CoChat
Sign Up
Cited by 197Protein Structure and DynamicsReceptor Mechanisms and SignalingMachine Learning in Bioinformatics

The DOI matches its Crossref record; 52 of the 52 references it cites were found in a registry or an index.

High-throughput prediction of protein conformational distributions with subsampled AlphaFold2

Open access
Monteiro da Silva, Gabriel, Cui, Jennifer Y., Dalgarno, David C., Lisi, George P., Rubenstein, Brenda M.·2024·Nature Communications
DOI 10.1038/s41467-024-46715-9PMC PMC10973385
Open this paper in CoChat — free accountOpen PDF

Abstract

This paper presents an innovative approach for predicting the relative populations of protein conformations using AlphaFold 2, an AI-powered method 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.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41467-024-46715-9

Cites

52 of 52 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Summary of the subsampled AF2 workflow employed in this study. A Traditionally, AF2 predicts the structure of a target by using a multiple sequence alignment (MSA). When running AF2 with standard parameters, the predicted structures are often similar to each other even with a large number of independent predictions (seeds). B In this study, we show that subsampling deep MSAs causes AF2 to output predictions that occupy different conformations of the same protein, and the predicted frequency of each conformation based upon a range of random seeds strongly correlates with its experimentally-determined relative state population. Figure Created with BioRender.com.
Fig. 1 Summary of the subsampled AF2 workflow employed in this study. A Traditionally, AF2 predicts the structure of a target by using a multiple sequence alignment (MSA). When running AF2 with standard parameters, the predicted structures are often similar to each other even with a large number of independent predictions (seeds). B In this study, we show that subsampling deep MSAs causes AF2 to output predictions that occupy different conformations of the same protein, and the predicted frequency of each conformation based upon a range of random seeds strongly correlates with its experimentally-determined relative state population. Figure Created with BioRender.com.CC BY · PubMed Central
Fig. 2 AF2’s multiple sequence alignment (MSA) clustering heuristic. A An MSA of arbitrary length is built from a target sequence and passed to AF2, which randomly selects a number of sequences (defined by max_seq) from the input MSA. Each of the selected sequences becomes a cluster center around which the sequences not selected in the previous step are distributed. The target sequence is always selected as a cluster center. The clusters obtained through this process are featurized and relevant statistics are calculated. B All of the previously discussed elements are used by AF2 for inference. Cluster features and a number of random non-cluster-center sequences (defined by extra_seq) are processed and passed to the Main Evoformer Stack, while the MSA containing the cluster centers is processed, passed to the comparatively expensive row/col attention track, and then finally passed to the Main Evoformer Stack as well. Figure Created with BioRender.com.
Fig. 2 AF2’s multiple sequence alignment (MSA) clustering heuristic. A An MSA of arbitrary length is built from a target sequence and passed to AF2, which randomly selects a number of sequences (defined by max_seq) from the input MSA. Each of the selected sequences becomes a cluster center around which the sequences not selected in the previous step are distributed. The target sequence is always selected as a cluster center. The clusters obtained through this process are featurized and relevant statistics are calculated. B All of the previously discussed elements are used by AF2 for inference. Cluster features and a number of random non-cluster-center sequences (defined by extra_seq) are processed and passed to the Main Evoformer Stack, while the MSA containing the cluster centers is processed, passed to the comparatively expensive row/col attention track, and then finally passed to the Main Evoformer Stack as well. Figure Created with BioRender.com.CC BY · PubMed Central
Fig. 3 Summary of Abl1 kinase core ensemble prediction results using subsampled AlphaFold2. A Top: Models of the Abl1 kinase core in its active and inactive conformations. Relevant structural elements are represented in different colors. Cyan: activation loop; red: phosphate-binding loop; yellow: C-helix; green: hinge; and pink: DFG motif. Models are taken from the PDB (6XR6 for Ground, 6RXG for I2)23. Top Right: Ensemble of 160 models of the Abl1 kinase core as predicted by subsampled AF2, with the activation loop highlighted in different colors for each prediction. B Projection of 160 Abl1 kinase core predictions from subsampled AF2 plotted according to their backbone RMSDs relative to the Abl1 kinase core active (6XR6) and inactive states (6XRG). Data points are colored according to their average predicted local distance difference test score (pLDDT) as calculated by AF2, which is a metric of the confidence of AF2 predictions11. Source data are provided as a Source Data file.
Fig. 3 Summary of Abl1 kinase core ensemble prediction results using subsampled AlphaFold2. A Top: Models of the Abl1 kinase core in its active and inactive conformations. Relevant structural elements are represented in different colors. Cyan: activation loop; red: phosphate-binding loop; yellow: C-helix; green: hinge; and pink: DFG motif. Models are taken from the PDB (6XR6 for Ground, 6RXG for I2)23. Top Right: Ensemble of 160 models of the Abl1 kinase core as predicted by subsampled AF2, with the activation loop highlighted in different colors for each prediction. B Projection of 160 Abl1 kinase core predictions from subsampled AF2 plotted according to their backbone RMSDs relative to the Abl1 kinase core active (6XR6) and inactive states (6XRG). Data points are colored according to their average predicted local distance difference test score (pLDDT) as calculated by AF2, which is a metric of the confidence of AF2 predictions11. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 4 Comparison between the I1 to I2 trajectory obtained using enhanced-sampling MD simulations of the Abl1 kinase core and representative AF2 predictions. A Structural elements relevant for Abl1 function and that are expected to shift significantly over the I1 to I2 transition. B Evolution of four relevant structural observables across the trajectory containing the I1 to I2 transition. Vertical lines indicate representative snapshots extracted from the trajectory for downstream comparisons with AF2 predictions (top). Three representative snapshots from the MD trajectory at 0 (I1), 8 (transition), and 24 (I2) ns, respectively (bottom). C Distribution of three observables relative to the MD snapshots for 160 subsampled AF2 predictions. Core and A-Loop RMSDs are defined as the backbone RMSDs of each AF2 prediction’s kinase core (residues 242–459) or activation loop (residues 379–395) vs. the kinase core or activation loop backbone of the MD snapshot selected at each time point. Distance deltas are defined as the difference in atom pair distances between each AF2 prediction and each respective MD snapshot. Distance 1 corresponds to the distance between the backbone oxygens of E377 and L409, and Distance 2 corresponds to the distance between the backbone oxygens of L409 and G457. D Comparison between six representative AF2 predictions and the six MD snapshots described above. Source data are provided as a Source Data file.
Fig. 4 Comparison between the I1 to I2 trajectory obtained using enhanced-sampling MD simulations of the Abl1 kinase core and representative AF2 predictions. A Structural elements relevant for Abl1 function and that are expected to shift significantly over the I1 to I2 transition. B Evolution of four relevant structural observables across the trajectory containing the I1 to I2 transition. Vertical lines indicate representative snapshots extracted from the trajectory for downstream comparisons with AF2 predictions (top). Three representative snapshots from the MD trajectory at 0 (I1), 8 (transition), and 24 (I2) ns, respectively (bottom). C Distribution of three observables relative to the MD snapshots for 160 subsampled AF2 predictions. Core and A-Loop RMSDs are defined as the backbone RMSDs of each AF2 prediction’s kinase core (residues 242–459) or activation loop (residues 379–395) vs. the kinase core or activation loop backbone of the MD snapshot selected at each time point. Distance deltas are defined as the difference in atom pair distances between each AF2 prediction and each respective MD snapshot. Distance 1 corresponds to the distance between the backbone oxygens of E377 and L409, and Distance 2 corresponds to the distance between the backbone oxygens of L409 and G457. D Comparison between six representative AF2 predictions and the six MD snapshots described above. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 5 Summary of experimental observations regarding the relative state populations of kinase cores along Abl1’s evolutionary line and in the Abl1 allelic series studied here. A Abl1, Anc-AS, and Src are part of the same evolutionary line and share significant homology, but Abl1 occupies the ground state significantly less frequently than Src, and Anc-AS occupies it slightly less than Src, 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 mutations 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 decrease the ground state population are shown as red spheres. The yellow sphere denotes phenylalanine 382, part of the DFG motif, which can be mutated to valine to slightly increase the ground state population, or to leucine to reduce it. While the effects of the E255V + T315I mutation on the ground state population were not directly reported in the literature, we used its cumulative effect on kinase activity and Imatinib binding55 to infer an increase in the ground state population. Figure Created with BioRender.com.
Fig. 5 Summary of experimental observations regarding the relative state populations of kinase cores along Abl1’s evolutionary line and in the Abl1 allelic series studied here. A Abl1, Anc-AS, and Src are part of the same evolutionary line and share significant homology, but Abl1 occupies the ground state significantly less frequently than Src, and Anc-AS occupies it slightly less than Src, 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 mutations 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 decrease the ground state population are shown as red spheres. The yellow sphere denotes phenylalanine 382, part of the DFG motif, which can be mutated to valine to slightly increase the ground state population, or to leucine to reduce it. While the effects of the E255V + T315I mutation on the ground state population were not directly reported in the literature, we used its cumulative effect on kinase activity and Imatinib binding55 to infer an increase in the ground state population. Figure Created with BioRender.com.CC BY · PubMed Central
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 populations of each mutation: red-tinted bars represent mutations that are known to decrease the ground state population, while blue-tinted bars represent mutations with the opposite effect23. Data are presented as mean values +/− standard error of the mean, and were calculated from three independent sets of predictions, each with 32 unique seeds and 5 models, totaling 160 predictions per replicate. Source data are provided as a Source Data file. Figure Created with BioRender.com.
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 populations of each mutation: red-tinted bars represent mutations that are known to decrease the ground state population, while blue-tinted bars represent mutations with the opposite effect23. Data are presented as mean values +/− standard error of the mean, and were calculated from three independent sets of predictions, each with 32 unique seeds and 5 models, totaling 160 predictions per replicate. Source data are provided as a Source Data file. Figure Created with BioRender.com.CC BY · PubMed Central
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.
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.CC BY · PubMed Central
Fig. 8 Subsampled AF2 results for GMCSF mutations. A Left: Annotated wild-type GMCSF structure in the ground (closed) conformation as predicted by AF2. Each colored element represents a region of putative high mobility according to chemical shift perturbations (CSPs) in the histidine triad mutants, identified by visual analysis of the CSP data. Specifically, residue ranges with significant perturbations and peak broadening were designated as regions of putative high mobility. Middle: Superposition of GMCSF in the ground conformation and a putative open conformation whose population is enriched in the AF2 predictions for H83 and H87 mutants, especially H83N. The inset compares distances between codon positions 15 and 83 in both conformations. Right: Superposition of GMCSF in the ground conformation and a tilted helix D conformation whose population is enriched in the AF2 predictions for H87 mutants (especially H87Y). Distances between flexible elements for both conformations are shown as yellow bars. B Per-residue chemical shift perturbations for GMCSF mutants, separated by codon position. Gold vertical bars represent the three mutation sites, and silver shaded areas correspond to the residue indices plotted on the x-axis in C. C Distributions of root mean square deviations of atomic positions for the backbone atoms of residues 80–90 (top) and 110–125 (bottom) superimposed upon the GMCSF wild-type ground state structure. The distributions span 480 independent predictions (32 unique seeds * 5 models * 3 replicates. The center of the inset box plot is defined as the median of each dataset, and corresponds to the following values for the R80–90 range: 2.11, 2.26, 2.41, 2.44, 2.04, 2.39, 2.01, 2.23, 2.20, 2.01; and to the following values for the R110-125 range: 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 Source Data file.
Fig. 8 Subsampled AF2 results for GMCSF mutations. A Left: Annotated wild-type GMCSF structure in the ground (closed) conformation as predicted by AF2. Each colored element represents a region of putative high mobility according to chemical shift perturbations (CSPs) in the histidine triad mutants, identified by visual analysis of the CSP data. Specifically, residue ranges with significant perturbations and peak broadening were designated as regions of putative high mobility. Middle: Superposition of GMCSF in the ground conformation and a putative open conformation whose population is enriched in the AF2 predictions for H83 and H87 mutants, especially H83N. The inset compares distances between codon positions 15 and 83 in both conformations. Right: Superposition of GMCSF in the ground conformation and a tilted helix D conformation whose population is enriched in the AF2 predictions for H87 mutants (especially H87Y). Distances between flexible elements for both conformations are shown as yellow bars. B Per-residue chemical shift perturbations for GMCSF mutants, separated by codon position. Gold vertical bars represent the three mutation sites, and silver shaded areas correspond to the residue indices plotted on the x-axis in C. C Distributions of root mean square deviations of atomic positions for the backbone atoms of residues 80–90 (top) and 110–125 (bottom) superimposed upon the GMCSF wild-type ground state structure. The distributions span 480 independent predictions (32 unique seeds * 5 models * 3 replicates. The center of the inset box plot is defined as the median of each dataset, and corresponds to the following values for the R80–90 range: 2.11, 2.26, 2.41, 2.44, 2.04, 2.39, 2.01, 2.23, 2.20, 2.01; and to the following values for the R110-125 range: 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 Source Data file.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 52 references

Research works printed but not linked: 1

  1. Palmer, A. G., Cavanagh, J., Wright, P. E. & Rance, M. Sensitivity improvement in proton-detected two-dimensional heteronuclear correlation nmr spectroscopy. J. Magn. Resonance 93, 151–170 (1991).

Cited by

Sorted by citations

See all 197

Related

Metadata

Authors
Monteiro da Silva, G., Cui, J. Y., Dalgarno, D. C., Lisi, G. P., & Rubenstein, B. M.
Affiliations
  • Brown University (US)
  • Brooke (GB)
Corresponding author
Rubenstein, Brenda M.
ORCID
Venue
Nature Communications
Publisher
Springer Science and Business Media LLC
Published
2024-03-27
Volume
15
Issue
1
Pages
2464-2464
ISSN
2041-1723 (ISSN-L)
DOI
10.1038/s41467-024-46715-9
PMC
PMC10973385
OpenAlex
W4393233076
Cited by
196
References
52
Citations by year
2026: 60 · 2025: 85 · 2024: 51
FWCI
26.40
Type
journal article
Language
en
Update notices
correction (2024-04-10)
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Funders
  • National Science Foundation (2027108)
  • U.S. Department of Health & Human Services | NIH | National Cancer Institute (R01GM144451)
  • Blavatnik Family Foundation
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Receptor Mechanisms and Signaling
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
Keywords
protein conformational landscape, AlphaFold2, protein structure prediction, ABL1 kinase, GM-CSF, point mutations, protein evolution
MeSH terms
Mutation, Protein Conformation, Sequence Alignment, Point Mutation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review