Use AF Cluster Online
Commercially Available AF Cluster No-Code Web Server
AF-Cluster: Predicting Alternative Protein Conformations with AlphaFold2 and MSA Clustering
AF-Cluster is a method for predicting multiple conformations of a protein by clustering its multiple sequence alignment (MSA) and running AlphaFold2 on each cluster. It was described by Hannah K. Wayment-Steele, Dorothee Kern and colleagues in Nature (2023) in the paper "Predicting multiple conformations via sequence clustering and AlphaFold2". The accompanying code repository is listed under an MIT license.
AlphaFold2 predicts a single, highly accurate structure for most proteins, yet many proteins function by switching between conformations, such as fold-switching and metamorphic proteins. Standard AlphaFold2 runs with a full MSA tend to return one dominant state, and the authors note that other sources of noise, such as dropout or random seeds, did not produce the fold-switched state of KaiB.
AF-Cluster uses the evolutionary signal itself. Sequences in the alignment are grouped by similarity so that each cluster can carry coevolutionary information about a different state of the protein.
Property | Detail from the paper |
|---|---|
Core idea | Cluster the MSA, then predict a structure from each sub-alignment with AlphaFold2 |
MSA generation | ColabFold MMseqs2 routine against UniRef30; sequences with more than 25% gaps removed |
Clustering | DBSCAN on edit distance; epsilon swept from 3 to 20 and chosen where cluster count peaks |
Ranking | Models ranked by pLDDT |
Journal | Nature, volume 625, pages 832-839 (2024); DOI 10.1038/s41586-023-06832-9 |
How AF-Cluster Samples Alternative Protein States
Build an MSA: an alignment is generated with ColabFold, and sequences with excessive gaps are filtered out.
Cluster by sequence similarity: DBSCAN groups sequences by edit distance, with epsilon selected where the number of clusters is greatest. Unassigned sequences are labeled as noise.
Predict per cluster: AlphaFold2 is run on each sub-alignment, and models are ranked by pLDDT.
Check coevolutionary support: MSA Transformer can be used to test whether a cluster carries couplings for a given state.
The repository also generates uniformly sampled control MSAs of 10 and 100 sequences. In the paper, uniform subsampling was far less effective: at an MSA size of 10, about 1 in 500 samples fell within 3 angstroms of the KaiB ground state, and at size 100 none did.
AF-Cluster Results on KaiB, RfaH and Other Metamorphic Proteins
Test | Result reported |
|---|---|
KaiB (108 residues) | Both the ground state and fold-switched state predicted; default ColabFold gives the fold-switched state, while the 50 closest sequences give the ground state |
KaiB variant validated by NMR | NMR structure differed from the AF-Cluster model by 1.89 +/- 0.13 angstroms RMSD |
KaiB point mutations (I68R, V83D, N84R) | Fold-switched population rose to 86% from 11% in wild type |
RfaH | Top-five average pLDDT of 84.2 (active) and 73.9 (autoinhibited); full MSA gave 68.6 for autoinhibited and missed the active state |
GA/GB point-mutation set | Correct most stable fold in 10 of 12 cases versus 8 of 12 for default AlphaFold2 |
Mpt53 screen | Six of ten DALI homologues sampled an analogous alternative fold |
The authors also screened 628 families, and report that open and closed states of MAD2 were predicted. They estimate that if 0.5 to 4% of proteins switch folds, this could correspond to roughly 1 to 8 million of the 214 million AlphaFold DB predictions.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with AF-Cluster on Tamarind Bio
Fold-switching discovery: screen proteins for candidate alternative states, as the authors did for Mpt53.
Mutation design: use predicted states to propose point mutations that shift the balance between conformations, as demonstrated for KaiB.
Hypothesis generation for experiments: prioritize NMR or biophysical tests using predicted alternative structures.
Studying evolution of conformational switching: compare states across sequence clusters, as in the 487 KaiB variants surveyed.
How to Use AF-Cluster on Tamarind Bio
Log in and open the tool: sign in to tamarind.bio and select AF Cluster.
Provide the target sequence: enter the amino acid sequence of the protein of interest, or an MSA if the tool accepts one.
Set clustering parameters: if exposed, adjust MSA filtering and the DBSCAN setting.
Run the job: submit and wait for per-cluster AlphaFold2 predictions.
Download the outputs: retrieve predicted structures and pLDDT values for each cluster.
Compare states: cluster models by RMSD to identify distinct conformations, and inspect those with high pLDDT.
Validate: follow up with experiments or other methods before drawing conclusions.
Things to Keep in Mind
pLDDT is not a free-energy measure; in KaiB the disfavored fold-switched state scored higher than the ground state.
The number of models per state reflects sequence abundance, not Boltzmann weight.
States absent from the evolutionary signal may require other methods.
Oligomeric transitions, such as those of selecase, lymphotactin and CLIC1, were not captured.
Source: Wayment-Steele H. K., Ojoawo A., Otten R., et al., Kern D. "Predicting multiple conformations via sequence clustering and AlphaFold2." Nature 625, 832-839 (2024), DOI 10.1038/s41586-023-06832-9, https://www.nature.com/articles/s41586-023-06832-9. Code: https://github.com/HWaymentSteele/AFCluster.