SyntheMol: Generative AI for Drug Discovery

SyntheMol is a generative AI method for designing structurally novel and diverse drug candidates with predicted bioactivity that are easy to synthesize.

SyntheMol consists of either a reinforcement learning model (SyntheMol-RL) or a Monte Carlo tree search (SyntheMol-MCTS) to explore a combinatorial chemical space consisting of molecular building blocks and chemical reactions. In both cases, SyntheMol is guided by a bioactivity prediction AI model, such as a graph neural network or multilayer perceptron. SyntheMol uses two chemical spaces which jointly contain over 46 billion molecules:

Enamine REAL Space

• 137,656 molecular building blocks
• 70 chemical reactions
• 30,330,025,259 molecules

WuXi GalaXi

• 14,977 molecular building blocks
• 36 chemical reactions
• 16,146,071,436 molecules

Notably, SyntheMol can be easily adapted to use any set of building blocks and reactions.

SyntheMol-RL is described in the following paper, where we applied SyntheMol-RL to design novel antibiotic candidates for the Gram-positive bacterium Staphylococcus aureus.

Swanson, K., Liu, G., Catacutan, D. B., McLellan, S., Arnold, A., Tu, M., Brown, E. D., Zou, J., Stokes, J. M. SyntheMol-RL: a flexible reinforcement learning framework for designing novel and synthesizable antibiotics. bioRxiv, 2025.

SyntheMol-MCTS is described in the following paper, where we applied SyntheMol to design novel antibiotic candidates for the Gram-negative bacterium Acinetobacter baumannii.

Swanson, K., Liu, G., Catacutan, D. B., Arnold, A., Zou, J., Stokes, J. M. Generative AI for designing and validating easily synthesizable and structurally novel antibiotics. Nature Machine Intelligence, 2024.

Full details for reproducing the SyntheMol results in both papers are provided in the docs directory.

Table of contents

• Installation
• Combinatorial chemical space
• Bioactivity prediction model
  • Train model
  • Pre-compute building block scores
• Generate molecules
• Filter generated molecules
  • Novelty
  • Bioactivity
  • Diversity

Installation

SyntheMol can be installed in < 3 minutes on any operating system using pip (optionally within a conda environment). SyntheMol can be run on a standard laptop (e.g., 16 GB memory and 8-16 CPUs), although a GPU is useful for faster training and prediction of the underlying bioactivity prediction model (Chemprop).

Optionally, create a conda environment.

conda create -y -n synthemol python=3.12
conda activate synthemol

Install SyntheMol via pip.

pip install synthemol

Alternatively, clone the repo and install SyntheMol locally.

git clone https://github.com/swansonk14/SyntheMol.git
cd SyntheMol
pip install -e .

If there are version issues with the required packages, create a conda environment with specific working versions of the packages as follows.

SyntheMol-RL

pip install -r requirements_rl.txt
pip install -e .

SyntheMol-MCTS

pip install -r requirements_mcts.txt
pip install -e .

Note: If you get the issue ImportError: libXrender.so.1: cannot open shared object file: No such file or directory, run conda install -c conda-forge xorg-libxrender.

Combinatorial chemical space

An alternate combinatorial chemical space can optionally be used by replacing the building blocks and chemical reactions as follows.

Building blocks: Create a building blocks file similar to synthemol/resources/real/building_blocks.csv with a custom file containing the building blocks. The file should be a CSV file with a header row and two columns: smiles and reagent_id. The smiles column should contain the SMILES string for each building block, and the reagent_id column should contain a unique ID for each building block.

Chemical reactions: In SyntheMol/reactions/custom.py, set CUSTOM_REACTIONS to a list of Reaction objects similar to the REAL_REACTIONS list in SyntheMol/reactions/real.py. Then specify --chemical_spaces custom when running SyntheMol.

Bioactivity prediction model

SyntheMol requires a bioactivity prediction model to guide its generative process. SyntheMol is designed to use one of these types of models:

1. Chemprop: a message passing neural network from https://github.com/chemprop/chemprop
2. Chemprop-RDKit: Chemprop augmented with 200 RDKit molecular features
3. MLP-RDKit: a feed-forward neural network using 200 RDKit molecular features
4. Random forest: a scikit-learn random forest model trained on 200 RDKit molecular features (SyntheMol-MCTS only)

Train model

All model types can be trained using Chemprop, which is installed along with SyntheMol. All model types can be trained on either regression or binary classification bioactivities. Full details are provided in the Chemprop README. Below is an example for training a Chemprop model on a binary classification task. By default, training is done on a GPU (if available).

Data file

# data/data.csv
smiles,activity
Br.CC(Cc1ccc(O)cc1)NCC(O)c1cc(O)cc(O)c1,0
CC[Hg]Sc1ccccc1C(=O)[O-].[Na+],1
O=C(O)CCc1ccc(NCc2cccc(Oc3ccccc3)c2)cc1,0
...

Train Chemprop

chemprop_train \
    --data_path data/data.csv \
    --dataset_type classification \
    --save_dir models/chemprop

Pre-compute building block scores

After training, use the model to pre-compute scores of building blocks to accelerate the SyntheMol generation process. Below is an example using the trained Chemprop model. By default, prediction is done on a GPU (if available).

chemprop_predict \
    --test_path "$(python -c 'import synthemol; print(str(synthemol.constants.BUILDING_BLOCKS_PATH))')" \
    --preds_path models/chemprop/building_blocks.csv \
    --checkpoint_dir models/chemprop

Generate molecules

SyntheMol uses the bioactivity prediction model a generative model (RL or MCTS) to generate molecules. SyntheMol-RL can be powered either by a Chemprop-based RL model (RL-Chemprop) or an MLP-based RL model (RL-MLP). Below is an example using SyntheMol-RL (RL-Chemprop version) to generate molecules with a trained Chemprop model for 10,000 rollouts using the Enamine REAL Space (the default).

synthemol \
    --score_model_paths models/chemprop \
    --score_types chemprop \
    --chemical_spaces real wuxi \
    --building_blocks_score_columns activity \
    --save_dir generations/chemprop \
    --n_rollout 10000 \
    --search_type rl \
    --rl_model_type chemprop \
    --rl_model_fingerprint_type rdkit \
    --rl_model_paths models/chemprop/fold_0/model_0/model.pt \
    --rl_prediction_types classification

Note: The building_blocks_score_columns must match the column name in the building blocks file that contains the building block scores. When using chemprop_train and chemprop_predict, the column name will be the same as the column that contains target activity/property values in the training data file (e.g., activity).

For more complex examples with multiparameter optimization using both the RL-Chemprop and RL-MLP versions of SyntheMol-RL, please see docs/rl/antibiotics.md.

Filter generated molecules

Optionally, the generated molecules can be filtered for structural novelty, predicted bioactivity, and structural diversity. These filtering steps use chemfunc, which is installed along with SyntheMol. Below is an example for filtering the generated molecules.

Novelty

Filter for novelty by comparing the generated molecules to a set of active molecules (hits) from the training set or literature and removing similar generated molecules.

Hits file

# data/hits.csv
smiles,activity
CC[Hg]Sc1ccccc1C(=O)[O-].[Na+],1
O=C(NNc1ccccc1)c1ccncc1,1
...

Compute Tversky similarity between generated molecules and hits.

chemfunc nearest_neighbor \
    --data_path generations/chemprop/molecules.csv \
    --reference_data_path data/hits.csv \
    --reference_name hits \
    --metric tversky

Filter by similarity, only keeping molecules with a nearest neighbor similarity to hits of at most 0.5.

chemfunc filter_molecules \
    --data_path generations/chemprop/molecules.csv \
    --save_path generations/chemprop/molecules_novel.csv \
    --filter_column hits_tversky_nearest_neighbor_similarity \
    --max_value 0.5

Bioactivity

Filter for predicted bioactivity by keeping the molecules with the top 20% highest predicted bioactivity.

chemfunc filter_molecules \
    --data_path generations/chemprop/molecules_novel.csv \
    --save_path generations/chemprop/molecules_novel_bioactive.csv \
    --filter_column score \
    --top_proportion 0.2

Diversity

Filter for diversity by clustering molecules based on their Morgan fingerprint and only keeping the top scoring molecule from each cluster.

Cluster molecules into 50 clusters.

chemfunc cluster_molecules \
    --data_path generations/chemprop/molecules_novel_bioactive.csv \
    --num_clusters 50

Select the top scoring molecule from each cluster.

chemfunc select_from_clusters \
    --data_path generations/chemprop/molecules_novel_bioactive.csv \
    --save_path generations/chemprop/molecules_novel_bioactive_diverse.csv \
    --value_column score