Skip to main content
NCBI home page
ACS AuthorChoice logoLink to ACS AuthorChoice
. 2025 Jul 27;65(15):8229–8237. doi: 10.1021/acs.jcim.5c00906

Structure Prediction of Alternate Frame Folding Systems with AlphaFold3

Gonzalo Jiménez-Osés †,‡,*, Francesca Peccati †,‡,*
PMCID: PMC12344692  PMID: 40785361

Abstract

AlphaFold has proven to be a valuable tool for predicting protein structures with unprecedented speed and accuracy. Extensive research from multiple groups has demonstrated that manipulating the multiple sequence alignment used for structure prediction can enhance AlphaFold2’s ability to explore protein conformational landscapes, yielding reliable models of proteins capable of switching between alternative conformations. The release of the thoroughly reengineered AlphaFold3, which promises even greater prediction accuracy and efficiency, raises the question of whether such alternative conformational states can be modeled–either natively or by tuning the multiple sequence alignment used for prediction. In this work, we use a family of green fluorescent proteins engineered through alternate frame folding to assess AlphaFold3′s prediction accuracy and uncover an unexpected role of disordered regions in driving the conformational preferences of the models.

graphic file with name ci5c00906_0008.jpg

graphic file with name ci5c00906_0006.jpg

Introduction

The development of artificial intelligence-based (AI) methods for protein structure prediction is profoundly transforming molecular biology by providing unprecedented access to high-quality, full-atom structural models of apo and holo proteins, with potentially far-reaching implications for drug discovery. With few exceptions, most AI-based tools map monomeric protein sequences to a single conformation, often closely resembling a Protein Data Bank (PDB) structure used for training. However, since the PDB contains multiple conformations for the same sequences, these tools are inherently exposed to conformational diversity to an unspecified extent. On the other hand, significant efforts are underway to achieve robust predictions of alternative conformations. For AlphaFold2 (AF2), resampling the multiple sequence alignment (MSA) used for structure prediction has enhanced the predicted conformational diversity in several protein families. However, the extent to which this increased diversity stems from memorization of training data or generalizes across different types of alternative conformations remains debated. −

In this work, we investigate the ability of AlphaFold3 (AF3) to predict alternative conformations in alternate frame folding systems. Alternate frame folding is a protein engineering technique that involves circular permutation and the duplication of a protein segment. In this approach, duplicated segments–typically differing by a single mutation–are appended to the opposite termini of the sequence and can fold into the native protein conformation in a mutually exclusive manner. That is, one of the segments completes the fold of the unduplicated protein and the other, the orphan segment, remains mostly unstructured. Switching between the two possible states–where either one of the duplicated segments folds while the other remains disordered–entails a dramatic conformational change, involving two coupled disordered-to-ordered and ordered-to-disordered transitions. Assessing the ability of AF3 to predict alternate frame folding conformations is valuable not only for designing molecular switches based on this principle but also for understanding how large-scale conformational rearrangements are handled. As structure prediction relies on MSAs generated from databases primarily composed of natural proteins with unpermuted and unduplicated sequences, alternate frame folding systems are intrinsically challenging. Moreover, prediction confidence depends on the extent of ordered secondary structure, which is inherently limited in this case, as only one of the duplicated segments can be fully ordered. In a previous work, we employed a family of circularly permuted GFP variants with alternative strands developed by Boxer and co-workers to test AF2’s ability to predict alternative conformations.

GFPs are ubiquitously used, genetically encoded fluorescent reporters, with applications in living cells and tissue imaging. They feature an 11-stranded β-barrel and an internal α-helix. Spontaneous cyclization and oxidation of a Thr-Tyr-Gly tripeptide on this α-helix leads to chromophore formation, in a maturation process that occurs only in the folded protein. While the chromophore is not fluorescent in solution, it is fluorescent within the β-barrel, with surrounding residues tuning its pK a and, importantly, its color. GFP tolerance to mutations has allowed for extensive engineering aimed at tuning color, photoswitching, expression, chromophore maturation rate, photoconversion, photoactivation, and stability. For instance, protein supercharging has enabled the development of GFPs that withstand extreme thermal and chemical denaturation conditions, with applications ranging from nucleic acid delivery in mammalian cells to lanthanide detection. As another example, split GFPs–based on the separation of a circularly permuted GFP into two fragments that reconstitute a native-like structure through formation of a noncovalently bound fluorescent heterodimer–have advanced the detection of protein-protein interactions.

GFPs engineered through alternate frame folding, the focus of this work, have proven highly versatile, enabling the development of diverse sensors-including a thrombin sensor that leverages proteolytic activity and a rapamycin sensor that exploits ligand binding-induced folding. The alternative conformations and common topology of the alternate frame folding GFPs considered here are shown in Figure A.

1.

1

Open in a new tab

(A) General topology of the seven constructs developed by Boxer and co-workers and analyzed in this work: a circularly permuted GFP with two alternative s10 strands at the opposite termini. The N-terminal s10 strand contains a tyrosine at position 203, specifying yellow fluorescence, while the C-terminal s10 strand contains a threonine at the same position, specifying green fluorescence. All constructs express as a mixture of two conformations: the N-bound conformation behaves like a yellow fluorescent protein (YFP), and the C-bound conformation behaves like a green fluorescent protein (GFP). (B) Population of the N-bound conformation for different constructs as a function of the variable loop length connecting the C-terminal s10 strand to the barrel, reproduced from ref .

These constructs are engineered to contain a duplicated s10 strand, with one located at the N-terminus and the other at the C-terminus. The β-barrel can be completed either by incorporating the N-terminal strand (N-bound conformation) or the C-terminal strand (C-bound conformation, Figure A). The two conformations exhibit distinct spectroscopic signatures, as the alternative s10 strands are not identical copies but differ at a single position, residue 203 (standard GFP numbering), which determines fluorescence color. The N-terminal strand specifies yellow fluorescence (Y203), while the C-terminal strand specifies green fluorescence (T203). Therefore, the N-bound conformation exhibits the spectroscopic behavior of a yellow fluorescent protein (YFP), while the C-bound conformation behaves as a green fluorescent protein (GFP). This enables a straightforward measurement of the population ratio between the two conformations by fitting the absorption spectrum to a linear combination of YFP and GFP base spectra. Additionally, GFP constructs include a purification His-tag at the N-terminus, which cannot be ignored in structure prediction, as it may influence population ratios and is expected to be unstructured (Figure A).

Populations can be controlled by leveraging loop closure entropy through variations in the length of the loop connecting the C-terminal s10 strand to the barrel (Figure A). This study focuses on seven constructs with loop lengths of 0, 1, 2, 4, 6, 10, and 14 residues (Figure B), spanning a wide range of population ratios. The construct with the shortest loop (loop0) is expressed as 5% in the N-bound conformation, while the construct with the longest loop (loop14) is 92% N-bound. Between these two extremes, a finely distributed range of populations is obtained, which is monotonic but not linear. This distribution provides an ideal benchmark for evaluating AlphaFold’s ability not only to detect both conformations but also to predict their relative population.

Analyzing AF2’s performance on these systems, we found that standard AF2 predictions fail to capture the alternative conformations, either entirely missing or severely underestimating the population of the N-bound conformation. This limitation arises from an imbalance in the information content of the MSA used for prediction. Specifically, the deeper MSA for the strand specifying green fluorescence leads to its overwhelmingly preferential insertion, preventing AF2 from accurately capturing the equilibrium between the two conformations. To address this shortcoming, we proposed targeted column masking of the MSA to expand AF2’s conformational sampling. Applying this correction significantly altered prediction outcomes, enabling not only the identification of both alternative conformations for all constructs but also a qualitative estimation of population shifts as a function of C-terminal loop length.

In this work, we use the same family of proteins to benchmark AF3′s ability to predict alternative conformations. This study is motivated by key differences between AF2 and AF3 architectures. Notably, AF3 de-emphasizes the role of the MSA, and replaces the evoformer module with a simpler pairformer module, suggesting that MSA manipulation and the use of templates would be less effective in modulating conformational diversity for monomer predictions compared to AF2. Additionally, unlike AF2, AF3 generates atomic coordinates using a diffusion model, which may yield hallucinations (spurious structuring of disordered regions). Our benchmark system–which contains a His-tag and where the alternative s10 strands are expected to be structured only upon insertion into the barrel–is well-suited for investigating this behavior. Finally, AF3 can be run with various sampling seeds to generate alternative solutions, with accuracy shown to improve as the number of seeds increases. Additionally, since AF3 employs a diffusion model to generate atomic coordinates, a batch of diffusion samples is produced for each seed. Consequently, both the number of model seeds and the number of diffusion samples can influence the structural diversity of the predicted solutions.

Methods

AF3 Structure Prediction

AF3 structure predictions were performed with codebase version 3.0.0. For all systems, we performed an initial prediction with one seed to generate the multiple sequence alignment and template informations using a standard input (an example is provided in the ESI for loop0). The job input JSON file containing the MSA and template data (<job_name>_data.json) was used to generate distinct sets of predictions (Figure S1) by varying the use of MSA and templates. The first set, corresponding to a standard AF3 prediction, retained the unaltered MSA and template information. A second set was generated using a modified MSA, where the columns corresponding to the N-terminal and C-terminal s10 strands were masked (cmAF3, see below). A third set was produced by removing all template information (tfAF3, see below). Finally, a fourth set was performed with both templates removed and MSA columns corresponding to the N-terminal and C-terminal s10 strands masked (tfcmAF3).

Furthermore, we explored the effect of seeds and samples on structural diversity by running the simulations with two sets of specifications: one set with 200 samples and 5 seeds per sample, and the other one with 10 samples and 100 seeds per sample. Finally, we analyzed the effect of the His-tag used for purification, which is present in all constructs at the N-terminus. To do so, we performed structure predictions with and without the His-tag (sequence MGHHHHHHSSGG). Figure S1 summarizes all the tested GFP prediction setups, for a total of 112,000 AF3 models.

MSA Column Masking

MSA columns corresponding to the N-terminal and C-terminal s10 strands of the GFPs (sequences LPDNHYLSYQTVLSKDPNE and LPDNHYLSTQTVLSKDPNE, respectively) were masked using a Python script. This script reads the MSA in A3M format from the < job_name>_data.json file and replaces the amino acid identities corresponding to the N-terminal and C-terminal s10 strands in the query sequence with gaps (-). Positions to be masked in the query sequence are provided as a list of column indexes. The edited < job_name>_data.json can directly be used as input for AF3 prediction. As recommended in the AF3 release documentation for monomer prediction (https://github.com/google-deepmind/alphafold3/blob/main/docs/input.md) the modified MSA is provided as the “unpairedMsa” field and the “pairedMsa” field is set to an empty string. The script used to perform MSA masking is provided in the ESI. The information content for the alternative s10 strands in the AF3 MSA is shown in Figure S2.

Template-free Prediction

Template removal in AF3 is straightforward and was performed following the release instructions by providing an empty list to the “templates” field in the input. The Python script used for this operation is provided in the Supporting Information For the prediction set that is both template-free and includes column masking, inputs were generated by running the two scripts sequentially.

Analysis

AF3 returns structural models in CIF format without hydrogen positions information. For easier analysis and visualization, we converted the models to PDB format using a two-step protocol: first, files were converted to PDB using Open Babel, and then hydrogens were added using tleap from AmberTools. The commands used for these operations are provided in the ESI. AF3 models were classified into one of three categories: N-bound conformation, C-bound conformation, or unassigned (for geometries that could not be unambiguously assigned to either conformation), using a Python script provided in the ESI. Finally, structural similarity to a precyclized GFP crystallographic structure (GFP R96M, PDB 2AWJ) was computed using a Python script, also provided in the ESI.

Contact probability maps were extracted from the confidences.json file generated during the AF3 structure prediction. Scalar contact probabilities between duplicated segments were computed by summing the matrix elements corresponding to contacts between the two segments. Scripts to perform these operations are provided in the ESI.

Results

AlphaFold3 Predicts Overlapping Insertion of the Alternative Strands with Large Clashes

We performed AF3 predictions for the seven full-length constructs (including the His-tag) using 200 seeds and 5 samples per seed, generating a total of 1000 models per construct. Four simulation setups were tested: standard prediction (sAF3), column masking (cmAF3), template-free prediction (tfAF3), and a combination of column masking and template-free prediction (tfcmAF3, see Methods). Analysis of these structures reveals a high prevalence of steric clashes across all models except for the shortest construct (loop0) (Figure A). This behavior is reminiscent of the extreme stereochemical violations observed in homomers, which have been highlighted as potential inherent limitations of AF3. For sAF3, the number of clash-free models fluctuates with C-terminal loop length, with loop6 and loop10 exhibiting the highest number of clashes: fewer than one-fifth of their models are clash-free. Column masking, template removal, and their combination have an inconsistent impact on the percentage of clash-free models, which remains below 50% in all cases for loop10. Analysis of the models containing steric clashes reveals a common underlying cause: the incorrect prediction of both alternative strands being inserted into the barrel simultaneously with overlapping positions (Figure B). This geometry violation occurs in the majority of loop6 and loop10 predictions. In fewer cases, AF3 predicts 12-stranded instead of 11-stranded β-barrels, which was also observed in a minority of AF2 structures, where the two strands are accommodated side by side (Figure S3). In contrast, the overlapping insertion responsible for all observed clashes places both strands in virtually the same spatial position, rendering the structure physically meaningless.

2.

2

Open in a new tab

(A) Percentage of clash-free AF3 structures of the seven GFP constructs (with His-tag) obtained requesting 1000 models (200 seeds, 5 samples/per seed) with standard prediction (sAF3), column masking (cmAF3), template-free prediction (tfAF3), and combination of column masking and template-free prediction (tfcmAF3). (B) Representative AF3 model of loop6 showing the origin of the clashes, i.e. the overlapping insertion of the two alternative strands (N-ter in yellow and C-ter in green, respectively) in the barrel. (C) Percentage of N-bound population in AF3 structures obtained in the four prediction setups (sAF3, cmAF3, tfAF3, and tfcmAF3). Population percentages are computed considering only models that can be assigned to either the N-bound or C-bound conformation, removing structures containing clashes and those that cannot be unambiguously assigned to either conformation. Experimentally determined populations (expt) and populations predicted with AF2 with column masking (cmAF2) are shown for comparison.

The origin of this overlap is likely related to the conditional diffusion-based generation of atomic coordinates in AF3, which does not explicitly penalize geometry violations or steric clashes, except for enforcing the positioning of covalently bound ligands and glycans. As a result, conflicting predictions may arise in alternate frame folding systems, such as this GFP family, where folding ambiguity is intrinsic due to the near-identical duplication of a peptide sequence. Since the duplicated sequences are expected to exhibit comparable contact probabilities with the rest of the protein, they are likely to be diffused into the same spatial position (vide infra). Similarly clashed structures were obtained running AF3 predictions on the AlphaFold server (https://alphafoldserver.com/) and are available as Supporting Information

Population of the N-Bound Conformation is Underestimated with AlphaFold3

We computed the relative populations of the N-bound and C-bound conformations using the subset of clash-free AF3 models. The population of the N-bound conformation as a function of C-terminal loop length is shown in Figure C, with spectroscopically determined reference values provided for comparison.

For standard AF3, as with standard AF2, the population of the N-bound conformation is underestimated for all constructs except for the largest one (loop14). Specifically, it is predicted to be at most 1% up to loop6, whereas experimental data place it at 50%, 58%, 65%, and 80% for loop1, loop2, loop4, and loop6, respectively. While AF2 predictions could be improved by targeted column masking of the MSA corresponding to the two alternative strands (blue bars in Figure C), the same MSA column masking, template removal, and their combination have a weaker and inconsistent effect on AF3 populations, leading to less accurate estimates for these systems. This behavior is expected given AF3′s reduced reliance on MSA processing compared to AF2 and the replacement of the evoformer module with the simpler pairformer module. While the reengineered architecture undoubtedly improves AF3′s accuracy and efficiency for predicting large natural proteins and complexes, it may complicate the prediction of alternate frame folding systems. Similar distributions of clash-free geometries and N-bound populations were obtained for a different combination of model seeds and number of samples per seed (10 seeds, 100 samples per seed, Figure S4), showing that modifying these parameters does not change the global outcome of the prediction.

Removing the His-Tag Reduces Overlapping Insertion of the Alternative Strands

Since the N-terminal s10 strand is preceded by an unstructured His-tag, we investigated whether its presence influences the conditional diffusion-based generation of AF3 structures, particularly regarding the overlapping of alternative strands. Indeed, the absence of the His-tag has a significant impact on AF3 geometries, substantially reducing the number of models containing steric clashes and effectively reducing structural perplexity during inference. Figure A shows the percentage of clash-free models obtained for the seven His-tag-less constructs under standard prediction (sAF3), column masking (cmAF3), template-free prediction (tfAF3), and the combination of column masking and template-free prediction (tfcmAF3).

3.

3

Open in a new tab

(A) Percentage of clash-free AF3 structures of the seven GFP constructs without His-tag obtained requesting 1000 models (200 seeds, 5 samples/per seed) with standard prediction (sAF3), column masking (cmAF3), template-free prediction (tfAF3), and combination of column masking and template-free prediction (tfcmAF3). (B) Percentage of N-bound population in AF3 structures obtained in the four prediction setups (sAF3, cmAF3, tfAF3, and tfcmAF3). Population percentages are computed considering only models that can be assigned to either the N-bound or C-bound conformation, removing structures containing clashes and those that cannot be unambiguously assigned to either conformation. Experimentally determined populations (expt) and populations predicted with AF2 with column masking (cmAF2) are shown for the corresponding constructs with His-tag for comparison. (C) Average pTM scores for clash-free models and models containing steric clashes, obtained from the four prediction setups (sAF3, cmAF3, tfAF3, and tfcmAF3) for the seven constructs with and without the His-tag. The error bars indicate 2s, where s is the standard deviation. A dashed line marks a pTM score threshold of 0.8.

In contrast to what is observed in the presence of the His-tag, over 90% of sAF3 models are clash-free for all constructs except the shortest (loop0). This suggests that unstructured regions spatially close to potentially ambiguous segments may play a nontrivial role in determining the atomic coordinates of AF3 models. Additionally, the absence of the His-tag makes the predicted N-bound population more responsive–although still inconsistently–to column masking and the use of structure templates in the prediction (Figure B). This results in an increased population of the N-bound conformation with the growing size of the C-terminal loop. However, these predicted populations do not necessarily match the experimental values, as the latter were determined only for His-tagged constructs. Despite the absence of experimental data for His-tag-less constructs, the pronounced difference in predicted populations between His-tagged and His-tag-free constructs is likely spurious, given that predictions in both cases are guided by essentially the same templates and MSA information (Tables S1 and S2). These findings suggest that the proximity of an unstructured region directly attached to one of the duplicated segments can be a significant source of error in AF3 predictions of alternative conformations in alternate frame folding systems. Unstructured regions should therefore be neglected when aiming to quantify population ratios. Similar results were obtained for the corresponding models generated using 10 seeds and 100 samples per seed (Figure S4), again demonstrating that these parameters do not determine the prediction outcome in this family of proteins.

Figure C shows the average pTM scores of AF3 models, classified by simulation setup, presence or absence of the His-tag, and presence or absence of steric clashes. pTM scores are distributed within a narrow range. sAF3 calculations in the absence of the His-tag yield higher pTM values (greater confidence) than their counterparts with the His-tag, which is expected since the disordered His-tag region lowers the overall prediction confidence. The removal of structural templates slightly reduces the prediction confidence of His-tag-less models. Importantly, we find no significant differences in pTM scores between clash-free models and those containing steric clashes, indicating that prediction confidence cannot be used to detect this type of inaccurate behavior. Likewise, analysis of the per-atom pLDDT score of the two s10 strands, a local confidence metric, shows overlapping populations for clashed and clash-free models (Figure S5).

To assess structure prediction accuracy of clash-free models, we calculated the root-mean-square deviation (RMSD) of the central α-helix relative to the crystallographic structure of a precyclized R96M GFP variant (PDB 2AWJ). A previous study demonstrated that this geometric descriptor can distinguish chromophore-forming proteins, such as GFP, from nonchromophore-forming ones, where high structural similarity (low RMSD) between AlphaFold models and the crystallographic reference indicates high prediction accuracy. All GFP models exhibit RMSD values between 0.4 and 0.8 Å, indicating high accuracy comparable to that of AF2 models. (Figures S6–S9).

Overlap of Alternatively Folding Regions is Reflected in Contact Probabilities

We further investigated the origin of the overlapping insertion of alternative strands by analyzing the contact probability matrix generated in sAF3 structure predictions. These matrices reflect the inherent ambiguity of alternate frame folding systems, as the contact probabilities for the alternative s10 β-strands are mirror images of each other. That is, both strands exhibit nonzero interaction probabilities with the same protein regions (the two flanking β-strands and the central α-helix), with the only difference being higher probabilities for the C-terminal strand specifying green fluorescence (Figures S10–S23). A second observation is that the alternative strands exhibit nonzero mutual contact probabilities, despite such interactions being “impossible by design” in alternate frame folding systems. This mutual contact probability can be interpreted as a measure of perplexity in AF3. A recent systematic analysis of deep learning-based contact matrices in two-state proteins has shown that these matrices contain structural information on both states, consistent with our observations. Interestingly, for constructs predicted with the His-tag, there is an almost perfect correlation between the sum of per-residue contact probabilities between the two strands and the number of models containing steric clashes (Pearson’s coefficient ρ = 0.97, Figure A). In contrast, for His-tag-less constructs, this correlation weakens significantly mostly due to loop0 being an outlier (Pearson’s coefficient ρ = 0.45, Figure B). Removing loop0, Pearson’s coefficient raises to 0.97 despite the drastic reduction in clashed models without the His-tag (Figure S24).

4.

4

Open in a new tab

(A) Percentage of clashed models predicted with sAF3 versus the sum of contact probabilities between the two alternative strands for the seven GFP constructs with the His-tag. (B) Percentage of clashed models predicted with sAF3 versus the sum of contact probabilities between the two alternative strands for the seven GFP constructs without the His-tag. ρ represents Pearson’s correlation coefficient. Contact probabilities are averaged over all models, and the x-axis error bars indicate 2s, where s is the standard deviation.

The simultaneous encoding of both conformations–consistently observed across all constructs and irrespective of the presence of the His-tag–motivated us to further investigate whether AF3 predictions can lead to unphysical overlaps in alternate frame folding systems beyond this GFP family. To explore this, we analyzed the performance of sAF3 in predicting the alternative conformations of an engineered molecular switch for Ca2+ sensing.

A Calcium-dependent Conformational Switch Shows a Minority of Incorrect Predictions

Calbindin-based molecular switches designed by Loh and co-workers represent the first example of sensors based on alternate frame folding (Figure A). , The protein calbindin D9k is 75 amino acids long and binds two Ca2+ ions through a pair of EF-hand domains (helix-loop-helix motifs) with an affinity of ∼0.1–1 μM. The crystallographic structure of the wild-type protein (PDB 3ICB) is shown in Figure B, highlighting the relevant Ca2+ cation involved in fold-switch engineering and the coordinating Glu65 side chain, which completes a roughly pentagonal bipyramid coordination geometry (see Figure S25 for more details).

5.

5

Open in a new tab

(A) Sequence representation of calbindin D9k and its two derived alternate frame folding constructs, E65Q-1 and E65Q-2. The N-terminal duplicated segment (residues 44–75 of the original protein) is shown in pink, while the C-terminal 44–75 segment is shown in blue. Glu65, which coordinates a Ca2+ ion, is highlighted. (B) Crystallographic structure of wild-type calbindin D9k (PDB 3ICB), color-coded according to sequence position as shown in panel A. The side chain of Glu65 is represented as sticks, and Ca2+ ions are shown as spheres. (C) Left to right: sample sAF3 models of E65Q-1 and E65Q-2, and a clashed sAF3 model of E65Q-1 showing partial overlap of the two alternative segments. Structures are color-coded according to their position in the permuted and duplicated frame, matching panel A. The side chains of residues at position 65 in the two alternative segments are shown as sticks, and Ca2+ ions are represented as spheres. (D) Distance distributions between Glu65 in E65Q-1 and E65Q-2 and the nearest Ca2+ ion in AF3 models. Bars are color-coded to match panels A and C, indicating which alternative segment is involved in Ca2+ binding. Distances are measured from Ca2+ to the Cδ atom of Glu65. A dashed line marks a distance of 4 Å, indicating direct coordination or close contact.

Alternate frame folding sensing constructs are generated by duplicating the C-terminal segment spanning residues 44 to 75 and appending it to the N-terminus via a short linker. This permuted and duplicated architecture is then subjected to mutagenesis, yielding constructs E65Q-1 and E65Q-2. In E65Q-1, the Glu65 residue in the C-terminal duplicated segment is mutated to Gln, while in E65Q-2, the Glu65 residue in the N-terminal segment is mutated to Gln (Figure A). As with all alternate frame folding systems, the folding of the duplicated segments is mutually exclusive, i.e. only one can fold to complete the native structure at a given time, while the other remains partially disordered. In the absence of Ca2+, the equilibrium between the two forms is governed solely by the folding thermodynamics of the segments. However, this equilibrium can be shifted by the addition of Ca2+. Since Glu65 coordinates Ca2+ through its side chain, its mutation to Gln reduces the affinity by a factor of 105. Consequently, in the presence of Ca2+, the construct undergoes a conformational change, folding the segment that retains Glu65 and is competent for Ca2+ coordination, i.e. the N-terminal segment in E65Q-1 and the C-terminal segment in E65Q-2 (Figure A).

Unlike AF2, AF3 can predict protein structures in the presence of cofactors, including metal ions such as Ca2+. To assess whether AF3 (i) responds to the preference for folding the Glu65-containing segment in the presence of calcium and (ii) exhibits unphysical overlap of alternative segments, we performed sAF3 structure predictions for E65Q-1 and E65Q-2. Predictions followed the same methodology described for GFPs, using 200 seeds and 5 samples per seed, the only difference being the inclusion of two Ca2+ ions. A sample AF3 input is provided in the ESI.

The majority of AF3 models adopt the intended geometry, where the segment containing Glu65 binds a Ca2+ ion, completing the calbindin fold. This is exemplified by the models shown in Figure C (left and center), which represent the fold observed in over 80% of predicted structures (Figure D). Most of the remaining models misassign the segment binding the Ca2+ ion, placing the latter in contact with Gln65 (on the C-terminal segment for E65Q-1 and on the N-terminal segment for E65Q-2; see Figures S26 and S27). The higher percentage of correctly aligned structures in E65Q-2 compared to E65Q-1 is tentatively attributed to the fact that E65Q-2 retains Glu65 in the C-terminal segment, maintaining higher sequence similarity to the wild-type protein.

Notably, both alternative segments are consistently predicted as folded, regardless of their interaction with Ca2+ or with the nonduplicated region (residues 1–43 in wild-type numbering). However, the structuring of the orphan segment is at least partially spurious, as NMR data suggest that it is neither well-folded nor completely unstructured. This spurious structuring is reflected in the contact probability maps, which, as in the case of GFPs, show nearly identical interaction patterns for the two alternative segments (Figures S28 and S29). Additionally, as observed in GFPs, a nonzero contact probability is detected between the alternative segments. In only three models (<1%), we identify significant steric clashes resulting from the overlapping folding of the alternative segments, exemplified in the right-hand panel of Figure C. The lower incidence of clashed models in this case may originate from the type of fold and the absence of unstructured regions in proximity of duplicated segments.

Discussion

Overall, these results indicate that unphysical overlap of duplicated regions in alternate frame folding systems can occur in AF3 structure predictions. This arises from the combination of (i) the intrinsic ambiguity of a single sequence capable of adopting two dramatically different conformations but an overall identical fold and (ii) the absence of an explicit penalty term for geometry violations in AF3′s model quality evaluation. Clashed structures represent a minority of the predicted calbindin and His-tag-less GFP models (27% of loop0 models and <7% in the remaining systems) but are the majority of predictions in some His-tagged GFPs. While the limited availability of accurate population data for alternate frame folding systems prevents us from drawing general conclusions, the two families of fold switches presented here illustrate that the accuracy of AF3 ensemble predictions is indeed case-dependent, with a higher number of clashed predictions observed for GFPs. Previous studies on fold-switching proteins that adopt two distinct conformations with differences in secondary and/or tertiary structure have suggested that AF predictions of conformational ensembles may reflect memorization of training-set structures rather than an understanding of underlying protein energetics. In our case, direct memorization of the alternate frame folding constructs is not possible, since their crystallographic structures have not been determined, but predictions may still be influenced by memorization of the parent, unpermuted and unduplicated proteins. This is particularly relevant for GFP, given that the PDB contains over 600 GFP-related entries, compared to only 15 calbindin-related entries.

It has also been shown that MSAs used for prediction can bias conformational preferences toward structures whose signals are more strongly represented, thereby obscuring conformations with weaker MSA signals. Since MSAs are constructed by querying sequence databases predominantly composed of natural proteins (i.e., unpermuted and unduplicated), they provide uneven coverage of the duplicated segments in the two classes of conformational switches analyzed here (Figures S2 and S30). This bias affects population prediction and likely represents a general limitation in the application of AF3 to alternate frame folding systems. Similarly, permutation and duplication can influence the outcome of template searches, thereby affecting the structural information used in prediction. For GFPs, AF3 templates are limited to two structures (Table S1), both corresponding to the ratiometric FRET-based calcium sensor Twitch-2B, which consists of two GFP units flanking a calcium-binding domaina completely different architecture from alternate frame folding GFPs. Weak structural alignment with such templates, combined with unbalanced MSAs, may represent a general source of error in predicting alternate frame folding systems. This could explain the unexpected increase in prediction accuracy in template-free and column-masked predictions for His-tag-less GFPs (which are not biased by MSA or template information, Figure B), relative to other prediction setups. These limitations are not an inherent deficiency of AF3, which is designed to map natural protein sequences to their three-dimensional structures. However, they are worth highlighting for future studies that may attempt to use this tool to predict the structures of proteins engineered through alternate frame folding.

Conclusions

In this work, we assess the ability of AlphaFold3 to predict alternative conformations of proteins engineered through alternate frame folding. Compared to AlphaFold2, we find that AlphaFold3 predictions of alternative conformations in engineered GFPs are less responsive to modifications of the multiple sequence alignment and generally do not match experimentally determined populations. We show that extreme geometry violations (large steric clashes), similar to those sporadically reported for homomers, systematically affect engineered GFP models when a disordered His-tag is connected to one of the alternatively folding strands. The same phenomenon is observed, albeit at a lower frequency, in His-tag-less GFPs and calbindin-based switches.

We discuss how alternative conformations are encoded in the contact probabilities associated with AlphaFold3 structure prediction and how steric clashes arise naturally from the intrinsic ambiguity of alternate frame folding sequences in the absence of an explicit penalty term for atomic clashes. In light of these findings, AlphaFold2’s structure module architecture and its higher responsiveness to multiple sequence alignment modifications make it a more suitable tool for predicting the conformations of alternate frame folding systems.

Supplementary Material

ci5c00906_si_001.pdf (40.1MB, pdf)

Acknowledgments

This research has been funded by MCIN/AEI/10.13039/501100011033 (grants RYC2022-036457-I and EUR2023-143462 to FP, PID2021-125946OB-I00 to GJO).

The following data are available through the Zenodo repository https://zenodo.org/records/15224509; doi 10.5281/zenodo.15224508: AlphaFold3 models of GFP- and calbidin-based constructs. No restrictions on data availability apply. Sample code used for analysis is provided in the ESI. The AlphaFold3 inference pipeline 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.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.5c00906.

  • Scheme of the AlphaFold3 simulation setups employed; sequence logo representation of AlphaFold3 multiple sequence alignments for the seven GFP constructs; representation of AlphaFold3 GFP construct models; population of clashed, N-bound, and C-bound models for GFP constructs obtained with 10 seeds and 100 samples per seed; average per-atom pLDDT scores of the two alternative s10 strands; distribution of RMSD values of AlphaFold3 models relative to a precyclized GFP crystallographic structure; representation of GFP constructs contact probability matrices; representation of calbindin calcium coordination spheres; representation of calbindin constructs AlphaFold3 models; representation of calbindin constructs contact probability matrices; constructs sequences; sample AlphaFold3 inputs; sequence logo representation of calbindin-based constructs; MSA and templates employed in AlphaFold3 prediction of GFP structures; Python scripts used for analysis (PDF)

Both authors contributed to the design and implementation of the research, to the analysis of the results and to the writing of the manuscript.

The authors declare no competing financial interest.

References

  1. 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]
  2. Bryant P., Noé F.. Structure prediction of alternative protein conformations. Nat. Commun. 2024;15:7328. doi: 10.1038/s41467-024-51507-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. 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]
  4. Wallner B.. AFsample: improving multimer prediction with AlphaFold using massive sampling. Bioinformatics. 2023;39:btad573. doi: 10.1093/bioinformatics/btad573. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Stein R. A., Mchaourab H. S.. SPEECH_AF: Sampling protein ensembles and conformational heterogeneity with Alphafold2. PLoS Comput. Biol. 2022;18:e1010483. doi: 10.1371/journal.pcbi.1010483. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. 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]
  7. Brown B. P., Stein R. A., Meiler J., Mchaourab H. S.. Approximating Projections of Conformational Boltzmann Distributions with AlphaFold2 Predictions: Opportunities and Limitations. J. Chem. Theory Comput. 2024;20:1434–1447. doi: 10.1021/acs.jctc.3c01081. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. del Alamo D., Sala D., Mchaourab H. S., Meiler J.. Sampling alternative conformational states of transporters and receptors with AlphaFold2. eLife. 2022;11:e75751. doi: 10.7554/elife.75751. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Chakravarty D., Lee M., Porter L. L.. Proteins with alternative folds reveal blind spots in AlphaFold-based protein structure prediction. Curr. Opin. Struct. Biol. 2025;90:102973. doi: 10.1016/j.sbi.2024.102973. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Soltani A.. Sampling the conformational states of P2X receptors with AlphaFold3. J. Proteins Proteom. 2025;16:55–86. doi: 10.1007/s42485-025-00177-9. [DOI] [Google Scholar]
  11. Vats S., Bobrovs R., Söderhjelm P., Bhakat S.. AlphaFold-SFA: Accelerated sampling of cryptic pocket opening, protein-ligand binding and allostery by AlphaFold, slow feature analysis and metadynamics. PLoS One. 2024;19:e0307226. doi: 10.1371/journal.pone.0307226. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Casadevall G., Duran C., Osuna S.. AlphaFold2 and Deep Learning for Elucidating Enzyme Conformational Flexibility and Its Application for Design. JACS Au. 2023;3:1554–1562. doi: 10.1021/jacsau.3c00188. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Monteiro da Silva G., Cui J. Y., Dalgarno D. C., Lisi G. P., Rubenstein B. M.. High-throughput prediction of protein conformational distributions with subsampled AlphaFold2. Nat. Commun. 2024;15:2464. doi: 10.1038/s41467-024-46715-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Schafer J. W., Porter L. L.. AlphaFold2’s training set powers its predictions of some fold-switched conformations. Protein Sci. 2025;34:e70105. doi: 10.1002/pro.70105. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Ha, J.-H. ; Loh, S. N. . Synthetic Protein Switches Methods and Protocols; Springer New York: New York, NY, 2017; pp 27–41. [Google Scholar]
  16. Abramson 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]
  17. 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]
  18. Do K., Boxer S. G.. GFP Variants with Alternative β-Strands and Their Application as Light-driven Protease Sensors: A Tale of Two Tails. J. Am. Chem. Soc. 2013;135:10226–10229. doi: 10.1021/ja4037274. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Chudakov D. M., Matz M. V., Lukyanov S., Lukyanov K. A.. Fluorescent Proteins and Their Applications in Imaging Living Cells and Tissues. Physiol. Rev. 2010;90:1103–1163. doi: 10.1152/physrev.00038.2009. [DOI] [PubMed] [Google Scholar]
  20. Romei M. G., Boxer S. G.. Split Green Fluorescent Proteins: Scope, Limitations, and Outlook. Annu. Rev. Biophys. 2019;48:19–44. doi: 10.1146/annurev-biophys-051013-022846. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Lawrence M. S., Phillips K. J., Liu D. R.. Supercharging Proteins Can Impart Unusual Resilience. J. Am. Chem. Soc. 2007;129:10110–10112. doi: 10.1021/ja071641y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Doerr A.. Supercharging through the cell membrane. Nat. Methods. 2009;6:322. doi: 10.1038/nmeth0509-322. [DOI] [Google Scholar]
  23. Huang K. Y., Cardenas L., Ellington A. D., Walker D. J. F.. Supercharged fluorescent proteins detect lanthanides via direct antennae signaling. Nat. Commun. 2024;15:9200. doi: 10.1038/s41467-024-53106-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. John A. M., Sekhon H., Ha J.-H., Loh S. N.. Engineering a Fluorescent Protein Color Switch Using Entropy-Driven β-Strand Exchange. ACS Sens. 2022;7:263–271. doi: 10.1021/acssensors.1c02239. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. O’Boyle N. M., Banck M., James C. A., Morley C., Vandermeersch T., Hutchison G. R.. Open Babel: An open chemical toolbox. J. Cheminf. 2011;3:33. doi: 10.1186/1758-2946-3-33. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Case D. A.. et al. AmberTools. J. Chem. Inf. Model. 2023;63:6183–6191. doi: 10.1021/acs.jcim.3c01153. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Wood T. I., Barondeau D. P., Hitomi C., Kassmann C. J., Tainer J. A., Getzoff E. D.. Defining the Role of Arginine 96 in Green Fluorescent Protein Fluorophore Biosynthesis. Biochemistry. 2005;44:16211–16220. doi: 10.1021/bi051388j. [DOI] [PubMed] [Google Scholar]
  28. Hartley S., Tiernan K., Ahmetaj G., Cretu A., Zhuang Y., Zimmer M.. AlphaFold2 and RoseTTAFold predict posttranslational modifications. Chromophore formation in GFP-like proteins. PLoS One. 2022;17:e0267560. doi: 10.1371/journal.pone.0267560. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Li J., Wang L., Zhu Z., Song C.. Exploring the Alternative Conformation of a Known Protein Structure Based on Contact Map Prediction. J. Chem. Inf. Model. 2024;64:301–315. doi: 10.1021/acs.jcim.3c01381. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Stratton M. M., Mitrea D. M., Loh S. N.. A Ca2+-Sensing Molecular Switch Based on Alternate Frame Protein Folding. ACS Chem. Biol. 2008;3:723–732. doi: 10.1021/cb800177f. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. DeGrave A. J., Ha J.-H., Loh S. N., Chong L. T.. Large enhancement of response times of a protein conformational switch by computational design. Nat. Commun. 2018;9:1013. doi: 10.1038/s41467-018-03228-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Szebenyi D. M., Moffat K.. The refined structure of vitamin D-dependent calcium-binding protein from bovine intestine. Molecular details, ion binding, and implications for the structure of other calcium-binding proteins. J. Biol. Chem. 1986;261:8761–8777. doi: 10.1016/S0021-9258(19)84447-2. [DOI] [PubMed] [Google Scholar]
  33. Stratton M. M., McClendon S., Eliezer D., Loh S. N.. Structural Characterization of Two Alternate Conformations in a Calbindin D9k-Based Molecular Switch. Biochemistry. 2011;50:5583–5589. doi: 10.1021/bi102040g. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Thestrup T.. et al. Optimized ratiometric calcium sensors for functional in vivo imaging of neurons and T lymphocytes. Nat. Methods. 2014;11:175–182. doi: 10.1038/nmeth.2773. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

ci5c00906_si_001.pdf (40.1MB, pdf)

Data Availability Statement

The following data are available through the Zenodo repository https://zenodo.org/records/15224509; doi 10.5281/zenodo.15224508: AlphaFold3 models of GFP- and calbidin-based constructs. No restrictions on data availability apply. Sample code used for analysis is provided in the ESI. The AlphaFold3 inference pipeline 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.

Articles from Journal of Chemical Information and Modeling are provided here courtesy of American Chemical Society

RESOURCES