<h1 align="center">CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation</h1> <p align="center"> <a href="https://openreview.net/forum?id=FSlTEObdLl"><img src="https://img.shields.io/badge/OpenReview-ICML'25.16227-b31b1b.svg" alt="Paper"></a> <a href="https://icml.cc/media/PosterPDFs/ICML%202025/45885.png?t=1752232241.6172879"> <img src="https://img.shields.io/badge/Poster-grey?logo=airplayvideo&logoColor=white" alt="Poster"></a> <a href="./assets/ICML2025-CombiMOTS_Slides.pdf"> <img src="https://img.shields.io/badge/Slides-grey?&logo=MicrosoftPowerPoint&logoColor=white" alt="Slides"></a> </p> Official implementation of CombiMOTS for Fragment-based Monte Carlo Tree Search for Dual-Inhibitors Molecular Graph Generation.

Refer to Poster or Slides for a more in-depth overview of our work!

<p align="center"><img src="./assets/overview.png" width=80%></p> <p align="center">Project overview.</p>

Broader Applications

• We release a pretrained ensemble ChemProp ClinTox model in models/clintox that can be used for Toxicity Optimization as described in our original manuscript.
• We created another repository (<a href="https://github.com/Tibogoss/KinSel"> <img src="https://img.shields.io/badge/KinSel-grey?&logo=MicrosoftPowerPoint&logoColor=white" alt="KinSel"> </a>) using CombiMOTS for the Selective Molecular Generation (using CDK7 as the target). Our main manuscript also discusses motivation background and implementation details.

Baseline papers

Activity-aware fragments are obtained with Graph Information Bottleneck - Adapted from https://arxiv.org/abs/2310.00841

Our Pareto MCTS pipeline is adapted from SyntheMol https://www.nature.com/articles/s42256-024-00809-7

The 13 Enamine (https://enamine.net/) REAL Space and corresponding reactions are also provided by the work above.

To accelerate molecular docking simulation, we utilize QuickVina-GPU-2.1 from https://pubmed.ncbi.nlm.nih.gov/39320991/

Install Environment

Implementation was originally conducted with Python3.10 and CUDA11.7 on a single NVIDIA RTX A6000 GPU or CPU.

conda create -n combimots
conda activate combimots

conda install -c bioconda mgltools -y
conda install python=3.10 -y
conda install -c nvidia/label/cuda-11.7.0 cuda-nvcc -y
conda install -c nvidia cuda-opencl -y
conda install -c conda-forge ocl-icd-system -y

conda install -c conda-forge boost=1.77.0 boost-cpp=1.77.0 pdbfixer openbabel openmm -y

export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

####### Installation of QuickVina-GPU-2.1
# In combimots/pmcts/docking, install and compile QuickVina-GPU-2.1 following https://github.com/DeltaGroupNJUPT/Vina-GPU-2.1

# Once installed, modify the DOCKING_PATH_PREFIX in 
# [combimots/pmcts/docking/docking_utils.py @l.16]
# [6-precompute_docking_scores.py @l.16]

pip install torch==2.0.1+cu117 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch-scatter torch-sparse torch-cluster -f https://data.pyg.org/whl/torch-2.0.1+cu117.html
pip install torch-geometric==2.0.4

pip install -r requirements.txt
pip install -e combimots/. # setup combimots in-line command

Note to the user

The next sections describe all pre-processing steps (running scripts 0 to 8).

If you only want to run generation and evaluation, we provide processed data and model checkpoints. You may skip these steps and directly go to the generation section.

Pipeline

In /data you may place a .csv file containing:

• smiles
• {target1}_activity
• {target2}_activity

For demonstration, we provide data for the GSK3B-JNK3, EGFR-MET and PIK3CA-mTOR target pairs.

This data is curated from ExCAPE-DB v2 (https://jcheminf.biomedcentral.com/articles/10.1186/s13321-017-0203-5)

Train Chemprop (D-MPNN) Property Predictor

chemprop_train --data_path {data_path} \
--dataset_type classification \
--split_type cv \
--num_folds 10 \
--seed 42 \
--gpu 0 \
--save_dir models/${model_name}

# $ chemprop_train --data_path data/GSK3B_JNK3.csv --dataset_type classification --split_type cv --num_folds 10 --seed 42 --gpu 0 --save_dir models/gsk3b_jnk3

Data Preparation

Search Space Reduction

Fragment-based Graph Information Bottleneck

Process the .csv to a .pt:

python utils_fgib/data.py --csv_path $data/{YOUR_DATA.csv} --target ${activity}

# $ python utils_fgib/data.py --csv_path data/GSK3B_JNK3.csv --target gsk3b_activity
# $ python utils_fgib/data.py --csv_path data/GSK3B_JNK3.csv --target jnk3_activity

Train the modules:

python 1-train_fgib.py -g ${gpu_id} --target ${target_activity}

# $ python 1-train_fgib.py -g 0 --target gsk3b_activity
# $ python 1-train_fgib.py -g 0 --target jnk3_activity

Extract and save the building blocks for both targets:

python 2-get_frags.py -g ${gpu_id} -t ${target_activity} -m ${target_pt_path} -v ${frags_path}

# $ python 2-get_frags.py -g 0 -t gsk3b_activity -m ckpt/gsk3b_activity_10.pt -v data/gsk3b.txt
# $ python 2-get_frags.py -g 0 -t jnk3_activity -m ckpt/jnk3_activity_10.pt -v data/jnk3.txt

Clean and merge both targets' building blocks:

python 3-frags_to_blocks.py ${frags_path1} ${frags_path2} ${fragments_path}

# $ python 3-frags_to_blocks.py data/gsk3b.txt data/jnk3.txt data/fgib_frags.csv

Map FGIB fragments to Enamine's REAL Space Building Blocks

python 4-get_similar_blocks.py --custom_path ${fragments_path} \
--real_path combimots/pmcts/resources/real/building_blocks.csv \
--output_path models/${model_name}/${blocks_path} \
--threshold ${tanim_thresh} --batch_size ${bs}

# $ python 4-get_similar_blocks.py --custom_path data/fgib_frags.csv --real_path combimots/pmcts/resources/real/building_blocks.csv --output_path models/gsk3b_jnk3/similar.csv --threshold 0.4 --batch_size 2500

Remove Salts:

chemfunc canonicalize_smiles --data_path models/${model_name}/${blocks_path} \
--save_path models/${model_name}/${blocks_path} \
--remove_salts --delete_disconnected_mols

# $ chemfunc canonicalize_smiles --data_path models/gsk3b_jnk3/similar.csv --save_path models/gsk3b_jnk3/similar.csv --remove_salts --delete_disconnected_mols

Remove Br, Si and Li atoms for QuickVina-GPU-2.1 compatibility:

python 5-remove_B_Si_Li_blocks.py models/${model_name}/${blocks_path} models/${model_name}/${blocks_path}

# $ python 5-remove_B_Si_Li_blocks.py models/gsk3b_jnk3/similar.csv models/gsk3b_jnk3/similar.csv

Precompute activities and docking scores

Bioactivities using Chemprop (D-MPNN):

chemprop_predict --test_path models/${model_name}/${blocks_path} \
--preds_path models/${model_name}/${blocks_path} \
--checkpoint_dir models/${model_name}

# $ chemprop_predict --test_path models/gsk3b_jnk3/similar.csv --preds_path models/gsk3b_jnk3/precompute.csv --checkpoint_dir models/gsk3b_jnk3

Docking Scores using QuickVina-GPU-2.1 This step is very important as docking oracles are the most expensive components during generation.

python 6-precompute_docking_scores.py models/${model_name}/${blocks_path} models/${model_name}/${blocks_path} --target_pair ${target_pair}

# $ python 6-precompute_docking_scores.py models/gsk3b_jnk3/precompute.csv models/gsk3b_jnk3/final_blocks.csv --target_pair gsk3b_jnk3
# target_pair: str= ["gsk3b_jnk3", "egfr_met", "pik3ca_mtor", "dhodh_rorgt"]

Map building blocks to Enamine's reactions

# Map building blocks -> .pkl
python 7-map_bbs_to_search_space.py --input models/${model_name}/${blocks_path} \
--real_path combimots/pmcts/resources/real/reaction_to_building_blocks.pkl \
--save_path combimots/pmcts/resources/real/${target_pair}.pkl \
--smiles_column smiles

# $ python 7-map_bbs_to_search_space.py --input models/gsk3b_jnk3/final_blocks.csv --real_path combimots/pmcts/resources/real/reaction_to_building_blocks.pkl --save_path combimots/pmcts/resources/real/gsk3b_jnk3.pkl --smiles_column smiles

# Filter non-matching BBs w.r.t. the provided reacions
python 8-filter_reactions_to_blocks.py \
--reaction_to_building_blocks_path combimots/pmcts/resources/real/${target_pair}.pkl \
--save_path combimots/pmcts/resources/real/${target_pair}.pkl

# $ python 8-filter_reactions_to_blocks.py --reaction_to_building_blocks_path combimots/pmcts/resources/real/gsk3b_jnk3.pkl --save_path combimots/pmcts/resources/real/gsk3b_jnk3.pkl

Generation: Pareto Monte-Carlo Tree Search

pmcts --model_path models/${model_name} --save_dir generations/${model_name}/ \
--target_activities ${activity1, activity2} \
--target_pair ${target_pair} \
--building_blocks_path models/${model_name}/final_blocks.csv \
--n_rollout 5000

# $ pmcts --model_path models/${model_name} --save_dir generations/${model_name}/ \
--target_activities gsk3b_activity jnk3_activity \
--target_pair gsk3b_jnk3 \
--building_blocks_path models/gsk3b_jnk3/final_blocks.csv \
--n_rollout 5000

Evaluation

Filter out the molecules predicted as dual actives

python 9-filter_dual_actives.py generations/${model_name}/pareto_molecules.csv generations/${model_name}/pareto_dual_actives.csv

# $ python 9-filter_dual_actives.py generations/gsk3b_jnk3/pareto_molecules.csv generations/gsk3b_jnk3/pareto_dual_actives.csv

Optionally, re-docking simulations have to be run separately.

For all other metrics (Validity, Uniqueness, Novelty, Diversity, Avg.QED, Avg.SA), we provide evaluate.py:

python 10-evaluate.py --model models/${model_name} \
--generation generations/${model_name}/pareto_dual_actives.csv \
--training {path_to_dual_positives_of_training_set_csv}

# $ python 10-evaluate.py --generation generations/gsk3b_jnk3/pareto_dual_actives.csv --training data/GSK3B_dual_actives.csv

If you find our paper/repo useful or use it for personal projects/research, please cite our original paper: -->

@inproceedings{
southiratn2025combimots,
title={Combi{MOTS}: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation},
author={