A Genetic Algorithm for Navigating
Synthesizable Molecular Spaces

This repository contains the SynGA algorithm and code to reproduce the paper: A Genetic Algorithm for Navigating Synthesizable Molecular Spaces (ICLR 2026).

• Paper: https://openreview.net/forum?id=OvMtGGaFUT
• ArXiv: https://arxiv.org/abs/2509.20719

---

Installation

To install requirements:

conda env create -f environment.yml 

We include our precise conda environment in environment_full.yml too. After activating your conda environment, install UniDock tools (v1.1.2) with:

pip install git+https://github.com/dptech-corp/Uni-Dock.git@1.1.2#subdirectory=unidock_tools

Data

To reproduce our main results, you need:

1. A building block set. In our work, we start with ChemProjector's block set which was obtained from the Enamine Building Blocks catalog. Follow their download instructions and extract the blocks from their matrix.pkl file into a text file of SMILES strings.
2. A reaction set (data/libs/hb.txt is used by default).

First, initialize a chemspace synthesis library by processing the downloaded blocks and reactions:

python -m data.libs.setup --name=chemspace --blocks=[BLOCK_PATH] --num_workers=16

This creates a new directory data/libs/chemspace with the processed blocks and some caches. The name chemspace is used as a default argument in subsequent commands.

Usage

Analog Search

For analog search, we include the datasets used in our experiments under data/test:

chembl.txt:         Random 1k ChEMBL molecules
chembl_small.txt:   Random 100 ChEMBL molecules 
designs.csv:        Filtered structure-based and goal-directed designs for LIT-PCBA and GuacaMol

Download the checkpoint and run analog search:

python -m src.analog \
   --dataset=[DATASET_PATH] \
   --seed=0 --num_workers=100 \
   --logger.project=[WANDB_PROJECT] --log_analogs=10 \
   --optimizer=SynthesisGA --optimizer.founder_size=5000 --budget=10000 \
   --objective='{"count": True, "murcko": False}' \
   --bbfilter.checkpoint=[CHECKPOINT_PATH]

The "count" field in the objective argument toggles whether to use count or bit fingerprints when computing the fitness function, and the "murcko" field controls whether to use 0.9Morgan + 0.1Murcko as the fitness function (as opposed to just Morgan). The 10 best analogs will be uploaded to WandB.

Practical Molecular Optimization

SynGA. Run SynGA on a task from PMO:

python -m src.optimize \
   --seed=0 --trials=5 --num_workers=5 \
   --optimizer=SynthesisGA \
   --logger.project=[WANDB_PROJECT] \
   --objective=[TASK] 

Under the hood, this runs 5 trials sequentially by incrementing the seed each time. In practice, one can run the same command 5 times with trials=1 and different seeds.

SynGBO. Run SynGBO on a task from PMO:

python -m src.optimize \
   --seed=0 --trials=5 --num_workers=0 \
   --optimizer=SynthesisGBO \
   --optimizer.device=cuda --optimizer.synga_num_workers=20 \
   --logger.project=[WANDB_PROJECT] \
   --objective=[TASK]

Here, num_workers are the workers used to parallelize oracle calls, whereas synga_num_workers are the workers used to parallelize the SynGA inner loop.

LIT-PCBA Docking

Download the LIT-PCBA dataset from RxnFlow into the data/LIT-PCBA folder. As a preprocessing stop, we need to convert the receptor PDB files into PDBQT files by running the script at scripts/litpcba/0_setup.py.

SynGA. Run SynGA on a receptor for 1 trial:

python -m src.optimize_dock \
   --seed=0 --num_workers=50 --budget=16000 \
   --optimizer=SynthesisGA --optimizer.maxatoms=50 \
   --optimizer.founder_size=1000 --optimizer.population_size=5000 --optimizer.offspring_size=100 \
   --logger.project=[WANDB_PROJECT] --log_every_n_calls=200 \
   --receptor=[RECEPTOR]

SynGBO. Run SynGBO on a receptor for 1 trial:

python -m src.optimize_dock \
   --seed=0 --num_workers=20 --budget=16000 \
   --optimizer=SynthesisGBO --optimizer.maxatoms=50 \
   --optimizer.device=cuda --optimizer.synga_num_workers=20 \
   --optimizer.initial_size=20 --optimizer.propose_size=20 \
   --logger.project=[WANDB_PROJECT] --log_every_n_calls=200 \
   --receptor=[RECEPTOR]

Scripts

The scripts directory contains one-off scripts for data processing, metric collection, and visualization. It is organized into the following subfolders

bbfilter          Training the MLP block filter
project_designs   Projecting molecules from generative models
nam               Ablating the NAM 
pmo               PMO benchmark 
litpcba           LIT-PCBA docking benchmark