A structure predictor returns a complete set of atomic coordinates for any sequence you hand it, including sequences that adopt no stable structure at all. The network never declines. What separates a usable model from a confident-looking artefact is not the coordinates but the confidence arrays shipped alongside them, and for peptides — which sit at the edge of the regime these methods were built and benchmarked on — reading those arrays correctly is most of the work.

AlphaFold2 can regularly predict protein structures with atomic accuracy even in cases where no similar structure is known, and was validated in the CASP14 assessment at accuracy competitive with experimental structures in a majority of cases [1]. That validation was run on protein targets. A 14-residue linear peptide, a disulfide-rich scaffold, a designed macrocycle and a short motif bound in a receptor groove are four different problems, and the confidence metrics behave differently on each.

What pLDDT and PAE measure

pLDDT is the model's predicted per-residue score on the lDDT-Cα metric: a self-assessment of local geometric accuracy, reported residue by residue, and intended for judging confidence within a single domain [2]. It says nothing about whether two confidently predicted segments are correctly placed relative to one another. That is the job of the predicted aligned error. PAE at position (x, y) is the expected positional error at residue x if the predicted and actual structures are aligned on residue y, which makes it the array to consult when you want to know how far to trust the relative position and orientation of two parts of a model [2].

The bands are calibrated and worth holding precisely. In the prediction covering 98.5% of human proteins, 58% of residues were predicted confidently at pLDDT above 70 — a cut-off corresponding to a generally correct backbone — and 36% of all residues fell into the highest band above pLDDT 90, the cut-off above which AlphaFold χ1 rotamers were 80% correct [3]. Read operationally, that gives three regimes: above 90 you may argue about side-chain rotamers; between 70 and 90 you may argue about backbone topology and little else; below 70 the geometry is not evidence for anything.

Two failure modes follow from the split between the two metrics. A model can be locally excellent and globally wrong — every segment at pLDDT 95, the arrangement between them invented — and only the PAE will show it. And a model can carry low pLDDT for reasons that are not network failure at all. The subtleties involved in evaluating model quality with both metrics are a large part of why a predicted model should be used to generate testable hypotheses rather than treated as a de facto ground-truth structure [4].

Low pLDDT on a short peptide is a statement about the peptide

For disordered sequence, pLDDT is not simply a failure flag. Assessed against traditional sequence-based disorder predictors, pLDDT provides an excellent measure of residue-wise disorder [5]. The same assessment found that the naive alternative — reading the predicted secondary structure and calling everything that is not helix, strand or hydrogen-bond-stabilised turn disordered — produces a dramatic overestimation of disorder [5]. The confidence array, not the cartoon, is the disorder read-out.

The effect is systematic rather than incidental. Intrinsically disordered regions carry systematically lower prediction confidence than either fold-switching or canonical single-fold proteins, and they do so regardless of sequence conservation; disordered-region alignments show low conservation, whereas fold-switching regions are conserved at rates statistically similar to single-fold proteins [6]. So a short synthetic peptide returning pLDDT in the 30s is not necessarily reporting a failed run. It is often reporting that the sequence offers no single dominant conformation for the network to commit to, which is itself a result — and one to test on the bench rather than by rerunning with more seeds.

The inverse reading is where the errors are expensive. Confident is not the same as unique. AlphaFold2 assigns confident structures to nearly 15% of human intrinsically disordered regions, and comparison with experimental NMR data for regions known to conditionally fold shows that it often predicts the structure of the conditionally folded state, identifying conditionally folding disordered regions at a precision as high as 88% at a 10% false positive rate [7]. A confident prediction for a motif that is disordered in isolation may therefore be a prediction of its bound or conditionally folded form, not a claim that it is folded in your tube.

What the peptide-specific benchmark shows

A systematic test run on peptides rather than proteins compared AlphaFold2 predictions for 588 peptide structures between 10 and 40 amino acids against experimentally determined NMR structures as the reference [8]. Accuracy was high for α-helical, β-hairpin and disulfide-rich peptides, and at least as good as methods developed specifically for peptide structure prediction [8]. That is the encouraging half.

The cautionary half is more useful. The same benchmark reported shortcomings in predicted Φ/Ψ angles and in disulfide bond patterns, and — the finding that should change how you use the output — the lowest-RMSD structures failed to correlate with the pLDDT-based ranking of structures [8]. Within a set of models for one peptide, the highest-confidence model is not reliably the closest to the NMR reference. Ranking by confidence selects a model; it does not identify the correct one, and the authors' own conclusion is that additional steps may be necessary to analyse and validate the results [8].

One structure per seed is not an ensemble

On a set of 98 fold-switching proteins, which adopt at least two distinct yet stable secondary and tertiary structures, 94% of AlphaFold2 predictions captured one experimentally determined conformation but not the other, while estimated confidences were moderate-to-high for 74% of fold-switching residues [6]. The reading offered is that high confidence for fold switchers indicates the network using pattern recognition to search for one most probable conformer rather than protein biophysics to model a structural ensemble [6]. The disorder literature reaches the same place from the other side: AlphaFold2 predictions do not reveal functionally relevant structural plasticity within disordered regions and cannot offer realistic ensemble representations of conditionally folded ones [7].

This is the line that separates structure prediction from conformational sampling. A molecular dynamics trajectory propagates a potential in time and yields populations; a prediction yields one structure per seed plus a confidence score, and that score is not a population. Deliberately degrading the input makes the point: models generated by the default pipeline are conformationally homogeneous and nearly identical to one another, and only stochastic subsampling of the input multiple sequence alignment generated accurate models in multiple conformations, with models at the extremes of those conformational distributions among the most accurate at an average template modeling score of 0.94 for transporters and receptors absent from the training set [9]. The same work is explicit that predicting ensembles of biophysically relevant states is a requirement for the next generation of algorithms, not a property of this one [9].

Peptide–protein complexes and the alignment-depth problem

Complex prediction is the application most peptide researchers actually want, and it survives the missing-alignment problem better than the monomer case suggests. A simple implementation of AlphaFold2 generates peptide–protein complex models without requiring multiple sequence alignment information for the peptide partner, and can handle binding-induced conformational changes of the receptor [10]. That dissolves the most common objection to using the method on a synthetic or designed peptide: the peptide has no evolutionary record and does not need one, because the receptor's alignment already constrains the binding site the peptide is docked into.

Benchmarked head to head, AlphaFold-Multimer produced peptide–protein complex models of acceptable or better quality (DockQ ≥ 0.23) for 66 of 112 complexes, 25 of them high quality (DockQ ≥ 0.8), against 23 or 47 acceptable and only four or eight high-quality models from energy-based docking and interaction-template approaches respectively [11]. Randomly perturbing the network weights to force wider sampling raised acceptable models from 66 to 75 and the median DockQ of first-ranked models from 0.47 to 0.55, while the best achievable DockQ rose from 0.58 to 0.72 — read by the authors as evidence that selecting the best model from a generated set is still a challenge [11]. Used as a screen rather than a modeller, the same method recovered 26% of the possible interactions at 85% precision with a 1% false positive rate [11].

Where alignment strategy genuinely bites is delimitation. On a dataset of protein–peptide complexes involving intrinsically disordered regions, deliberately chosen to be non-redundant with the structures used in AlphaFold2 training, submitting full-length sequences gave AlphaFold2-Multimer only a 40% success rate at identifying the correct site and structure of the interface; delineating the interaction region into fragments of decreasing size and combining different strategies for integrating evolutionary information raised that success rate to as high as 90% [12]. What you submit is a modelling decision, and handing the model two full-length proteins to locate a ten-residue motif is the worst of the available choices.

The same study examined whether the AlphaFold2 confidence score can discriminate between alternative binding partners, and reported both advantages and limitations, noting that the task is particularly challenging for small interaction motifs [12]. Treat interface confidence as a ranking aid within one target, not as an affinity surrogate across targets. Nothing in the output is a binding constant, and the experimental methods that produce one — surface plasmon resonance, isothermal titration calorimetry, fluorescence polarisation — remain the only route to that number.

What is not in the model

AlphaFold predicts structure from amino acid sequence alone [1], and the predictions do not take into account ligands, covalent modifications or other environmental factors [13]. There is no bound metal ion, no cofactor, no lipid, no membrane and no buffer in the calculation. When predictions were compared directly against experimental crystallographic maps, many matched remarkably closely, but in other cases even very high-confidence predictions differed from the maps on a global scale through distortion and domain orientation, and on a local scale in backbone and side-chain conformation [13]. The recommendation attached to that comparison is specific: weigh the confidence when interpreting a prediction, and carry out experimental structure determination to verify structural details, particularly those that involve interactions not included in the prediction [13].

For a peptide that catalogue is long — a chelated zinc, a phosphorylated residue, an acyl chain, or a micelle or detergent the conformation depends on. Newer architectures narrow it. AlphaFold 3 uses a substantially updated diffusion-based architecture capable of predicting the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues, with far greater accuracy for protein–ligand interactions than state-of-the-art docking tools [14]. Entities outside that set remain absent, and adding a ligand does not change the epistemic status of the result: a modelled ligand pose is one more prediction, not an observation.

Sequence perturbation deserves separate warning, because analogue work runs straight into it. AlphaFold2 predicts structures based on those available in the Protein Data Bank rather than from the fundamental driving forces of protein folding, and has been found largely unable to predict when a point mutation causes defective protein folding [15]. If you substitute one residue in an analogue series and the predicted structure barely moves, that is close to uninformative about whether the substitution is structurally tolerated. The covalent-detail version of the same limitation appears in the peptide benchmark's difficulty with disulfide bond patterns [8].

Designed and cyclic sequences

Designed peptides are the extreme case of the missing-alignment problem, and also the clearest demonstration that it can be engineered around. The development of deep learning methods for accurately designing small cyclic peptides has been slow, mostly for lack of sufficiently large training sets [16]. An AlphaFold2-based approach to cyclic peptide structure prediction, sequence redesign and de novo hallucination nonetheless yielded over 10,000 structurally diverse designs predicted to fold into the designed structures with high confidence, and X-ray crystal structures for eight tested de novo designed sequences matched the design models closely, at RMSD below 1.0 Å [16]. The hallucinated scaffolds were then used as starting points for binders with nanomolar IC50 against two protein targets [16].

Two things follow. Predicted confidence for a designed sequence with no natural homologues can be meaningful, so the network is not merely retrieving from its training set. But the atomic-accuracy claim was settled by eight crystal structures, not by the confidence score, and those eight were a small fraction of the ten thousand confident designs. Confidence selected the candidates; crystallography decided them.

Validating a prediction

The methods that settle a predicted peptide structure are the ones already on the bench, and predicted models can be integrated with experimental data in the form of SAXS, solution NMR, cryo-EM and X-ray diffraction [4]. The pairings are specific. Solution NMR is the reference the peptide benchmark itself used, across 588 peptides of 10 to 40 residues [8], and NMR data is what established that confident predictions for conditionally folding disordered regions correspond to the folded state [7]. Crystallographic maps are what exposed distortion, domain-orientation and side-chain errors in high-confidence models [13].

Before a predicted peptide model enters a figure or drives a synthesis decision, three checks are worth making explicitly. State which class of claim the per-residue confidence licenses — rotamer, backbone, or none [3]. Read the PAE before asserting anything about how two segments sit relative to each other [2]. And name the entity you have omitted — ion, cofactor, lipid, covalent modification — then ask whether the conformation you are about to reason from depends on it [13]. A prediction that survives all three is still a hypothesis; it is just a hypothesis worth the cost of testing.