CryoStructNet

Deep 3D Protein Reconstruction from Unaligned Cryo-EM Projections

---

CryoStructNet is a deep learning framework for reconstructing high fidelity 3D protein density volumes directly from unaligned 2D Cryo-Electron Microscopy (Cryo-EM) projections. Inspired by Neural Radiance Fields (NeRF), our system learns to predict projection poses, fuse multi-view image features, and generate full 3D structures all without traditional alignment or averaging steps.

Highlights

• View-conditioned volume prediction using NeRF-style positional encodings
• Pose estimation network to infer projection orientation from raw images
• Uncertainty-aware volume output with confidence maps
• Trained on real Cryo-EM datasets (EMPIAR/EMDB) and synthetic simulations

Project Structure

CryoStructNet/
├── data/                  # Raw + processed Cryo-EM data
├── models/                # Pose estimator + volume decoder
├── training/              # Training scripts, loss functions
├── utils/                 # Metrics, visualization, I/O
├── notebooks/             # Demo + experiment logs
├── environment.yaml       # Conda environment specification
└── README.md              # Project overview

Setup

Using Conda (recommended)

git clone https://github.com/yourusername/CryoStructNet.git
cd CryoStructNet
conda env create -f environment.yaml
conda activate cryostructnet

Note: If you plan to use simulated projection data, you may need to manually install ASTRA Toolbox for tomographic projection support.

Data

We use datasets from:

• EMPIAR – raw Cryo-EM images (e.g. EMPIAR-10061)
• EMDB – corresponding 3D density volumes
• Simulated projections from known 3D volumes using ASTRA Toolbox

Training

python training/train.py --config training/config.yaml

Output includes:

• Reconstructed 3D volume
• Voxel-wise confidence map
• Pose predictions per projection

Evaluation

• Voxel-wise MSE
• SSIM between prediction and ground truth
• 3D correlation coefficient
• Fourier Shell Correlation (FSC)
• Visual inspection using ChimeraX or PyMOL

Results

I will put them up once I am done

TODO

• Pose estimation module
• NeRF-style decoder
• Volume + confidence head
• Weak supervision on real Cryo-EM datasets
• Paper-ready visualizations and metrics

Citation (Planned)

@article{rohith2025cryostructnet,
  title={CryoStructNet: Deep 3D Protein Reconstruction from Unaligned Cryo-EM Projections},
  author={Parahmesh, Rohith},
  journal={bioRxiv (preprint)},
  year={2025}
}

License

MIT License.

Acknowledgements

• CryoDRGN (UCSF)
• EMPIAR / EMDB datasets
• NeRF authors
• ASTRA Toolbox (tomographic projection simulation)