When can AlphaFold predict the oligomeric states of proteins?

INTRODUCTION

Between 30% and 50% of all proteins are thought to self‐associate into stable complexes consisting of multiple copies of themselves (Levy & Teichmann,2013; Schweke et al.,2024). This oligomerisation process is important for the structural and functional properties of these proteins, as it enables the formation of repetitive structural elements, catalytic interfaces in enzymes and the central pore in many membrane proteins. Establishing the biologically relevant oligomeric state of proteins and the structures they adopt is thus critical for understanding the way in which they perform their biological roles.

A variety of biochemical and biophysical techniques such as native mass spectrometry, size exclusion chromatography and native gel electrophoresis are available to characterize the oligomeric states of proteins (Chetri et al.,2022). However, these experimental techniques can sometimes be expensive, time‐consuming and limited by technical constraints. Existing computational tools to characterize protein oligomerisation states include sequence similarity‐based methods which predict the oligomerisation state using related proteins with resolved structures (Baek et al.,2017). Once the number of subunits is known, a model of the oligomer can be built using a structural templating technique if a relevant template exists or ab initio docking where this is not possible. However, this method relies on the existence of oligomeric state information from a related protein. More recently developed deep learning tools trained on large datasets of curated subunit information from experimental data predict oligomeric states with relatively high accuracy but do not provide structural information (Avraham et al.,2023; Deng et al.,2024). Currently, for yet to be characterized proteins, the AlphaFold Protein Structure Database (AFDB) provides predicted monomeric structures but lacks predictions for functionally relevant multimeric states which may provide key insights into their physiological behavior (Varadi et al.,2024).

Since its release, AlphaFold2 Multimer has found wide application—including to identify plant–pathogen interactions across two different species (Homma et al.,2023), to predict protein–protein complexes in yeast (Humphreys et al.,2021), and capture interactions between intrinsically disordered protein regions (Omidi et al.,2024). Additionally, recent proteome wide predictions using AlphaFold2 highlighted the high proportion of proteins which form homomers and highlight the potential for structural prediction tools to model protein homooligomerization (Schweke et al.,2024). AlphaFold3 provides further improvements in accuracy for the modeling of protein–protein interactions, in addition to the ability to model structures of other biological molecules (Abramson et al.,2024). More recently, AlphaFold and AlphaFold based tools have also been applied to predict oligomeric states and model structures of homomeric complexes across different datasets (Luo et al.,2025; Madaj et al.,2025; Shor et al.,2024).

Here, we investigate whether AlphaFold‐Multimer v2.3 (AF2‐M) and AlphaFold3 (AF3) can correctly predict both the oligomeric states of proteins and the corresponding multimeric structures. Using a curated benchmark set of 40 proteins, we show that the interface predicted TM (ipTM) score reported by both methods is a reliable indicator of oligomeric state, and that the resulting multimeric structures closely resemble experimentally resolved assemblies.

We then extend this approach to a dataset of over 1000 proteins and observe a similar overall level of accuracy in oligomeric state prediction following filtering to exclude low‐confidence structural predictions. However, we find that proteins lacking structures or close homologues in the AlphaFold training set are substantially less likely to be predicted in the correct oligomeric state, even when their predicted structures are of high confidence. Based on this analysis, we identify prediction strategies that minimize computational cost while maintaining accuracy for proteins with unknown oligomeric states. To increase the diversity of protein folds within our dataset, we extend our analysis to a further 3560 proteins and observe comparable accuracy across a diversity of protein folds, with a median success rate of 86.7% across 801 unique protein X Groups. Finally, we apply this technique to a select set of membrane proteins and show that small differences in ipTM scores between multiple oligomeric states can indicate proteins that may adopt more than one biologically relevant assembly.

RESULTS

First, we asked whether the ipTM score reported by AF2‐M, which measures the accuracy of the predicted relative positions of subunits within a protein–protein complex, can be used to correctly predict known protein oligomeric states.

Using the dataset curated from the UniProt Database by Deng et al. (2024), for their study on oligomeric state prediction using a deep learning approach (Deng et al.,2024), we extracted a sample of 40 proteins with known oligomeric states (TableS1). This sample was selected to contain eight examples each of proteins known to have oligomeric states of 1, 2, 3, 4 and 5 and contained a mix of soluble and membrane proteins (6 single‐pass and 10 multi‐pass). This set contains only human proteins and cannot represent the diversity of possible protein structures but provides an initial test of the potential for AF2 to predict oligomeric states. For each protein, we used AF2‐M to predict their structures as homodimers, homotrimers, homotetramers and homopentamers. We used three recycles and predicted 20 structures in each oligomeric state for a total of 80 structures produced per protein.

For the 32 multimeric proteins in our dataset, the distribution of ipTM scores for all structures predicted in the correct oligomeric state (e.g., a known dimeric protein predicted with two subunits) differs considerably from the distribution found for all structures predicted in an incorrect oligomeric state (Figure1a). In general, higher ipTM scores are observed (peak at ipTM 0.75–1) for proteins predicted in their annotated oligomeric states compared to an incorrect state (peak at ipTM 0–0.32). While there exists some overlap between the distributions, the presence of distinct differences between these peaks suggests the ipTM score reported by AF2‐M can distinguish between correctly and incorrectly predicted oligomeric states. For the eight known monomeric proteins in our test set, predicting them as homodimers, homotrimers, homotetramers, or homopentamers consistently yielded low ipTM scores (0–0.4). The distribution of their ipTM scores differs substantially to that of the known 32 known multimeric proteins predicted in any oligomeric state (Figure1b), suggesting that AF2‐M can also distinguish between proteins likely to be monomeric and those forming higher order protein assemblies.

ipTM scores from AlphaFold2 Multimer correctly predict oligomeric states. (a) Probability distribution plot of ipTM scores for 32 known multimeric proteins predicted in the correct (green) and incorrect (pink) oligomeric states. (b) Probability distribution plot of ipTM scores for 8 known monomeric proteins predicted with 2,3,4,5 subunits (blue) and 32 known multimeric proteins predicted with 2,3,4,5 subunits (pink). The gray line indicates the cut‐off ipTM value of 0.32. (c) Number of correctly assigned oligomeric states (out of 40) vs. cut‐off value selected for 4 different methods to distinguish between monomeric and multi‐subunit proteins. The gray line indicates the optimal cut‐off ipTM value of 0.32. (d) Box plots showing individual ipTM scores for different folded oligomeric states across the test set of 40 proteins with the correct oligomeric state highlighted in yellow. The dashed line indicates the cut‐off ipTM value of 0.32. UniProt IDs are noted in green for correct predictions and red for incorrect ones.

These data indicate that we may be able to predict a protein's most likely oligomeric state by selecting the state which yields the highest mean or max ipTM score (calculated across the 20 structures predicted for each protein) if we can firstly define a cut‐off value to assign a protein as monomeric, since monomeric predictions do not have an associated ipTM score. A cut‐off value that is too low would lead to incorrect assignment of some monomeric proteins as multimeric and a too high cut‐off value could lead to incorrect assignment of some multimeric proteins as monomeric. We optimized the cut‐off value across four methods – taking the mean of all ipTM scores for a protein (mean), using the mean of the highest ipTM value for each oligomeric state (mean max), using the maximum of the mean ipTM value for each oligomeric state (max mean), and taking the maximum ipTM score for the protein (max) (Figure1c). We found that taking the mean of the maximum ipTM score of each oligomeric state for the protein led to the highest number of correct states assigned when we optimize this cut‐off value to 0.32 and use the max ipTM to assign the oligomeric state if it is not a monomer.

Using this method, we find that 35/40 proteins in our dataset are correctly assigned. Individual ipTM distributions across the 40 proteins highlight trends across the different types of oligomeric states (Figure1d). Across monomeric proteins, predictions at all oligomeric states yield low ipTM scores. In this set of proteins, we find that all monomeric, homotrimeric, and homopentameric proteins were assigned to the correct oligomeric state. Three homodimers were incorrectly assigned – as a monomer (UniProt ID:P06727), tetramer (UniProt ID:O94804), and trimer (UniProt ID:Q2M3C6). Additionally, two homotetramers were incorrectly assigned as trimers (UniProt IDs: O7549,O43827). Interestingly, all incorrectly assigned proteins had low ipTM scores across all oligomeric states, even those that were not ultimately assigned as monomeric, suggesting that the absolute values of the ipTM scores may be used to gauge confidence in the prediction if a protein is predicted as a homooligomer.

We note that the annotated oligomeric states for each protein typically have smaller ipTM ranges when compared to other oligomeric states (FigureS1a). Additionally, the average pLDDT (a per‐residue measure of local confidence) cannot be used to distinguish between annotated and other oligomeric states, as the monomeric form usually has the highest pLDDT score (FigureS1b). ipTM values are not statistically significantly impacted by the random seed, indicating that the oligomeric state assignment is not dependent on specific individual structures produced by AF2.

Next, we examined whether the structures of the oligomeric assemblies predicted resembled the monomeric structures in the AFDB and experimentally determined structures. Snapshots of the top ranked structural predictions in the correct oligomeric state were produced for each protein after aligning to the monomeric AFDB structure (Figure2a).

Multimeric structures predicted by AlphaFold2 resemble experimentally determined structures. (a) Snapshots of the top ranked structural predictions in the correct oligomeric state for each multimeric protein (ribbon, colored by chain), aligned with the AFDB prediction of the protein as a monomer (ribbon, dark green). The presence of an experimentally resolved structure in the PDB or AlphaFold2 training set is indicated with an icon near each UniProt ID. The structures of proteins predicted in an incorrect oligomeric state are highlighted in red. (b) Histogram showing the length of the longest helix in the corresponding AFDB monomer structure for proteins predicted in the correct (green) and incorrect (pink) oligomeric states. (c) Probability distribution plot of average monomeric pLDDT values (using the AFDB structure) for proteins predicted with the correct (green) and incorrect (pink) oligomeric states. (d) Two‐way table showing the distribution of proteins with oligomeric states predicted correctly and incorrectly based on whether an experimentally resolved structure exists within the AlphaFold2 training set. (e) TM score between the top ranked AlphaFold2 predicted structure and experimentally resolved structure (if available) plotted against the TM score between the top ranked AlphaFold2 predicted structure and the monomeric structure from the AFDB, colored by annotated oligomerisation state. (f) ipTM scores of the annotated (green) and incorrect (red, various line styles) states across 0, 1, 2, 3, 5, and 10 recycles for incorrectly assigned dimers (left) and incorrectly assigned tetramers (right)

Proteins where the oligomeric state was incorrectly predicted by AF2‐M all have abnormally long helices in both the monomeric and oligomeric structures (Figure2a,b), with the longest helices spanning 64–173 residues. We examined also whether the average monomeric pLDDT, calculated using the structure present in the AFDB, indicates whether the correct oligomeric state will be predicted (Figure2c). In this sample set, if the pLDDT value was >90 (14 structures), then the correct oligomeric state was assigned. If the pLDDT value was <50 (1 structure), then an incorrect oligomeric state was predicted.

The presence of an experimentally resolved structure in the AlphaFold2.3 training set (prior to 30 September 2021) did not have a large impact on whether the correct oligomeric state was predicted in this set of proteins (Figure2d). 28/31 of structures present in the training set were correctly predicted compared to 7/9 structures absent from the training set being correctly predicted.

Next, we quantified the similarity between the predicted multimeric structures to both experimentally resolved structures and to AF2 monomeric predictions found in the AFDB in the cases for which the oligomeric state was correctly predicted and an experimentally resolved structure exists (Figure2eand TableS2). The TM score between the computationally predicted and experimentally determined structures is generally very high across all oligomerisation states, with most predicted structures having TM scores of >0.95 to experimentally resolved structures. This suggests that predicted structures of higher order oligomers are likely to be accurate. In many cases, the predicted structures of subunits in the multimer showed significant differences to the AF2 predicted monomeric structure (Figure2e), suggesting changes to the individual subunit structure imposed by oligomer formation. Visual analysis of these cases (e.g., UniProt ID: P5O591) indicates that this is usually due to changes in the relative orientations of protein domains and highlights a limitation of using predicted monomeric structures from the AFDB.

Taken together, this suggests that AF2‐M can predict the oligomeric states and structures of homomeric proteins accurately, independent of whether the protein structure is present in the training set. The oligomeric state prediction is likely to be accurate if the monomeric pLDDT is high and there are no abnormally long alpha helices in the monomeric structure.

For the five proteins which were predicted incorrectly using three recycles and 20 structures, increasing the number of recycles to 5 or 10 does not correct the prediction and ipTM scores for structures predicted in both the incorrect and annotated oligomeric states do not further increase after five recycles (Figure2f). The use of the alternative interface metric ipSAE rescued one of these proteins but incorrectly predicted an additional five (FigureS2a,b).

While use of the raw ipTM metric was relatively successful for an initial benchmark set of 40 proteins, we next evaluated its performance at predicting the correct oligomeric state with a substantially larger and more challenging dataset. While some proteins in the benchmark set lacked experimentally resolved structures, a similarity search with Foldseek (van Kempen et al.,2024) revealed that highly similar homologous structures were present in the AlphaFold training set for these proteins (TableS1), raising the question of how strongly accurate oligomeric state prediction depends on the presence of close structural homologues in the training data. To explicitly test this, we randomly selected 1006 proteins from the DeepSub database with annotated homo‐oligomeric states, none of which had experimental structures present in the AlphaFold training set, reducing the possibility of template memorisation. Each protein was modeled as a dimer, trimer, tetramer, and pentamer using AF2‐M, and oligomeric state prediction was assessed based on relative ipTM values across these assemblies.

Proteins were divided into two classes based on the Foldseek determined TM score between the AFDB monomeric structure and experimentally determined structures in the AlphaFold training set. Of the 1006 proteins, 424 had at least one homologous structure with a Foldseek TM score >0.5, while 582 proteins showed no detectable similarity to any PDB structure (TM score ≤0.5), corresponding to proteins for which AlphaFold has no close structural reference.

Compared to the initial benchmark set, the overall proportion of proteins predicted correctly by AF2‐Multimer was slightly lower for both TM score >0.5 and TM score ≤0.5 proteins (80.7% and 66.4%, respectively). However, filtering predictions by pLDDT, using either the monomeric structure or the predicted oligomeric complex, improved accuracy for both groups, with maximum accuracy occurring for structures where the pLDDT of the oligomeric complex was above 90 (88.5% for TM score >0.5 proteins, and 82% for TM score ≤0.5 proteins) (Figures3aandS4). In contrast, the Foldseek TM score for each protein strongly influenced the likelihood of correct oligomeric state prediction, with proteins that had TM scores <0.3 showing the lowest success rates (~30–75% predicted correctly) regardless of filtering by pLDDT (Figure3b). A similar pattern is seen when plotting accuracy against the sequence similarity to structures in the PDB (Figure3c). We also observe a strong relationship between accuracy and the monomeric pLDDT score (Figure3d). The most common error in the predictions was assigning dimeric proteins to be monomers (Figure3e). As the difference between the ipTM scores of the top and second ranked oligomeric states is sometimes very small (see Figure1d), we conducted bootstrap analysis using the 20 structures and calculated the ipTM margin across the TM set to evaluate whether these metrics can provide information on whether the prediction is likely to be correct (FigureS3). We find that neither the bootstrap confidence (what proportion of time the same oligomeric state is assigned when one structure at each oligomeric state is selected over 1000 trials) nor the margin between the best and next best ipTM scores (of different oligomeric states) represents clear indicators of whether a prediction is correct. However, we note that a larger proportion of correct structures can be found at a bootstrap confidence of 1.0 and ipTM margin >0.3.

Oligomeric state prediction is less accurate for proteins that lack structural similarity to proteins in the AlphaFold training set. (a) Proportion of proteins predicted correctly by AF2‐M from a set of approximately 500 proteins Foldseek TM score <0.5 to any protein in the AlphaFold training set (pink), and approximately 500 proteins with Foldseek TM score >0.5 to at least one protein in the AlphaFold training set (blue), filtered by different pLDDT cutoffs for the predicted oligomeric complexes. For monomeric proteins, the pLDDT value used corresponds to the folded oligomeric state with the highest mean ipTM. (b) Proportion of proteins predicted correctly as a function of Foldseek TM score. Data are filtered by the pLDDT of the predicted oligomeric complexes and shown as separate lines for each pLDDT cutoff. A purple square is added to highlight proteins with a TM score of <0.3. Marker size is proportional to the number of proteins per bin. The number of proteins per bin is reported in FigureS4e. (c) Proportion of proteins predicted correctly as a function of MM2Seq Sequence similarity to proteins in the AF2 training set. A purple square is added to highlight proteins with a sequence similarity <0.25. Data filtering and point size as in “b.” (d) Proportion of proteins predicted correctly as a function of monomer and complex pLDDT. Data filtering and point size as in “b.” (e) Confusion matrix of proteins that were assigned to the incorrect oligomeric state using AF2‐M. Rows represent the annotated oligomeric state of each protein, and columns represent the oligomeric state that the protein got incorrectly assigned to. Darker cells indicate higher protein counts. (f) Proportion of proteins with Foldseek TM scores <0.3 to any protein in the AlphaFold training set that were predicted correctly using different interface confidence metrics, including ipSAE, ipSAE d0 chn (ipSAE d0c), ipSAE d0 dom (ipSAE d0d), ipTM d0 chn (ipTM d0c), pDockQ, pDockQ2, LIS, and ipTM. (g) Heatmap showing the proportion of correct predictions as a function of the number of structures and number of recycles used. (h) Number of correct oligomeric state predictions (using monomeric cut off value of 0.32) vs. % of max compute cost across 0, 1, 2, 3 recycles and 1, 5, 10, 20 structures predicted. The dashed gray line indicates the highest number of correct predictions across all parameters.

Recent studies have suggested that ipTM alone may be insufficient for assessing the quality of AlphaFold‐predicted complexes and have proposed alternative interface‐focused metrics such as ipSAE, ipSAE d0 chn, ipSAE d0 dom, ipTM d0 chn (Dunbrack,2025), pDockQ (Bryant et al.,2022), pDockQ2 (Zhu et al.,2023), and LIS (Kim et al.,2024). To determine whether these metrics improved oligomeric state prediction for structurally novel proteins, we evaluated prediction accuracy using each metric for proteins with TM scores <0.3. However, none of the alternative metrics substantially outperformed ipTM for this subset (Figure4c). Similarly, ipTM also outperformed other metrics for identifying the correct oligomeric state for our full datasets of TM score ≤0.5 and TM score >0.5 proteins (FigureS4a,b). Together, these results indicate that proteins lacking structural similarity to the AlphaFold training set are intrinsically more difficult to classify correctly, and that this limitation is not overcome by current interface confidence metrics. Notably, ipTM also performed best for both TM score ≤0.5 and TM score >0.5 proteins, even with filtering for pLDDT (FigureS4).

Oligomeric state prediction across a diverse set of protein folds. (a) Proportion of proteins predicted correctly by AF2‐M from a set of 3560 proteins representing diverse X Groups. Results are shown for all proteins, as well as for each oligomeric state. (b) Proportion of proteins predicted correctly by AF2‐M averaged for each X Groups. X Groups with 5 or more proteins are shown by blue points and those with less proteins in gray. Data shown as max to min box and whiskers plots with midline showing median and box representing interquartile range with the mean indicated by the dashed green line. (c) Confusion matrix of proteins that were assigned to the incorrect oligomeric state using AF2‐M. Rows represent the annotated oligomeric state of each protein, and columns represent the oligomeric state that the protein got incorrectly assigned to. Darker cells indicate higher protein counts.

To optimize the computational cost associated with making these predictions, which can be considerable for larger proteins, we next assessed how the chosen number of recycles and number of predicted structures influences the accuracy of oligomeric state prediction. We carried out predictions using combinations of 0, 1, 2, 3 recycles and 1, 5, 10, 20 structures per condition using the same set of proteins and assessed how these parameters altered the ipTM scores of the structures predicted in the annotated and incorrect oligomeric states. Increasing the number of structures from 1 to 5, 10, or 20 only yielded a small increase in the accuracy of the predictions; however, increasing the number of recycles had a more significant effect, with no recycles yielding the worst performance (Figure3g). A comparison between compute cost and number of proteins predicted in the correct oligomeric state suggests that using 1 structure and 1–3 recycles is a computationally efficient set of parameters (predicting ~78% of proteins in the correct oligomeric state) (Figure3h), although using more structures may be beneficial when resources allow.

To test the general applicability of this approach across diverse protein folds, we applied it to a further set of 3560 proteins. This set covers 801 of the 808 unique X Groups in the DeepSub database (as defined in the Evolutionary Classification of Protein Domains database; Schaeffer et al.,2025) with oligomeric states from monomers to pentamers, containing proteins from 1189 organisms. The missing 7 X Groups contained only large proteins >2550 amino acids long, which exceeded memory constraints. We limited representation of any individual X group to at most 21 proteins (<1% of all proteins in the X Group set of 3560) and produced only one structure per protein for computational efficiency. The overall success rate of our predictions was 70.1%, with the success rate for monomers slightly higher than oligomers (Figure4a). Averaging the results within X Groups shows a median success rate of 86.7% with 358 X Groups having all members assigned to the correct oligomeric state (Figure4band TablesS6andS7). While predictions were good for many X Groups, the mean success rate per X Group was 66%, brought down by X Groups that have few or no members predicted correctly. This included 11X Groups containing 5 or more proteins that had no correct predictions. Some of these, such as the Protozoan pheromone proteins and ATP‐grasp_6 proteins, may be misannotated in the DeepSub/UniProt databases, with literature suggesting that at least some of these proteins adopt different oligomeric states. Protozoan pheromone proteins with the UniProt IDsP26887(Luginbühl et al.,1996),P26886(Ottiger et al.,1994),P10774(Anderson et al.,1996), andP12350(Brown et al.,1993) are annotated as homodimers in both UniProt and DeepSub, while solution structures have only been resolved as monomers and AF2‐M assigns these proteins as monomers (TablesS6andS7). ATP‐grasp_6 proteins with the UniProt IDsQ9CM00andQ8DXM9are annotated as monomers in the UniProt and DeebSub databases, while structures exist as homodimers (Vergauwen et al.,2006) and AF2‐M assigns these proteins as homodimers (TablesS6andS7).

Results from this X Group set of 3560 proteins showed no relationship between proportion correct and sequence length (FigureS6b) and a similar trend in the proportion correct vs. monomer pLDDT (FigureS6c), Foldseek TM‐score (FigureS6d) and MMseqs2 similarity (FigureS6e) to that seen for the TM set of 1006 proteins. Overall, the most common error in the predictions was assigning dimeric proteins as monomers (Figure4c). To check if this could be improved, we recalculated the ipTM cutoff for monomer/oligomer discrimination but found only a minor increase in accuracy (by 1%) occurred using a cutoff of 0.25 compared to the previously determined 0.32 (FigureS6f). While it is possible to optimize the cutoff within each X Group to increase the overall mean and median success rates to 78.7% and 100%, respectively (FigureS6g,hand TableS7), this risks overfitting small datasets and would be challenging in practice for X Groups where oligomeric states are completely unknown. For example, an X Group annotated as monomers can always be predicted correctly if the cutoff is increased enough; while an X Group can be forced not to be predicted as a monomer by setting the cutoff to 0.

While availability of the source code for AF2‐M allowed us to generate structural predictions for many proteins without restriction, we wondered whether the online AlphaFold3 server, which provides an accessible user interface to predict structural complexes, could also be used to predict oligomeric states. To assess this, we carried out structural predictions of the initial 40 proteins as dimers, trimers, tetramers, and pentamers as above using the default AF3 parameters of 10 recycles, 5 structures.

As with AF2‐M, the ipTM scores of predicted oligomer structures can distinguish between proteins predicted in the annotated state versus an incorrect oligomeric state (Figure5a), as well as between monomeric and multimeric proteins (Figure5b). By optimizing for the monomeric cutoff value to 0.24 as we did when using AF2‐M, 33/40 of the proteins in our test set were correctly assigned (Figures4candS7andS8).

ipTM scores from the AlphaFold3 online server also accurately predict oligomeric states. (a) Probability distribution plot of ipTM scores for 32 known multimeric proteins predicted in the correct (pink) and incorrect (green) oligomeric states. (b) Probability distribution plot of ipTM scores from AF3 for eight known monomeric proteins predicted with 2, 3, 4, 5 subunits (blue) and 32 known multimeric proteins predicted with 2, 3, 4, 5 subunits (pink). The gray line indicates the cut‐off ipTM value of 0.24. (c) Number of correctly assigned oligomeric states (out of 40) vs. cut‐off value selected for four different methods to distinguish between monomeric and multi‐subunit proteins. The gray line indicates the optimal cut‐off ipTM value of 0.24. (d) Two‐way table showing the distribution of proteins with oligomeric states predicted correctly and incorrectly based on whether an experimentally resolved structure exists within the training set. (e) TM score between the top ranked AlphaFold3 predicted structure and experimentally resolved structure (if available) plotted against the TM score between the top ranked AlphaFold3 predicted structure and the monomeric structure from the AFDB, colored by annotated oligomerisation state. (f) Comparison of AlphaFold2 and AlphaFold3 prediction accuracy by annotated oligomerisation state. (g) Snapshots of the top ranked structural predictions generated by AF2‐M (top row) and AF3 (bottom row) in the correct oligomeric state for proteins that were assigned to an incorrect oligomeric state by AF2‐M, AF3, or both AF2‐M and AF3 (ribbon, colored by chain). For monomeric proteins, the AFDB monomeric structure is shown in dark green. The presence of an experimentally resolved structure in the PDB or AlphaFold2 training set is indicated with an icon near each UniProt ID. The correct oligomerisation state is indicated by each structure in green and the incorrect prediction by the associated method in red.

As AF3 uses PDB templates by default where available, we find most proteins (29/31) where a structure was present in the training set is predicted in the correct oligomeric state (Figure5d). In contrast, 5/9 structures which are not present in the training set were predicted in the incorrect oligomeric state.

Similar to AF2‐M, the majority of predicted oligomeric structures had high TM scores to the experimentally resolved structures where available, but a lower correlation of TM scores to the predicted monomeric structures (Figure5e). Similar TM scores were obtained for the same proteins across AF2‐M and AF3, and proteins which were predicted with lower TM scores using AF2‐M were also less like the experimental structures when predicted using AF3.

Compared to AlphaFold2, AlphaFold3 assigned more dimeric and tetrameric proteins correctly (Figure5f), although the significance of this is hard to assess given the size of the data set. Structures predicted in an incorrect oligomeric state by AF2‐M but had a PDB available were predicted in the correct oligomeric state by AF3 (Figure5g). This is because AF2‐M predictions were run without templates, while the results from the AF3 server use templating by default. These settings reflect their default configurations and likely usage by non‐specialist users, complicating any direct comparison of accuracy of the two approaches. Interestingly, 5/7 of the incorrect state assignments using AF3 were due to difficulty distinguishing between monomeric and homooligomeric proteins. Three monomeric proteins were predicted to form higher order oligomers (2 dimers, 1 trimer), and one dimeric and one trimeric protein was incorrectly assigned as monomeric.

Next, we sought to apply this technique of oligomerisation state and structure prediction to membrane proteins. As many pore forming membrane proteins are multimeric, their oligomeric state is crucial for understanding their physiology. We curated a set of 14 membrane proteins with a diversity of known oligomeric states (monomeric to heptameric) of interest in our lab and used these to assess AF2‐M's ability to correctly predict their oligomeric states (Figure6aand TableS3).

Application to membrane proteins. (a) Box plots showing individual ipTM scores for different folded oligomeric states (2–7) across a curated set of 14 membrane proteins with the correct oligomeric state highlighted in yellow. The dashed line indicates the previously determined AF2‐M cut‐off ipTM value of 0.32. UniProt IDs and common protein names are shown. (b) Structural predictions of select membrane proteins in various oligomeric states. Protein names and oligomeric states shown are given below each set of structures. (c) Box plots showing the difference in ipTM (ΔipTM) between the top two folded oligomeric states with the highest mean ipTM for each protein, grouped by family. Each dot in the box plot represents a single protein from the OSCA, K2P, XPR, ASIC, AQP, Kv, TRPV, PANX, pLGIC, FNT, or UreI families. Above the AQP boxplot, points are annotated for HsAQP1 (blue) and HsAQP11 (pink). (d) Boxplots of ipTM scores for each folded oligomerisation state for HSAQP11 and HSAQP1. The correct oligomeric states are shown in pink and blue, respectively. The dashed line represents the cutoff of 0.32. (e) Structural predictions of HsAQP11 and HsAQP1 as trimers and tetramers.

In this curated set, all proteins which form higher order oligomers (dimers and above) were predicted in the correct oligomeric states. However, two proteins which we expected to be assigned as monomeric (HsTMEM63a, UniProt ID:O94886and HsSERT, UniProt ID:P31645) were assigned as trimers and dimers, respectively, using our monomeric cut‐off value of 0.32. Interestingly, we note that for HsSERT, the top ranked structures of the dimeric form of the protein (as shown in Figure6b) have the highest ipTM score (Figure6a). Dimerisation and formation of higher order oligomers for HsSERT and other SLC6 members have been previously reported, hinting that the oligomerisation state predicted here could be biologically relevant (Jayaraman et al.,2021). Additionally, the top ranked predicted dimeric structure of HsSERT resembles one dimer conformation sampled in previously conducted molecular dynamics simulations probing HsSERT self‐assembly (Periole et al.,2018).

The transient receptor potential vanilloid‐3 channel TRPV3 (UniProt ID:Q8NET8) was correctly predicted to be tetrameric. In addition, we noticed the generally high ipTM scores and low ipTM range of structures predicted in the trimeric and pentameric states, which both had ipTM scores only slightly lower than the tetramers. The predicted trimeric structures resembled the tetramers with one missing subunit (Figure5b), highlighting a case where visual examination is necessary to understand whether a predicted oligomeric state is plausible. Interestingly, however, the pentameric structures resembled the cryo‐EM structure of TRPV3 recently elucidated (after the AF2.3 training date cutoff) as a pentamer, thought to be a transient state of the protein with a dilated pore (Lansky et al.,2023).

Similarly, we observed that the mechanosensitive channel of large conductance MscL, whose structure has been determined in both the tetrameric and pentameric states (Chang et al.,1998; Liu et al.,2009), had similar ipTM score distributions for these two states. However, we note in this case that both structures were determined prior to the cut‐off date for training AF2. Finally, structural predictions of pannexin 2 suggest that the hexameric and heptameric oligomerisation states are both probable, as they have similar high ipTM scores. Our results are consistent with a recent report that the closely related pannexin 1 channels are active in both the hexameric and heptameric states (Gupta et al.,2025). This result was surprising given that all structures of pannexins thus far have been resolved as heptamers (Deng et al.,2020; Jin et al.,2020; Mou et al.,2020; Qu et al.,2020; Ruan et al.,2020).

Interestingly, the ipTM score distributions and structural predictions for HpUreI (a proton‐gated urea channel inH. pylori) (Strugatsky et al.,2012), PfFNT (a lactate/proton transporter inP. falciparum) (Lyu et al.,2021; Peng et al.,2021) and HsGlyR (the human glycine receptor) (Zhu & Gouaux,2021) suggest that these proteins may also oligomerise with a different number of subunits in addition to their currently known annotated states.

As HsTRPV3, HsPANX2, PfFNT, HsGlyR and HPUreI all had at least two oligomeric states with similar ipTM values, we extended our investigations with AF2‐M to additional proteins from the same families (TRPV, PANX, FNT, pLGIC, and UreI, respectively) to see if a similar trend would be observed in the range of ipTM values (Figure6cand TableS4). For comparison, we also folded various proteins from the OSCA, XPR, ASIC, AQP, K2P, and Kv families. For each folded protein, the difference in ipTM (ΔipTM) between the top two folded oligomeric states with the highest mean ipTM was calculated. Interestingly, proteins from the TRPV, PANX, FNT, pLGIC, UreI, and Kv families generally had lower ΔipTM values than those from the OSCA, XPR, ASIC, AQP, and K2P families. While HsTRPV3 is known to adopt different oligomeric states and HsPANX2 has been hypothesized to adopt different oligomeric states, and both proteins have low ΔipTM values, it is plausible that comparing ΔipTM between folded oligomeric states may suggest if a protein is likely to adopt different oligomeric states. However, we note the presence of apparent false positives, including the case of the Kv family, which shows low ΔipTM values between tetramers and pentamers. Although one study has reported a pentameric structure of the soluble T1 domain of Kv2.1 (Xu et al.,2022), full‐length Kv channels have only been resolved as tetramers.

Recently, a structure of a homotrimeric HsAQP11 was resolved—distinguishing it from other homotetrameric members of the AQP family (Suzuki et al.,2026). Using our AF2‐M oligomeric state prediction pipeline, HsAQP11 was correctly predicted as a homotrimer, with a larger ΔipTM between folded oligomeric states than the other members of the AQP family that were included in our predictions (Figure6c–e). These findings suggest that this approach may also be useful for identifying proteins that adopt divergent oligomeric states within otherwise structurally conserved protein families.

DISCUSSION

Many proteins undergo oligomerisation to form their biologically relevant states, and predicting both stoichiometry and structure is critical for understanding protein function and regulation. Here, we show that AlphaFold2‐Multimer and AlphaFold3 can be used to predict oligomeric states for a subset of proteins with high confidence, while also revealing clear limitations when predictions are extended to larger and more structurally diverse datasets. Using over 4700 proteins, we identify the conditions under which oligomeric state prediction is reliable and those under which it remains challenging. As the benchmark dataset contained only human proteins, whereas the large‐scale expanded datasets included proteins from many species, indicating that AlphaFold‐based oligomeric state prediction may be largely species agnostic.

The small number of proteins in our test set and differences in parameters used mean that it is not possible to say here whether AF2‐M or AF3 is more likely to predict the correct oligomeric state. However, it appears that within this set of proteins, proteins with generally high pLDDT scores and without abnormally long helices in the AFDB structure have their oligomeric states correctly predicted by AF2‐M. It is less clear what determines whether a protein will be predicted in the correct oligomeric state by AF3, although it seems to perform less well for proteins without an available structure in the PDB and has some trouble distinguishing between monomeric and non‐monomeric proteins. We note that in 30/31 cases, if both AF2‐M and AF3 predict the same oligomeric state for a given protein, then this is the annotated state, suggesting that use of AF2‐M and AF3 concurrently could increase confidence in the oligomeric state prediction.

While average monomeric pLDDT is a strong indicator of oligomeric state prediction accuracy in the small, curated dataset, analysis of the larger dataset of 1006 proteins demonstrates that high pLDDT alone is not sufficient to guarantee correct oligomeric state prediction. This limitation is particularly apparent for proteins that lack similarity to any structures present in the AlphaFold training set, highlighting a challenge for extending oligomeric state prediction to proteins that are more structurally diverse and evolutionarily distant.

To determine whether this limitation arises from the confidence metrics used rather than the underlying structural predictions, we examined alternative interface‐based scoring approaches. The ipTM metric used to identify the most likely oligomeric state is reported by default for each structure in both AF2‐M and AF3 and performs better than pLDDT values which are not expected to be strongly influenced by the interchain contacts. However, it has been noted that the ipTM scores can be impacted by the presence of disordered protein regions (Bret et al.,2024; Dunbrack,2025; Martin,2024). The effect of these regions can be reduced by removing them if they are known not to be involved in the interaction, restricting calculations to only residues near the interaction site (Varga et al.,2024), or to interchain residue pairs that have well predicted aligned error distances (Dunbrack,2025). While these approaches may improve ipTM‐based assessments in specific cases, our large‐scale benchmarking indicates that more advanced interface confidence metrics, including ipSAE (Dunbrack,2025), pDockQ (Bryant et al.,2022), pDockQ2 (Zhu et al.,2023), and LIS (Kim et al.,2024), do not substantially improve oligomeric state prediction for proteins for which AlphaFold has no close structural reference. This suggests that the primary limitation lies not in the choice of interface metric, but in the absence of sufficient evolutionary or structural context for AlphaFold to predict accurate inter‐chain interactions.

Despite these limitations, ipTM distributions can still provide valuable information about protein oligomeric states. For several membrane proteins which can be physiologically found in different oligomeric states (TRPV3, MscL, and Pannexins), this information can be captured in the ipTM distributions. This suggests the value of this approach for identifying proteins where multiple physiologically relevant oligomerization states may be plausible and warrant further investigation.

Benchmarking across the TM set of 1006 proteins suggests that producing one structure with one to three recycles is sufficient to obtain an accuracy of 77–79% while minimizing computational costs.

We further extended this analysis to proteins from the OSCA, K2P, XPR, ASIC, AQP, Kv, TRPV, PANX, pLGIC, FNT, and UreI families. Proteins belonging to families containing at least one member previously reported to adopt multiple oligomeric states exhibited small differences in ipTM values between two or more folded oligomeric states. This pattern suggests that low ΔipTM between folded oligomeric states may be able to suggest proteins capable of adopting multiple biologically relevant oligomeric forms. However, we note that this method is not completely accurate, as members of the Kv family also showed low ΔipTM values despite full‐length channels not having been reported to adopt different oligomeric states.

These findings also highlight a limitation of relying on existing database annotations for knowledge of protein oligomeric state, which may sometimes be inaccurate or outdated. For two membrane proteins we examined (HsXPR1 (Lu et al.,2024; Wang et al.,2025; Zhang et al.,2025) and NavAb (Payandeh et al.,2011)), oligomeric states determined via structural biology were not annotated in the UniProt database. HsAQP11 is still annotated as homodimer and homotetramer in UniProt despite there now being a homotrimeric structure (Suzuki et al.,2026). Additionally, for PANX1, which has recently been shown to be functional as both hexamers and heptamers (Gupta et al.,2025), only the heptameric state is annotated, suggesting that interaction annotations in UniProt may sometimes provide an incomplete picture of physiologically relevant oligomerisation. This is further reflected for proteins capable of polymerization, for which UniProt typically annotates the minimal repeating interface rather than the biologically relevant assembly.

While we benchmarked the approach across a wide range of proteins with differing similarity to structures in the training data and different folds, it is challenging to completely represent protein space. For example, the 808 X Groups that can form monomers to pentamers contained within the annotated Deepsub database is much smaller than the 19,014 distinct X Groups in the ECOD database, although how many of those form monomers to pentamers is unclear. While our data contain proteins from 1189 organisms (FigureS6a), we have not explicitly covered organism or functional diversity.

Despite its benefits, it is also key to consider the computational feasibility of this approach. The high computational cost of modeling four or more copies of larger proteins may prohibit this technique from being easily usable for proteins where one subunit contains more than 1000 amino acids. Additionally, due to computational constraints, we did not consider hexameric and higher‐order assemblies in either the curated benchmark set of 40 proteins or the expanded dataset of 1006 proteins, with the exception of a small number of hexameric and heptameric membrane proteins shown in Figure5.

Future studies will therefore be needed to assess the performance of this technique for higher‐order oligomeric complexes to determine whether this approach generalizes beyond pentamers, as it may be harder to distinguish higher order oligomers with more similar angles between subunits. Notably, in the case of CLCC1, homooligomerization is supported experimentally, yet assemblies of different sizes (e.g., 14‐mer and 16‐mer) yield comparable ipTM scores (Dai et al.,2025), underscoring this ambiguity. Further extensions of this work may also include investigating whether this technique could provide insight into stoichiometric ratios for heteromeric complexes in instances where the identity of interacting proteins is known but the number of each subunit is not.

Beyond oligomeric state prediction, an additional advantage of using AlphaFold is that it provides structural models of proteins in their predicted assemblies, which may enable further functional insight. To increase confidence in these models, plausible oligomeric structures could be combined with short, physics‐based molecular dynamics simulations to assess their stability in solution or membrane environments, alongside experimental validation.

Although comparisons with other sequence‐ and template‐based methods are complicated by differences in training sets, benchmark datasets, and reported performance metrics, AF‐M (80.7% accuracy for proteins with Foldseek TM >0.3 and 66.4% for TM <0.3, median accuracy across X Groups of 86.7%) is competitive with other approaches. Deep learning approaches such as DeepSub, QUEEN and POST achieve F1‐scores of 0.558, 0.63 and 0.73, respectively, across a set of 1146 proteins (although DeepSub achieves 0.917 on its own test set). Seq2Symm reports an area‐under‐the‐curve precision–recall of 0.47. The AF2‐based Combfold reports 57–64% of tested proteins yield accurate oligomeric structures. We note that AlphaFold‐based methods provide structural models, albeit at higher computational cost, whereas deep learning approaches are faster but predict only oligomeric state.

Ultimately, AlphaFold2‐Multimer and AlphaFold3 represent powerful tools for predicting oligomeric states and generating structural models, particularly for proteins with fold‐level similarity to known structures. However, our results demonstrate that accurate oligomeric state prediction remains challenging for structurally novel proteins, even when predicted structures are of high confidence and evaluated using advanced interface metrics. This indicates that the performance of these predictions is contingent on the breadth and quality of the training data. Together, these findings define both the promise and the current limitations of AlphaFold‐based oligomeric state prediction and highlight the continued importance of experimental validation for proteins for which there are currently no structurally similar representatives in the AlphaFold training set.

METHODS

Selection of oligomeric protein sequences

The DeepSub database (https://github.com/tibbdc/DeepSub/blob/main/DATA/Dataset_0724_new.csv) containing proteins from the UniProt database annotated with experimentally determined oligomeric states was used to randomly select 40 human proteins (5 each of known monomeric, homodimeric, homotrimeric, homotetrameric, and homopentameric proteins), detailed in TableS1(Deng et al.,2024).

Structural prediction using AlphaFold2 multimer

Selected sequences were extracted from the UniProt Database and used to generate multiple sequence alignments for structure prediction by ColabFold implementing AlphaFold‐Multimer v.3 (Jumper et al.,2021; Mirdita et al.,2022). 20 structures containing 2, 3, 4, 5 copies of each protein were predicted with three recycles without subsequent relaxation nor modification to the multiple sequence alignment.

To extend our assessment of the accuracy of using AF2‐M to predict the oligomeric states of proteins, we randomly selected 1006 proteins from the DeepSub database for which the oligomeric state was known, but no structures were available. To ensure that none of the test proteins had homologues represented in the AlphaFold2 (AF2) training set, proteins from the DeepSub database were first screened to confirm the absence of experimentally determined structures.

After excluding proteins with known experimental structures, monomeric AFDB models were downloaded for the remaining proteins, and structural similarity searches were performed against all PDB entries using Foldseek (van Kempen et al.,2024) with easy‐search and alignment‐mode 1. These proteins were divided almost equally into two categories TM score ≥0.5 and TM score <0.5 based on the Foldseek determined TM score between the AFDB monomeric structure and experimentally determined structures in the AlphaFold training set. The pLDDT values were analyzed both from monomeric AFDB structures and, for oligomeric proteins, from predicted oligomeric complexes. Both groups contained at least 40 monomers and approximately 100 dimers, trimers, tetramers, and pentamers. Each protein was then modeled as a dimer, trimer, tetramer, and pentamer using AF2 with three recycles, generating 20 structural models for each prediction. Sequence similarity scores to proteins in the PDB were also obtained using MMseqs2 (Steinegger & Söding,2017) using easy‐search. Information for all 1006 proteins from the TM set is detailed in TableS5.

To evaluate oligomeric‐state predictions, we compared raw ipTM scores from AF2 with a set of advanced interface‐quality metrics. This included the ipSAE, ipSAE‐d0chn, ipSAE‐d0dom, ipTM‐d0chn (Dunbrack,2025), pDockQ (Bryant et al.,2022), pDockQ2 (Zhu et al.,2023), and LIS scores (Kim et al.,2024). These metrics were computed using code obtained from the Dunbrack lab's repository (https://github.com/dunbracklab/IPSAE). Cut‐offs for oligomeric state assignment were optimized for each metric using the method we described previously for the ipTM cutoff.

To assess whether confidence in predictions across this set could be derived from the distributions of the ipTM scores, we conducted bootstrapping analysis in which we assigned the oligomeric state based on one randomly selected structure from each oligomeric state across 1000 trials, and calculated the proportion of times this state agreed with the (correct or incorrect) state assigned using all 20 produced structures. We also calculated the difference in ipTM between the oligomeric states with the highest and second highest ipTM scores.

To assess the performance of AF2‐M for oligomeric state prediction across different protein architectures, an additional set of 3560 proteins from 801 distinct X Groups was chosen. X Group annotations were obtained for all proteins in the DeepSub Database by searching the Evolutionary Classification of Protein Domains (ECOD) Databse (Schaeffer et al.,2025). Proteins were randomly sampled from the DeepSub database such that no more than 21 proteins were present from the same X Group (<1% of the overall number of proteins in the X Group dataset). In cases where proteins were assigned to more than one X Group, the protein was annotated as belonging to the X Group that covered the most residues of that protein. Each protein was folded as a dimer, trimer, tetramer, and pentamer using AF2. Structural predictions were made using three recycles and one structure. Information for all 3560 proteins from the X Group set is detailed in TableS6.

AF2‐M was used to fold an additional 98 proteins from the OSCA, K2P, XPR, ASIC, AQP, Kv, TRPV, PANX, pLGIC, FNT, and UreI families as dimers, trimers, tetramers, pentamers, hexamers, and heptamers. All predictions were run using three recycles and 20 structures. For each protein, the top two folded oligomeric states with the highest mean ipTM were determined. A mean ipTM value was calculated for the second highest state and was subtracted from each ipTM value from the highest state, and these values were then averaged to calculate the mean ΔipTM per protein. For each protein in the TM set of 1006 proteins, the X Group set of 3560 proteins and the set of 98 membrane proteins, structural and sequence similarity to proteins in the PDB were calculated using Foldseek (van Kempen et al.,2024) and MMseqs2 (Steinegger & Söding,2017), respectively. Information for these proteins is given in TableS4.

For all structural predictions, excluding the X Group set of 3560 proteins, multiple sequence alignments (MSAs) were generated locally using colabfold_search with db‐load‐mode 1, using sequence information from the UniRef30, BFD, and ColabFoldDB databases. To reduce computational cost for running predictions for the X Group set, precalculated MSAs were sourced from the AFDB or were generated locally when the AFDB MSA was not available.

Structural prediction using AlphaFold3

The AlphaFold3 online server was used to predict multimeric complexes (2, 3, 4, 5 copies) of each protein, using sequences extracted from the UniProt Database (Abramson et al.,2024). By default, 10 recycles were performed and 5 structures generated. As above, for two large proteins (UniProt IDs:P35670,P53396), we were unable to predict their structures as tetramers/pentamers due to computational restrictions.

Analysis of structures

Structures and output files produced by AlphaFold2‐Multimer and AlphaFold3 were analyzed using python scripts written using the pandas, scipy, matplotlib, seaborn, MDAnalysis, and MDTraj libraries (Hunter,2007; McGibbon et al.,2015; Mckinney,2010; Michaud‐Agrawal et al.,2011; Virtanen et al.,2020; Waskom,2021). Structures were visualized and figures produced using Visual Molecular Dynamics (VMD) (Humphrey et al.,1996).

For each protein, the predicted monomeric structure from the AFDB and the top ranked structure produced for the correct oligomeric state were oriented along the Z principal axis using the GMX editconf princ function within GROMACS (v.2023) (Bauer et al.,2022). For oligomeric structures predicted using AlphaFold2‐Multimer, the top ranked structure was chosen from those generated from 3 recycles and 20 structures. The α‐carbon atoms of the monomeric structure and the first chain of the predicted oligomeric structure were aligned within VMD to generate the structural images.

The average monomeric pLDDT for each protein was calculated using each associated structure in the AFDB by averaging the per‐residue pLDDT for all residues in each structure.

The secondary structure was calculated using the DSSP algorithm implemented in MDTraj (Kabsch & Sander,1983; McGibbon et al.,2015). For each structure, the length of the longest helix was assigned to be the longest continuous stretch of residues calculated to have a helical ('H') secondary structure.

The template modeling (TM) score for the structures of each protein predicted in the correct oligomeric state were calculated following sequence and structural alignment to the predicted monomeric structure in the AFDB and, where available, to valid experimentally resolved structures using US‐align (Zhang et al.,2022). All experimentally resolved structures associated with the UniProt ID of each structure were assessed for validity. Monomeric structures were removed from comparison to our oligomeric predictions. Structures with <95% sequence similarity between the UniProt sequence and PDB sequence were discarded, as were structures containing only fragments (<50% of the UniProt sequence) of the protein. Using options ‐mm 1 and ‐ter 0, US‐Align was used to calculate the TM score after alignment of the two multi‐chain oligomeric structures. As some proteins of interest were resolved with accessory proteins and additional chains were present in some PDBs, structures were manually inspected to identify the correct chains for comparison where necessary. For the five proteins with a large number (>10) of eligible PDBs (TableS3), if a PDB with the correct number of chains yielded a TM score above 0.95, we did not calculate further TM scores for PDBs resolved with additional chains. TM scores to the monomeric structure were also calculated using US‐Align, normalizing by the length of the monomeric protein.

AUTHOR CONTRIBUTIONS

Ben Corry:Conceptualization; funding acquisition; writing – review and editing; writing – original draft; supervision.Yiechang Lin:Conceptualization; investigation; funding acquisition; writing – original draft; methodology; visualization; writing – review and editing; formal analysis; data curation.Ciara Wallis:Conceptualization; investigation; methodology; visualization; writing – original draft; writing – review and editing; formal analysis; data curation.

ACKNOWLEDGMENTS

This research was supported by services and resources provided by the National Computational Infrastructure (NCI), which was funded by the Australian Government, under project g15 accessed through the national Merit Allocation Scheme (NCMAS) and Australian National University Merit Allocation Scheme (ANUMAS). We acknowledge funding from the Australian Research Council (DP200100860), and National Health and Medical Research Council (APP2020565). Ciara Wallis acknowledges support from an Australian Government Research Training Program (RTP) Ph.D. Scholarship. Open access publishing facilitated by Australian National University, as part of the Wiley ‐ Australian National University agreement via the Council of Australasian University Librarians

Lin Y, Wallis C, Corry B. When can AlphaFold predict the oligomeric states of proteins? Protein Science. 2026;35(7):e70686. 10.1002/pro.70686

Contributor Information

Yiechang Lin, Email: yiechang.lin@anu.edu.au.

Ben Corry, Email: ben.corry@anu.edu.au.

DATA AVAILABILITY STATEMENT

The data that supports the findings of this study are available in the supplementary material of this article.

Associated Data

Data Availability Statement

The data that supports the findings of this study are available in the supplementary material of this article.

References

1. Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630:493–500. doi.org/10.1038/s41586-024-07487-w
2. Anderson DH, Weiss MS, Eisenberg D. A challenging case for protein crystal structure determination: the mating pheromone Er‐1 from Euplotes raikovi.urn:issn:0907‐444952, pp. 469–480. 1996. doi.org/10.1107/S0907444995014235
3. Avraham O, Tsaban T, Ben‐Aharon Z, Tsaban L, Schueler‐Furman O. Protein language models can capture protein quaternary state. BMC Bioinformatics. 2023;24:433. doi.org/10.1186/s12859-023-05549-w
4. Baek M, Park T, Heo L, Park C, Seok C. GalaxyHomomer: a web server for protein homo‐oligomer structure prediction from a monomer sequence or structure. Nucleic Acids Res. 2017;45:W320–W324. doi.org/10.1093/nar/gkx246
5. Bauer P, Hess B, Lindahl E. GROMACS 2022 Source code. 2022, 10.5281/ZENODO.6103835 doi.org/10.5281/ZENODO.6103835
6. Bret H, Gao J, Zea DJ, Andreani J, Guerois R. From interaction networks to interfaces, scanning intrinsically disordered regions using AlphaFold2. Nat Commun. 2024;15:597. doi.org/10.1038/s41467-023-44288-7
7. Brown LR, Mronga S, Bradshaw RA, Ortenzi C, Luporini P, Wüthrich K. Nuclear magnetic resonance solution structure of the pheromone Er‐10 from the ciliated protozoan Euplotes raikovi. J Mol Biol. 1993;231:800–816. doi.org/10.1006/jmbi.1993.1327
8. Bryant P, Pozzati G, Elofsson A. Improved prediction of protein‐protein interactions using AlphaFold2. Nat Commun. 2022;13:1265. doi.org/10.1038/s41467-022-28865-w
9. Chang G, Spencer RH, Lee AT, Barclay MT, Rees DC. Structure of the MscL homolog from mycobacterium tuberculosis: a gated mechanosensitive ion channel. Science. 1998;282:2220–2226. doi.org/10.1126/science.282.5397.2220
10. Chetri PB, Khan H, Tripathi T. Methods to determine the oligomeric structure of proteins. In: Tripathi T, Dubey VK, editors. Advances in Protein Molecular and Structural Biology Methods. Academic Press; 2022. p. 49–76. 10.1016/B978-0-323-90264-9.00005-2 doi.org/10.1016/B978-0-323-90264-9.00005-2
11. Dai B, Sperl A, Polack L, et al. ER protein CLCC1 promotes nuclear envelope fusion in herpesviral and host processes. Nat Commun. 2025;16:10256. doi.org/10.1038/s41467-025-65115-1
12. Deng R, Wu K, Lin J, et al. DeepSub: utilizing deep learning for predicting the number of subunits in homo‐oligomeric protein complexes. Int J Mol Sci. 2024;25:4803. doi.org/10.3390/ijms25094803
13. Deng Z, He Z, Maksaev G, Bitter RM, Rau M, Fitzpatrick JAJ, et al. Cryo‐EM structures of the ATP release channel pannexin 1. Nat Struct Mol Biol. 2020;27:373–381. doi.org/10.1038/s41594-020-0401-0
14. Dunbrack RL. Res ipSAE loquunt: What's wrong with AlphaFold's ipTM score and how to fix it. bioRxiv 2025.02.10.637595. 2025. 10.1101/2025.02.10.637595 doi.org/10.1101/2025.02.10.637595
15. Gupta S, Chiu YH, Manjegowda MC, Desai BN, Ravichandran KS, Bayliss DA. Distinct properties and activation of hexameric and heptameric pannexin 1 channel concatemers. J Gen Physiol. 2025;157:1–14. doi.org/10.1085/jgp.202413676
16. Homma F, Huang J, van der Hoorn RAL. AlphaFold‐multimer predicts cross‐kingdom interactions at the plant‐pathogen interface. Nat Commun. 2023;14:6040. doi.org/10.1038/s41467-023-41721-9
17. Humphrey W, Dalke A, Schulten K. VMD: visual molecular dynamics. J Mol Graph. 1996;14:33–38. doi.org/10.1016/0263-7855(96)00018-5
18. Humphreys I, Pei J, Baek M, et al. Computed structures of core eukaryotic protein complexes. Science. 2021;374:eabm4805. doi.org/10.1126/science.abm4805
19. Hunter J. Matplotlib: a 2D graphics environment. Comput Sci Eng. 2007;9:90–95.
20. Jayaraman K, das AK, Luethi D, Szöllősi D, Schütz GJ, Reith MEA, et al. SLC6 transporter oligomerization. J Neurochem. 2021;157:919–929. doi.org/10.1111/jnc.15145
21. Jin Q, Zhang B, Zheng X, Li N, Xu L, Xie Y, et al. Cryo‐EM structures of human pannexin 1 channel. Cell Res. 2020;30:449–451. doi.org/10.1038/s41422-020-0310-0
22. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589. doi.org/10.1038/s41586-021-03819-2
23. Kabsch W, Sander C. Dictionary of protein secondary structure: pattern recognition of hydrogen‐bonded and geometrical features. Biopolymers. 1983;22:2577–2637. doi.org/10.1002/bip.360221211
24. Kim A‐R, Hu Y, Comjean A, Rodiger J, Mohr SE, Perrimon N. Enhanced protein‐protein interaction discovery via AlphaFold‐multimer. bioRxiv. 2024.02.19.580970. 10.1101/2024.02.19.580970 doi.org/10.1101/2024.02.19.580970
25. Lansky S, Betancourt JM, Zhang J, Jiang Y, Kim ED, Paknejad N, et al. A pentameric TRPV3 channel with a dilated pore. Nature. 2023;621:206–214. doi.org/10.1038/s41586-023-06470-1
26. Levy ED, Teichmann SA. Structural, evolutionary, and assembly principles of protein oligomerization. In: Giraldo J, Ciruela F, editors. Progress in Molecular Biology and Translational Science. Vol. 117. Academic Press; 2013. p. 25–51. doi.org/10.1016/B978-0-12-386931-9.00002-7
27. Liu Z, Gandhi CS, Rees DC. Structure of a tetrameric MscL in an expanded intermediate state. Nature. 2009;461:120–124. doi.org/10.1038/nature08277
28. Lu Y, Yue C, Zhang L, et al. Structural basis for inositol pyrophosphate gating of the phosphate channel XPR1. Science. 2024;386:eadp3252. doi.org/10.1126/science.adp3252
29. Luginbühl P, Wu J, Zerbe O, Ortenzi C, Luporini P, Wüthrich K. The NMR solution structure of the pheromone Er‐11 from the ciliated protozoan Euplotes raikovi. Protein Sci. 1996;5:1512–1522. doi.org/10.1002/pro.5560050807
30. Luo Y, Wu H, Wei H, Peng Z, Yang J. Predicting the oligomeric state of proteins using multiple templates detected by complementary alignment methods. Proteins. 2025;93:2138–2149. doi.org/10.1002/prot.70017
31. Lyu M, Su C, Kazura JW, Yu EW. Structural basis of transport and inhibition of the plasmodium falciparum transporter PfFNT. EMBO Rep. 2021;22:1–12. doi.org/10.15252/embr.202153596
32. Madaj R, Martinez‐Goikoetxea M, Kaminski K, Ludwiczak J, Dunin‐Horkawicz S. Applicability of AlphaFold2 in the modeling of dimeric, trimeric, and tetrameric coiled‐coil domains. Protein Sci. 2025;34:e5244. doi.org/10.1002/pro.5244
33. Martin J. AlphaFold2 predicts whether proteins interact amidst confounding structural compatibility. J Chem Inf Model. 2024;64:1473–1480. doi.org/10.1021/acs.jcim.3c01805
34. McGibbon RT, Beauchamp KA, Harrigan MP, Klein C, Swails JM, Hernández CX, et al. MDTraj: a modern open library for the analysis of molecular dynamics trajectories. Biophys J. 2015;109:1528–1532. doi.org/10.1016/j.bpj.2015.08.015
35. Mckinney W. Data structures for statistical computing in python. Proc. of the 9th python in science conf. 2010.
36. Michaud‐Agrawal N, Denning EJ, Woolf TB, Beckstein O. MDAnalysis: a toolkit for the analysis of molecular dynamics simulations. J Comput Chem. 2011;32:2319–2327. doi.org/10.1002/jcc.21787
37. Mirdita M, Schütze K, Moriwaki Y, Heo L, Ovchinnikov S, Steinegger M. ColabFold: making protein folding accessible to all. Nat Methods. 2022;19:679–682. doi.org/10.1038/s41592-022-01488-1
38. Mou L, Ke M, Song M, et al. Structural basis for gating mechanism of pannexin 1 channel. Cell Res. 2020;30:452–454. doi.org/10.1038/s41422-020-0313-x
39. Omidi A, Møller MH, Malhis N, Bui JM, Gsponer J. AlphaFold‐multimer accurately captures interactions and dynamics of intrinsically disordered protein regions. Proc Natl Acad Sci U S A. 2024;121:e2406407121. doi.org/10.1073/pnas.2406407121
40. Ottiger M, Szyperski T, Luginbühl P, Ortenzi C, Luporini P, Bradshaw RA, et al. The NMR solution structure of the pheromone Er‐2 from the ciliated protozoan Euplotes raikovi. Protein Sci. 1994;3:1515–1526. doi.org/10.1002/pro.5560030917
41. Payandeh J, Scheuer T, Zheng N, Catterall WA. The crystal structure of a voltage‐gated sodium channel. Nature. 2011;475:353–359. doi.org/10.1038/nature10238
42. Peng X, Wang N, Zhu A, et al. Structural characterization of the plasmodium falciparum lactate transporter PfFNT alone and in complex with antimalarial compound MMV007839 reveals its inhibition mechanism. PLoS Biol. 2021;19:1–19. doi.org/10.1371/journal.pbio.3001386
43. Periole X, Zeppelin T, Schiøtt B. Dimer Interface of the human serotonin transporter and effect of the membrane composition. Sci Rep. 2018;8:5080. doi.org/10.1038/s41598-018-22912-7
44. Qu R, Dong L, Zhang J, Yu X, Wang L, Zhu S. Cryo‐EM structure of human heptameric pannexin 1 channel. Cell Res. 2020;30:446–448. doi.org/10.1038/s41422-020-0298-5
45. Ruan Z, Orozco IJ, Du J, Lü W. Structures of human pannexin 1 reveal ion pathways and mechanism of gating. Nature. 2020;584:646–651. doi.org/10.1038/s41586-020-2357-y
46. Schaeffer RD, Medvedev KE, Andreeva A, Chuguransky SR, Pinto BL, Zhang J, et al. ECOD: integrating classifications of protein domains from experimental and predicted structures. Nucleic Acids Res. 2025;53:D411–D418. doi.org/10.1093/nar/gkae1029
47. Schweke H, Pacesa M, Levin T, Goverde CA, Kumar P, Duhoo Y, et al. An atlas of protein homo‐oligomerization across domains of life. Cell. 2024;187:999–1010.e15. doi.org/10.1016/j.cell.2024.01.022
48. Shor B, Schneidman‐Duhovny D, Rachel T, Benin S. CombFold: predicting structures of large protein assemblies using a combinatorial assembly algorithm and AlphaFold2. Nat Methods. 2024;21:477–487. doi.org/10.1038/s41592-024-02174-0
49. Steinegger M, Söding J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat Biotechnol. 2017;35:1026–1028. doi.org/10.1038/nbt.3988
50. Strugatsky D, McNulty, R , Munson K, et al. Structure of the proton‐gated urea channel from the gastric pathogen helicobacter pylori. Nature. 2012;493:255. doi.org/10.1038/nature11684
51. Suzuki S, Kamegawa A, Kozai D, Nishikawa K, Irie K, Fujiyoshi Y. Cryo‐EM structure of human AQP11 reveals a trimeric architecture with a large pore. Sci Adv. 2026;12:eaeb5769. doi.org/10.1126/sciadv.aeb5769
52. van Kempen M, Kim SS, Tumescheit C, Mirdita M, Lee J, Gilchrist CLM, et al. Fast and accurate protein structure search with Foldseek. Nat Biotechnol. 2024;42:243–246. doi.org/10.1038/s41587-023-01773-0
53. Varadi M, Bertoni D, Magana P, Paramval U, Pidruchna I, Radhakrishnan M, et al. AlphaFold protein structure database in 2024: providing structure coverage for over 214 million protein sequences. Nucleic Acids Res. 2024;52:D368–D375. doi.org/10.1093/nar/gkad1011
54. Varga JK, Ovchinnikov S, Schueler‐Furman O. actifpTM: a refined confidence metric of AlphaFold2 predictions involving flexible regions. bioRxiv 2025.02.10.637595. 2024;41:btaf107. doi.org/10.1093/bioinformatics/btaf107
55. Vergauwen B, De Vos D, Van Beeumen JJ. Characterization of the bifunctional γ‐glutamate‐cysteine ligase/glutathione synthetase (GshF) of Pasteurella multocida. J Biol Chem. 2006;281:4380–4394. doi.org/10.1074/jbc.M509517200
56. Virtanen P, Gommers R, Oliphant T, et al. SciPy 1.0: fundamental algorithms for scientific computing in python. Nat Methods. 2020;17:261–272. doi.org/10.1038/s41592-019-0686-2
57. Wang X, Bai Z, Wallis C, Wang H, Han Y, Jin R, et al. KIDINS220 and InsP8 safeguard the stepwise regulation of phosphate exporter XPR1. Mol Cell. 2025;85:3209–3224.e8. doi.org/10.1016/j.molcel.2025.08.003
58. Waskom ML. seaborn: statistical data visualization. J Open Source Softw. 2021;6:3021.
59. Xu Z, Khan S, Schnicker NJ, Baker S. Pentameric assembly of the Kv2.1 tetramerization domain. Acta Crystallogr D Struct Biol. 2022;78:792–802. doi.org/10.1107/S205979832200568X
60. Zhang C, Shine M, Pyle AM, Zhang Y. US‐align: universal structure alignments of proteins, nucleic acids, and macromolecular complexes. Nat Methods. 2022;19:1109–1115. doi.org/10.1038/s41592-022-01585-1
61. Zhang W, Chen Y, Guan Z, et al. Structural insights into the mechanism of phosphate recognition and transport by XPR1. Nat Commun. 2025;16:1–10. doi.org/10.1038/s41467-024-55471-9
62. Zhu H, Gouaux E. Architecture and assembly mechanism of native glycine receptors. Nature. 2021;599:513–517. doi.org/10.1038/s41586-021-04022-z
63. Zhu W, Shenoy A, Kundrotas P, Elofsson A. Evaluation of AlphaFold‐multimer prediction on multi‐chain protein complexes. Bioinformatics. 2023;39:btad424. doi.org/10.1093/bioinformatics/btad424

---

Republished from the open web under CC-BY. Authors: Lin Y, Wallis C, Corry B. Read the original.