Detailed protocol for generating closed-, open-, and inactivated-state hERG channel models using ColabFold with tailored structural templates

Abstract

Voltage-gated ion channels such as hERG (K_V11.1) adopt multiple conformational states that are critical for ion conduction and drug binding, yet experimentally resolved structures often capture only a limited subset of these states. Here, we describe a protocol for generating hERG channel models in closed, open, and inactivated conformations using ColabFold AlphaFold2 notebook with template-guided sampling (1). The approach uses carefully constructed structural templates derived from homologous proteins or intermediate model predictions to bias AlphaFold2 (2) toward specific conformational states. The workflow includes template construction in ChimeraX (3), multimeric ColabFold prediction, clustering of predicted models, manual identification of alternative conformations, and refinement using Rosetta relaxation with membrane energy function (4). This method enables reproducible generation of state-specific models and can be adapted to other membrane proteins where conformational heterogeneity is functionally important.

Background

Ion channels are dynamic membrane proteins that transition between multiple conformational states to regulate ion conduction and cellular excitability. These states, including closed, open, and inactivated conformations, are central to channel function and strongly influence interactions with small molecules. For many channels, including hERG (K_v11.1), drug binding is highly state-dependent, which makes accurate structural representation of multiple states important for mechanistic studies and drug safety evaluation. However, experimentally resolved structures often capture only one or a limited number of conformations, leaving key functional states underrepresented.

AlphaFold2 has significantly improved the accuracy of protein structure prediction, but in most cases it converges to a single dominant conformation. To access alternative states, additional guidance is required. One effective approach is to introduce structural templates that encode features of a desired state, allowing the prediction to be biased toward specific conformations while still permitting reconstruction of the remaining regions.

In this workflow, two complementary strategies are used to generate alternative conformational states (5). The first is direct template bias, in which structural elements from experimentally resolved proteins in known conformations are incorporated into the template to guide prediction. The second is iterative sampling, where an initial round of prediction is performed with minimal structural constraints to explore conformational diversity. Candidate models exhibiting alternative features are then used to construct new templates for a second round of prediction that enriches the desired state.

A critical aspect of this approach is the design of the template itself. The choice of which structural fragments are included in the template model, how those fragments are aligned relative to one another, and how much of each fragment is retained can strongly influence the outcome. Templates that are too restrictive may prevent exploration of alternative conformations, whereas templates that are too minimal may fail to bias the prediction. Careful alignment of homologous domains and iterative refinement of template boundaries are therefore essential for successfully generating distinct conformational states.

This protocol provides a practical framework for applying template-guided AlphaFold predictions to capture multiple conformational states and can be adapted to other ion channels and membrane proteins.

Software and Datasets

• Dryad dataset (input template models and output predicted models):
  https://datadryad.org/dataset/doi:10.5061/dryad.18931zd5x
• Structural model clustering and dihedral angle analysis scripts:
  https://github.com/k-ngo/AlphaFold_Analysis
• ColabFold notebook:
  https://colab.research.google.com/github/sokrypton/ColabFold/blob/v1.5.2/AlphaFold2.ipynb
• Rosetta software suite (tested on 3.14):
  https://downloads.rosettacommons.org/downloads/academic/

• ChimeraX:

  https://www.cgl.ucsf.edu/chimerax/download.html

Procedure

I. Open-state hERG model

a) Prepare initial structure

     1. Download hERG cryo-EM structure and open in ChimeraX:

PDB: 5VA2 | Cryo-EM structure of the human Ether-à-go-go-Related Gene (hERG) K⁺ channel in a putative open state: https://www.rcsb.org/structure/5VA2

2. Rebuild missing extracellular loops using ChimeraX “Model Loops” (https://www.cgl.ucsf.edu/chimerax/docs/user/tools/modelloops.html) or Rosetta “LoopRemodel”. In the ChimeraX GUI, first, go to Molecules Display -> Sequence to initiate the sequence alignment. 

3. After the sequence is displayed, you can now access the model loops feature. To do so, navigate to Tools -> Structure Editing -> Model Loops

4. Select the chain and the missing residue interval. Use the default anchor residues unless there is a reason to expand the rebuilt region slightly into neighboring resolved residues. For short loops, default settings are usually sufficient. Finally, click “Ok” to initiate loop modeling. A MODELLER license key is required, which is free for academic usage. 

b) Template preparation and structural prediction

This step is used to confirm that AlphaFold reproduces the open state.

1. In ChimeraX, retain only the following regions:

• Voltage-sensing domain: residues W398–V549
• Selectivity filter + pore-lining S6 helix: I607–T634
• Partial S4 helix: L520–V549
• Cytosolic domain: S660–R863

This can be done by executing the following commands in ChimeraX to first select regions of interest, then second delete everything outside of the selected regions: sel #1:398-549,607-634,660-863; delete ~sel;

     2. Save template using ChimeraX “Save As” option in CIF format and name it 5va2.cif. Keep all four chains (A–D).

     3. Run ColabFold (see Section IV) using 5va2.cif from the previous step as the template and generate 100 models (num_seeds = 20).

     4. The top ranked model (rank_001) by predicted Local Distance Difference Test) (pLDDT) will be selected as the representative open-state model.

II. Closed-state hERG model

a) Template preparation and structural prediction

1. Download and open the following structures in ChimeraX:

PDB: 5VA2 | Cryo-EM structure of the hERG K⁺ channel in a putative open state: https://www.rcsb.org/structure/5VA2

PDB: 8EP1 | Eag K_v channel in a closed state with voltage sensor in the down conformation: https://www.rcsb.org/structure/8EP1

     2. Superimpose the models on top of each other by entering “matchmaker” command in ChimeraX. Example command below will align model 1’s chain A with model 2’s chain A:

           mm #1/A to #2/A

     3. Extract open-state selectivity filter structure from PDB 5VA2 and deactivated-state voltage-sensing domain structure from PDB 8EP1:  

sel #1:607-634  #2:208-343

     4. Delete all other regions:

delete ~sel

    5. Combine fragments:

combine

     6. Save only the combined model using ChimeraX “Save As” option in CIF format and name it 5va2.cif. Keep all four chains (A–D).

     7. Run ColabFold (see Section IV) using 5va2.cif from the previous step as the template and generate 100 models (num_seeds = 20).

b) Cluster models

Use clustering script “clustering_pdb_AF_models.py” provided in https://github.com/k-ngo/AlphaFold_Analysis

The script is designed to help cluster AlphaFold-predicted PDB models based on the structural similarity of user-specified residues. The script calculates the all-atom root-mean-square deviation (RMSD) between each pair of models (by averaging the values over specified chains) and then clusters the models using the DBSCAN algorithm. It outputs detailed cluster information and generates a bar plot showing the average pLDDT (prediction confidence score stored in the B-factor column of the PDB file) for each cluster.

For the closed state, one needs to cluster using the entire model, because the goal is to compare overall closed-state architecture rather than only a local region. In the original analysis, closed-state models were clustered across the full channel with an RMSD threshold of 0.75 Å.

AlphaFold models start at residue 1 rather than canonical numbering. Before clustering, use ChimeraX to renumber all AlphaFold-predicted models to canonical hERG residue numbering to ensure consistency across structures, particularly if trimming or template-based modeling has altered residue indices. For instructions, see https://www.cgl.ucsf.edu/chimerax/docs/user/tools/renumber.html.

Example Python command, assuming the AlphaFold-predicted files are stored in the “alphafold_models” folder:

python clustering_pdb_AF_models.py \

--pdb_pattern "alphafold_models/*rank*.pdb" \

-c A,B,C,D \

--rmsd_threshold 0.75 \

--min_cluster_size 3 \

--results_file closed_cluster_results.json \

--overwrite

c) Final model selection

1. Identify the cluster with the highest average pLDDT. This is defined as Cluster 1.
2. Select the top-ranked model in that cluster, that is, the model with the highest average pLDDT according to the clustering script output.
3. Open candidate models in ChimeraX and reject any model with obvious steric clashes (ChimeraX: Tools -> Structure Analysis -> Clashes), such as overlapping side chains or backbone atoms.

III. Inactivated-state hERG model

a) First-round sampling (conformational space exploration)

This step is designed to increase conformational diversity and allow AlphaFold to explore alternative selectivity filter conformations.

1. Download and open the following structure in ChimeraX:

PDB: 5VA2 | Cryo-EM structure of the hERG K⁺ channel in a putative open state: https://www.rcsb.org/structure/5VA2

     2. Prepare template containing only cytosolic domain (residues S660–R863) by entering the following commands in ChimeraX:

sel #1:660-863; delete ~sel;

     3. Save template using ChimeraX “Save As” option in CIF format and name it 5va2.cif. Keep all four chains (A–D).

     4. Run ColabFold (see Section IV) using “5va2.cif” from the previous step as the template and generate 5 models (num_seeds = 1).

b) Identify candidate conformation

This first round is intentionally exploratory and uses a minimal template so that the transmembrane region, including the selectivity filter, can be resampled more freely.

1. Open all models in ChimeraX.
2. Align the models together using matchmaker (example command to align model 1 to 2: mm #1 to #2). 
3. Inspect selectivity filter (residues S624–G628) visually. Identify models showing alternative selectivity filter conformations that deviate from open-state selectivity filter (such as selectivity filter residue V625 carbonyl flipping)
4. This selected model is used to build the second-round inactivated-state template. Ensure that it is free of any obvious steric clashes (ChimeraX: Tools -> Structure Analysis -> Clashes), then save it as “5va2.cif”
5. If no suitable candidate is identified, repeat part (a) step 4 using a reduced max_msa, fewer num_recycles, and an increased num_seeds.

c) Construct second-round template

1. Open structures in ChimeraX

#1 = selected first-round model from step b (with selectivity filter deformation, e.g. V625 backbone carbonyl group flip)

#2 = PDB 5VA2 structure from Protein Data Bank

     2. Align subunit A of both models so domains are spatially consistent using ChimeraX command:

           mm #1/A to #2/A

     3. From selected first-round model (#1), select selectivity filter region Y607–T634:

           sel #1:607-634

     4. From 5VA2 (#2), select:

• Voltage-sensing domain: residues W398–V549 
• Cytosolic domain: residues S660–R863 

Use the following ChimeraX command:

sel add #2:398-549; sel add #2:660-863;

     5. Delete everything else. This keeps only the selected regions across both models:

           delete ~sel

     6. Combine both models into one template:

           combine

        This creates a single merged model containing:

• Selectivity filter from first-round model 
• Voltage-sensing domain + cytosolic domain from PDB: 5VA2  

     7. Save only the combined model using ChimeraX “Save As” option in CIF format and name it 5va2.cif. Keep all four chains (A–D).

d) Second-round ColabFold sampling

1. Run ColabFold with full settings (Section IV) using 5va2.cif provided in part (c) as the template.
2. Generate 100 models (num_seeds = 20).

e) Cluster models

Use the same clustering script “clustering_pdb_AF_models.py” from https://github.com/k-ngo/AlphaFold_Analysis

For the inactivated state, cluster only on the selectivity filter residues S624-G628, because inactivation primarily alters the selectivity filter and nearby pore geometry rather than the entire channel. In the original analysis, inactivated-state models were clustered using an RMSD threshold of 0.35 Å over this region. 

Before clustering, use ChimeraX to renumber all AlphaFold-predicted models to canonical hERG residue numbering to ensure consistency across structures (for instructions, see https://www.cgl.ucsf.edu/chimerax/docs/user/tools/renumber.html)

Example Python command, assuming the AlphaFold-predicted files are stored in the “alphafold_models” folder:

python clustering_pdb_AF_models.py \
--pdb_pattern "alphafold_models/*rank*.pdb" \
--residues 624,625,626,627,628 \
-c A,B,C,D \
--rmsd_threshold 0.35 \
--min_cluster_size 3 \
--results_file inactivated_cluster_results.json \
--overwrite

f) Select final model

1. Use the clustering results generated from the previous step to identify candidate clusters. Clusters are ranked automatically by the script based on average pLDDT score (extracted from the B-factor column of the PDB file)
2. Identify candidate clusters corresponding to plausible non-conductive (inactivated-like) conformations. In practice, priority should be given to clusters enriched in noncanonical selectivity filter conformations, such as V625 backbone carbonyl flipping and constriction at the level of G626 residues.
3. Extract candidate models from the selected cluster. This can be done by using the clustering output (JSON file and printed summary) to identify model filenames belonging to the cluster. 
4. (Optional but recommended) Further evaluate candidate models using analysis script “plot_dihedral_angles_AF_models.py” available in https://github.com/k-ngo/AlphaFold_Analysis

The script is designed to extract backbone dihedral angles (φ or ψ) from AlphaFold-predicted PDB models. Optionally, it can also incorporate reference PDB models for comparison. The script processes a user-specified set of residues and generates histogram plots showing the distribution of the chosen dihedral angle.

Use dihedral angle analysis to quantify backbone conformations in the selectivity filter region using the following Python command:

python plot_dihedral_angles_AF_models.py \
-p alphafold_models/*rank*.pdb \
--residues 624,625,626,627,628 \
--angle_type phi

This allows identification of models occupying distinct conformational states at key residues 

     5. Select the final model. This can be done by picking the model with the highest average pLDDT within the selected cluster while ensuring that the model displays structural deformations likely to impair ion conduction (such as backbone carbonyl flipping or constriction)

     6. Perform final validation by visual inspection in ChimeraX: 

• Reject models with steric clashes (e.g., overlapping backbone or side chains) using ChimeraX’s built in clashes locator: Tools -> Structure Analysis -> Clashes.
• Confirm consistency of the conformation across all four subunits

IV. ColabFold configuration

Use ColabFold v1.5.2 AlphaFold2 notebook:
https://colab.research.google.com/github/sokrypton/ColabFold/blob/v1.5.2/AlphaFold2.ipynb

Modify the following settings (everything else set to default):

max_msa = 256:512

num_seeds = 20 

use_dropout = True

num_recycles = 20

recycle_early_stop_tolerance = 0.5

Input:

• The hERG construct used corresponds to residues W398–R863 (resembling the hERG 1b isoform). Format the sequence as a homotetramer by concatenating the sequence four times (to model four subunits), separated by colons (“:”): SEQUENCE:SEQUENCE:SEQUENCE:SEQUENCE
• Below is the single-subunit sequence used (W398–R863):

WTILHYSPFKAVWDWLILLLVIYTAVFTPYSAAFLLKETEEGPPATECGYACQPLAVVDLIVDIMFIVDILINFRTTYVNANEEVVSHPGRIAVHYFKGWFLIDMVAAIPFDLLIFGSGSEELIGLLKTARLLRLVRVARKLDRYSEYGAAVLFLLMCTFALIAHWLACIWYAIGNMEQPHMDSRIGWLHNLGDQIGKPYNSSGLGGPSIKDKYVTALYFTFSSLTSVGFGNVSPNTNSEKIFSICVMLIGSLMYASIFGNVSAIIQRLYSGTARYHTQMLRVREFIRFHQIPNPLRQRLEEYFQHAWSYTNGIDMNAVLKGFPECLQADICLHLNRSLLQHCKPFRGATKGCLRALAMKFKTTHAPPGDTLVHAGDLLTALYFISRGSIEILRGDVVVAILGKNDIFGEPLNLYARPGKSNGDVRALTYCDLHKIHRDDLLEVLDMYPEFSDHFWSSLEITFNLR:WTILHYSPFKAVWDWLILLLVIYTAVFTPYSAAFLLKETEEGPPATECGYACQPLAVVDLIVDIMFIVDILINFRTTYVNANEEVVSHPGRIAVHYFKGWFLIDMVAAIPFDLLIFGSGSEELIGLLKTARLLRLVRVARKLDRYSEYGAAVLFLLMCTFALIAHWLACIWYAIGNMEQPHMDSRIGWLHNLGDQIGKPYNSSGLGGPSIKDKYVTALYFTFSSLTSVGFGNVSPNTNSEKIFSICVMLIGSLMYASIFGNVSAIIQRLYSGTARYHTQMLRVREFIRFHQIPNPLRQRLEEYFQHAWSYTNGIDMNAVLKGFPECLQADICLHLNRSLLQHCKPFRGATKGCLRALAMKFKTTHAPPGDTLVHAGDLLTALYFISRGSIEILRGDVVVAILGKNDIFGEPLNLYARPGKSNGDVRALTYCDLHKIHRDDLLEVLDMYPEFSDHFWSSLEITFNLR:WTILHYSPFKAVWDWLILLLVIYTAVFTPYSAAFLLKETEEGPPATECGYACQPLAVVDLIVDIMFIVDILINFRTTYVNANEEVVSHPGRIAVHYFKGWFLIDMVAAIPFDLLIFGSGSEELIGLLKTARLLRLVRVARKLDRYSEYGAAVLFLLMCTFALIAHWLACIWYAIGNMEQPHMDSRIGWLHNLGDQIGKPYNSSGLGGPSIKDKYVTALYFTFSSLTSVGFGNVSPNTNSEKIFSICVMLIGSLMYASIFGNVSAIIQRLYSGTARYHTQMLRVREFIRFHQIPNPLRQRLEEYFQHAWSYTNGIDMNAVLKGFPECLQADICLHLNRSLLQHCKPFRGATKGCLRALAMKFKTTHAPPGDTLVHAGDLLTALYFISRGSIEILRGDVVVAILGKNDIFGEPLNLYARPGKSNGDVRALTYCDLHKIHRDDLLEVLDMYPEFSDHFWSSLEITFNLR:WTILHYSPFKAVWDWLILLLVIYTAVFTPYSAAFLLKETEEGPPATECGYACQPLAVVDLIVDIMFIVDILINFRTTYVNANEEVVSHPGRIAVHYFKGWFLIDMVAAIPFDLLIFGSGSEELIGLLKTARLLRLVRVARKLDRYSEYGAAVLFLLMCTFALIAHWLACIWYAIGNMEQPHMDSRIGWLHNLGDQIGKPYNSSGLGGPSIKDKYVTALYFTFSSLTSVGFGNVSPNTNSEKIFSICVMLIGSLMYASIFGNVSAIIQRLYSGTARYHTQMLRVREFIRFHQIPNPLRQRLEEYFQHAWSYTNGIDMNAVLKGFPECLQADICLHLNRSLLQHCKPFRGATKGCLRALAMKFKTTHAPPGDTLVHAGDLLTALYFISRGSIEILRGDVVVAILGKNDIFGEPLNLYARPGKSNGDVRALTYCDLHKIHRDDLLEVLDMYPEFSDHFWSSLEITFNLR

• Upload template CIF file. Ensure filename contains 4-letter PDB code (e.g., “5va2.cif”)

General notes and troubleshooting

After selecting the final model, it is recommended to perform additional structural and functional validation to ensure that the model is physically reasonable and consistent with the intended conformational state.

a) Structural quality assessment

1. Open the selected model in ChimeraX. 
2. Identify steric clashes using Tools -> Structure Analysis -> Clashes 
3. Inspect: 

• Backbone continuity 
• Side-chain packing 
• Inter-subunit interfaces 
• Selectivity filter geometry 

Reject models that exhibit:

• Severe atomic overlaps 
• Distorted backbone geometry 
• Broken secondary structure

b) Structural relaxation

Structural relaxation can be used to resolve minor clashes and improve local geometry.

Available options include:

• ColabFold Amber/OpenMM relaxation notebook
  https://colab.research.google.com/github/sokrypton/ColabFold/blob/main/beta/relax_amber.ipynb
• ROSIE web server (Rosetta-based tools)
  https://rosie.graylab.jhu.edu/

Relaxation should preserve the overall conformation while improving local structural quality.

c) Functional validation

To confirm that the selected model represents the intended state, perform downstream analyses, such as: 

Molecular dynamics simulations:

• Assess structural stability over simulation time (tens to hundreds of ns) 
• Evaluate conformational flexibility 
• Test ion conduction properties 
• For hERG: open-state models should support K⁺ permeation, while inactivated-state models should disrupt conduction at the selectivity filter 

Drug docking studies:

• Evaluate state-dependent binding 
• Identify interactions with key residues (e.g., Y652 and F656 on the pore-lining S6 helices) 
• Compare drug binding poses across conformational states  
• Docking can be performed using Rosetta-based tools (e.g., Rosetta GALigandDock, RosettaLigand, or online webserver like ROSIE) or other molecular docking platforms (e.g., AutoDock Vina - https://vina.scripps.edu/)

d) Interpretation

Final models should satisfy both:

1. Structural criteria 
   • No major steric clashes 
   • Stable protein fold after relaxation 
2. Functional criteria 
   • Expected ion conduction behavior 
   • Consistent drug-binding characteristics 

Models that meet both criteria are suitable for further mechanistic or pharmacological studies. Model selection should not rely solely on pLDDT or clustering. Final validation requires structural inspection + physical plausibility + functional behavior.

e) Troubleshooting

If ColabFold does not produce models with the desired conformational state, the most common cause is template design. The placement, length, and identity of template segments strongly influence the predicted structure. Adjust the template by modifying which protein regions are included, especially around functionally critical domains such as the voltage-sensing domain or selectivity filter. Use ChimeraX Matchmaker tool to align homologous domains from structures with known conformational states, ensuring that the inserted template fragment is correctly positioned relative to the rest of the protein.

Another important factor is template boundary selection. Including too large template can overconstrain the model and bias it toward the original structure, while including too little may fail to guide the prediction. Test alternative boundaries for helices, loops, or pore regions to improve sampling of the desired state.

If insufficient conformational diversity is observed, adjust ColabFold settings. Increasing stochasticity through dropout, reducing MSA depth, or using fewer structural constraints can promote exploration of alternative conformations. Conversely, if results are too variable, increasing template guidance or sampling depth (e.g., more seeds) can help stabilize predictions.

Finally, consider sequence length and construct design. Removing poorly resolved or flexible regions (e.g., large cytosolic domains or termini) can improve prediction quality and allow better sampling of the regions of interest. Iterative workflows, where an initial prediction is used to seed a second round of template-guided modeling, are often effective for capturing alternative conformational states.

References

1.             M. Mirdita, et al., ColabFold: making protein folding accessible to all. Nat. Methods 19, 679–682 (2022).

2.             J. Jumper, et al., Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).

3.             E. F. Pettersen, et al., UCSF ChimeraX: Structure visualization for researchers, educators, and developers. Protein Sci. Publ. Protein Soc. 30, 70–82 (2021).

4.             C. A. Rohl, C. E. M. Strauss, K. M. S. Misura, D. Baker, “Protein Structure Prediction Using Rosetta” in Methods in Enzymology, Numerical Computer Methods, Part D., (Academic Press, 2004), pp. 66–93.

5.             K. Ngo, P.-C. Yang, V. Yarov-Yarovoy, C. E. Clancy, I. Vorobyov, Harnessing AlphaFold to reveal hERG channel conformational state secrets. eLife 13, RP104901 (2025).