Holographic-VAE

Installation

Create the hvae conda environment by running the following

conda env create -f env.yml

to install the necessary dependencies. Then run

pip install .

to install the holographic_vae package.

If you're going to make edits to the holographic_vae package, run

pip install -e .

so you can test your changes.

Overview

Holographic-(V)AE

Code containing building blocks for building expressive SO(3)-equivariant networks using spherical harmonics projections. The code is heavily based on the e3nn package. We decided to re-implement certain operations to work with Dictionary inputs (index by $\ell$) instead of flat Tensor inputs, as we observed speed gains during early development, as well as more flexibility to build custom normalization layers. The architecture of H-(V)AE is found in holographic_vae/models/h_vae.py.

Experiments

MNIST-on-the-Sphere

Code replicating (most of) our experiments on the MNIST-on-the-Sphere dataset is found in experiments/mnist. Crucially, this folder contains general functions and scripts for forward and inverse spherical fourier transforms, helpful to preprocess any spherical image dataset to be used by H-(V)AE.

Protein Neighborhoods

Code to use H-(V)AE on protein neighborhoods. We provide code to use our pre-trained models to extract embeddings from a series of PDBs, as well as code to train new models. The code is found in experiments/protein_neighborhoods.

Ligand Binding Affinity

Code to replicate our experiments on the ATOM3D Ligand Binding Affinity task, where our method achieves state-of-the-art performance, as shown in our paper.
Note: this necessitates having a trained H-(V)AE model, either provided by us, or user-trained. Checkout the README in experiments/protein_neighborhoods to learn how to download a pre-trained model, or train a new one with your data.

---

@article{PhysRevResearch.6.023006,
  title = {Holographic-(V)AE: An end-to-end SO(3)-equivariant (variational) autoencoder in Fourier space},
  author = {Visani, Gian Marco and Pun, Michael N. and Angaji, Arman and Nourmohammad, Armita},
  journal = {Phys. Rev. Res.},
  volume = {6},
  issue = {2},
  pages = {023006},
  numpages = {21},
  year = {2024},
  month = {Apr},
  publisher = {American Physical Society},
  doi = {10.1103/PhysRevResearch.6.023006},
  url = {https://link.aps.org/doi/10.1103/PhysRevResearch.6.023006}
}