# Introduction

CryoDRGN-AI is an open-source software package for *ab initio* heterogeneous reconstruction in cryo-EM and cryo-ET.

## Latest News!

Note: We are currently merging the standalone cryoDRGN-AI [github repository](https://github.com/ml-struct-bio/drgnai) into the [cryodrgn](https://github.com/ml-struct-bio/cryodrgn) software package. CryoDRGN-AI will be available in the [cryodrgn](https://github.com/ml-struct-bio/cryodrgn) software's next major release.

* June 2025: CryoDRGN-AI is now published in [Nature Methods](https://doi.org/10.1038/s41592-025-02720-4)!
* June 2025: Version 0.3.2-beta release with usability improvements. See the [github](https://github.com/ml-struct-bio/drgnai) for more updates.
* April 2025: Updated name from DRGN-AI to cryoDRGN-AI in our [preprint](https://www.biorxiv.org/content/10.1101/2024.05.30.596729v3).
* Dec 2024: Ab initio reconstruction of cryo-ET subtomograms is now described in our preprint.

## Inputs/Outputs

CryoDRGN-AI takes a set of particle images with estimated CTF parameters and reconstructs a distribution of 3D density maps, modeling the conformational space of the biomolecular complex. The output includes a set of low-dimensional *latent embeddings* (one for each particle) and a neural network that associates this latent representation to 3D density maps.

<figure><img src="https://3886634198-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FmQzLIojwQvJKwWQnBP0u%2Fuploads%2FxH5eTknL9jnrokvZ7y0P%2Finputs_outputs_gitbook.png?alt=media&#x26;token=cce634d3-476c-4c29-a8d5-c534612d7a21" alt=""><figcaption></figcaption></figure>

## Method Overview

CryoDRGN-AI aims to jointly infer the pose of each particle in the dataset while reconstructing the structural variability of the molecule. It combines the flexibility of implicit neural representations with a fast and scalable pose estimation strategy. Poses are optimized using a hierarchical search strategy and subsequently refined through stochastic gradient descent. Conformations are independently optimized in an autodecoding fashion. The whole model (neural field + conformations + poses) is optimized by gradient descent in order to minimize the difference between reconstructed images and real observations.

<figure><img src="https://3886634198-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FmQzLIojwQvJKwWQnBP0u%2Fuploads%2Fs7NZcTHS9PpfHchEwMSa%2Fmethod_overview_2_short.png?alt=media&#x26;token=82de2830-1363-4bbe-86a4-d0650b531fb3" alt=""><figcaption></figcaption></figure>

## Tutorial Outline

The user guide is split into 3 parts:

1. **Installation**: details on how to install cryoDRGN-AI.
2. **Running a job**: tutorial explaining how to run cryoDRGN-AI, monitor your job and analyze the outputs.
3. **Full API documentation**: details on all the cryoDRGN-AI parameters.

## Citing this Work

```
@article{cryodrgnai,
  title = "CryoDRGN-AI: neural ab initio reconstruction of challenging cryo-EM and cryo-ET datasets",
  author = "Levy, Axel and Raghu, Rishwanth and Feathers, Ryan and Grzadkowski, Michal and Poitevin, Frederic and
            Johnston, Jake D, Vallese, Francesca and Clarke, Oliver B and Wetzstein, Gordon and Zhong, Ellen D",
  journal = "Nature Methods",
  doi = 10.1038/s41592-025-02720-4,
  url = nature.com/articles/s41592-025-02720-4,
  month = jun,
  year = 2025,
}
```