CryoPhold: CryoEM meets AlphaFold and molecular simulation to reveal protein dynamics

CryoPhold is a computational pipeline that performs Bayesian reweighting of AlphaFold predicted structural ensembles against experimental cryo-EM data. The pipeline determines optimal weights for each AlphaFold structure and identifies the minimal subset that best reproduces experimental density maps, enabling the elucidation of protein dynamics through the integration of AI-predicted structures and experimental data.

🧬 Scientific Background

CryoPhold implements a maximum entropy approach (BioEN algorithm) to:

1. Generate simulated density maps from each AlphaFold structure
2. Optimize structural weights to match experimental cryo-EM data
3. Posterior structural ensemble while preserving fit quality and conformational heterogeneity

🚀 Key Features

• Bayesian Ensemble Reweighting: Optimally combines AlphaFold structures using experimental cryo-EM constraints
• Structure Selection: Identifies minimal structural subsets that maintain fit quality
• Comprehensive Analysis: Provides correlation analysis, Fourier Shell Correlation, RMSF, and PCA
• Automated Pipeline: End-to-end workflow from AlphaFold generation to final analysis

📦 Installation

git clone git@github.com:strauchlab/cryoPhold.git
cd cryoModule
chmod +x setup.sh
./setup.sh
conda activate cryophold

Test Installation

python ./cryoModule/cryoPhold.py -h

Boom! You should be ready to go! 🎉

🔧 Usage Workflow

Step 1: Generate AlphaFold Structural Ensemble

Check if colabfold is installed in the cryophold enviornment using:

colabfold_batch -h

Generate a conformational ensemble using ColabFold with dropout to sample structural diversity:

colabfold_batch \
  --num-recycle 3 \
  --recycle-early-stop-tolerance 0.5 \
  --num-ensemble 5 \
  --model-type auto \
  --templates \
  --use-gpu-relax \
  --amber \
  --max-seq 8 \
  --max-extra-seq 16 \
  --num-seeds 16 \
  --use-dropout \
  --relax-max-iterations 200 \
  ./input/ \
  ./results/

Quick start: Use the provided script ./scripts/colabfold.sh

Step 2: Align Ensemble to Cryo-EM Map

Align the AlphaFold generated conformational ensemble to your experimental cryo-EM map using Situs:

• Use the Situs standalone tool: https://situs.biomachina.org
• Run the alignment script: ./scripts/align.sh
• Important: Update the script with your specific:
  • Map resolution
  • Path to cryo-EM map
  • Number of processors

Alternative: If you have a corresponding PDB, align the AlphaFold ensemble using PyMOL or UCSF Chimera and save as combined.pdb. Transform the combined.pdb to a .xtc format using gmx, mdtraj or mdanalysis

Step 3: Align and Combine Structures

Process the aligned structures:

# Refine aligned structures
./scripts/refine.sh

# Visually inspect alignment quality with cryo-EM map
# Combine all aligned PDBs into trajectory format
python ./scripts/combine-pdb.py --remove-hydrogens  # if needed

This generates combined.xtc for the next step.

We have already provided AF2 and Boltz generated structural ensemble as examples.

Step 4: Bayesian Reweighting

Run the main CryoPhold pipeline:

python cryoPhold.py \
  --path 'path_to_directory_containing_combined.xtc' \
  --threshold 2.95 \
  --resolution 3.05 \
  --mask sim

Parameters:

• --path: Directory containing AlphaFold output (combined.xtc, prot-masses.pdb, reference.map)
• --threshold: Cryo-EM map density threshold (default: 2.95)
• --resolution: Map resolution in Å (default: 3.05)
• --mask: Masking strategy - 'exp' (experimental) or 'sim' (simulated, default)

📊 Output Files

Primary Results

File	Description
map_posterior.mrc	Bayesian reweighted density map
map_prior.mrc	Uniform weighted (prior) map
weights.dat	Optimal weights for each AlphaFold structure
best.pdb	Best single structure with RMSF as B-factors
statistics.dat	Comprehensive analysis statistics

Analysis Plots

File	Description
lcurve.svg	L-curve showing entropy vs χ² trade-off
weights.svg	Structure weight distribution
FSC.svg	Fourier Shell Correlation analysis
RMSF.svg	Root Mean Square Fluctuation comparison
PCA.svg	Principal Component Analysis

Iterative Refinement (iter/ directory)

The pipeline automatically identifies the minimal structural subset:

• iter/best_frames.pdb - Final optimized structural ensemble
• iter/map_posterior_iter.mrc - Refined posterior map
• iter/statistics_iter.dat - Iteration-specific statistics
• Updated plots for each refinement iteration

Featurize CryoPhold generated PDB and ML Module

python ./scripts/dihedral.py -h

Use:

python dihedral.py iter/best_frames.pdb --angles phi psi chi1

This will create a folder dihedral_features

Go to mlmodule and install the software train the features using ML and it will create a free energy surface, residue hotspots and plumed.dat file. You can use the plumed.dat file to run Enhanced sampling calculations

Simulation Module

Now you want to use the best_frames.pdb generated by cryoPhold to launch new simulations.

Step 1: Split the PDBs

python ./scripts/split_pdb.py -h
usage: split_pdb.py [-h] -i INPUT [-o OUTDIR]

Split a multi‐MODEL PDB into one file per MODEL block.

options:
  -h, --help            show this help message and exit
  -i INPUT, --input INPUT
                        Path to input multi‐model PDB
  -o OUTDIR, --outdir OUTDIR
                        Directory to write individual PDBs

Step 2: Prepare the PDBs to launch simulations

Install Gromacs, AmberTools and acpype. Go to simmodule and run the process.sh script to prepare the PDB files.

You are ready to launch simulations from cryoPhold ensemble! 🎉 Try launching multiple independent metadynamics simulations from each structures using plumed.dat file! 🎉

Step 3: Featurize the MD data

python ./scripts/dihedral.py -h

Use:

python dihedral.py --directory ./trajectories  --angles phi psi chi1

Step 4: Train ML models on the MD data

Create a new conda enviornment using MDML package (https://github.com/svats73/mdml) or use the mlmodule to train the featurized data using ML model.

Boom! Build MSM, capture Hotspots, perform Clustering and find allosteric pockets! 🎉

🤝 Contributing

We welcome contributions! Please feel free to submit issues and pull requests.

📄 License

Please refer to the repository license for usage terms.

💜 Builders

Soumendranath Bhakat, Shray Vats, Andreas Mardt and Eva M. Strauch

🙏 Citation

@article {Bhakat2025.09.12.675912,
	author = {Bhakat, Soumendranath and Vats, Shray and Mardt, Andreas and Strauch, Eva M.},
	title = {CryoPhold: CryoEM meets AlphaFold and molecular simulation to reveal protein dynamics},
	elocation-id = {2025.09.12.675912},
	year = {2025},
	doi = {10.1101/2025.09.12.675912},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2025/09/13/2025.09.12.675912},
	eprint = {https://www.biorxiv.org/content/early/2025/09/13/2025.09.12.675912.full.pdf},
	journal = {bioRxiv}
}

Bridge the gap between AI-predicted structures and cryo-EM? Start your CryoPhold journey today! ✨