Compressing biology

Table of Contents

1. Project overview
2. Folder structure
3. Key scripts and notebooks
4. Steps

Project overview

This repository contains the code and some resources used for experiments involving biological image compression and analysis related to the paper:

Compressing Biology: Evaluating the Stable Diffusion VAE for Phenotypic Drug Discovery

Accepted at the 3rd Workshop on Imageomics: Discovering Biological Knowledge from Images Using AI, held at the 39th Conference on Neural Information Processing Systems (NeurIPS 2025).

https://arxiv.org/abs/2510.19887

Folder structure

• cpg0000-jump-pilot/

Contains metadata and scripts required to run experiments with the CPJUMP1 dataset. Includes scripts for downloading and preprocessing the images.

• lsun/

Includes partition_names.txt for parallelizing inference on the LSUN dataset.

• dataset/

Contains files for defining Dataset objects used in inference.

• models/

Holds definitions of all pre-trained models and their associated inference functions.

Key scripts and notebooks

• inference_cpg0000.py

Inference of pre-trained models on the CPJUMP1 dataset.

• inference_lsun.py

Inference of pre-trained models on the LSUN dataset.

• notebook.sh

Sets up the environment for running Jupyter notebooks on a cluster.

• phenotypic_activity.ipynb

Computes the phenotypic activity (FR) on the CPJUMP1 dataset.

• plots.ipynb

Generates plots for evaluation metrics including MAE, SSIM, EMD, KLD, and FID.

• requirements.txt

List of all libraries installed in the virtual environment using pip3 freeze.

Steps

1. Download CPJUMP1 images

Run cpg0000-jump-pilot/download.sh to download the dataset.

2. Preprocess images

Use processing.sh to preprocess the downloaded images.

3. Download LSUN dataset

Follow instructions or use provided scripts to obtain the LSUN dataset.

4. Run inference

Execute inference_cpg0000.sh and inference_lsun.sh to perform inference on the respective datasets.

5. Analyze results

• Use phenotypic_activity.ipynb to compute phenotypic activity.

• Use plots.ipynb to visualize evaluation metrics.