ChemProjector

:dart: Projecting Molecules into Synthesizable Chemical Spaces (ICML 2024)

[Paper]

⚠️ Update

ChemProjector and its derivatives (e.g. SynFormer) have been deprecated. Please refer to our latest model PrexSyn: Efficient and Programmable Exploration of Synthesizable Chemical Space:

• Code: https://github.com/luost26/prexsyn
• Paper: https://arxiv.org/abs/2512.00384
• Documentation: https://prexsyn.readthedocs.io/

Install

Clone Repo

Please clone the repository with the --recurse-submodules flag to include the third-party submodules.

git clone --recurse-submodules https://github.com/luost26/ChemProjector.git

Environment

# Install conda environment
conda env create -f env.yml -n chemprojector
conda activate chemprojector

# Install ChemProjector package
pip install -e .

The default CUDA version is 11.8. If you have to use a different version, please modify the env.yml file accordingly.

Building Block Data

We provide preprocessed building block data. You can download it from here and put it in the data directory.

However, the data is derived from Enamine's building block catalog, which are available only upon request. Therefore, you should first request the data from Enamine here and download the US Stock catalog into the data directory. Then run the following script which will check whether you have a copy of the Enamine's catalog and unarchive the preprocessed data for you:

python unarchive_wizard.py

You may also process the building block data by yourself. Please refer to the scripts/preprocess_data directory for more details.

Trained Weights

You can download the trained weights from here and put them in the data/trained_weights directory.

Summary of Downloads

Item	Link	Destination
Preprocessed building blocks and reaction templates (password protected)	https://drive.google.com/file/d/1scui0RZ8oeroDAafnw4jgTi3yKtXxXpe/view?usp=drive_link	data/processed.zip
Enamine building block US stock catalog (to unzip the password protected preprocessed files)	https://enamine.net/building-blocks/building-blocks-catalog	data/Enamine_*.sdf
Trained weights	https://drive.google.com/drive/folders/1T9f9MsEAsAjPV8GR0pXimHKCvq97SIzs?usp=drive_link	data/trained_weights/original_default.ckpt, data/trained_weights/original_split.ckpt

Usage

Project Molecules

You can create a list of SMILES strings in CSV format (example: data/example.csv) and run the following command to project them into the synthesizable chemical space.

python sample.py \
    --model-path data/trained_weights/original_default.ckpt \
    --input data/example.csv \
    --output results/example.csv

Model Evaluation

Bottom-Up Synthesis Planning (Chemical Space Projection)

Test set	Command
Test split	./scripts/synthesis_planning_test_split.sh
ChEMBL	./scripts/synthesis_planning_chembl.sh
Enamine REAL	./scripts/synthesis_planning_enamine.sh

Structure-Based Drug Design

Please refer to the scripts/sbdd directory for details.

Goal-Directed Generation

Please refer to the scripts/goal_directed directory for details.

Train

python train.py ./configs/original_default.yml

Reference

@inproceedings{luo2024chemprojector,
  title={Projecting Molecules into Synthesizable Chemical Spaces},
  author={Shitong Luo and Wenhao Gao and Zuofan Wu and Jian Peng and Connor W. Coley and Jianzhu Ma},
  booktitle={Forty-first International Conference on Machine Learning},
  year={2024}
}