Use CombFold Online
Commercially Available CombFold No-Code Web Server
CombFold: Predicting Structures of Large Protein Assemblies with AlphaFold2
CombFold is a combinatorial assembly algorithm that predicts the structures of large protein complexes by combining pairwise AlphaFold-Multimer predictions. It was developed by Ben Shor and Dina Schneidman-Duhovny at the School of Computer Science and Engineering, The Hebrew University of Jerusalem, and described in Nature Methods in 2024.
Many of the cell's key machines, such as ribosomes, transporters and signaling assemblies, are large multi-protein complexes. Experimental structures of them are hard to obtain, and the authors note that end-to-end AlphaFold-Multimer struggles as complexes grow. CombFold sidesteps this limit by predicting small subsets of subunits and then assembling the full structure from them.
Property | Detail |
|---|---|
Task | Structure prediction of large protein assemblies |
Inputs | Subunit sequences; optional distance restraints (for example crosslinking mass spectrometry or FRET); optional stoichiometry |
Outputs | Assembled complex structures ranked by a confidence score |
Scale | Up to 128 subunits in the current implementation |
Building block | AlphaFold-Multimer run through ColabFold, without templates |
How CombFold Assembles Large Protein Complexes
The method has three stages:
Pairwise and subset prediction: AlphaFold-Multimer predicts structures for pairs of subunits and for subsets of three to five subunits, with subsets capped at a total of 1,800 residues.
Representative selection: representative subunit structures are chosen by highest mean pLDDT, and pairwise transformations between subunits are computed and scored from the predicted aligned error (PAE).
Combinatorial hierarchical assembly: the transformations are combined step by step into full complexes, which are ranked by a confidence score.
Optional restraints and stoichiometry
Distance restraints from crosslinking mass spectrometry or FRET can be added to guide assembly. When the number of copies of each subunit is unknown, candidate stoichiometries can be enumerated and compared by confidence.
CombFold Benchmark Results for Large Heteromeric and Homomeric Assemblies
Across two datasets of 60 large asymmetric assemblies, 72% of complexes were predicted with TM-score above 0.7 in the top 10 models, with structural coverage 20% higher than the corresponding PDB entries.
Benchmark | Result |
|---|---|
35 heteromers (5-20 chains) | Top-1 success 60% (21/35); top-10 74% |
25 heteromers (5-30 chains) | Top-1 64%; with simulated crosslinks 76%; end-to-end AlphaFold-Multimer v3 top-1 36% |
153 mostly homomers | Top-1 57% (87/153) |
CASP15 targets over 3,000 residues | Six of seven assembled successfully |
In one example, the eIF2B:eIF2 assembly was covered over 6,114 of 7,486 residues at pLDDT above 50, versus 4,680 in the cryo-EM structure. For TIM22, 12 crosslink restraints raised the TM-score to 0.85 from 0.57 without restraints.
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 CombFold on Tamarind Bio
Modeling large complexes: build models of assemblies with many subunits that are hard to predict end to end.
Integrative structural biology: combine predictions with crosslinking mass spectrometry or FRET restraints, which the authors show can improve accuracy.
Testing stoichiometry: compare confidence across candidate copy numbers, as the authors did for ATP synthase subunit c and the PelC dodecamer.
Extending partial structures: increase structural coverage beyond what is resolved in experimental models.
How to Use CombFold on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and select CombFold.
Provide subunit sequences: enter the sequence of each subunit in the complex, as the method defines its input.
Set stoichiometry: specify copy numbers if known, or enumerate candidates, if exposed.
Add restraints: supply crosslinking or FRET distance restraints if you have them and the option is exposed.
Run the job: submit and wait for the pairwise predictions and assembly to finish.
Download outputs: retrieve the ranked assembled structures.
Interpret confidence: prefer higher-ranked models and check interfaces carefully, since the paper reports interface accuracy is lower than global accuracy.
Use downstream: visualize, validate against experimental data or refine the models.
Things to Keep in Mind
Accuracy is limited by the success of AlphaFold-Multimer pairwise predictions, cited at roughly 60%.
Interface contact similarity is lower than TM-score, so interfaces are less accurate than overall folds.
Stoichiometry must be known or enumerated, which the authors describe as the main bottleneck.
Compositional and conformational heterogeneity are not addressed directly.
Some predictions do not match experiment, for example the Mrp complex (TM-score 0.67).
Source: Ben Shor and Dina Schneidman-Duhovny, "CombFold: predicting structures of large protein assemblies using a combinatorial assembly algorithm and AlphaFold2", Nature Methods 21, 477-487 (2024), DOI 10.1038/s41592-024-02174-0, https://www.nature.com/articles/s41592-024-02174-0.