Abstract
Beyond predicting unique native globular protein structures, AlphaFold has been used to sample conformational diversity in globular proteins. Here, we asked whether this capacity can be transferred to amyloid fibrils formed by short peptides, despite the fundamental difficulties posed by their polymorphism, sensitivity to environmental variables such as pH and cofactors, limited evolutionary constraints, and underrepresentation of amyloids in structural databases. Using a comprehensive set of short amyloidogenic peptides, we show that sets of AlphaFold3 (AF3) models (pseudo-ensembles) are accurate predictors of parallel versus antiparallel β-sheet organization, correctly assigning 112 of the 117 peptides for which a clear preference is predicted. Moreover, the simultaneous sampling of both organizations provides a signature of structural diversity, as shown for the polymorphic peptides MVGGVVIA and LYIQWL. Modeling of the precise steric-zipper architecture is less reliable and biased toward compact arrangements, while prediction in the presence of cosolvents and metal ions does not reproduce solvent-mediated interfaces or noncanonical coordination geometries. Thus, although AF3 does not fully capture the physicochemical complexity of amyloid assemblies, its pseudo-ensembles provide a rapid and useful approach for predicting β-sheet organization and identifying amyloidogenic sequences with potential structural diversity.


Introduction
The transformative impact of AlphaFold and related AI-based structure prediction methods on structural bioinformatics lies in their unprecedented ability to infer the native structures of globular proteins directly from sequence, with a speed and accuracy that surpass those of classical computational approaches. − AlphaFold predictions depend on evolutionary information in the form of residue covariation encoded in multiple-sequence alignments (MSAs) and assume a one-to-one sequence–structure relationship consistent with Anfinsen’s thermodynamic hypothesis. According to this hypothesis, for a given sequence, the native state of a globular protein is the unique, kinetically accessible conformation corresponding to the global free-energy minimum, and this is what the model is trained to infer. This concept raises the question of whether AlphaFold has implicitly learned a scoring function analogous to free energy: that is, a function with a critical point corresponding to the native structure, and whose value for a given protein conformation correlates with the probability of observing the target sequence in that conformation. Equivalently, this amounts to asking to what extent the scoring function learned by AlphaFold reflects thermodynamic principles.
There is convincing evidence that this connection exists in AlphaFold2. In a systematic analysis of a large data set of native structures and decoys, coevolutionary information was shown to be critical for identifying the general region of the native conformation, which is then refined to yield a highly accurate model of the global minimum. This concept is not limited to protein structure prediction: a recent study explored this thermodynamic connection in a protein docking diffusion model, showing qualitative agreement between the learned potential and the energy funnels generated with physics-based energy functions.
Given the evidence that the AlphaFold-predicted structure corresponds to a thermodynamic minimum, a natural next step in protein structure prediction is to interrogate systems that do not obey a one-to-one sequence–structure relationship. These include, for instance, globular proteins that can populate alternative conformations, corresponding to distinct free-energy minima. Since coevolutionary signals from the MSA have been suggested to provide the initial structural guess, attenuating these signals is expected to increase the conformational diversity of the predicted structures, potentially enabling the exploration of alternative conformations. Along these lines, several approaches based on MSA subsampling schemes have been developed, e.g., AF-cluster, SF-cluster, and AFsample2. Although it remains debated whether this conformational sensitivity reflects genuine modeling ability rather than memorization of alternative conformations present in the training set, heuristic strategies can successfully exploit this imperfect capacity for conformational sampling. In this context, we have shown that collections of AlphaFold structures can, in specific cases, behave as pseudoconformational ensembles (hereafter, pseudo-ensembles) by leveraging the stochastic character of AlphaFold predictions. This approach has enabled the quantification of the effects of missense mutations on the thermal stability of enzymes and the estimation of the relative populations of alternative conformations in alternate-frame folding systems. ,
Given the potential of AlphaFold to generate alternative conformations in globular proteins, we next ask whether it can also provide predictive value in the context of the more conformationally heterogeneous amyloid state. Amyloids constitute a paradigmatic case in which AlphaFold prediction should fail: a single sequence populates multiple structurally distinct states, including disordered monomers, partially folded intermediates, oligomers, and fibrils. Moreover, amyloid fibrils are polymorphic and form under kinetic, rather than purely thermodynamic, control, in contrast to the thermodynamic hypothesis generally applicable to globular proteins. The resulting conformational ensemble is further modulated by environmental factors such as pH and cofactors. , In this context, the intrinsic polymorphism of amyloids may be regarded as an extreme manifestation of the ability of a single amino acid sequence to adopt multiple alternative conformations. In addition, because the aggregated states of amyloidogenic proteins are typically associated with disease (e.g., α-synuclein, tau, amyloid-β, and prion protein), they are generally not subject to positive evolutionary selection. This eliminates one of the principal advantages conferred by AlphaFold’s reliance on evolutionary information. Finally, amyloid structures represent only a small fraction of the experimentally determined high-resolution structures available for training AI-based models, rendering the amyloid state a major blind spot in protein structure prediction.
However, despite the limited availability of training examples, a recent study has shown that AlphaFold can qualitatively recapitulate the morphology and compactness of fibers, micelles, and nanotubes formed by self-assembling peptides, as well as the secondary-structure preferences of the amino acids within these assemblies. In this work, we evaluate the usefulness of AlphaFold3 (AF3) for characterizing the amyloid state by assessing the ability of AF3 pseudo-ensembles to predict the aggregated organization of short peptides corresponding to aggregation-prone regions (APRs) of amyloidogenic proteins such as the islet amyloid polypeptide (IAPP), amyloid-β, tau, the prion protein, and α-synuclein.
These short motifs, for example 306VQIVYK311 in tau, are critical determinants of aggregation: they seed primary nucleation and form β-sheets that pack into typically dry steric zippers, thereby stabilizing fibrillar aggregates across distinct polymorphs. In tau, this mechanism was recently demonstrated by time-resolved cryo-EM, which showed that formation of a “first intermediate amyloid,” in which the APR-containing segment 302GGGSVQIVYKPVDLS316 constitutes the sole ordered core, is the initial and rate-determining step of assembly, and is shared across distinct polymorphs. Following this primary nucleation event, the intrinsic polymorphism of this region drives continuous conformational remodeling, giving rise to a succession of structurally distinct polymorphic intermediates and end states. A similar behavior during assembly has also been reported for a disease-associated variant of human IAPP, suggesting that APRs play a central role in navigating the complex energy landscape of fibril maturation. , Rather than assembling directly into a single rigid structure, growing filaments sample a wide range of packing modes, reflecting the polymorphism of APRs, corresponding to distinct structural arrangements of these motifs. Time-resolved structural analyses indicate that this continuous remodeling is functionally important, as it prevents kinetic trapping in highly frustrated dead-end states and thereby enables sustained progression toward mature fibrils.
Given this central role of APRs, the ability to predict the potentially polymorphic aggregated organization of APRs, and to determine how that organization is modulated by mutations and cofactors, is of considerable interest for understanding the molecular basis of amyloid diseases, − as well as for the design of artificial amyloids as nanomaterials. − For instance, metal cations, including Cu2+ and Zn2+, have been implicated in amyloid diseases because their interactions with amyloidogenic proteins can substantially alter aggregation pathways, oligomer/fibril stability, and toxicity. Metal ions can also play a functional role in catalytic amyloids, where fibrillar assemblies organize metal-binding residues into metalloenzyme-like active sites, enabling reactions such as Zn2+-dependent ester hydrolysis.
Polymorphism in short peptides can occur at different structural levels. At the first level, peptide units mostly adopt a β-strand conformation and assemble in either parallel or antiparallel β-sheets, the dominant structural element of amyloids. , The next level of organization concerns how β-sheets are arranged with respect to each other; this gives rise to a typically compact, commonly hydrophobic and solvent-excluded interface composed of two β-sheets, known as a steric zipper (Figure A). ,
1.

(A) Representation of the ten parallel and antiparallel steric-zipper classes. Green and purple are used to distinguish the two opposite faces of each β-strand, while color intensity indicates the relative position of each β-sheet with respect to the reader, with paler shades denoting the more distant sheet in the top view. Adapted from ref under the Creative Commons Attribution 4.0 International License (CC BY 4.0). (B) Crystal structures of the polymorphic peptides MVGGVVIA and LYIQWL. Amino acid residues are shown as sticks and colored by residue number; water oxygen atoms are shown as spheres; and cosolvent molecules are shown as thin black sticks. (C) Size distribution (number of amino acids) of the database of 146 short peptides used in this work. (D) Distribution of GRAVY (Grand Average of Hydropathy) index values for the same peptides. Negative values indicate hydrophilic sequences, whereas positive values indicate hydrophobic sequences.
Homosteric zippers (i.e., those formed by the same peptide on both opposing β-sheets) can organize into the ten classes shown in Figure A, six of which originate from parallel (classes 1–4, 9, and 10) and four of which from antiparallel (classes 5–8) β-sheets. To the best of our knowledge, classes 9 and 10 remain hypothetical, as no steric-zipper structures belonging to these classes have yet been experimentally determined. In addition, at least one experimentally resolved structure exhibits a more complex topology that does not conform to any of the ten established classes.
Inter- and intra-β-sheet interactions exert conflicting demands on the global packing organization. From the point of view of backbone–backbone hydrogen bonds within individual β-sheets, the antiparallel organization is more stable, yielding a larger number of hydrogen bonds. However, steric zippers deriving from antiparallel β-sheets have less intersheet interdigitation, which leads to poorer side-chain–side-chain contacts at the interface. The preferred organization is therefore determined by a subtle energetic balance between these effects as a function of the composition of the peptide and environmental conditions. This is exemplified by the polymorphic 35MVGGVVIA42 segment of amyloid-β, which crystallizes into two forms, one of which is parallel (class 2) and the other antiparallel (class 7), shown in Figure B. An even more extreme example of polymorphism is that of the LYIQWL peptide from Tc5b. As a function of pH and cosolvents (acetonitrile and ethanol), this peptide can explore a wide variety of polymorphs, some of which are shown in Figure B. These polymorphs span multiple steric-zipper classes (among them, 1, 4, and 8), diverse shifts along the spine axis, and different degrees of “wetting” of the interface by solvent molecules.
Further complexity can arise at a higher level of organization, namely, in how peptides pack at the nanoscale level, through the interactions among the many β-sheets held together by zippers (dry or partially wet) and through the interactions of the peptide termini. Of note, the available high-resolution structural information mostly comes from X-ray crystallography, i.e., from conditions in which peptides pack in a pseudoinfinite 3D arrangement that may not exactly match the organization in solution or in the biological environment. For the polymorphic LYIQWL peptide, a strong topological mirroring has been demonstrated between the solution state and the mature crystal phase. Key similarities include their identical response to environmental triggerssuch as the transition from antiparallel to parallel β-sheets when ethanol is addedand the detection of stabilizing motifs like Gln ladders and aromatic stacking in solution. On this basis, the available crystallographic structures of short peptides are used here as the reference for assessing the preferred organization of peptides in the biological environment.
In this work, we performed AF3 predictions on a comprehensive database of 146 short peptides for which experimentally determined structures are available (113 from X-ray crystallography, 32 from electron crystallography, and one from solid-state NMR). This database is largely derived from the one recently reported by Perczel and co-workers and covers peptides containing from 4 to 11 residues, with hexapeptides, by far the most widely studied, representing over 80 entries (Figure C). Peptides cover a wide range of polarity: most sequences are hydrophobic and lead to classic, tightly packed, solvent-excluded interfaces (positive GRAVY index in Figure D), but polar sequences (negative GRAVY index in Figure D), which can lead to the formation of solvated interfaces, are also present. We show that, despite the well-established caveats surrounding the use of AF3 in amyloid systems, AF3 pseudo-ensembles of short amyloidogenic peptides reliably discriminate between parallel and antiparallel β-sheet organization (112 of the 117 peptides with a clear predicted preference are correctly assigned). This is likely because AF3 captures fundamental packing principles learned from globular proteins. In contrast, AF3 pseudo-ensembles are less effective at predicting steric-zipper architectures and only qualitatively recapitulate interactions with metal ions. Importantly, AF3 pseudo-ensembles also capture parallel/antiparallel polymorphism, highlighting their value as predictors of structural diversity.
Methods
AlphaFold3 Structure Prediction
AF3 structure predictions were performed using version 3.0.1. For each peptide, predictions were run over a range of numbers of copies of the peptide: 4, 6, 8, 10, 12, 14, 16, 18, and 20. The rationale for this choice is that 4 is the minimal number of units that can lead to a zipper-like assembly (i.e., two β-sheets of 2 peptides each), and progressively increasing the number of peptides can allow the system to explore different, potentially polymorphic arrangements. For each number of peptide copies, 100 models were generated using 20 seeds and five diffusion samples per seed, resulting in a total of 900 models for each peptide. Predictions were run without MSA calculation because the MSA carries evolutionary information on the globular state and constrains conformational diversity, as discussed in the Introduction, and because a short peptide can be spuriously matched against multiple sequences of longer globular proteins, which can lead to contradictory information that can degrade the prediction quality. For the same reason, the search for structural templates was also skipped. Sample input files for prediction are provided in the Supporting Information. Under these conditions, the generation of 100 models for the largest assembly (20 units) of a hexapeptide takes ∼6 min on an A6000 GPU. All predicted structures are available at https://zenodo.org/records/21510614. All analyses were performed with custom Python scripts described below. Molecular graphics were prepared with PyMOL.
Peptide Crystallographic Data
The peptide database is based on the study by Perczel and co-workers. The reference crystallographic structures and steric-zipper class assignments were taken from the same work.
Secondary Structure Assignment
The parallel or antiparallel organization of β-strand pairs was assigned using a Python script operating directly on the CIF coordinates generated by AF3. For every pair of peptide chains, all possible relative residue registers were evaluated independently for parallel and antiparallel arrangements. In the parallel search, residues were matched while preserving the N-to-C direction of both chains, whereas in the antiparallel search, the residue order of the second chain was reversed. Candidate registers containing fewer than four matched residues were discarded. For each remaining register, an alignment score was calculated as the mean Euclidean distance between the Cα atoms of the matched residues, and the lowest-scoring register was retained independently for the parallel and antiparallel orientations. A chain pair was assigned only if three conditions were simultaneously satisfied: (i) the best register contained at least four aligned residues; (ii) its mean Cα distance was ≤5.3 Å; and (iii) its score was at least 0.75 Å lower than that of the best register obtained for the alternative orientation. Chain pairs not satisfying these criteria were left unassigned rather than being forced into either class. Parallel and antiparallel pair counts were subsequently obtained for each structure and averaged over the 100 structures constituting each AF3 pseudo-ensemble.
The script is available in the Supporting Information and uses the standard-library modules argparse, csv, itertools, and pathlib, as well as the third-party packages numpy and Bio (Biopython).
An analysis of the robustness of the classification with respect to the Cα distance cutoff is included in the Supporting Information (Figures S1 and S2).
RMSD Calculation
To quantify the local structural agreement between AF3 models and the corresponding experimentally determined amyloid structures, we calculated a symmetry-aware RMSD for adjacent pairs of β-strands. The peptide sequence to be analyzed was identified from the AF3 models and was required to match exactly a peptide chain present in the experimental structure. When the experimental asymmetric unit did not contain the complete local β-sheet arrangement, crystallographic symmetry operations defined in the CIF file were applied to generate neighboring peptide copies. For every adjacent strand pair identified in an AF3 model, the pair was compared against all valid experimental neighboring-strand templates. Structural superposition was performed by least-squares Kabsch alignment using the backbone atoms N, Cα, C, and O of both peptide chains simultaneously. Both possible assignments of the two AF3 chains onto the two experimental chains were tested, and the assignment giving the lower RMSD was retained. For each AF3 adjacent pair, the experimental template giving the best local structural match was retained. Finally, the median backbone RMSD over all valid adjacent strand pairs in a given AF3 model was used as the model-level structural similarity metric. Models for which no valid adjacent strand pair could be identified were reported as having no RMSD value.
The script is available in the Supporting Information and uses the Python standard-library modules argparse, glob, itertools, math, sys, collections, dataclasses, pathlib, and typing, together with the third-party packages Gemmi, NumPy, and pandas.
Zipper Class Assignment
Steric-zipper classes were assigned from discrete relative-orientation operators describing the two interacting β-sheets using Stroud’s classification of steric zippers. First, a graph was constructed in which each peptide chain represents a node and two chains are connected when at least one interchain backbone N–O distance is <4.0 Å. Each connected component therefore represents a backbone hydrogen-bond-connected β-sheet. The components were ranked by strand count, and the two largest were selected as the two sheets forming the steric-zipper interface; each was required to contain at least two strands.
For each strand, a local orthonormal frame (x, y, z) was defined, with z along the N-to-C direction, x describing the alternating side-chain face, and y = z × x. Relative orientations were encoded as
| 1 |
Within each sheet, the dominant intrasheet signature (s z , s x ) was required to represent at least 80% of hydrogen-bonded strand pairs. This criterion excluded β-sheets with a high mix of parallel and antiparallel strands from the analysis. Across the interface, each strand was compared with the closest strand in the opposite sheet within 15 Å, yielding a dominant intersheet signature (s x , s y , s z ) that was likewise required to represent at least 80% of the evaluated pairs. The final steric-zipper class was assigned according to the mapping shown in Table , where (s z , s x ) denotes the dominant intrasheet signature and (s x ,s y , s z ) denotes the dominant intersheet signature.
1. Mapping of Intra- and Intersheet Orientation Signatures to Steric-Zipper Classes.
| (s z , s x ) | (s x , s y , s z ) | Steric-zipper class |
|---|---|---|
| (+,+) | (−,+,−) | 1 |
| (+,+) | (+,+,+) | 2 |
| (+,+) | (−,–,+) | 3 |
| (+,+) | (+,–,−) | 4 |
| (−,+) | (−,+,−) or (−,–,+) | 5 |
| (−,+) | (+,+,+) or (+,–,−) | 6 |
| (−,−) | (+,+,+) or (−,+,−) | 7 |
| (−,−) | (+,–,−) or (−,–,+) | 8 |
| (+,−) | (+,+,+) or (−,–,+) | 9 |
| (+,−) | (+,–,−) or (−,+,−) | 10 |
Here, each sign indicates whether the corresponding local axes of two strands point in the same (+1) or opposite (−1) directions. For the intrasheet descriptors, s z = +1 and −1 identify parallel and antiparallel strand organization, respectively, because the local z axis follows the N-to-C direction of each strand. Similarly, s x = +1 indicates that the local x axes of neighboring strands point in the same direction, corresponding to the same side-chain-face orientation, whereas s x = −1 indicates opposite x-axis orientations and therefore alternating side-chain faces (antiparallel classes 7 and 8 as well as hypothetical parallel classes 9 and 10, face = back). For the intersheet descriptors, s x , s y , and s z indicate whether the corresponding x, y, and z axes of paired strands from opposite sheets are aligned (+1) or reversed (−1). Thus, s z directly reports whether their N-to-C directions are the same or opposite, while s x and s y describe the relative orientation of the side-chain-face and spine directions, respectively. For example, classes 7 and 8 both correspond to antiparallel, face = back steric zippers, but differ in the relative orientation of the two sheets: class 7 is up–up, with the corresponding sheet edges oriented in the same direction (s y = +1), whereas class 8 is up–down, with one sheet reversed relative to the other (s y = −1).
The script is available in the Supporting Information and uses the Python standard-library modules glob and collections, together with the third-party packages Gemmi, NumPy, and natsort.
The following peptides were excluded from this analysis. NFVFGT, corresponding to residues 98–103 of the human immunoglobulin light-chain variable domain (PDB ID 6DIX), was excluded because it adopts an organization that does not correspond to any of the ten canonical steric-zipper classes. IIHFGS, corresponding to residues 138–143 of the human prion protein (PDB ID 3NVF), the equivalent MIHFGN segment from the mouse prion protein (PDB ID 3NVG), and MIHFGND, corresponding to residues 137–143 of the mouse prion protein (PDB ID 3NVH), were excluded because the presence of a glycine at position 142 (human prion protein numbering) introduces a kink in the β-sheet that precludes automated zipper assignment of the AF3 models with the script described above (Figure S3).
Zn2+ Coordination Sphere Analysis
For each of the 900 AF3 models, Zn coordination was analyzed, considering the ND1 and NE2 atoms of histidine residues as potential binding atoms. A Zn–N contact was classified as coordinated when the interatomic distance was ≤2.6 Å. The coordination number of each Zn center was defined as the total number of histidine ND1/NE2 atoms satisfying this criterion. For Zn centers coordinated by n nitrogen atoms, all n(n – 1)/2 unique N–Zn–N angles were calculated and classified as ND1–ND1, ND1–NE2, or NE2–NE2. Individual Zn–N distances were also retained and classified by nitrogen type. All 20 structures in the experimental ensemble and 900 AF3 models were used to generate the ensemble distributions. Mean Zn coordination numbers were reported together with the corresponding sample standard deviations, while Zn–N distance and N–Zn–N angle distributions were represented as violin plots. These distributions combine 680 Zn–N distances and 640 N–Zn–N angles from the experimental ensemble, and 21,402 Zn–N distances and 13,220 N–Zn–N angles from the AF3 models.
The script is available in the Supporting Information and uses the Python standard-library modules glob, warnings, and pathlib, together with the third-party packages Matplotlib, NumPy, pandas, seaborn, and Biopython.
Results and Discussion
AF3 Is a Strong Predictor of β-Sheet Organization
Of the 146 peptides in the data set, 88 experimentally determined structures are parallel, and 56 are antiparallel, while two peptides, MVGGVVIA and LYIQWL, have been resolved in both parallel and antiparallel organizations. The vast majority of AF3 models of peptide assemblies adopt β-sheet conformations (ϕ ∼ −135° and ψ ∼ +135°; upper-left quadrant of the Ramachandran plot shown in Figure S4) that maximize β-sheet hydrogen bonding. This is consistent with the tendency of AF3 to generate idealized, highly regular protein geometries, which is advantageous in this context because classical amyloidogenic short peptides likewise adopt highly regular β-sheet architectures. The high prediction accuracy can therefore be partially attributed to this effect and is not expected to extrapolate to more complex amyloid folds arising from the aggregation of longer peptide sequences.
Averaging over the AF3 models for the 146 peptides shows that 117 peptides have a clear preference for either parallel (82) or antiparallel (35) organization, with the preferred arrangement accounting for at least 90% of the chain pairs. The remaining 29 peptides are predicted to form oligomers containing a mixture of parallel and antiparallel strands (Figure S5). Focusing on the peptides with a clear preference, AF3 models accurately reproduce the parallel versus antiparallel preferences, as shown in Figure A, with only five peptides incorrectly assigned.
2.

(A) Heatmap comparing the predicted and crystallographic parallel versus antiparallel β-sheet organization of the 117 peptides with a clear predicted parallel or antiparallel preference. (B) Crystallographic versus predicted organizations of the correctly predicted GAVVTGVTAVA peptide and the incorrectly predicted GSTSTA peptide. Amino acid residues are shown as sticks and colored by residue number. In the overlays, the crystallographic structure is shown in yellow and the AF3 model in blue.
An example of a correctly predicted model is shown in Figure B for the 68GAVVTGVTAVA78 segment of α-synuclein, also known as NACore. This peptide crystallizes into nearly fully extended, in-register β-sheets, forming a compact class 1 steric zipper characterized by extensive interdigitation of threonine and valine residues across a tight and mostly dry interface. Both residues have high β-sheet propensity, and valine is an aggregation-prone amino acid frequently observed in APRs. The AF3 model shown in Figure B displays a nearly identical organization, differing only at the N-terminal glycine residue. A further plausible reason for this prediction accuracy is the ability of AF3 to recapitulate the secondary-structure propensities of individual amino acids, together with the strong bias of its learned scoring function toward local interactions that optimize side-chain interdigitation at the steric-zipper interface.
A measure of the local structural agreement between AF3 models and the corresponding crystal structures is provided by the median backbone RMSD values calculated between adjacent chain pairs, as shown in Figure S6 (see Methods for details). Of the 117 peptides analyzed, 98 have an RMSD distribution mode below 1 Å and 105 below 2 Å, indicating that AF3 accurately recapitulates the β-sheet organization of these peptides.
In all five failure cases, AF3 incorrectly predicts peptides with known antiparallel organization as having parallel organization. These are the IAPP 16LVHSSN21 and 15FLVHSSNNFGA25 segments, , the ice-nucleation protein InaZ 707GSTSTA712 segment (Figure B), the glycine-rich Keratin-8 58SGMGGIT64 segment, and the CsgA curli protein 129TASNSS134 segment. The experimentally determined crystal structures of these peptides are shown in Figure S7 and cover different classes of antiparallel steric zippers, namely, classes 6, 7, and 8. These sequences are rich in amino acids generally considered not aggregation-prone but rather aggregation-neutral, such as Gly, Asn, Ser, and Ala. Furthermore, Asn is found in APRs predominantly in parallel β-sheets, where it is involved in the formation of asparagine ladders. , These unusual sequence traits might partially explain the prediction failure. To gain further insight, we analyzed the predicted steric-zipper organization across all AF3 models.
AF3 Is a Weak Predictor of Steric-Zipper Organization
For all five incorrectly assigned peptides, AF3 models predominantly converge on parallel class 1 arrangements (Figure A). This systematic misassignment may reflect an intrinsic preference for highly compact interfaces. Analyses of crystallographic steric-zipper structures have shown that parallel interfaces, especially the highly represented class 1 interfaces, generally have greater shape complementarity than antiparallel interfaces. , Thus, AF3 may identify plausible, tightly packed solutions that nevertheless differ from the experimentally determined organizations.
3.

(A) Representative AF3 hexamer models of the five peptides that are misclassified and incorrectly predicted to form class 1 steric zippers. Amino acid residues are shown as sticks and colored by residue number. (B) Heatmap comparing the predicted and crystallographic steric-zipper organizations of 113 peptides with a clear predicted parallel or antiparallel preference.
Extending the steric-zipper analysis to all peptides with a clear predicted parallel or antiparallel preference reveals a broader AF3 bias toward class 1 architecture (Figure B). Of the 48 peptides in this set that crystallize as class 1 zippers, 47 are correctly identified by AF3. However, AF3 also assigns class 1 architecture to 23 peptides that crystallize as class 2 or class 4 zippers as well as to 5 antiparallel peptides. In addition to the intrinsically higher compactness of class 1 interfaces, this bias may be reinforced by their greater abundance in the PDB.
Overall, 62% of the steric-zipper organizations are predicted correctly. However, this percentage reflects both the intrinsically higher abundance of class 1 zippers in the benchmark set and AF3′s tendency to overpredict highly compact structures. The class-specific recall for the seven steric-zipper classes represented in the crystallographic data set is reported in Table . The macro-averaged recall, which assigns equal weight to each experimentally represented class, is 45%, indicating that AF3 cannot be considered a reliable predictor of steric-zipper organization. Specific tools are available for predicting overall fibril organization. −
2. Experimental and Predicted Steric-Zipper Class Distributions and Class-Specific Recall.
| Zipper class | n exp | n pred | n correct | Recall | Exp. fraction | Pred. fraction |
|---|---|---|---|---|---|---|
| 1 | 48 | 75 | 47 | 0.98 | 0.42 | 0.66 |
| 2 | 15 | 4 | 3 | 0.20 | 0.13 | 0.04 |
| 3 | 0 | 0 | - | - | - | - |
| 4 | 11 | 0 | 0 | 0.00 | 0.10 | 0.00 |
| 5 | 2 | 2 | 1 | 0.50 | 0.02 | 0.02 |
| 6 | 11 | 12 | 9 | 0.82 | 0.10 | 0.11 |
| 7 | 15 | 20 | 10 | 0.67 | 0.13 | 0.18 |
| 8 | 11 | 0 | 0 | 0.00 | 0.10 | 0.00 |
| 9 | 0 | 0 | - | - | - | - |
| 10 | 0 | 0 | - | - | - | - |
On the one hand, AF3 appears to capture fundamental packing principles, likely learned primarily from globular proteins, well enough to globally recapitulate the trade-offs between intra- and inter-β-sheet interactions that determine parallel or antiparallel organization. On the other hand, it fails to capture the finer structural determinants that define the exact steric-zipper architecture. This limitation may also arise, at least in part, from the intrinsic difference between finite oligomeric models and the pseudoinfinite three-dimensional packing achieved in crystal structures. For instance, electrostatic interactions between free amino and carboxy termini or charged side chains can tune the kinetic versus thermodynamic control of aggregation as a consequence of the lateral stacking of multiple steric zippers. , Similarly, staggering between interdigitated β-sheets also depends on the three-dimensional crystal arrangement. In this sense, AF3 models are expected to resemble aggregates with a dominant one-dimensional axial cross-β repeat more closely than mature crystalline assemblies.
Additionally, pH conditions and cosolvents in the crystallization medium can alter the energetic trade-offs between quasi-degenerate organizations, effects that AF3 predictions fail to capture when modeling dry oligomers, as in the cases discussed so far. This point will be explored in the following sections. Despite these limitations, the results indicate that AF3 pseudo-ensembles are reliable predictors of parallel versus antiparallel β-sheet organization and may serve as a screening tool for the rational design of functional amyloids or for evaluating the structural impact of mutations.
AF3 Predicts Kinked β-Sheet Structures in LARKS
In light of the observed tendency of AF3 to favor overly compact structures, it is particularly relevant to assess its predictions for segments from glycine-rich, low-complexity domains of FUS, nucleoporins, and heterogeneous nuclear ribonucleoproteins (hnRNPs). We therefore determined whether these sequences are predicted to adopt classical, tightly interdigitated β-sheet architectures or the less compact, kinked β-sheet conformations characteristic of LARKS (low-complexity aromatic-rich kinked segments). To this end, we analyzed AF3 models for the 37SYSGYS42, 54SYSSYGQS61, and 77STGGYG82 segments from FUS; the 243GYNGFG248 and 234GGGYGGS240 segments from human hnRNPA1; and the 116GFGNFGTS123 segment from nucleoporin Nup98. All of these segments are LARKS except GGGYGGS, which was included as a control because it crystallizes in an essentially flat peptide conformation (ϕ, ψ ∼180°).
The distributions of Ramachandran angles obtained from the AF3 models of these systems are shown in Figure A, together with the corresponding crystallographic values for each peptide. The predictions are sensitive to peptide composition even in the absence of MSA information and structural templates, as evidenced by the occupancy of regions outside the canonical β-sheet conformational space (ϕ ∼ −135° and ψ ∼ +135°). Overall, the predicted backbone dihedral angles populate approximately the same regions of conformational space as the corresponding crystallographic structures, indicating that AF3 models can provide qualitative information on backbone architecture even for noncanonical assemblies.
4.

(A) Scatter plot of the ϕ and ψ angles across 900 AF3 models for each peptide adopting a nonconventional β-sheet conformation. Data points are color-coded according to peptide identity, and the corresponding crystallographic values are shown as crosses of the same color. (B) Crystallographic structure (yellow) and AlphaFold models (blue) of oligomers containing 6 and 20 SYSGYS peptide units. The intramolecular hydrogen bond between the Tyr38 hydroxyl group and the Tyr41 backbone carbonyl is highlighted by orange and green spheres, respectively. (C) Histogram of the intramolecular Tyr38 hydroxyl–Tyr41 backbone carbonyl distance across the 900 SYSGYS AF3 models. The corresponding crystallographic value (2.69 Å) is indicated by a vertical dashed line.
This behavior is exemplified in Figure B for the 37SYSGYS42 segment from FUS. The crystallographic structure of this peptide (yellow) displays a Gly40-centered kink stabilized by an intramolecular hydrogen bond between the Tyr38 hydroxyl group and the Tyr41 backbone carbonyl (shown as orange and green spheres, respectively). AF3 models of oligomers containing 6 and 20 peptide units (blue) reproduce the same structural feature, resulting in strong similarity between the predicted and crystallographic structures at the level of β-sheet organization (backbone RMSD < 1 Å in Figure S6). In contrast, AF3 samples a range of different intersheet arrangements rather than converging on the specific packing observed in the crystal structure (see overlays). This heterogeneity in intersheet packing is consistent with the comparatively weak interactions between opposing β-sheets in LARKS.
Unlike the tightly packed steric zippers of classical amyloid fibrils, the kinked sheets of LARKS interact more weakly through polar and aromatic side-chain contacts. Estimates of the energy required to separate paired β-sheets, obtained by applying atomic solvation parameters to the crystal structures, give a mean of 0.567 ± 0.556 kcal mol–1 β-strand–1 for LARKS interfaces, compared with 1.431 ± 0.685 kcal mol–1 β-strand–1 for 75 classical amyloid steric zippers, i.e., approximately 40% of the latter. This weaker energetic constraint may explain why AF3 models do not converge on a unique topology of intersheet contacts. By contrast, the local Tyr38–Tyr41 interaction that stabilizes the backbone kink is consistently reproduced across predictions with different numbers of peptide chains, as illustrated by the corresponding distance distribution in Figure C. Overall, these results show that AF3 models can capture features of the β-sheet organization of noncanonical amyloid assemblies.
AF3 Pseudo-Ensembles Capture Parallel/Antiparallel Polymorphism in Peptides MVGGVVIA and LYIQWL
The premise of this work is that the learned scoring function of AF3 should enable the sampling of alternative organizations in polymorphic peptides and potentially the identification of cases in which parallel and antiparallel organizations are both possible. While the majority of peptides are confidently assigned to well-defined parallel or antiparallel preferences, a fraction of them is predicted with models that mix parallel and antiparallel β-sheets. These include the peptides MVGGVVIA (predicted with overall 67% parallel and 33% antiparallel β-sheet content) and LYIQWL (predicted with overall 14% parallel and 86% antiparallel β-sheet content), which are known to crystallize in both parallel and antiparallel forms (Figure B). In terms of individual structures, AF3 models sample different arrangements, including fully parallel and fully antiparallel zippers, as well as models mixing parallel and antiparallel strands in the same assembly (Figure ).
5.

Representative AF3 hexamer models of the MVGGVVIA and LYIQWL peptides illustrating how a mixture of parallel and antiparallel content can be predictive of polymorphism. Amino acid residues are shown as sticks and colored by residue number.
While it is not expected that these models reproduce the experimentally determined arrangements, which are highly sensitive to the crystallization conditions (e.g., cosolvents, as in the case of LYIQWL), the diversity sampled in the AF3 pseudo-ensemble gives qualitative information on the polymorphism propensity of these segments. The parallel and antiparallel fractions reported above are descriptors of the pseudo-ensemble and should not be interpreted as physical populations of the corresponding polymorphs. However, despite this caveat, -AF3 pseudoensembles may be used to guide artificial peptide design, both to predict the preferred secondary-structure elements and to detect potential polymorphisms. Other peptides that are predicted with both parallel and antiparallel content have one form experimentally determined, but can plausibly adopt both. Examples include the 35MVGGVV40 segment of amyloid-β, closely related to MVGGVVIA, for which polymorphic antiparallel structures have been reported (PDB IDs 2ONA, 2OKZ). Another example is the 113AGAAAA118 segment from the human prion protein (PrP). The latter has been reported with an antiparallel structure (PDB ID 6PQ5), although full-length PrP fibrils all show a parallel, in-register organization. These mixed organizations might also reflect intrinsic polymorphism along the aggregation process, as discussed in the Introduction. For instance, it has been suggested that antiparallel structure generally dominates at the first stages of aggregation, being progressively replaced by parallel organization at large length scales.
Prediction in the Presence of Cosolvents and Cocrystallized Reagents Does Not Shift Secondary Structure Preferences
LYIQWL exhibits pronounced solvent-induced polymorphism (Figure B), adopting either parallel or antiparallel arrangements depending on the crystallization conditions and cocrystallizing reagents. It preferentially crystallizes in a parallel arrangement at high ethanol concentrations (70% water/30% ethanol; PDB entry 8ANG). By contrast, it predominantly adopts an antiparallel organization in pure water (PDB entry 8ANM), at lower ethanol concentrations in the presence of trifluoroacetic acid (90% water/10% ethanol; PDB entry 8ANI), and at high acetonitrile concentrations in the presence of trifluoroacetic acid (66% water/33% acetonitrile; PDB entry 8ANH). All four cocrystallized species are well represented in the PDB (Chemical Component Dictionary (CCD) codes HOH for water, EOH for ethanol, TFA for trifluoroacetic acid, and CCN for acetonitrile). Because AF3 enables protein structure prediction in the presence of small molecules, we investigated whether predicting LYIQWL oligomers in the presence of equimolar or higher stoichiometric ratios of these cosolvents or cocrystallizing reagents affects the distribution of their secondary-structure content.
The average parallel and antiparallel contents under the different coprediction conditions are shown in Figure A. Overall, the presence and stoichiometric ratio of water or other cosolvents have only a modest effect on the predicted parallel content, which remains below 30% in all cases. The most noticeable effect is a change in parallel secondary structure content for the smaller oligomers comprising four or six chains.
6.

(A) Heatmap showing the percentage of parallel β-sheet content in AF3 pseudo-ensembles generated for the LYIQWL peptide predicted as 4-, 6-, 8-, 10-, 12-, 14-, 16-, 18-, and 20-mers under the following conditions: peptide alone (LYIQWL); peptide with water (HOH) at 1:1, 1:2, and 1:5 peptide-to-solvent ratios; peptide with ethanol (EOH) at 1:1, 1:2, and 1:5 ratios; and peptide with equimolar trifluoroacetic acid (TFA) or acetonitrile (CCN). (B) Representative AF3 models of LYIQWL hexamers predicted in the presence of HOH and EOH at a 1:5 ratio, and of TFA and CCN at a 1:1 ratio. Amino acid residues are shown as sticks and colored by residue number; water oxygen atoms are shown as spheres; and EOH, TFA, and CCN are shown as thin black sticks.
Analysis of the AF3 models generated at high water-to-peptide or ethanol-to-peptide stoichiometric ratios reveals a propensity to form compact peptide assemblies surrounded by water or ethanol molecules. The solvent molecules neither insert between β-sheets nor form specific interactions with polar groups (Figure B). Consequently, these models are generally more compact than those predicted in the absence of cosolvent molecules, as indicated by the narrower radius-of-gyration distributions shown in Figure S8. Similar behavior is observed for trifluoroacetic acid and acetonitrile (Figure B). These results suggest that AF3 and related models, although they may have learned interaction patterns involving small ligand molecules and functional groups and may generalize them to some extent, , remain unable to recapitulate how (co)solvents and cocrystallizing reagents fine-tune peptide-oligomer architecture. Instead, they favor compact peptide–peptide assemblies over small-molecule-mediated peptide networks. In light of these results, we next evaluated whether AF3 can recapitulate the more specific interactions of amyloid fibrils with metal ions.
Prediction of Zn2+ Coordination Spheres Does Not Generalize to Unusual Binding Patterns Observed in Amyloid Fibrils
To assess AF3′s ability to predict metal-ion coordination, we considered the peptide IHVHLQI, which forms zinc-binding catalytic fibrils. The peptide adopts a class 1 steric zipper (PDB ID 5UGK, NMR ensemble), with a hydrophobic interface stabilized by the apolar side chains of valine, leucine, and isoleucine. Its polar surface displays aligned histidine residues that coordinate Zn2+ ions, forming a zigzag metal–histidine chain along the fibril axis (Figure A).
7.

(A) Left: representation of the solid-state NMR zinc-binding structure of the catalytic amyloid formed by self-assembly of the IHVHLQI peptide. Right: AF3 model of the same peptide in the presence of equimolar Zn2+ ions (dodecamer). Zn2+ ions are shown as yellow spheres, and the coordinating histidine residues at positions 2 and 4 are shown as sticks and colored green and purple, respectively. (B) Distribution of Zn–N distances (left) and N–Zn–N angles (right) in zinc-binding IHVHLQI amyloid fibrils, shown as a side-by-side comparison between the solid-state NMR ensemble (20 structures) and the AF3 pseudo-ensembles (900 structures). The data sets are further divided according to the type of coordinating nitrogen (Nδ versus Nϵ).
Each Zn2+ ion is coordinated by two bidentate histidine ligands that bridge adjacent strands, while a third histidine completes the coordination sphere, leaving one coordination site available for water binding and activation. This metal-mediated bridging between strands stabilizes the fibril and produces a Zn2+ coordination geometry distinct from the classical motifs involving cysteine, histidine, and aspartate/glutamate residues found in catalytic, cocatalytic, structural, and protein-interface zinc-binding environments in globular proteins. , Because AF3 is expected to have learned Zn2+ coordination geometries predominantly from globular proteins, which constitute the majority of structures in the PDB, it is of interest to assess whether these general zinc-binding principles can be extrapolated to amyloid fibrils.
AF3 pseudo-ensembles were generated for this peptide in the presence of equimolar Zn2+ ions. In all models, AF3 correctly predicted the parallel class 1 steric-zipper organization of the peptide and placed Zn2+ ions in proximity to histidine side chains. However, the most frequently predicted coordination geometry did not reproduce the experimentally observed zigzag metal–histidine arrangement. Instead, the Zn2+ ions adopted a linear organization parallel to the fibril axis (Figure A). In a minority of models, dinuclear Zn2+ clusters surrounded by four histidine side chains were predicted (Figure S9A). These models exhibit unphysically short Zn–Zn and Zn–N distances and are responsible for the low-distance tail of the Zn–N distribution in Figure B. Spurious clashes and unphysical atomic overlaps have previously been reported in AF3 predictions. , Moreover, although AF3 generally places metal ions near experimentally observed binding sites, substantially distorted first-shell coordination geometries, including anomalously short metal–ligand distances relative to crystallographic structures, have also been reported.
In both arrangements, histidine coordination was strictly monodentate. These results suggest that AF3 is unable to capture the bidentate coordination mode of histidine in this system and instead favors the monodentate mode commonly observed in globular proteins. This preference is suboptimal in the present context and reduces the average Zn2+ coordination number from ∼3 to ∼2 (Figure S9B).
In addition to the lower predicted coordination number, a global comparison of metal-coordination geometries in the NMR and AF3 ensembles, based on Zn–N distances and N–Zn–N angles, revealed broader distributions of Zn–N coordination distances in the AF3 models, particularly for coordination through Nδ (Figure B). The predicted N–Zn–N angles also deviated more strongly from the ideal tetrahedral value of 109°. Nevertheless, Zn2+ coordination spheres are highly flexible, and the NMR ensemble itself shows a partial population of Nδ−Zn–Nϵ angles of approximately 130°. Overall, these results show that, although AF3 correctly identifies histidine residues as the Zn2+-binding sites in this amyloid system, it fails to reproduce the unusual coordination geometry characteristic of this amyloid fibril.
Conclusions
We have shown that conformational sampling of oligomeric assemblies of short peptides using pseudo-ensembles of AF3 structures provides a robust approach for predicting β-sheet organization in short amyloidogenic peptides, with 112 of the 117 peptides that have a clear predicted preference correctly assigned. In particular, AF3 is a strong predictor of parallel versus antiparallel organization and β-sheet structure, which are governed by a subtle energetic balance between intra- and intersheet noncovalent interactions. This success likely originates from the implicit energy function learned by AF3, which effectively optimizes side-chain rotamers and packing. The short computational time required for these predictions, on the order of minutes, together with the straightforward input format, makes AF3 pseudo-ensembles an attractive tool for the efficient screening of peptide libraries comprising hundreds of sequences, potentially in combination with conventional or enhanced-sampling molecular dynamics simulations to further refine promising initial models.
Prediction of the packing of β-sheets, namely, the preferred steric-zipper architecture, is less reliable. This limitation likely reflects the uneven representation of steric-zipper classes in the PDB, the propensity of AF3 to generate overly compact structures (favoring the more abundant class 1 steric zipper), and its bias toward local interactions. Moreover, structure prediction in the presence of cosolvents and metal ions shows that AF3 fails to reproduce the “wet” interfaces characteristic of some amyloids formed by polar peptide sequences and describes Zn2+ binding only qualitatively, occasionally predicting unphysically short metal coordination distances. The increasing availability of structural data for metal-associated amyloids in the coming years should enable a more quantitative assessment of these limitations.
These findings are consistent with a “protein-first” bias, that is, a preference for the packing and coordination patterns learned from globular proteins. Because AF3 was trained predominantly on globular protein structures with densely packed, solvent-excluded cores, it tends to maximize side-chain packing within fibril cores rather than intercalating solvent molecules, even when solvent-mediated interactions are a major determinant of the preferred fibril architecture. Similarly, AF3 does not reproduce unusual metal-coordination spheres characteristic of the amyloid state that have not been encountered during training on globular proteins.
Despite these limitations, AF3 can serve as a useful tool for artificial amyloid design, given the conformational diversity captured by its pseudo-ensembles. In particular, the presence of mixed parallel and antiparallel secondary structure content in a pseudo-ensemble may serve as a qualitative indicator of polymorphism, as observed for the known polymorphic peptides MVGGVVIA and LYIQWL, rather than as a representation of physical populations.
Supplementary Material
Acknowledgments
This research has been funded by MCIN/AEI/10.13039/501100011033 grants RYC2022-036457-I and PID2025-174578NA-I00.
The following data are available through the Zenodo repository https://zenodo.org/records/21510614: AlphaFold3-predicted structures of oligomeric assemblies for 146 short peptides. AlphaFold3 open-source code can be downloaded from https://github.com/google-deepmind/alphafold3. PyMOL open-source code can be downloaded from https://github.com/schrodinger/pymol-open-source. Sample code used for analysis is provided in the Supporting Information.
The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.6c02489.
Summary of crystallographic and modeled structures; benckmark of secondary structure assignment scripts(XLSX)
Ramachandran plots of AlphaFold3-predicted oligomeric assemblies of 146 short peptides; distributions of parallel and antiparallel content in the AlphaFold3-predicted structures; distribution of RMSD values between AlphaFold3 models and crystallographic references; representations of peptide crystal structures incorrectly predicted by AlphaFold3; crystal structures of peptides exhibiting unusual steric-zipper organizations; distributions of the radius of gyration of AlphaFold3-predicted LYIQWL oligomers as a function of the presence of solvent molecules; representations of AlphaFold3-predicted IHVHLQI oligomers in the presence of zinc ions, together with the corresponding average zinc coordination numbers; example input files for AlphaFold3 structure prediction; and Python scripts used to analyze the AlphaFold3-predicted structures (PDF)
The author declares no competing financial interest.
References
- Jumper J., Evans R., Pritzel A.. et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589. doi: 10.1038/s41586-021-03819-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Abramson J., Adler J., Dunger J.. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630:493–500. doi: 10.1038/s41586-024-07487-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Terwilliger T. C., Liebschner D., Croll T. I., Williams C. J., McCoy A. J., Poon B. K., Afonine P. V., Oeffner R. D., Richardson J. S., Read R. J., Adams P. D.. AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination. Nat. Methods. 2024;21:110–116. doi: 10.1038/s41592-023-02087-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Roney J. P., Ovchinnikov S.. State-of-the-Art Estimation of Protein Model Accuracy Using AlphaFold. Phys. Rev. Lett. 2022;129:238101. doi: 10.1103/PhysRevLett.129.238101. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Sarma, S. ; Truscott, H. H. ; Xu, D. ; Reid, K. ; Chu, L.-S. ; Chen, J. ; Gray, J. J. . Can We Extract Physics-like Energies from Generative Protein Diffusion Models? bioRxiv 2025, 10.1101/2025.11.28.690021. [DOI]
- Wayment-Steele H. K., Ojoawo A., Otten R., Apitz J. M., Pitsawong W., Hömberger M., Ovchinnikov S., Colwell L., Kern D.. Predicting multiple conformations via sequence clustering and AlphaFold2. Nature. 2024;625:832–839. doi: 10.1038/s41586-023-06832-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Cao, H. ; Gao, Z. ; Gu, C. ; Liu, G. ; Heng, P. A. ; Chatterjee, P. . SF-Cluster: Frustration-Guided MSA Subsampling for Alternative Protein Conformation Recovery arXiv 2026. arXiv:2607.00180. arXiv.org e-Printarchive. 10.48550/arXiv.2607.00180. [DOI]
- Kalakoti Y., Wallner B.. AFsample2 predicts multiple conformations and ensembles with AlphaFold2. Commun. Biol. 2025;8:373. doi: 10.1038/s42003-025-07791-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Chakravarty D., Schafer J. W., Chen E. A., Thole J. F., Ronish L. A., Lee M., Porter L. L.. AlphaFold predictions of fold-switched conformations are driven by structure memorization. Nat. Commun. 2024;15:7296. doi: 10.1038/s41467-024-51801-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Peccati F., Alunno-Rufini S., Jiménez-Osés G.. Accurate Prediction of Enzyme Thermostabilization with Rosetta Using AlphaFold Ensembles. J. Chem. Inf. Model. 2023;63:898–909. doi: 10.1021/acs.jcim.2c01083. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Núñez-Franco R., Muriel-Olaya M. M., Jiménez-Osés G., Peccati F.. AlphaFold2 Predicts Alternative Conformation Populations in Green Fluorescent Protein Variants. J. Chem. Inf. Model. 2024;64:7135–7140. doi: 10.1021/acs.jcim.4c01388. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Jiménez-Osés G., Peccati F.. Structure Prediction of Alternate Frame Folding Systems with AlphaFold3. J. Chem. Inf. Model. 2025;65:8229–8237. doi: 10.1021/acs.jcim.5c00906. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Pinheiro F., Santos J., Ventura S.. AlphaFold and the amyloid landscape. J. Mol. Biol. 2021;433:167059. doi: 10.1016/j.jmb.2021.167059. [DOI] [PubMed] [Google Scholar]
- Chen, S. ; Huh, J. ; Sklar, C. ; Gray, J. J. . Predicting Supramolecular Self-Assembly of Peptide Structures with AlphaFold3 bioRxiv 2026. 10.64898/2026.04.28.720402. [DOI]
- Chen D., Drombosky K. W., Hou Z., Sari L., Kashmer O. M., Ryder B. D., Perez V. A., Woodard D. R., Lin M. M., Diamond M. I., Joachimiak L. A.. Tau local structure shields an amyloid-forming motif and controls aggregation propensity. Nat. Commun. 2019;10:2493. doi: 10.1038/s41467-019-10355-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lövestam S., Li D., Wagstaff J. L., Kotecha A., Kimanius D., McLaughlin S. H., Murzin A. G., Freund S. M. V., Goedert M., Scheres S. H. W.. Disease-specific tau filaments assemble via polymorphic intermediates. Nature. 2024;625:119–125. doi: 10.1038/s41586-023-06788-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Wilkinson M., Xu Y., Thacker D., Taylor A. I. P., Fisher D. G., Gallardo R. U., Radford S. E., Ranson N. A.. Structural evolution of fibril polymorphs during amyloid assembly. Cell. 2023;186:5798–5811. doi: 10.1016/j.cell.2023.11.025. [DOI] [PMC free article] [PubMed] [Google Scholar]
- van der Kant R., Louros N., Schymkowitz J., Rousseau F.. Thermodynamic analysis of amyloid fibril structures reveals a common framework for stability in amyloid polymorphs. Structure. 2022;30:1178–1189. doi: 10.1016/j.str.2022.05.002. [DOI] [PubMed] [Google Scholar]
- Duran-Romaña R., Schymkowitz J., Rousseau F., Louros N.. Energetic profiling reveals thermodynamic principles underlying amyloid fibril maturation. Nat. Commun. 2025;16:10081. doi: 10.1038/s41467-025-65035-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Madine J., Jack E., Stockley P. G., Radford S. E., Serpell L. C., Middleton D. A.. Structural insights into the polymorphism of amyloid-like fibrils formed by region 20–29 of amylin revealed by solid-state NMR and X-ray fiber diffraction. J. Am. Chem. Soc. 2008;130:14990–15001. doi: 10.1021/ja802483d. [DOI] [PubMed] [Google Scholar]
- Tycko R., Wickner R. B.. Molecular structures of amyloid and prion fibrils: Consensus versus controversy. Acc. Chem. Res. 2013;46:1487–1496. doi: 10.1021/ar300282r. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Irizarry B. A., Davis J., Zhu X., Boon B. D. C., Rozemuller A. J. M., Van Nostrand W. E., Smith S. O.. Human cerebral vascular amyloid contains both antiparallel and parallel in-register Aβ40 fibrils. J. Biol. Chem. 2021;297:101259. doi: 10.1016/j.jbc.2021.101259. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Zhaliazka K., Kurouski D.. Nanoscale characterization of parallel and antiparallel β-sheet amyloid beta 1–42 aggregates. ACS Chem. Neurosci. 2022;13:2813–2820. doi: 10.1021/acschemneuro.2c00180. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Knowles T. P. J., Mezzenga R.. Amyloid fibrils as building blocks for natural and artificial functional materials. Adv. Mater. 2016;28:6546–6561. doi: 10.1002/adma.201505961. [DOI] [PubMed] [Google Scholar]
- Wang X., Zhang S., Zhang J., Wang Y., Jiang X., Tao Y., Li D., Zhong C., Liu C.. Rational design of functional amyloid fibrillar assemblies. Chem. Soc. Rev. 2023;52:4603–4631. doi: 10.1039/D2CS00756H. [DOI] [PubMed] [Google Scholar]
- Dürvanger Z., Bencs F., Menyhárd D. K., Horváth D., Perczel A.. Solvent induced amyloid polymorphism and the uncovering of the elusive class 3 amyloid topology. Commun. Biol. 2024;7:968. doi: 10.1038/s42003-024-06621-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Horváth D., Dürvanger Z., Menyhárd D. K., Sulyok-Eiler M., Bencs F., Gyulai G., Horváth P., Taricska N., Perczel A.. Polymorphic amyloid nanostructures of hormone peptides involved in glucose homeostasis display reversible amyloid formation. Nat. Commun. 2023;14:4621. doi: 10.1038/s41467-023-40294-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Robang A. S., Wong K. M., Leisen J., Liu R., Radford W. L., Rao Sudarshan T., Hudalla G. A., Paravastu A. K.. Parallel β-Sheet Structure and Structural Heterogeneity Detected within Q11 Self-Assembling Peptide Nanofibers. J. Phys. Chem. B. 2024;128:5387–5396. doi: 10.1021/acs.jpcb.4c00825. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Faller P., Hureau C., Berthoumieu O.. Role of Metal Ions in the Self-assembly of the Alzheimer’s Amyloid-β Peptide. Inorg. Chem. 2013;52:12193–12206. doi: 10.1021/ic4003059. [DOI] [PubMed] [Google Scholar]
- Lee M., Wang T., Makhlynets O. V., Wu Y., Polizzi N. F., Wu H., Gosavi P. M., Stöhr J., Korendovych I. V., DeGrado W. F., Hong M.. Zinc-binding structure of a catalytic amyloid from solid-state NMR. Proc. Natl. Acad. Sci. U.S.A. 2017;114:6191–6196. doi: 10.1073/pnas.1706179114. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Sawaya M. R., Sambashivan S., Nelson R., Ivanova M. I., Sievers S. A., Apostol M. I., Thompson M. J., Balbirnie M., Wiltzius J. J. W., McFarlane H. T., Madsen A. O., Riekel C., Eisenberg D.. Atomic structures of amyloid cross-β spines reveal varied steric zippers. Nature. 2007;447:453–457. doi: 10.1038/nature05695. [DOI] [PubMed] [Google Scholar]
- Colletier J.-P., Laganowsky A., Landau M., Zhao M., Soriaga A. B., Goldschmidt L., Flot D., Cascio D., Sawaya M. R., Eisenberg D.. Molecular basis for amyloid-β polymorphism. Proc. Natl. Acad. Sci. U.S.A. 2011;108:16938–16943. doi: 10.1073/pnas.1112600108. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Kyte J., Doolittle R. F.. A simple method for displaying the hydropathic character of a protein. J. Mol. Biol. 1982;157:105–132. doi: 10.1016/0022-2836(82)90515-0. [DOI] [PubMed] [Google Scholar]
- Stroud J. C.. The Zipper Groups of the Amyloid State of Proteins. Acta Crystallogr., Sect. D:Biol. Crystallogr. 2013;69:540–545. doi: 10.1107/S0907444912050548. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Brumshtein B., Esswein S. R., Sawaya M. R., Rosenberg G., Ly A. T., Landau M., Eisenberg D. S.. Identification of two principal amyloid-driving segments in variable domains of Ig light chains in systemic light-chain amyloidosis. J. Biol. Chem. 2018;293:19659–19671. doi: 10.1074/jbc.RA118.004142. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Perczel A., Hudáky P., Pálfi V. K.. Dead-End Street of Protein Folding: Thermodynamic Rationale of Amyloid Fibril Formation. J. Am. Chem. Soc. 2007;129:14959–14965. doi: 10.1021/ja0747122. [DOI] [PubMed] [Google Scholar]
- Sulyok-Eiler M., Harmat V., Perczel A.. Unravelling the Complexity of Amyloid Peptide Core Interfaces. J. Chem. Inf. Model. 2024;64:8628–8640. doi: 10.1021/acs.jcim.4c01479. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Apostol M. I., Wiltzius J. J. W., Sawaya M. R., Cascio D., Eisenberg D.. Atomic Structures Suggest Determinants of Transmission Barriers in Mammalian Prion Disease. Biochemistry. 2011;50:2456–2463. doi: 10.1021/bi101803k. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Rodriguez J. A., Ivanova M. I., Sawaya M. R.. et al. Structure of the toxic core of α-synuclein from invisible crystals. Nature. 2015;525:486–490. doi: 10.1038/nature15368. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Fujiwara K., Toda H., Ikeguchi M.. Dependence of α-helical and β-sheet amino acid propensities on the overall protein fold type. BMC Struct. Biol. 2012;12:18. doi: 10.1186/1472-6807-12-18. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Deng, Z. ; Hou, R. ; Xie, N. ; Tyers, M. ; Koziarski, M. . Discontinuous Epitope Fragments as Sufficient Target Templates for Efficient Binder Design arXiv 2025. arXiv:2509.25479. arXiv.org e-Printarchive. 10.48550/arXiv.2509.25479. [DOI]
- Krotee P., Rodriguez J. A., Sawaya M. R.. et al. Atomic structures of fibrillar segments of hIAPP suggest tightly mated β–sheets are important for cytotoxicity. eLife. 2017;6:e19273. doi: 10.7554/eLife.19273. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Soriaga A. B., Sangwan S., Macdonald R., Sawaya M. R., Eisenberg D.. Crystal Structures of IAPP Amyloidogenic Segments Reveal a Novel Packing Motif of Out-of-Register Beta Sheets. J. Phys. Chem. B. 2016;120:5810–5816. doi: 10.1021/acs.jpcb.5b09981. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Zee C.-T., Glynn C., Gallagher-Jones M., Miao J., Santiago C. G., Cascio D., Gonen T., Sawaya M. R., Rodriguez J. A.. Homochiral and racemic MicroED structures of a peptide repeat from the ice-nucleation protein InaZ. IUCrJ. 2019;6:197–205. doi: 10.1107/S2052252518017621. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Murray K. A., Hughes M. P., Hu C. J., Sawaya M. R., Salwinski L., Pan H., French S. W., Seidler P. M., Eisenberg D. S.. Identifying amyloid-related diseases by mapping mutations in low-complexity protein domains to pathologies. Nat. Struct. Mol. Biol. 2022;29:529–536. doi: 10.1038/s41594-022-00774-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Perov S., Lidor O., Salinas N., Golan N., Tayeb-Fligelman E., Deshmukh M., Willbold D., Landau M.. Structural Insights into Curli CsgA Cross-β Fibril Architecture Inspire Repurposing of Anti-amyloid Compounds as Anti-biofilm Agents. PLoS Pathog. 2019;15:e1007978. doi: 10.1371/journal.ppat.1007978. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Kurt T. D., Aguilar-Calvo P., Jiang L., Rodriguez J. A., Alderson N., Eisenberg D. S., Sigurdson C. J.. Asparagine and glutamine ladders promote cross-species prion conversion. J. Biol. Chem. 2017;292:19076–19086. doi: 10.1074/jbc.M117.794107. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lawrence M. C., Colman P. M.. Shape Complementarity at Protein/Protein Interfaces. J. Mol. Biol. 1993;234:946–950. doi: 10.1006/jmbi.1993.1648. [DOI] [PubMed] [Google Scholar]
- Louros N., Orlando G., De Vleeschouwer M., Rousseau F., Schymkowitz J.. Structure-based machine-guided mapping of amyloid sequence space reveals uncharted sequence clusters with higher solubilities. Nat. Commun. 2020;11:3314. doi: 10.1038/s41467-020-17207-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Wojciechowski J. W., Kotulska M.. PATH – Prediction of Amyloidogenicity by Threading and Machine Learning. Sci. Rep. 2020;10:7721. doi: 10.1038/s41598-020-64270-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Tabatabaei Ghomi H., Topp E. M., Lill M. A.. Fibpredictor: a computational method for rapid prediction of amyloid fibril structures. J. Mol. Model. 2016;22:206. doi: 10.1007/s00894-016-3066-1. [DOI] [PubMed] [Google Scholar]
- Wojciechowski J. W., Szczurek W., Szulc N., Szefczyk M., Kotulska M.. PACT – Prediction of amyloid cross-interaction by threading. Sci. Rep. 2023;13:22268. doi: 10.1038/s41598-023-48886-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Park J., Kahng B., Hwang W.. Thermodynamic Selection of Steric Zipper Patterns in the Amyloid Cross-β Spine. PLoS Comput. Biol. 2009;5:e1000492. doi: 10.1371/journal.pcbi.1000492. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Iadanza M. G., Jackson M. P., Hewitt E. W., Ranson N. A., Radford S. E.. A new era for understanding amyloid structures and disease. Nat. Rev. Mol. Cell Biol. 2018;19:755–773. doi: 10.1038/s41580-018-0060-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Riek R.. The three-dimensional structures of amyloids. Cold Spring Harb. Perspect. Biol. 2017;9:a023572. doi: 10.1101/cshperspect.a023572. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Hughes M. P., Sawaya M. R., Boyer D. R., Goldschmidt L., Rodriguez J. A., Cascio D., Chong L., Gonen T., Eisenberg D. S.. Atomic structures of low-complexity protein segments reveal kinked β sheets that assemble networks. Science. 2018;359:698–701. doi: 10.1126/science.aan6398. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Gui X., Luo F., Li Y.. et al. Structural basis for reversible amyloids of hnRNPA1 elucidates their role in stress granule assembly. Nat. Commun. 2019;10:2006. doi: 10.1038/s41467-019-09902-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Glynn C., Sawaya M. R., Ge P., Gallagher-Jones M., Short C. W., Bowman R., Apostol M., Zhou Z. H., Eisenberg D. S., Rodriguez J. A.. Cryo-EM structure of a human prion fibril with a hydrophobic, protease-resistant core. Nat. Struct. Mol. Biol. 2020;27:417–423. doi: 10.1038/s41594-020-0403-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Zanjani A. A. H., Reynolds N. P., Zhang A., Schilling T., Mezzenga R., Berryman J. T.. Amyloid evolution: Antiparallel replaced by parallel. Biophys. J. 2020;118:2526–2536. doi: 10.1016/j.bpj.2020.03.023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Bryant P., Kelkar A., Guljas A., Clementi C., Noé F.. Structure prediction of protein-ligand complexes from sequence information with Umol. Nat. Commun. 2024;15:4536. doi: 10.1038/s41467-024-48837-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Laitaoja M., Valjakka J., Jänis J.. Zinc Coordination Spheres in Protein Structures. Inorg. Chem. 2013;52:10983–10991. doi: 10.1021/ic401072d. [DOI] [PubMed] [Google Scholar]
- Putignano V., Rosato A., Banci L., Andreini C.. MetalPDB in 2018: a database of metal sites in biological macromolecular structures. Nucleic Acids Res. 2018;46:D459–D464. doi: 10.1093/nar/gkx989. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Nam K. H.. Evaluation of AlphaFold3 prediction for post-translational modification, oligomeric assembly, and quenchable metal binding of fluorescent proteins. J. Mol. Graphics Modell. 2026;142:109169. doi: 10.1016/j.jmgm.2025.109169. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The following data are available through the Zenodo repository https://zenodo.org/records/21510614: AlphaFold3-predicted structures of oligomeric assemblies for 146 short peptides. AlphaFold3 open-source code can be downloaded from https://github.com/google-deepmind/alphafold3. PyMOL open-source code can be downloaded from https://github.com/schrodinger/pymol-open-source. Sample code used for analysis is provided in the Supporting Information.
