. It says nothing about whether two distant parts of the chain are positioned correctly relative to each other, which is the job of PAE. ### PAE: Predicted Aligned Error The predicted aligned error, PAE, is a pairwise measure reported in Ångströms, each Ångström being one ten-billionth of a meter. For every pair of residues i and j, the score estimates the expected position error at residue i if the predicted and true structures were aligned on residue j. The output is an N by N matrix for a protein of N residues, usually displayed as a heatmap in which low error is dark and high error is pale. The pattern in the matrix carries the message. Confidently packed regions form dark blocks along the diagonal, one block per structured domain. When two domains sit in a fixed relative orientation, the off-diagonal block connecting them is dark as well. When their relative orientation is uncertain, that off-diagonal block stays pale even if each domain is internally confident. A residue can carry pLDDT of 95 and still show a PAE of 25 Å to a domain at the other end of the chain, which tells you the local geometry is excellent while the global arrangement is a guess. A practical rule of thumb in the community is that inter-domain PAE values below roughly 5 Å support the modeled domain arrangement, while values in the tens of Å mark the relative placement as unresolved. These are heuristics rather than hard cutoffs, and they work best when read together with the block structure of the matrix. ### pTM and ipTM in Brief pTM is a scalar estimate of the TM-score, a length-normalized measure of global topological similarity between the model and the true structure. By long-standing convention, a TM-score above 0.5 indicates the same overall fold, so a pTM in that range supports the claim that the prediction has the right topology. ipTM extends the idea to interfaces and is the relevant number when predicting complexes with AlphaFold-Multimer or AlphaFold 3. For single-chain questions about loops, active sites, and domain packing, pLDDT and PAE do the heavy lifting. ### The Scores at a Glance | Score | What it measures | Scale and format | Read it for | Most often confused with | |--|--|--|--|--| | pLDDT | Expected local backbone accuracy at each residue | 0 to 100, one value per residue, stored in the B-factor column | Deciding which regions to trust, flagging disorder | A probability that the whole model is correct | | PAE | Expected position error between each pair of residues | Ångströms, N by N matrix | Judging domain packing and interface geometry | Per-residue accuracy | | pTM | Expected fold-level similarity of the entire chain | 0 to 1 scalar | Asking whether the overall topology is right | Loop-level detail | | ipTM | Expected quality of an interface in a complex | 0 to 1 scalar | Multimer and complex predictions | Monomer accuracy | | B-factor field | The file location where pLDDT is stored | Same numeric scale as pLDDT | Coloring models in molecular viewers | Experimental crystallographic uncertainty | ## Why Non-Model Organisms Are a Special Case ![Chart of how a shallow MSA can lead to underconfidence on a correct fold or to overconfidence on a wrong one.](/article-images/illustrations/alphafold-confidence-scores-non-model-organisms-sparse-msa-failure.webp) AlphaFold learns from experimental structures in the [Protein Data Bank](/blog/guides/protein-data-bank) and from evolutionary information in sequence databases, and neither resource is distributed evenly across the tree of life. Structures and deep sequence families concentrate on model organisms, human disease targets, and agriculturally important species. A protein from a poorly studied lineage often enters prediction with a shallow [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) (MSA), the record of related sequences the network relies on, and sometimes no structural template at all. Confidence generally falls as evolutionary distance from well-covered families grows, but the relationship is not simple. Ancient, near-universal proteins, including core metabolic enzymes and photosynthetic reaction center subunits, retain deep cross-lineage alignments, so predictions can stay strong even in distant species. The trouble spots are lineage-specific expansions and rapidly evolving multigene families, which are common in non-model genomes, particularly for gene families tied to ecology such as the odorant receptors of insects [2]. A contrast makes the point concrete. A human kinase prediction inherits alignments containing thousands of relatives spanning vertebrates, plus dozens of experimental templates, so high confidence is expected. A receptor from a newly assembled beetle genome may find its closest relatives in other insects, with alignment depth depending on how well those genomes are annotated. The same network and the same score scales sit on completely different evidence bases. Sparse evolutionary information creates two distinct failure modes. The first is underconfidence, where the model finds the correct fold but reports modest scores because it lacks supporting evidence. The second, rarer and more dangerous, is overconfidence on a wrong answer. With very sparse alignments the network can assemble plausible-looking but incorrect helical bundles, and the warning sign is usually a distorted or featureless PAE matrix rather than a depressed pLDDT. This is why MSA depth and template status belong in every interpretation, alongside the two headline scores. ## Reading pLDDT in Practice ![Decision path that starts from a pLDDT value and ends at the kind of claim the band supports, from mechanism-level to probable disorder.](/article-images/illustrations/alphafold-confidence-scores-non-model-organisms-plddt-reading-path.webp) Every major route to an AlphaFold model exposes pLDDT. The public AlphaFold [Protein Structure](/blog/guides/protein-structure) Database, which holds over 200 million precomputed monomer models, colors each model by pLDDT by default and provides downloadable per-residue values. ColabFold, a widely used notebook and command-line implementation, writes pLDDT into the output coordinate file and plots the per-residue profile automatically. In a local AlphaFold2 run, the ranked model files carry pLDDT in the B-factor column. Band by band, the reading rules look like this: - **Above 90.** Safe for mechanism-level claims, active-site geometry, and detailed comparative modeling of that segment. - **70 to 90.** Safe for fold recognition, domain boundaries, and most comparative analyses. Be careful with side-chain specifics and fine contacts. - **60 to 70.** Use for the presence and rough arrangement of secondary structure, not for contact details. - **Below 60.** Treat as hypothesis. Check whether the region is a known flexible linker or a candidate disordered segment. - **Below 50.** Probable intrinsic disorder. The presence of disorder is itself a finding worth reporting. A short example shows the reading in action. Take a 450-residue [protein kinase](/knowledge/molecular-biology/protein-kinase) annotated from a fungal genome with no structural data. A typical profile shows a dark blue core spanning the N-lobe and C-lobe, light blue around the hinge, yellow patches in the activation loop, and an orange C-terminal tail of roughly 40 residues. The interpretation follows directly: the kinase fold and nucleotide-binding core are well determined, the activation loop is uncertain in conformation, and the tail is probably disordered. None of that required an experimental structure, and all of it came from the score profile. Across large surveys of non-model proteomes the same pattern recurs. Scores run highest where structure is constrained by function and lowest in loops and termini where the chain is mobile, so the confidence profile doubles as a first-pass map of order and disorder. ## Reading PAE in Practice PAE arrives as a matrix image or a JSON file alongside the coordinates. Read it in blocks. Label the axes with residue numbers and candidate domain boundaries, then ask three questions. Are the diagonal blocks dark, meaning each domain is internally consistent? Are the off-diagonal blocks between neighboring domains dark, meaning their arrangement is determined? Is there a pale strip connecting one structured region to the rest, meaning that segment floats? A two-domain enzyme from a non-model eukaryote illustrates the division of labor between the scores. Suppose both domains show pLDDT above 90 internally, but the off-diagonal PAE block between them sits near 20 Å. The correct conclusion is that each domain is reliable while the hinge orientation is unknown. Docking the domains or reasoning about a substrate tunnel that crosses the interface would outrun the evidence. If the inter-domain block is dark instead, the relative geometry can be used, including for designing constructs that connect the domains. PAE also helps when domain boundaries are unclear from the sequence alone. Dark diagonal blocks separated by pale valleys outline the domains directly from the data, which is useful for choosing construct boundaries before expressing a protein. For predicted complexes, PAE between chains is the primary evidence for or against an interface. High pLDDT on both chains paired with a pale cross-chain block means the proteins were predicted together but not placed relative to each other with confidence. This single check prevents a large share of overconfident complex models. ## How Reliable Are the Scores Themselves? Confidence scores are calibrated in the statistical sense, meaning the network was trained so that its stated errors match the errors observed against experimental structures. In blind benchmark assessments, well-covered targets reached backbone accuracies comparable to the disagreement between independent experimental determinations of the same protein, and the reported scores tracked those accuracies closely. Calibration is strongest for globular domains of the kind that dominate the training data. Calibration weakens at the edges. Novel folds with no structural relatives, long disordered regions, proteins whose fold depends on bound cofactors, and unusual chemistries all sit far from the training distribution, and scores in those regimes deserve extra skepticism. The honest reading is directional. High scores are strong evidence when the input resembles the training world. Low scores are strong evidence that the model itself is uncertain, which is still useful information. ## Running Predictions: From the AlphaFold Server to a Local Pipeline The scores are outputs of the network, not post-hoc statistics. During prediction, the model refines its internal representation through several recycling rounds, and separate output heads estimate per-residue local error for pLDDT and pairwise aligned error for PAE. In AlphaFold2 pipelines these values are computed on the raw model before the optional relaxation step that tidies stereochemistry, so relaxing a structure does not improve its scores. AlphaFold 3 reports pLDDT and PAE along with additional error measures, and its PAE covers the modified residues and ligands included in the prediction. Four routes cover most needs: 1. **AlphaFold Server.** The web interface to AlphaFold 3 accepts interacting proteins, [nucleic acids](/blog/guides/nucleic-acids), ions, and selected ligands, and it is free for non-commercial use subject to daily quotas. For a handful of sequences from a non-model organism, it is the fastest route to a model with full confidence output. 2. **AlphaFold Protein Structure Database.** Check here before spending compute, because the proteome may already be predicted. Entries include the pLDDT-colored model, the PAE matrix, and downloadable per-residue confidence values. 3. **ColabFold.** A notebook and command-line implementation that builds MSAs with MMseqs2 and runs AlphaFold2. It handles batches, which suits non-model genomes with hundreds of targets. A minimal batch run reads `colabfold_batch queries.fasta results/`, and the output directory contains coordinate files with pLDDT in the B-factor column along with PAE output. 4. **Local AlphaFold2 or AlphaFold-Multimer.** Full control over MSA construction, templates, and sampling seeds, at the cost of GPU hardware with generous memory. Whatever the route, calibrate your eye first. Walk through an AlphaFold tutorial on a protein with a well-known experimental structure and compare the score profile against the solved coordinates. Seeing what 90, 70, and 50 look like on a known answer teaches more than any written rule, and it costs an afternoon. Record the software version, database versions, MSA depth, and template settings with every run. Scores shift subtly between releases, and a prediction without provenance is difficult to defend later. ## A Step-by-Step Workflow for Interpreting AlphaFold Confidence Scores for Non-Model Organisms The following sequence turns a prediction folder into a defensible interpretation. 1. **Curate the input sequence.** Confirm the gene model, remove frameshift artifacts, and decide which isoform matters. Input errors degrade scores in ways that have nothing to do with the biology. 2. **Identify the family.** Search public protein databases for homologs, map domain boundaries, and find the closest well-studied relative. The family tells you what accuracy to expect and which regions are conserved. 3. **Check for an existing model.** A database hit saves compute and provides an immediate score profile to inspect. 4. **Generate the prediction** through the route that fits the scale, one sequence on the AlphaFold Server, a batch through ColabFold, or a local installation. Log the number of sequences in the MSA and whether templates were used. 5. **Read pLDDT as a profile,** not a single number. Note the fraction of the chain in each band, then map the bands onto domains, motifs, termini, and predicted transmembrane segments. 6. **Read PAE for architecture.** Mark domain boundaries on the matrix and classify each inter-domain block as determined or undetermined. 7. **Overlay independent biology.** Place catalytic residues, conserved motifs, cysteines, and transmembrane helices onto the model, and check that functional sites fall in high-confidence regions. Agreement between independent evidence and the confidence profile is the strongest available validation short of experiment. 8. **Compare with homologs.** Align the model to an experimental structure of a relative when one exists, and confirm that conserved cores superpose. Regions that are both sequence-divergent and low in confidence should be reported as unresolved. 9. **Classify every region.** For each part of the chain, record reliable, plausible but flexible, probable disorder, or unresolved. Only the reliable class supports detailed structural claims. ## Worked Examples from Two Non-Model Systems Two published studies show how the scores behave at scale in organisms far from the training mainstream. ### Photosystem II D1 Across Photosynthetic Lineages The D1 protein forms the reaction center core of photosystem II and is conserved across oxygenic photosynthesis. A survey published in 2025 used AlphaFold to predict D1 structures across a broad range of photosynthetic organisms and obtained high-confidence models throughout [1]. Because D1 is ancient and near-universal in phototrophs, each prediction drew on a deep cross-lineage alignment, and the confidence profile tracked biology closely. Regions that structural alignment identified as conserved across D1 groups coincided with high pLDDT regions, while varying pLDDT in the DE loop and the terminal regions appeared to reflect different degrees of structural flexibility or disorder [1]. The same study moved from single models to evolution. A structural phylogeny built with Foldtree agreed with earlier sequence-based trees and supported a scenario in which the far-red light-adapted D1 form and the insertion-bearing D1INT form descended from an ancestral G4-type D1 [1]. The lesson for non-model work is that when a protein family is deep and conserved, confidence scores can be read almost directly as structural biology, and the models can support downstream evolutionary analysis. ### Odorant Receptors in Two Longhorn Beetles Insect odorant receptors are seven-transmembrane proteins, and in beetles they form large multigene families that evolve quickly. A genomic case study examined the predicted structures of odorant receptors from two longhorn beetles with divergent life histories, the highly specialized red milkweed beetle (*Tetraopes tetrophthalmus*) and the broadly polyphagous Asian longhorned beetle (*Anoplophora glabripennis*) [2]. Unsupervised clustering of pairwise protein comparisons found a strong correlation between structural and sequence similarity, supported the established evolutionary classification of receptor groups, and brought previously unrecognized subclusters into view [2]. The study also identified specific receptors with substantial structural divergence despite relatively low sequence divergence from their paralogs, meaning duplicated genes within the same species, which marked them as candidates for unusual evolutionary trajectories [2]. The lesson differs from the D1 case. In a young, rapidly expanding family, scores matter most as comparative tools across paralogs and genomes, and structure-aware clustering can recover biology that sequence similarity alone would miss. ## Using Confidence Scores Comparatively Across Species and Families The soundest comparison is within a family. Align orthologs, meaning the same gene across species, extract the per-residue pLDDT profile for each, and compare position by position. Positions where confidence drops in one lineage but not others often mark lineage-specific insertions, accelerated substitution, or disorder gained or lost on one branch. In practice this looks like a table of orthologs with a column per species and a row per alignment position, where a stretch of 85 to 95 across all species marks the conserved core and a dip below 50 in one species flags a lineage-specific change. Handled this way, the scores stop being a quality label and become an evolutionary signal. Survey-mode comparison across families and genomes answers a different question, namely where in a proteome the predictions are strong enough for downstream modeling. The beetle study took this approach when it compared confidence scores across odorant receptor families and between the two genomes [2]. Results at this level guide resource allocation and flag the families that need experimental attention first. Treat absolute values with care when the comparison crosses family boundaries. Calibration depends on family properties such as length, secondary structure content, disorder content, and oligomeric state, so a pLDDT of 80 in a globular kinase and a pLDDT of 80 in a transmembrane receptor do not carry identical implications. Relative statements within a family are robust. Absolute statements across families need qualification. Predicted structures also unlock analyses that sequences alone cannot support, including structure-based phylogenies and clustering by structural similarity. Both examples above moved in this direction, one through a structural phylogeny of photosystem cores and the other through clustering of pairwise structural comparisons [1][2]. Confidence scores gate such analyses, because trees and clusters built from low-confidence regions inherit that uncertainty. ## Common Mistakes and Limitations ### Misconceptions Students Bring to AlphaFold Output 1. **"Low pLDDT means the model is wrong there."** Often the model is fine and the region is genuinely flexible or disordered, which is why termini and long loops turn orange across thousands of predictions. The score marks reduced positional certainty, not certain error. 2. **"pLDDT is the probability the structure is correct."** It is a predicted lDDT, an expected local similarity to the true structure, not a probability of correctness. 3. **"High pLDDT everywhere means the whole arrangement is solved."** Domain orientation lives in PAE, and a model can be locally excellent while globally ambiguous. 4. **"Scores are absolute and portable."** An 80 in one family is not the same claim as an 80 in another, and comparisons are safest within families. 5. **"Confidence validates function."** The scores describe geometric certainty. They say nothing about ligand identity, catalytic activity, or binding affinity. 6. **"Orange regions are junk to delete."** Low-confidence regions often carry the biology, including disorder-mediated regulation, and deleting them distorts downstream analyses. ### Practical Pitfalls in Analysis Pipelines - **Reading the B-factor column as experimental uncertainty.** Validation tools written for crystal structures will misreport AlphaFold output unless configured to treat B-factors as pLDDT. - **Averaging pLDDT into one number per protein and comparing those averages.** Means hide the profile, and the profile is where the information lives. - **Building docking or mutagenesis arguments on loops with pLDDT below 70.** Those regions move, and the model shows one snapshot among many. - **Forgetting what the model omits.** A standard monomer prediction describes the polypeptide chain. Bound cofactors, metals, post-translational modifications, membrane lipids, and alternative conformational states are outside it. A predicted photosystem II subunit model describes the protein scaffold, not the chlorophylls and quinones bound in the active complex. - **Letting distant templates bias the model.** Template hits from remote relatives can pull a divergent protein toward the wrong scaffold. Rerunning without templates is a cheap diagnostic. - **Mixing versions without labels.** AlphaFold2 and AlphaFold 3 differ in architecture and output details, and scores are not interchangeable across releases. Record versions in the methods. - **Treating a single state as the truth.** Database models represent one conformation. Proteins with documented alternative states need explicit modeling or experiment. ### What the Scores Cannot Tell You Confidence scores quantify geometric certainty. They do not establish function, identify ligands, resolve conformational ensembles, or prove oligomeric stoichiometry unless complexes are modeled explicitly. They also do not substitute for experimental validation. For any specific project, the sensible next step is targeted experiment or consultation with someone who knows the protein family, because predictions generate hypotheses and hypotheses need tests. ## Quick Review - pLDDT grades local backbone confidence per residue on a 0 to 100 scale, with standard bands above 90, 70 to 90, 60 to 70, and below 60, and values below 50 often indicating disorder. - PAE reports expected positional error in Ångströms for every residue pair and is the tool for judging domain packing and interfaces. - Low pLDDT usually means flexibility or disorder, not a wrong fold. - High pLDDT does not guarantee correct inter-domain geometry, so always check the PAE matrix. - Evolutionary context sets expectations. Deep conserved families predict well in distant species, while lineage-specific expansions are the weak spot. - Compare scores within families rather than across them, and log MSA depth and software version with every prediction. - Scores describe the chain, not its cofactors, modifications, partners, or alternative states, and they never replace experimental validation. ## Frequently Asked Questions ### What does pLDDT actually measure? It measures expected local accuracy. pLDDT is the model's prediction of the lDDT score at each residue, which reflects how well the local backbone geometry around that residue is expected to match the true structure, on a 0 to 100 scale. It is a per-residue grade for the neighborhood of each Cα position, not a statement about the whole protein. ### How high does pLDDT need to be before I can use a model? There is no universal cutoff, because the answer depends on the use. Backbone-level claims such as fold recognition and domain architecture are generally safe in the 70 to 90 band and above. Mechanistic claims about active-site geometry are best supported above 90. Below 60, treat the region as a hypothesis, and below 50 suspect disorder. ### Why are the termini and some loops always colored orange? Because those regions are usually mobile or disordered. Low pLDDT at termini and in long loops reflects genuine conformational variability rather than a defect in the prediction, and the pattern is consistent enough that low-confidence termini serve as a useful disorder flag. ### Can I compare confidence scores between two species or two protein families? Yes within a family, cautiously across families. Position-wise comparison of pLDDT profiles among aligned orthologs is a sound way to locate lineage-specific divergence. Comparing absolute scores across different families is weaker, because score calibration depends on properties such as length, disorder content, and oligomeric state. ### Do I need to install AlphaFold locally, or is the AlphaFold Server enough? For a few sequences, the AlphaFold Server is enough and is the fastest route. For proteome-scale batches or runs that need custom alignments and parameters, ColabFold or a local installation fits better. Check the AlphaFold Protein Structure Database first, because the model you need may already exist. ### Does high confidence mean the prediction is experimentally validated? No. Confidence scores estimate agreement with a hypothetical true structure based on evolutionary and structural patterns learned during training. They are self-assessments, not measurements. Experimental validation, or at minimum comparison with structures of homologs, remains necessary before treating any model as established fact. ## Sources 1. [Exploring the Structural Diversity and Evolution of the D1 Subunit of Photosystem II Using AlphaFold and Foldtree.](https://pubmed.ncbi.nlm.nih.gov/40401773/) 2. [Learning the Structural Diversity of Olfactory Receptors: A Genomic Case Study in Two Longhorn Beetles (Cerambycidae: Lamiinae).](https://pubmed.ncbi.nlm.nih.gov/42355320/)