Mol2Image: Improved Conditional Flow Models for Molecule to Image Synthesis

This is the accompanying code for the paper, "Mol2Image: Improved Conditional Flow Models for Molecule to Image Synthesis" (CVF).

Dataset

We use the subset of pre-processed images from the "Cell Painting Assay Dataset" provided by Hofmarcher et al., (2019). Their dataset can be directly accessed here: https://ml.jku.at/software/cellpainting/dataset. Download and unzip the images (in .npz format), and place them in a directory called data/images.

For the full cell painting dataset, see https://github.com/gigascience/paper-bray2017.

Dependencies

Python dependencies can be installed via conda from the environment.yml file:

conda env create -f environment.yml
conda activate mol2image

In addition to these dependencies, we also rely on an older version of chemprop (https://github.com/chemprop/chemprop). Clone this repository, checkout the required version, and install it as a package in the mol2image conda environment:

cd /path/to/chemprop
git checkout f9581c59483310b2eddae278b3507980c54249fa
pip install -e .

Usage

Generating Images

Download the pretrained model weights from Google Drive and place them in a directory called pretrained. To generate images corresponding to the molecules that were observed during training, run:

python generate.py --save-dir /path/to/results --val-metafile data/metadata/datasplit_gen_test_easy.csv

To generate images corresponding to the molecules that were held-out during training, run:

python generate.py --save-dir /path/to/results --val-metafile data/metadata/datasplit_gen_test_hard.csv

To generate images corresponding to the selected molecules for CellProfiler analysis, run:

python generate.py --save-dir /path/to/results --val-metafile data/metadata/datasplit_gen_test_easy_ext10.csv
python generate.py --save-dir /path/to/results --val-metafile data/metadata/datasplit_gen_test_hard_ext10_unique.csv

The generated and corresponding real images for the molecules will be saved to /path/to/results/images.

CellProfiler Evaluation

To evaluate the generated images using CellProfiler, follow the installation instructions here: https://github.com/CellProfiler/CellProfiler. Convert the generated .npz images to .png images (separate image for each channel) by running:

python convert_npz_to_png.py /path/to/results/images /path/to/results/png

Launch the CellProfiler GUI and open the pipeline from the file mol2image.cpproj. Then add the images in the directory /path/to/results/png to the pipeline and run.