Article https://doi.org/10.

1038/s41467-024-46715-9

High-throughput prediction of protein

conformational distributions with
subsampled AlphaFold2
Received: 3 August 2023 Gabriel Monteiro da Silva 1, Jennifer Y. Cui 1
, David C. Dalgarno 2
,
George P. Lisi 1,3 & Brenda M. Rubenstein 1,3
Accepted: 28 February 2024

This paper presents an innovative approach for predicting the relative popu-
Check for updates lations of protein conformations using AlphaFold 2, an AI-powered method
1234567890():,;
1234567890():,;

that has revolutionized biology by enabling the accurate prediction of protein

structures. While AlphaFold 2 has shown exceptional accuracy and speed, it is
designed to predict proteins’ ground state conformations and is limited in its
ability to predict conformational landscapes. Here, we demonstrate how
AlphaFold 2 can directly predict the relative populations of different protein
conformations by subsampling multiple sequence alignments. We tested our
method against nuclear magnetic resonance experiments on two proteins with
drastically different amounts of available sequence data, Abl1 kinase and the
granulocyte-macrophage colony-stimulating factor, and predicted changes in
their relative state populations with more than 80% accuracy. Our subsampling
approach worked best when used to qualitatively predict the effects of
mutations or evolution on the conformational landscape and well-populated
states of proteins. It thus offers a fast and cost-effective way to predict the
relative populations of protein conformations at even single-point mutation
resolution, making it a useful tool for pharmacology, analysis of experimental
results, and predicting evolution.

Proteins are essential biomolecules that carry out a wide range of The recent development of machine learning algorithms has
functions in living organisms. Understanding their three-dimensional significantly improved the speed of protein structure prediction9,10.
structures is critical for elucidating their functions and designing drugs One of the most remarkable achievements in this area is the AlphaFold
that target them1. Historically, experimental techniques such as X-ray 2 (AF2) engine developed by DeepMind, which uses a deep neural
crystallography, nuclear magnetic resonance (NMR) spectroscopy, network to predict ground state protein structures from amino acid
and electron microscopy have been used to determine protein sequences11,12. AlphaFold 2 was trained using large amounts of
structures2–4. However, these methods can be time-consuming, tech- experimental data and incorporates co-evolutionary information from
nically challenging, and expensive, and may not work for all proteins5. massive metagenomic databases11. Its accuracy has revolutionized the
To meet this challenge, ab initio structure prediction methods, which field of protein structure prediction11,13,14, opening up new possibilities
use computational algorithms to predict protein structures from their for drug discovery and basic research with clear consequences for
amino acid sequences, have been developed6. For many years, ab initio human health15,16.
structure prediction methods have relied on physics-based algorithms However, a series of studies have found that the default AF2
to predict stable protein structures7. Although successful, these algorithm is limited in its capacity to predict alternative protein con-
methods are challenged by larger and more complex proteins8. formations and the effects of sequence variants17,18. Although AF2’s

1
Brown University Department of Molecular and Cell Biology and Biochemistry, Providence, RI, USA. 2Dalgarno Scientific LLC, Brookline, MA, USA. 3Brown
University Department of Chemistry, Providence, RI, USA. e-mail: [email protected]

Nature Communications | (2024)15:2464 1

Article https://doi.org/10.1038/s41467-024-46715-9

inability to predict multiple conformations is unsurprising given its enhanced-sampling MD simulations. As a second example, we pre-
initial scope, the capacity to make predictions of different conforma- dicted changes in the conformational ensemble of GMCSF, a protein
tions would be as revolutionary as the capacity to accurately predict with minimal known homology, in response to point mutations. Our
ground states. Phenomena that involve different conformations such predictions strongly correlated with experimentally-determined NMR
as fold-switching and order-disorder transitions are ubiquitous across results, further showcasing subsampled AF2’s remarkable capacity to
the proteome19,20 and are directly tied to the activity of many decode signals pertaining to conformational changes even when
macromolecules21. Moreover, methods that can rapidly predict multi- sequence data is scarce. Altogether, these results highlight the strong,
ple conformations may have the potential to revolutionize drug dis- yet untapped potential of AF2 for predicting changes in the con-
covery by uncovering substantially more drug targets22. To fully realize formational ensembles of proteins, which will have substantial impacts
this potential, methods like AF2 will need to account for the relative on the fields of biophysics and drug discovery.
populations of different conformations (states) since the conforma-
tional equilibrium of drug receptors is directly related to their affinities Results
for drugs23,24. A prime example of this relationship is Imatinib25, a tyr- Optimizing MSA subsampling to predict kinase core con-
osine inhibitor whose exceptional selectivity for Abl1 kinase was found formational ensembles
to be caused by the enzyme’s significant preference for conforma- In recent years, multiple groups have observed that AF2 with different
tional states that facilitate Imatinib recognition and subsequent parameters and MSA depths is capable of predicting conformational
induced-fit binding26. changes based on sequence data alone27,28. These alternative
In an attempt to realize this potential, researchers have devised AF2 pipelines share the principle of subsampling MSAs to modulate
new ways of employing the AF2 method to detect conformational co-evolutionary signals at different structural domains27. In its stan-
changes, with significant success in a few test cases27–31. Although AF2 dard implementation, AF2 takes as input a target sequence and a
cannot conventionally predict conformational ensembles, researchers corresponding multiple sequence alignment. An arbitrary number
have found that sub-sampling the input multiple sequence alignments of sequences (defined by the max_seq parameter) are randomly
(MSAs) and increasing the number of predictions leads to structural selected from the master MSA (the target sequence is always selected),
ensembles that capture different physiologically-relevant conforma- and the remaining sequences are clustered around each of the selected
tions from the same sequence27. These predictions can be used as sequences using a Hamming distance. Both the cluster centers and
seeds in molecular dynamics (MD) simulations seeking to explore a sample from each cluster with a length of extra_seq are used by
larger swaths of the conformational space and the relative populations AF2 for inference (see Fig. 2). Previous works have shown that a sig-
of each predicted state28. Despite being a significant improvement nificant reduction in the values of max_seq and extra_seq from their
over methods that only predict ground states, methods such as these default values achieves ensemble prediction for a series of model
still rely on expensive MD simulations to infer most relative state systems29.
population information, which comes at a significant cost compared to Motivated by these observations, we started our work by sys-
simply running a prediction engine. tematically testing the accuracy of different AF2 parameter combina-
Here, we show that these MD simulation steps may be unneces- tions for predicting the Abl1 kinase core structural ensemble. We chose
sary if the goal is to discover major alternative conformations and their the Abl1 kinase core (residues 229–515) as our first test case due to this
relative populations in a high-throughput fashion, such as for con- protein’s extensively documented dynamics23. Abl1 is thought to
trasting differences in the dynamics of orthologs or allelic series of a occupy three major conformations with different populations. In
protein of interest. We take inspiration for this work from the obser- solution, Abl1 primarily exists in an active (ground) state. Infrequently,
vation that proteins from the same evolutionary line can have differ- Abl1 will switch to inactive state 1 (I1), and then to inactive state 2 (I2)23,
ences in relative state populations that are strongly correlated with the which strongly binds to Imatinib (Gleevec)26. While the change from
genetic distance between them26. Since AF2 works by decoding co- the ground to I1 state is subtle, the transition from the I1 to I2 state
evolutionary signals15 and previous works have suggested that sub- involves considerable backbone rearrangements: the activation loop
sampling MSAs leads to accurate predictions of different conforma- detaches from its resting position below the C-helix and folds on itself,
tions of the same protein27, it seems reasonable to hypothesize that a change that shifts the activation loop by over 15 Å from its original
some instructions for conformational sampling could be decoded position as shown in Fig. 3.
from sequence data alone. If this hypothesis holds true, AF2 and other To encourage AF2 to generate a full ensemble of Abl1 conforma-
AI-based methods could be capable of quantifying sequence-encoded tions, we started by compiling an extensive MSA spanning over
dynamic signals, which would make it possible to predict not only 600,000 sequences using the JackHMMR algorithm32 on the wild-type
alternative conformations of the same protein, but also changes in its Abl1 kinase core (residues 229–515). This algorithm builds the MSA by
relative state populations. querying sequences from the UniRef9033, Small BFD34, and MGnify35
With this as motivation, we show how subsampling multiple databases. To increase the statistical power of our results, we then ran
sequence alignments can generate ensembles of protein conforma- 32 predictions with independent seeds for each test, and enabled
tions (see Fig. 1 for an overview of our method) and systematically test dropouts during inference to sample from the uncertainty of the
AF2’s capacity to predict sequence-induced differences in the con- models. Dropout rates were the same as those found to improve
formational distributions of the Abl1 tyrosine kinase core and of the sampling in other studies (10% for the Evoformer module, and 25% for
granulocyte-macrophage colony-stimulating factor (GMCSF). Diver- the structural module)36. All other parameters were left in their default
ging from previous works, as a first example, we focus on detecting settings (3 recycles per prediction, 5 models per seed, a total of 160
changes in the active state population across the Src kinase to Abl1 predictions per run, 3 independent runs with unique seeds, 480 pre-
evolutionary line and test our ability to predict the effects of single and dictions per test).
double point mutations known or suspected to shift state distribu- As described in Supplementary Table 1, we find that a max_se-
tions. Crucially, we found that subsampled AF2 can qualitatively pre- q:extra_seq value of 256:512 leads to the most diverse results in terms of
dict both the positive and negative effects of mutations on the active activation loop conformations (see Supplementary Figs. 1 and 2).
state populations of kinase cores with up to eighty percent accuracy. Importantly, the ensemble of activation loop conformations predicted
We also found that AF2 predicts most of the activation loop inter- by AF2 with the above parameter set is distributed across the
mediate states in the active-to-inactive transition of the kinase ground state to I2 state transition in Abl1, with no predictions
core with an ensemble that is comparable to that obtained from falling significantly outside the boundaries of known activation loop

Nature Communications | (2024)15:2464 2

Article https://doi.org/10.1038/s41467-024-46715-9

Fig. 1 | Summary of the subsampled AF2 workflow employed in this study. AF2 to output predictions that occupy different conformations of the same protein,
A Traditionally, AF2 predicts the structure of a target by using a multiple sequence and the predicted frequency of each conformation based upon a range of random
alignment (MSA). When running AF2 with standard parameters, the predicted seeds strongly correlates with its experimentally-determined relative state popu-
structures are often similar to each other even with a large number of independent lation. Figure Created with BioRender.com.
predictions (seeds). B In this study, we show that subsampling deep MSAs causes

conformations and no blatantly unphysical or misfolded predictions. conformational distributions without the need for downstream MD
As a further test of the claim that we are actually predicting con- simulations. As a basic sanity check, we tested if the Abl1 prediction
formations along a transition, we compared the ensemble of 160 sub- results were actually the product of AF2 decoding co-evolutionary
sampled AF2 Abl1 predictions to representative snapshots extracted signal pertaining to relative state populations, or just a fortuitous
from a I1 to I2 trajectory generated with enhanced-sampling MD coincidence. Accordingly, we used the same subsampled AF2 protocol
simulations of apo Abl1 in solution. Specifics about the methodology outlined in the Supplementary Materials to predict the relative state
used to generate this trajectory are described in Supplementary populations of the wild-type Src kinase, which is known to occupy the
Figs. 3 and 4 and their accompanying discussion. Representative ground (active) state significantly more frequently than Abl126 (see
results from this comparison are illustrated in Fig. 4 and the results of Fig. 5), making it an attractive control case. If our hypothesis regarding
the entire analysis are illustrated in Supplementary Figs. 5–8. Although the potential of subsampled AF2 is indeed correct, we expect that the
we expected to observe a range of conformations, the coverage of the method will output significantly more predictions of ground state Src
activation loop transition is remarkable and suggests the possibility of than ground state Abl1. Accordingly, we built a large MSA for the Src
using AF2 to sample intermediate states and uncover pathways and kinase core (residues 235–497) sequence using the same procedure as
mechanisms. described for Abl1 and ran our implementation of subsampled AF2
with it as an input. We then measured the relative population of the Src
Predicting the conformations of members of a kinase evolu- kinase core ground and I2 states.
tionary line using subsampled AF2 Crucially, we found the vast majority of Src kinase core predic-
Given our success in predicting the relative population of the Abl1 tions from subsampled AF2 to be in the ground state (for more
kinase ground state, we next studied AF2’s potential for predicting information about how predictions were binned into different states,

Nature Communications | (2024)15:2464 3

Article https://doi.org/10.1038/s41467-024-46715-9

Fig. 2 | AF2’s multiple sequence alignment (MSA) clustering heuristic. A An MSA elements are used by AF2 for inference. Cluster features and a number of random
of arbitrary length is built from a target sequence and passed to AF2, which ran- non-cluster-center sequences (defined by extra_seq) are processed and passed to
domly selects a number of sequences (defined by max_seq) from the input MSA. the Main Evoformer Stack, while the MSA containing the cluster centers is pro-
Each of the selected sequences becomes a cluster center around which the cessed, passed to the comparatively expensive row/col attention track, and then
sequences not selected in the previous step are distributed. The target sequence is finally passed to the Main Evoformer Stack as well. Figure Created with
always selected as a cluster center. The clusters obtained through this process are BioRender.com.
featurized and relevant statistics are calculated. B All of the previously discussed

see Supplementary Fig. 1), with a predicted relative state population of be considered accurate, the relative population of the ground state in
97% compared to 89% for Abl1, as summarized in Fig. 6. Interestingly, the Anc-AS predictions should be in between the populations of the
none of the Src predictions were found to be in the I2 state, although same state for the Abl1 and Src predictions, as observed in experi-
the enzyme is known to infrequently occupy this conformation. This mental results. Once again, subsampled AF2 correctly replicated
suggests a resolution limitation in using AF2 to predict relative state experiments (see Fig. 6), as Anc-AS was predicted to be in the ground
populations: conformations with very low occupancy such as I2 in Src state 93% of the time, in-between the frequencies predicted for Src
might be missed by the algorithm in its current implementation. A (97%) and Abl1 (89%).
potential cause of this limitation is the fact that AF2’s prediction
models were trained on all structures deposited in the Protein Data Predicting the conformations of a kinase allelic series with AF2
Bank (PDB) up until 2018, and most of the structures of Src and its While the results we obtained for the Abl1 to Src evolutionary line are
orthologs within AF2’s training set are in the ground state37. We promising, an even more impactful application would be predicting
anticipate that fine-tuning or retraining AF2 and similar AI methods how state populations change across an allelic series. This is because
with significantly more diverse structural datasets representing dif- many point mutations in proteins are thought to lead to different
ferent conformational states of a given protein domain could be a phenotypes - such as drug resistance - by changing conformational
viable strategy for increasing the resolution and accuracy of predicting landscapes and relative state populations.
the conformational plasticity of that domain. Additionally, using dee- To measure the capacity of subsampled AF2 to fill this niche, we
per MSAs in the training could also improve prediction accuracy, in repeated our predictions using a series of Abl1 single and double
line with recent results that used deep MSAs to achieve higher pre- mutants with well-characterized and significant effects on the relative
diction accuracies than earlier methods38. populations of the ground and I2 states, and contrasted the results
Despite resolution caveats, subsampled AF2 correctly predicted with those obtained from the wild-type prediction. Specifically, we
the difference in conformational distributions between the Abl1 and tested the method on four mutations that are expected to decrease the
Src kinase cores, lending credence to its promise as a high-throughput population of the ground state (M290L, L301I, F382V, M290L + L301I),
method for predicting relative state populations. To shed further light and four mutations that are expected to increase it (E255V, T315I,
on AF2’s capabilities, we applied subsampled AF2 to make predictions F382L, E255V + T315I). The mutations tested, their locations in Abl1,
for the Anc-AS kinase core (residues 1–263), an Abl1 ancestor with a and their expected effects on ground state populations are summar-
known conformational state distribution and dynamics26, and com- ized in Fig. 5 (see Supplementary Table 2 for more details regarding the
pared the results to the Abl1 and Src cases. The sequences of Abl1, Anc- expected outcomes of the mutations in the relative state populations
AS, and Src used to generate the MSAs and as target sequences for AF2 of the Abl1 kinase core). The length of the kinase core sequence used as
are summarized in Supplementary Fig. 9. While there are other Abl1 input for this test (229–515) differs slightly from the previous one
ancestors that could be used in this test, there are experimental results (235–497) so as to match the length of the constructs tested in the
for Anc-AS, including a deposited structure in the Protein Databank literature. This difference caused a slight variation in the wild-type Abl1
(4UEU)26, justifying its choice. For the subsampled AF2 predictions to ground state predictions (84% vs. 89%).

Nature Communications | (2024)15:2464 4

Article https://doi.org/10.1038/s41467-024-46715-9

Predicting GMCSF dynamics with AF2

Considering the success of our Abl1 predictions, we sought to test if
the accuracy of these predictions was contingent on the wealth of
kinase sequence data, or if we could obtain similar results with sig-
nificantly less sequence data. To do so, we repeated our prediction
workflow but used the sequence of the human granulocyte-
macrophage colony-stimulating factor (GMCSF). GMCSF is a 14 kDa
monomeric glycoprotein that plays a central role in innate immunity,
stimulating a variety of cells in response to pathogens39. In contrast to
that for kinases, literature regarding GMCSF’s structure and dynamics
is sparse, and sequence data for homologs or orthologs is orders of
magnitude less abundant. Crucially, the MSA built for the wild-type
human GMCSF sequence is only 112 sequences long, while that for
human Abl1 kinase is over 600,000 sequences using the same para-
meters and methods (see Fig. 7). This stark contrast represents a useful
opportunity to study how the accuracy of subsampled AF2 is modu-
lated by the availability of sequence data.
Previous NMR results have shown that the dynamics of GMCSF’s
N-terminal helix A are involved in the binding of heparin and other
charged molecules40. It is thought that the position of helix A with
respect to the center of the protein is flexible, usually being tightly
packed against helix C through π-π interactions among histidines 15,
83, and 87. Experimental results suggest that this closed configuration
Fig. 3 | Summary of Abl1 kinase core ensemble prediction results using sub- is the most stable GMCSF conformation when packed as a crystal41. For
sampled AlphaFold2. A Top: Models of the Abl1 kinase core in its active and clarity, this closed conformation will be henceforth referred to as the
inactive conformations. Relevant structural elements are represented in different ground state. It is thought that the π-π interactions are eventually
colors. Cyan: activation loop; red: phosphate-binding loop; yellow: C-helix; green: weakened either through intrinsic breathing motions and resultant
hinge; and pink: DFG motif. Models are taken from the PDB (6XR6 for Ground, interactions with the solvent or via induction by other molecules, and
6RXG for I2)23. Top Right: Ensemble of 160 models of the Abl1 kinase core as helix A is thought to move away from the core. This opening motion
predicted by subsampled AF2, with the activation loop highlighted in different exposes part of the GMCSF core to solvent and creates a groove that is
colors for each prediction. B Projection of 160 Abl1 kinase core predictions from thought to be the binding site of heparin and other immune system
subsampled AF2 plotted according to their backbone RMSDs relative to the Abl1 modulators40. This binding-competent GMCSF state will be referred to
kinase core active (6XR6) and inactive states (6XRG). Data points are colored as the open conformation. Importantly, GMCSF binds heparin much
according to their average predicted local distance difference test score (pLDDT) as
more strongly in an acidic environment, suggesting that helix A
calculated by AF2, which is a metric of the confidence of AF2 predictions11. Source
dynamics are protonation-dependent40.
data are provided as a Source Data file.
Further NMR experiments (see Supplementary Figs. 11 and 12)
have shown that point mutations in the aforementioned histidine triad
AF2 predictions for this allelic series are summarized in Fig. 6. lead to significant chemical shift perturbations, similar to those caused
Strikingly, subsampled AF2 correctly predicted a change in the relative by reductions in pH40, hinting at increased occupancy of the open state
state population and its direction in over 80% of the tested cases. or other unknown conformations. Importantly, the amplitudes of the
Although promising, these results are not without significant caveats. changes to the backbone dynamics caused by these mutations vary
First, the M290L predictions are inaccurate. Specifically, the effects significantly depending on which histidine of the triad is mutated.
of the mutation on the ground state population are predicted to be Mutating H15 or H83 leads to more pronounced changes in chemical
the opposite of those seen in experiments. Second, the prediction shifts (see Fig. 8). This is expected as H15 and H83 are significantly
accuracy only applies in a qualitative sense, as double mutations that are more buried than H87, meaning that mutations at these positions are
known to significantly increase the ground state population such likely to be harder for the GMCSF backbone to accommodate.
as M290L + L301I are predicted by subsampled AF2 to increase it Although the magnitude and distribution of the 1H-15N peaks observed
only slightly more than single mutations such as M301I, which is known in Fig. 8 suggest that significant conformational changes involving
to cause a more subtle increase. We believe that this inaccuracy is backbone atoms may occur in GMCSF mutants (especially those
a direct consequence of the incorrect M290L prediction, which involving H15 and H83), we must not discard the possibility that the
should also result in the underestimation of the effects of double side chains surrounding the measured atoms might be inducing or
mutations including M290L. Furthermore, the statistical significance of involved in the conformational changes. To distinguish between con-
the results is reduced for the benchmarks in which the mutations tributions from the main and side chains, further studies exploring
are known to reduce ground-state populations. We hypothesize that GMCSF dynamics with methods that show ensemble averaging of
AF2 performed better at predicting decreases in the ground state backbone structures such as residual dipolar coupling might be
population because wild-type Abl1 occupies this state with near 90% necessary. Finally, mutations H83Y and H83R lead to larger-scale
occupancy, thus mutations that decrease it generally have a greater conformational changes (inferred from broadened peaks) than any
magnitude effect on relative state populations than mutations that other mutation in the test set, and considerably more than the H83N
further increase it. mutation, suggesting that specific substitutions at each position
Finally, considering that the version of AF2 used in this manu- induce more significant changes in the GMCSF backbone dynamics
script ships with five different models (see Supplementary Table 3), we (see Supplementary Fig. 11 for a detailed analysis of all mutant CSPs
asked how well these different models performed at estimating and broadened peaks). These mutations are therefore useful for
changes in the relative state populations of Abl1 states caused by dif- benchmarking as they represent a tiered challenge of predicting which
ferent point mutations. For a discussion of how different models fared specific amino acid substitution will lead to the largest changes in
at this challenge, please see Supplementary Fig. 10. GMCSF structure.

Nature Communications | (2024)15:2464 5

Article https://doi.org/10.1038/s41467-024-46715-9

Fig. 4 | Comparison between the I1 to I2 trajectory obtained using enhanced- RMSDs of each AF2 prediction’s kinase core (residues 242–459) or activation loop
sampling MD simulations of the Abl1 kinase core and representative AF2 (residues 379–395) vs. the kinase core or activation loop backbone of the MD
predictions. A Structural elements relevant for Abl1 function and that are expected snapshot selected at each time point. Distance deltas are defined as the difference
to shift significantly over the I1 to I2 transition. B Evolution of four relevant struc- in atom pair distances between each AF2 prediction and each respective MD
tural observables across the trajectory containing the I1 to I2 transition. Vertical snapshot. Distance 1 corresponds to the distance between the backbone oxygens of
lines indicate representative snapshots extracted from the trajectory for down- E377 and L409, and Distance 2 corresponds to the distance between the backbone
stream comparisons with AF2 predictions (top). Three representative snapshots oxygens of L409 and G457. D Comparison between six representative AF2 pre-
from the MD trajectory at 0 (I1), 8 (transition), and 24 (I2) ns, respectively (bottom). dictions and the six MD snapshots described above. Source data are provided as a
C Distribution of three observables relative to the MD snapshots for 160 sub- Source Data file.
sampled AF2 predictions. Core and A-Loop RMSDs are defined as the backbone

Nature Communications | (2024)15:2464 6

Article https://doi.org/10.1038/s41467-024-46715-9

Fig. 6 | Subsampled AF2 predictions of the percent of conformations not in the

ground state for proteins along the Src to Abl1 evolutionary pathway and Abl1
resistance-causing mutations. A Representation of the evolutionary history
connecting Abl1, Anc-As, and Src, and the relative population of the ground state
for each kinase core26. Right: Percentage of predictions that fall outside of the
ground state for each test case. B Results of subsampled AF2 for the Abl1 kinase
domain allelic series. Data are colored based on the expected relative state popu-
lations of each mutation: red-tinted bars represent mutations that are known to
Fig. 5 | Summary of experimental observations regarding the relative state decrease the ground state population, while blue-tinted bars represent mutations
populations of kinase cores along Abl1’s evolutionary line and in the Abl1 with the opposite effect23. Data are presented as mean values +/− standard error of
allelic series studied here. A Abl1, Anc-AS, and Src are part of the same evolu- the mean, and were calculated from three independent sets of predictions, each
tionary line and share significant homology, but Abl1 occupies the ground state with 32 unique seeds and 5 models, totaling 160 predictions per replicate. Source
significantly less frequently than Src, and Anc-AS occupies it slightly less than Src, data are provided as a Source Data file. Figure Created with BioRender.com.
suggesting that Abl1 evolution has directed it to be more flexible than Src26. B A
series of point mutations in wild-type Abl1 are known to increase or decrease the
relative population of the enzyme’s active (ground) state23. Residues with muta-
structures. We then repeated every previous step minus the parameter
tions known to increase the population of the ground state are shown as blue
spheres in the Abl1 structure on the left side, while those with mutations known to
optimization for each GMCSF mutant and quantified the ground state
decrease the ground state population are shown as red spheres. The yellow sphere populations from the resulting set of predictions. Although we were
denotes phenylalanine 382, part of the DFG motif, which can be mutated to valine still capable of sampling different conformational states even with
to slightly increase the ground state population, or to leucine to reduce it. While GMCSF’s shallow MSA, it is worth noting the occasional prediction of
the effects of the E255V + T315I mutation on the ground state population were not partially unfolded conformations with no experimental analogs. The
directly reported in the literature, we used its cumulative effect on kinase activity same did not happen with the Abl1 example. We posit that this loss of
and Imatinib binding55 to infer an increase in the ground state population. Figure resolution at aggressive subsampling levels could be a consequence of
Created with BioRender.com. the shallow input MSA.
To assess how mutations affected the conformational distribution
of GMCSF, we measured the RMSD of specific backbone atomic posi-
Accordingly, we sought to test if our subsampled AF2 method was tions of each predicted GMCSF structure with respect to the AF2
capable of predicting the expected changes in the backbone dynamics prediction that was most similar to the wild-type crystallographic
of each mutant with respect to the flexibility of helix A and other reference (PDB 1CSG)41. The regions mentioned above correspond to
structural elements in GMCSF. Our predictions were considered GMCSF elements known to show significant perturbations in NMR
accurate if they matched NMR experiments we performed along two experiments. The results of this analysis for the two regions with the
axes: if they predicted mutations to H15 and H83 to provoke larger most significant changes in our NMR results are described in Fig. 8.
changes in the distribution of conformations than mutations to H87; Additionally, we also measured the distance between the alpha car-
and if they correctly predicted the most significant mutations in H83 or bons of H15 and H83 as well as the overall backbone RMSD with respect
H87 as evidenced by CSPs or broadened peaks. After building the MSA to the ground state reference for each prediction with the goal of
using the wild-type human GMCSF sequence as a query and the identifying predictions that led to unexpected conformations or par-
JackHMMR method32, we determined the max_seq and extra_seq tially or completely unfolded structures. These additional measure-
parameters that led to the greatest diversity of GMCSF conformations ments as well as three examples of unexpected structures are
(see Supplementary Fig. 13). illustrated in Supplementary Fig. 14.
Employing these parameters while keeping all others the same Despite the paucity of sequence data, our subsampled AF2
from the Abl1 tests, we used AF2 to predict wild-type GMCSF method correctly identified residues H15 and H83 as the most sensitive

Nature Communications | (2024)15:2464 7

Article https://doi.org/10.1038/s41467-024-46715-9

Fig. 7 | Comparison between MSA length and per-position coverage for two protein systems whose conformational ensembles were predicted in this study.
A Abl1 kinase core. B GMCSF.

to mutation, as the range of the distribution of RMSDs of residues physiological functions. Although certainly promising, these experi-
80–90 and 110–125 is significantly larger for most of the mutations ments are beyond the scope of this study due to their complexity and
tested at both of these sites. Conversely and in line with the experi- cost. In the absence of ground-truth evidence, the facts that these
mental results, mutations to H87 led to significant changes in the predictions display structural rearrangements with amplitudes that
distribution of the residue 80–90 RMSDs, but to comparatively mod- seem to be comparable to those of the H83 mutations (as evidenced by
est changes for the RMSDs of residues 110–125. In addition to accu- the H83 NMR chemical shift perturbations and broadened peaks
rately predicting the differences in amplitude of backbone [Supplementary Fig. 11]) and that these conformations were predicted
rearrangements between mutants H15/83 and H87, subsampled AF2 more often for H83 mutants highlights the promise of using sub-
correctly estimated the significant impact of mutations H83R and sampled AF2 to help understand NMR results and derive novel
H83N on C-terminal conformations, while also accurately predicting hypotheses or mechanisms.
H83N, H83Y, and all three H87 mutants to have a large impact on the Finally, many of the alternative conformations not discussed
R80–90 RMSD distribution. above were found to be partially unfolded. Although the frequency of
As with the Abl1 allelic series example, our method did not achieve these was low compared to predictions binned to the ground or
perfect accuracy. Specifically, it failed to predict the significant chan- A1 states, the fact that they existed at all was surprising as running
ges in GMCSF’s N-terminus dynamics associated with the H83N subsampled AF2 for the Abl1 test case led to no unfolded predictions
mutation (see Supplementary Fig. 11). Additionally, our method failed whatsoever. One plausible hypothesis beyond the destabilizing effects
to replicate the significant changes in dynamics of residues 80–90 for of the mutations that may explain this observation is that the sub-
the H83Y mutations, which should be larger in amplitude than those stantially shorter length of the GMCSF MSA relative to Abl1’s drastically
for H83N and closer in amplitude to those for H83R. Similarly, we increases the uncertainty in AF2’s predictions leading to a wider range
expect the H87Y mutation to induce larger conformational changes in of predictions.
the region composed of residues 80–90 than the H87N mutation due
to the enrichment of broadened peaks in that region in the H87Y NMR Additional examples
results, but AF2 predicts H87Y to be significantly less influential than To shed light on how well our results on the Abl1 kinase core and
H87N in that context. In summary, AF2 performed exceptionally well at GMCSF generalize and how the accuracy of our predictions depend
predicting subtle conformational changes in specific loops for most of upon MSA depth and diversity, we recognize the need for more tests
our test cases but failed at replicating the comparatively larger on a wider array of protein systems. We have therefore curated an
expected effects of mutation H83Y. Refer to Supplementary Table 5 for additional test set composed of eight proteins diverse in size,
a measurement of the statistical power of comparisons between dynamics, function, and evolutionary history. The results of this chal-
ensembles of predictions belonging to wild-type or mutant GMCSF. lenge are thoroughly described in the Supplementary Appendix
Beyond these limitations, we also observed a set of alternative (Supplementary Table 4, Supplementary Figs. 16–27), and ultimately
conformations that are significantly different than both the ground support the notion that the approach described in this study can be
and open states. Clustering the results by the structural features applied for predicting the conformational landscapes of a wide array of
described in Supplementary Fig. 15 reveals that one alternative con- proteins and their variants.
formation, in particular, is significantly enriched, especially in predic-
tions of the H83 mutants. In this alternative conformation, henceforth Discussion
dubbed A1, the C helix has switched places with the B helix, placing In this work, we successfully modified AF2’s inputs and parameters to
H83 and H15 more than 10 Å away (Supplementary Fig. 14A). In this predict the conformational ensembles and relative state populations
state, helix B occupies the groove to which heparin is thought to bind, of two proteins, Abl1 kinase and human granulocyte-macrophase col-
which could be a mechanism for self-inhibition. We believe that further ony-stimulating factor (GMCSF), that have very different amounts of
NMR experiments such as chemical exchange saturation transfer known sequence homology. In studying these proteins, we focused on
(CEST) or ensemble studies via residual dipolar coupling could be the how well our subsampled version of AF2 can reproduce the effects of
best way to confirm if this is a metastable GMCSF conformation with evolution and mutations on dynamics. In addition to these two clearly

Nature Communications | (2024)15:2464 8

Article https://doi.org/10.1038/s41467-024-46715-9

Fig. 8 | Subsampled AF2 results for GMCSF mutations. A Left: Annotated wild- elements for both conformations are shown as yellow bars. B Per-residue chemical
type GMCSF structure in the ground (closed) conformation as predicted by AF2. shift perturbations for GMCSF mutants, separated by codon position. Gold vertical
Each colored element represents a region of putative high mobility according to bars represent the three mutation sites, and silver shaded areas correspond to the
chemical shift perturbations (CSPs) in the histidine triad mutants, identified by residue indices plotted on the x-axis in C. C Distributions of root mean square
visual analysis of the CSP data. Specifically, residue ranges with significant pertur- deviations of atomic positions for the backbone atoms of residues 80–90 (top) and
bations and peak broadening were designated as regions of putative high mobility. 110–125 (bottom) superimposed upon the GMCSF wild-type ground state structure.
Middle: Superposition of GMCSF in the ground conformation and a putative open The distributions span 480 independent predictions (32 unique seeds * 5 models * 3
conformation whose population is enriched in the AF2 predictions for H83 and H87 replicates. The center of the inset box plot is defined as the median of each dataset,
mutants, especially H83N. The inset compares distances between codon positions and corresponds to the following values for the R80–90 range: 2.11, 2.26, 2.41, 2.44,
15 and 83 in both conformations. Right: Superposition of GMCSF in the ground 2.04, 2.39, 2.01, 2.23, 2.20, 2.01; and to the following values for the R110-125 range:
conformation and a tilted helix D conformation whose population is enriched in the 1.84, 2.16, 2.52, 2.56, 1.72, 2.39, 2.65, 1.79, 1.73, 1.70. Source data are provided as a
AF2 predictions for H87 mutants (especially H87Y). Distances between flexible Source Data file.

contrasting examples, we also measured our approach’s capacity for ground state nine times more frequently than its I2 state, consistent
predicting alternative conformations and their relative populations with NMR observations. More importantly, subsampled AF2 correctly
(when known) for a diverse set of eight protein systems. predicted the relative state populations of two evolutionarily related
Our subsampled AF2 implementation predicted Abl1, a kinase for kinases, Anc-AS and Src kinase, leading to the correct correlation
which there is an abundance of sequence homology, to occupy its between their evolutionary distances and relative state populations.

Nature Communications | (2024)15:2464 9

Article https://doi.org/10.1038/s41467-024-46715-9

Using single and double Abl1 mutants with known effects on the remains to be shown whether it is capable of directly predicting
relative state populations of Abl1’s ground and I2 states, we moreover Boltzmann ratios of states quantitatively.
found that AF2 yielded surprisingly accurate state populations even for The diverse protein test set also allowed us to evaluate how our
single mutations. This is best evidenced in the results for the modified AlphaFold2 approach fared at predicting conformational
F382 substitutions: phenylalanine 382 is a codon known to slightly changes of different scales, both in terms of the number of atoms
increase the population of Abl1’s ground state if mutated to leucine, or involved and in the expected timescale of the change itself. From this
significantly reduce it if mutated to valine, an observation that is analysis, we observed that subsampled AlphaFold2 fared better at
accurately replicated by our prediction method. Furthermore, an predicting large and slow conformational changes, such as the LmrP or
unexpected but remarkable feature illustrated by all of the Abl1 test LAT1 channel openings that involve the correlated motions of hun-
cases is the capacity of subsampled AF2 to predict intermediate con- dreds of backbone atoms (Supplementary Figs. 17 and 20, respec-
formations spanning the transition from the ground state to I2 which tively). For these types of conformational changes, AlphaFold2
closely match intermediate conformations obtained from more costly predicted a variety of potentially intermediate conformations span-
MD simulations. ning the transition and both ground and high-energy states for a wide
We also obtained overwhelmingly accurate conformational range of subsampling values. For faster or less significant conforma-
ensembles that correlated with NMR experiments for GMCSF. Despite tional changes, such as the conversion between the Fyn-SH3 inter-
the lack of GMCSF sequence data which led to an input MSA of fewer mediate and its folded ground state (Supplementary Fig. 23),
than 120 sequences (versus more than 600,000 for Abl1), predictions subsampled AF2 predicted no potential transition conformations and
from subsampled AF2 were distributed with frequencies that matched only predicted both states under narrow subsampling conditions,
observed changes in dynamics for most GMCSF variants. These results suggesting a resolution limitation.
suggest that AF’s prediction engine is robust enough to decode some Additionally, subsampled AF2 performed better at predicting
conformational distribution information from relatively scarce data. alternative conformations when used on systems whose high-energy
The results we have obtained for the two test cases discussed states are more frequently populated. For example, the NMR-resolved
above confirm AF2’s potential for predicting conformational ensem- relative populations of wild-type Abl1’s and mutant Fyn-SH3’s higher-
bles and, more importantly, demonstrate unforeseen applications of energy states are 6% and 2%, respectively. While our approach suc-
AF2. In particular, we show that optimization of subsampling para- cessfully predicted wild-type Abl1’s (I2, inactive) high-energy state
meters allows users to accurately predict relative state populations using multiple subsampling conditions, our approach was only able to
and how they change in response to mutations with single-substitution predict the high-energy intermediate folding state of mutant Fyn-SH3
resolution. This feature is a significant advance over the previous state within a narrow range of subsampling conditions. Further, the Abl1
of the art, as it facilitates high-throughput applications such as the ensemble of predictions often included a small but significant (>5%)
design of fold-switching proteins, inference of mechanisms of population of putative in-between conformations, while the mutant
acquired drug resistance, and reweighting of binding affinity predic- Fyn-SH3 prediction ensemble did not include such structures (Sup-
tions. Additionally, our workflow generates conformational inter- plementary Fig. 23). These results indicate that there is some minimum
mediates, which has direct implications for discovering drugs that threshold for the relative population between 2 and 6% to be able to
bind alternative conformations and for improving our understanding detect higher energy states using subsampled AF2.
of biophysics in general. In addition to these immediate practical Although these results are overwhelmingly positive, they also
applications, the more fundamental observation that AF2 is decoding come with caveats that were not observed in the Abl1 example. Nota-
information regarding conformational distributions from sequence bly, our capacity to predict relative state populations without prior
data alone points to many other potential unforeseen uses of AF2 that knowledge was limited for a few protein systems within the test set,
can result in further methodological advances and discoveries. such as LAT1, for which the ground state population was dependent on
This said, our AF2 pipeline is not perfect; our workflow inaccu- the level of subsampling. Additionally, in some examples such as AkrB,
rately predicted the M290L mutation to significantly decrease the aggressive subsampling led to the prediction of unfolded conforma-
ground state population when that mutation is known experimentally tions unlikely to map to the functional states of the protein studied. In
to have the opposite effect. Interestingly, AF2 predicted the double the Supplementary Appendix, we discuss a few heuristics for mini-
mutation (M290L + L301I) to increase the ground state population mizing the incidence of these potential pitfalls, such as measuring the
more than the L301I mutation alone. As AF2 correctly predicted the coverage of the putative pathway between two presumed states and
relative state populations of Src which differs from Abl1 by dozens of estimating the frequency of unfolded predictions. Although it is not
mutations, this suggests a potentially more general trend that AF2 is within the scope of this work to offer a one-size-fits-all approach for
more accurate in predicting the effects of multiple mutations than discovering appropriate subsampling parameters for every protein
those of a single mutation. system, we believe that these suggestions can help orient future stu-
Furthermore, our pipeline also struggled to correctly predict all of dies seeking to accomplish similar objectives.
the structural elements expected to differ in each conformation. Despite these limitations, we believe that our results showcase
Specifically, while AF2 predicted an ensemble of activation loop con- promising unexpected applications of AF2. Using the high-throughput
formations, a few of the inactive-like predictions (activation loop pipeline we developed or one inspired by it could save significant time
closed) contained structural elements that are typically thought to when filtering large allelic series to identify mutations with significant
belong to the active state, such as the position of the phosphate- impacts on conformation for further study using NMR or other, more
binding loop and the dihedral angles of D381 and F382. Moreover, even expensive experimental or computational techniques. It could also be
when our pipeline predicts a change in a structural element from its used as a prediction engine for classifying arrays of drug-resistant
configuration in the ground (active) state, the amplitude of the change mutations based on their shared effects on the stability of a given
may not correspond to that seen in experimentally-resolved struc- conformation, thus facilitating polypharmacology. Finally, as demon-
tures. For example, the side-chain dihedral angle of the D381 residue in strated by AF2’s Abl1 activation loop predictions, our method could
Abl1 with the plane formed by the A380 side-chain ranges from −130 also be useful for identifying previously unknown, potentially meta-
(active state) to 40 (inactive state) degrees in structures in the Protein stable states of known proteins. While it remains to understand exactly
Data Bank (PDB)23, but ranges from −130 to −90 degrees in AF2 pre- how AF2 is gathering and interpreting signals about state populations
dictions. Last but not least, while our method is capable of predicting from sequence data, we hope that our work will motivate many future
whether mutations will increase or decrease state populations, it investigations.

Nature Communications | (2024)15:2464 10

Article https://doi.org/10.1038/s41467-024-46715-9

Table 1 | Summary of parameters used for enhanced-sampling 8.0. GMCSF was refolded by dilution via dropwise addition of the 10
molecular dynamics simulations of Abl1 inactivation mL elution into 100 mL of a refolding buffer containing 10mM Tris-
HCl, 100 mM sodium phosphate, and 750 mM arginine at pH 8.0. The
Box dimensions (112.4, 124.3, 118.0) refolded protein was dialyzed exhaustively against a buffer containing
Total number of atoms 77251 2 mM sodium phosphate at pH 7.4. GMCSF was concentrated to
Number of water molecules 724982 ~200 μM with an Amicon centrifugal device and stored at −20 °C.
Salt concentration 0.125M
NMR spectroscopy
Thermostat Noose-Hoover54
NMR samples were prepared by dialyzing 200 μ GMCSF and GMCSF
Barostat Monte Carlo barostat44
mutants against a buffer of 20 mM HEPES and 1 mM EDTA at pH 7.4.
Nonbonded cutoff 10.0 A NMR experiments were performed on a Bruker Avance NEO 600 MHz
Temperature 300 K spectrometer at 25 °C. NMR data were processed in NMRPipe45 and
Pressure 1 ATM analyzed in Sparky46. The 1H and 15N carrier frequencies were set to
Prod. Integration Timestep 2 fs water resonance and 120 ppm, respectively.
t (W.E. iteration length) 100 ps All NMR experiments were performed using a Bruker Avance NEO
600 MHz spectrometer equipped with a triple resonance inverse
detection cryoprobe. The data were collected through the companion
Methods software Topspin 4.0.3 using a 2D HN correlation via double inept
Protein structure visualization transfer heteronucelar single quantum coherence (HSQC) pulse pro-
To visualize the predicted structures and trajectories and calculate gram (hsqcetfpf3gpsi) at 298 K47–50. Spectral processing was per-
descriptors such as distances between atoms, RMSD to reference, formed using NMRFAM-SPARKY46 while data analysis was performed
dihedral angles, etc., we used PyMol (version 2.4.1) (Schrodinger using Microsoft Excel 202151 and GraphPad Prism 10.0.1 for MacOS52.
LLC, 2020). All GMCSF NMR samples were stored in 20mM HEPEs 1 mM EDTA, at
concentrations between 300–500 uM.
Multiple sequence alignments
The JackHMMR algorithm32 was used to generate multiple sequence Miscellaneous and data visualization
alignments (MSAs) for each protein of interest by querying sequences Data plotting and visualization were performed using Seaborn (ver-
from the UniRef9033, Mgnify35, and small BFD34 databases. sion 0.11.1).
Figures were composed with BioRender (BioRender.com, 2020).
Structure prediction
We used AlphaFold 211 within the localcolabfold colabfold-batch 1.5.0 Reporting summary
implementation42 to predict protein structures of Abl1, Src, ANC-AS, Further information on research design is available in the Nature
and of GMCSF. Portfolio Reporting Summary linked to this article.

Molecular dynamics simulations Data availability

Enhanced-sampling molecular dynamics simulations were conducted Sample MSAs and the resulting analysis of AlphaFold2 predictions and
on the proteins described using the WESTPA243 implementation within MD simulation data generated in this study have been deposited in the
the OpenMM molecular simulation engine44. The general parameters GitHub repository https://github.com/GMdSilva/gms_natcomms_
for the simulations are described in Table 1. For an extensive 1705932980_data53. Due to storage limitations, the repository neither
description of the MD methodology employed in this study, please includes the resulting PDB ensembles nor all of the MSAs generated
refer to the Supplementary Materials. and used in this study, although access can be obtained by contacting
the corresponding author. Source data are provided with this study.
Protein expression and purification
Plasmid DNA containing either wildtype GMCSF or GMCSF containing Code availability
a point mutation with an N terminal 6-His tag was cloned into a pET-15b Scripts used to assemble the MSAs and run AlphaFold2 with different
vector and transformed into BL21(DE3) cells in a manner described subsampling conditions have been deposited in the GitHub repository
elsewhere. Isotopically enriched GMCSF was expressed at 37 °C in M9 https://github.com/GMdSilva/gms_natcomms_1705932980_data53.
minimal medium containing CaCl2, MgSO4 and MEM vitamins with
15
NH4Cl as the sole nitrogen source. Small cultures of GMCSF were References
grown overnight in LB medium. The following morning, cloudy sus- 1. Robertson, A. D. & Murphy, K. P. Protein structure and the ener-
pensions were collected by centrifugation and resuspended in the final getics of protein stability. Chem. Rev. 97, 1251–1268 (1997).
M9 growth medium. Cultures of GMCSF were grown to an OD600 of 2. chen Bai, X., McMullan, G. & Scheres, S. H. How cryo-EM is revo-
0.8–1.0 before induction with 1 mM isopropyl β-D-1- thiogalactopyr- lutionizing structural biology. Trends Biochem. Sci. 40,
anoside (IPTG). Post induction, cells were kept at 18 °C while shaking. 49–57 (2015).
Cells were harvested after 18h and resuspended in a denaturing lysis 3. Su, X.-D. et al. Protein crystallography from the perspective of
buffer containing 10 mM Tris-HCl, 100 mM sodium phosphate, and 6 technology developments. Crystallogr. Rev. 21, 122–153 (2014).
M guanidine hydrochloride (GuHCl) at pH 8.0. Cells were lysed by 4. Wüthrich, K. Protein structure determination in solution by NMR
sonication and cell debris was removed by centrifugation. The result- spectroscopy. J. Biol. Chem. 265, 22059–22062 (1990).
ing supernatant was incubated and nutated with 10 mL of Ni-NTA 5. Slabinski, L. et al. The challenge of protein structure determination-
agarose beads for 30 min at room temperature before the Ni-NTA lessons from structural genomics. Protein Sci. 16,
slurry was packed into a gravity column. The column was washed with 2472–2482 (2007).
the initial lysis buffer, followed by a gradient of the same buffer 6. Leach, A. & Thomas, P. Protein structure prediction and homology
without GuHCl over 100 mL. Elution of GMCSF in its denatured form modeling. In Comprehensive Medicinal Chemistry III, 120–144
was performed with 1 column volume of a buffer containing 10 mM (Elsevier, 2017). https://doi.org/10.1016/b978-0-12-409547-2.
Tris- HCl, 100 mM sodium phosphate, and 250 mM imidazole at pH 12350-9.

Nature Communications | (2024)15:2464 11

Article https://doi.org/10.1038/s41467-024-46715-9

7. Ołdziej, S. et al. Physics-based protein-structure prediction using a 33. Suzek, B. E., Wang, Y., Huang, H., McGarvey, P. B. & and, C. H. W.
hierarchical protocol based on the UNRES force field: Assessment UniRef clusters: a comprehensive and scalable alternative for
in two blind tests. Proc. Natl Acad. Sci. 102, 7547–7552 (2005). improving sequence similarity searches. Bioinformatics 31,
8. Jothi, A. Principles, challenges and advances in ab initio protein 926–932 (2014).
structure prediction. Protein Peptide Lett. 19, 1194–1204 (2012). 34. Steinegger, M. & Söding, J. Clustering huge protein sequence sets
9. Torrisi, M., Pollastri, G. & Le, Q. Deep learning methods in protein in linear time. Nat. Commun. 9, https://doi.org/10.1038/s41467-
structure prediction. Comput. Struct. Biotechnol. J. 18, 018-04964-5 (2018).
1301–1310 (2020). 35. Richardson, L. et al. MGnify: the microbiome sequence data ana-
10. AlQuraishi, M. Machine learning in protein structure prediction. lysis resource in 2023. Nucl. Acids Res. 51, D753–D759 (2022).
Curr. Opin. Chem. Biol. 65, 1–8 (2021). 36. Wallner, B. AFsample: improving multimer prediction with Alpha-
11. Jumper, J. et al. Highly accurate protein structure prediction with Fold using massive sampling. Bioinformatics 39, https://doi.org/10.
AlphaFold. Nature 596, 583–589 (2021). 1093/bioinformatics/btad573 (2023).
12. Tunyasuvunakool, K. et al. Highly accurate protein structure pre- 37. Modi, V. & Dunbrack, R. L. Kincore: a web resource for structural
diction for the human proteome. Nature 596, 590–596 (2021). classification of protein kinases and their inhibitors. Nucl. Acids Res.
13. Baek, M. et al. Accurate prediction of protein–nucleic acid 50, D654–D664 (2021).
complexes using RoseTTAFoldNA. Nat. Methods 21, 117–121 (2024). 38. Lee, S. et al. Petascale homology search for structure prediction.
14. Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein bioRxiv, https://doi.org/10.1101/2023.07.10.548308 (2023).
structure with a language model. Science 379, 1123–1130 (2023). 39. Lee, K. M., Achuthan, A. A. & Hamilton, J. A. Gm-csf: A promising
15. Roney, J. P. & Ovchinnikov, S. State-of-the-art estimation of protein target in inflammation and autoimmunity. ImmunoTargets Therapy
model accuracy using AlphaFold. Phys. Rev. Lett. 129, https://doi. 9, 225–240 (2020).
org/10.1103/physrevlett.129.238101 (2022). 40. Cui, J. Y. et al. Mapping the structural and dynamic determinants of
16. Callaway, E. What’s next for AlphaFold and the AI protein-folding pH-sensitive heparin binding to granulocyte macrophage colony
revolution. Nature 604, 234–238 (2022). stimulating factor. Biochemistry 59, 3541–3553 (2020).
17. Chakravarty, D. & Porter, L. L. Alphafold 2 fails to predict protein fold 41. Walter, M. R. et al. Three-dimensional structure of recombinant
switching. Protein Sci. 31, https://doi.org/10.1002/pro.4353 (2022). human granulocyte-macrophage colony-stimulating factor. J. Mol.
18. Pak, M. A. et al. Using AlphaFold to predict the impact of single Biol. 224, 1075–1085 (1992).
mutations on protein stability and function. PLOS One 18, 42. Mirdita, M. et al. ColabFold: making protein folding accessible to all.
e0282689 (2023). Nat. Methods 19, 679–682 (2022).
19. Porter, L. L. & Looger, L. L. Extant fold-switching proteins are 43. Bogetti, A. T. et al. A suite of advanced tutorials for the WESTPA 2.0
widespread. Proc. Natl Acad. Sci. 115, 5968–5973 (2018). rare-events sampling software [article v2.0]. Living J. Comput. Mol.
20. Bryan, P. N. & Orban, J. Proteins that switch folds. Curr. Opin. Struct. Sci. 5, https://doi.org/10.33011/livecoms.5.1.1655 (2022).
Biol. 20, 482–488 (2010). 44. Eastman, P. et al. OpenMM 7: Rapid development of high perfor-
21. Kim, A. K. & Porter, L. L. Functional and regulatory roles of fold- mance algorithms for molecular dynamics. PLOS Comput. Biol. 13,
switching proteins. Structure 29, 6–14 (2021). e1005659 (2017).
22. Borkakoti, N. & Thornton, J. M. AlphaFold2 protein structure pre- 45. Delaglio, F. et al. NMRPipe: A multidimensional spectral processing
diction: Implications for drug discovery. Curr. Opin. Struct. Biol. 78, system based on UNIX pipes. J. Biomol. NMR 6, https://doi.org/10.
102526 (2023). 1007/bf00197809 (1995).
23. Xie, T., Saleh, T., Rossi, P. & Kalodimos, C. G. Conformational states 46. Lee, W., Tonelli, M. & Markley, J. L. NMRFAM-SPARKY: enhanced
dynamically populated by a kinase determine its function. Science software for biomolecular NMR spectroscopy. Bioinformatics 31,
370, https://doi.org/10.1126/science.abc2754 (2020). 1325–1327 (2014).
24. Michielssens, S., de Groot, B. L. & Grubmüller, H. Binding affinities 47. Palmer, A. G., Cavanagh, J., Wright, P. E. & Rance, M. Sensitivity
controlled by shifting conformational equilibria: Opportunities and improvement in proton-detected two-dimensional heteronuclear
limitations. Biophys. J. 108, 2585–2590 (2015). correlation nmr spectroscopy. J. Magn. Resonance 93, 151–170 (1991).
25. Iqbal, N. & Iqbal, N. Imatinib: A breakthrough of targeted therapy in 48. Kay, L., Keifer, P. & Saarinen, T. Pure absorption gradient enhanced
cancer. Chemotherapy Res. Pract. 2014, 1–9 (2014). heteronuclear single quantum correlation spectroscopy with
26. Wilson, C. et al. Using ancient protein kinases to unravel a modern improved sensitivity. J. Am. Chem. Soc. 114, 10663–10665 (1992).
cancer drug’s mechanism. Science 347, 882–886 (2015). 49. Schleucher, J. et al. A general enhancement scheme in hetero-
27. Wayment-Steele, H. K. et al. Predicting multiple conformations via nuclear multidimensional nmr employing pulsed field gradients. J.
sequence clustering and AlphaFold2. Nature 625, 832–839 (2024). Biomol. NMR 4, https://doi.org/10.1007/BF00175254 (1994).
28. Vani, B. P., Aranganathan, A., Wang, D. & Tiwary, P. AlphaFold2- 50. Grzesiek, S. & Bax, A. The importance of not saturating water in
RAVE: From sequence to boltzmann ranking. J. Chem. Theory protein nmr. application to sensitivity enhancement and noe mea-
Comput. https://doi.org/10.1021/acs.jctc.3c00290 (2023). surements. J. Am. Chem. Soc. 115, 12593–12594 (1993).
29. Meller, A., Bhakat, S., Solieva, S. & Bowman, G. R. Accelerating 51. Microsoft excel 2021 116.54, microsoft corporation, https://office.
cryptic pocket discovery using AlphaFold. J. Chem. Theory Comput. microsoft.com/excel (2021).
https://doi.org/10.1021/acs.jctc.2c01189 (2023). 52. Graph prism 10.0.1 for macos, graphpad software, www.graphpad.
30. Stein, R. A. & Mchaourab, H. S. SPEACH_AF: Sampling protein com (2023).
ensembles and conformational heterogeneity with alphafold2. 53. da Silva, G. M., Cui, J. Y., Dalgarno, D. C., Lisi, G. P. & Rubenstein, B.
PLOS Comput. Biol. 18, e1010483 (2022). M. High-throughput prediction of protein conformational distribu-
31. del Alamo, D., Sala, D., Mchaourab, H. S. & Meiler, J. Sampling tions with subsampled alphafold2. gms_natcomms_1705932980_-
alternative conformational states of transporters and receptors with data, https://doi.org/10.5281/zenodo.10621196 (2024).
AlphaFold2. eLife 11, e75751 (2022). 54. Evans, D. J. & Holian, B. L. The nose–hoover thermostat. J. Chem.
32. Finn, R. D., Clements, J. & Eddy, S. R. HMMER web server: interactive Phys. 83, 4069–4074 (1985).
sequence similarity searching. Nucleic Acids Res. 39, 55. Hoemberger, M., Pitsawong, W. & Kern, D. Cumulative mechanism
W29–W37 (2011). of several major imatinib-resistant mutations in abl kinase. Proc. Natl
Acad. Sci. 117, 19221–19227 (2020).

Nature Communications | (2024)15:2464 12

Article https://doi.org/10.1038/s41467-024-46715-9

Acknowledgements Peer review information Nature Communications thanks Arne Elofsson,

The authors thank Jacob Rosenstein, Joseph Larkin, F. Marty Ytreberg, Emmanuel Salawu and Vincenzo Venditti for their contribution to the
Jagdish S. Patel, Marcelo D. Polêto, Winston Y. Li, Haibo Li, and Kyle Lam peer review of this work. A peer review file is available.
for insightful conversations and support. G.M.d.S. and J.Y.C. were sup-
ported by a Blavatnik Family Fellowship award. G.M.d.S. and B.M.R. were Reprints and permissions information is available at
supported in part by the National Science Foundation under Grant No. http://www.nature.com/reprints
2027108. J.Y.C. and G.P.L.’s contributions were supported by NIH Grant
R01 GM144451. The computational aspects of this research were con- Publisher’s note Springer Nature remains neutral with regard to jur-
ducted using computational resources and services at the Center for isdictional claims in published maps and institutional affiliations.
Computation and Visualization, Brown University.
Open Access This article is licensed under a Creative Commons
Author contributions Attribution 4.0 International License, which permits use, sharing,
G.M.d.S. performed the simulations and analyzed the sequence and adaptation, distribution and reproduction in any medium or format, as
AlphaFold 2 data. J.C. performed NMR experiments and contributed long as you give appropriate credit to the original author(s) and the
NMR data. D.C.D., G.P.L., and B.M.R. provided direction and oversight. source, provide a link to the Creative Commons licence, and indicate if
G.M.d.S. and B.M.R. drafted the manuscript. All authors provided notes changes were made. The images or other third party material in this
and edits to the manuscript. article are included in the article’s Creative Commons licence, unless
indicated otherwise in a credit line to the material. If material is not
Competing interests included in the article’s Creative Commons licence and your intended
The authors declare no competing interests. use is not permitted by statutory regulation or exceeds the permitted
use, you will need to obtain permission directly from the copyright
Additional information holder. To view a copy of this licence, visit http://creativecommons.org/
Supplementary information The online version contains licenses/by/4.0/.
supplementary material available at
https://doi.org/10.1038/s41467-024-46715-9. © The Author(s) 2024, corrected publication 2024

Correspondence and requests for materials should be addressed to

Brenda M. Rubenstein.

Nature Communications | (2024)15:2464 13