ACEGEN is a comprehensive toolkit designed to leverage reinforcement learning (RL) techniques for generative chemistry tasks, particularly in drug design. ACEGEN harnesses the capabilities of TorchRL, a modern library for general decision-making tasks, to provide a flexible and integrated solution for generative drug design challenges.
The full paper can be found here.
-
Multiple Generative Modes: ACEGEN facilitates the generation of chemical libraries with different modes: de novo generation, scaffold decoration, and fragment linking.
-
RL Algorithms: ACEGEN offers task optimization with various reinforcement learning algorithms such as Proximal Policy Optimization (PPO), Advantage Actor-Critic (A2C), Reinforce, Reinvent, and Augmented Hill-Climb (AHC).
-
Other Algorithms: ACEGEN also includes Direct Preference Optimization (DPO) and Hill Climbing.
-
Pre-trained Models: ACEGEN contains pre-trained models including Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), GPT-2, LLama2 and Mamba.
-
Scoring Functions : ACEGEN defaults to MolScore, a comprehensive scoring function suite for generative chemistry, to evaluate the quality of the generated molecules. MolScore allows to train agents on single scoring functions, on entire benchmarks containing multiple scoring functions (e.g., MolOpt, GuacaMol), or using curriculum learning where the same agent is optimized on a sequence of different scoring functions.
-
Customization Support: ACEGEN provides tutorials for integrating custom models and custom scoring functions, ensuring flexibility for advanced users.
To create the conda / mamba environment, run
conda create -n acegen python=3.10 -y
conda activate acegen
pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121
To install Tensordict, run
git clone https://github.com/pytorch/tensordict.git
cd tensordict
python setup.py install
To install TorchRL, run
git clone https://github.com/pytorch/rl.git
cd rl
python setup.py install
To install ACEGEN, run (use pip install -e ./
for develop mode)
pip3 install tqdm wandb hydra-core
git clone https://github.com/Acellera/acegen-open.git
cd acegen-open
pip install ./
Unless you intend to define your own custom scoring functions, install MolScore by running
pip3 install MolScore
To use the scaffold decoration and fragment linking, install promptsmiles by running
pip3 install promptsmiles
To learn how to configure constrained molecule generation with AcGen and promptsmiles, please refer to this tutorial.
ACEGEN has multiple RL algorithms available, each in a different directory within the acegen-open/scripts
directory. Each RL algorithm has three different generative modes of execution: de novo, scaffold decoration, and fragment linking.
Each mode of execution has its own configuration file in YAML format, located right next to the script. To modify training parameters for any mode, edit the corresponding YAML file. For a breakdown of the general structure of our configuration files, refer to this tutorial.
While the default values in the configuration files are considered sensible, a default scoring function and model architecture are also defined so users can test the scripts out of the box. However, users might generally want to customize the model architecture or the scoring function.
To customize the model architecture, refer to the [Changing the model architecture](##Changing the model architecture) section. To customize the scoring function, refer to the [Changing the scoring function](##Changing the scoring function) section.
To run the training scripts for denovo generation, run the following commands:
python scripts/reinforce/reinforce.py --config-name config_denovo
python scripts/a2c/a2c.py --config-name config_denovo
python scripts/ppo/ppo.py --config-name config_denovo
python scripts/reinvent/reinvent.py --config-name config_denovo
python scripts/ahc/ahc.py --config-name config_denovo
python scripts/dpo/dpo.py --config-name config_denovo
python scripts/hill_climb/hill_climb.py --config-name config_denovo
To run the training scripts for scaffold decoration, run the following commands (requires installation of promptsmiles):
python scripts/reinforce/reinforce.py --config-name config_scaffold
python scripts/a2c/a2c.py --config-name config_scaffold
python scripts/ppo/ppo.py --config-name config_scaffold
python scripts/reinvent/reinvent.py --config-name config_scaffold
python scripts/ahc/ahc.py --config-name config_scaffold
python scripts/dpo/dpo.py --config-name config_scaffold
python scripts/hill_climb/hill_climb.py --config-name config_scaffold
To run the training scripts for fragment linking, run the following commands (requires installation of promptsmiles):
python scripts/reinforce/reinforce.py --config-name config_linking
python scripts/a2c/a2c.py --config-name config_linking
python scripts/ppo/ppo.py --config-name config_linking
python scripts/reinvent/reinvent.py --config-name config_linking
python scripts/ahc/ahc.py --config-name config_linking
python scripts/dpo/dpo.py --config-name config_linking
python scripts/hill_climb/hill_climb.py --config-name config_linking
Scripts are also available as executables after installation, but both the path and name of the config must be specified. For example,
ppo.py --config-path=<path_to_config_dir> --config-name=<config_name.yaml>
YAML config parameters can also be specified on the command line. For example,
ppo.py --config-path=<path_to_config_dir> --config-name=<config_name.yaml> total_smiles=100
To change the scoring function, the easiest option is to adjust the molscore
parameters in the configuration files. Modifying these parameters allows to switch betwewn different scoring modes and scoring objecitves.
Please refer to the molscore
section in the configuration tutorial for a more detailed explaination. Additionally, refer to the tutorials in the MolScore repository.
Alternatively, users can define their own custom scoring functions and use them in the ACEGEN scripts by following the instructions in this other tutorial.
We provide a variety of default priors that can be selected in the configuration file. These include:
-
A Gated Recurrent Unit (GRU) model
- pre-training dataset1 (default): ChEMBL
- pre-training dataset2: ZINC250k
- tokenizer: SMILESTokenizerChEMBL
- number of parameters: 4,363,045
- to select set the field
model
togru
in any configuration file
-
A Long Short-Term Memory (LSTM) model
- pre-training dataset: ChEMBL
- tokenizer: SMILESTokenizerChEMBL
- number of parameters: 5,807,909
- to select set the field
model
tolstm
in any configuration file
-
A GPT-2 model (requires installation of HuggingFace's
transformers
library)- pre-training dataset: REAL 350/3 lead-like, 613.86M cpds, CXSMILES
- tokenizer: SMILESTokenizerEnamine
- number of parameters: 5,030,400
- to select set the field
model
togpt2
in any configuration file
-
A Mamba model (requires installation of
mamba-ssm
library)- pre-training dataset: ChEMBL
- tokenizer: SMILESTokenizerChEMBL
- number of parameters: 2,809,216
- to select set the field
model
tomamba
in any configuration file
-
A Llama2 model (requires installation of HuggingFace's
transformers
library)- pre-training dataset: REAL Database, 6B cpds, CXSMILES
- tokenizer: AsciiSMILESTokenizer
- number of parameters: 5,965,760
- to select set the field
model
tollama2
in any configuration file
Users can also combine their own custom models with ACEGEN.
A detailed guide on integrating custom models can be found in this tutorial.
Results on the MolOpt benchmark
Algorithm comparison for the Area Under the Curve (AUC) of the top 100 molecules on MolOpt benchmark scoring functions. Each algorithm ran 5 times with different seeds, and results were averaged. The default values for each algorithm are those in our de novo configuration files. Additionally, for Reinvent we also tested the configuration proposed in the MolOpt paper.
Task | REINFORCE | REINVENT | REINVENT MolOpt | AHC | A2C | PPO | PPOD |
---|---|---|---|---|---|---|---|
Albuterol_similarity | 0.68 | 0.69 | 0.90 | 0.77 | 0.82 | 0.93 | 0.94 |
Amlodipine_MPO | 0.55 | 0.56 | 0.65 | 0.56 | 0.55 | 0.58 | 0.68 |
C7H8N2O2 | 0.83 | 0.82 | 0.90 | 0.76 | 0.84 | 0.89 | 0.89 |
C9H10N2O2PF2Cl | 0.70 | 0.70 | 0.76 | 0.68 | 0.69 | 0.66 | 0.79 |
Celecoxxib_rediscovery | 0.63 | 0.64 | 0.77 | 0.72 | 0.73 | 0.65 | 0.82 |
DRD2 | 0.98 | 0.97 | 0.99 | 0.98 | 0.98 | 0.99 | 0.99 |
Deco_hop | 0.63 | 0.63 | 0.67 | 0.64 | 0.62 | 0.62 | 0.66 |
Fexofenadine_MPO | 0.71 | 0.71 | 0.80 | 0.72 | 0.71 | 0.73 | 0.78 |
GSK3B | 0.84 | 0.84 | 0.92 | 0.82 | 0.85 | 0.90 | 0.92 |
JNK3 | 0.75 | 0.75 | 0.85 | 0.75 | 0.74 | 0.80 | 0.87 |
Median_molecules_1 | 0.26 | 0.24 | 0.36 | 0.24 | 0.31 | 0.33 | 0.35 |
Median_molecules_2 | 0.22 | 0.22 | 0.28 | 0.24 | 0.25 | 0.25 | 0.29 |
Mestranol_similarity | 0.60 | 0.55 | 0.85 | 0.66 | 0.69 | 0.75 | 0.89 |
Osimertinib_MPO | 0.82 | 0.82 | 0.86 | 0.83 | 0.81 | 0.82 | 0.84 |
Perindopril_MPO | 0.48 | 0.47 | 0.54 | 0.47 | 0.48 | 0.50 | 0.53 |
QED | 0.94 | 0.94 | 0.94 | 0.94 | 0.94 | 0.94 | 0.94 |
Scaffold_hop | 0.80 | 0.79 | 0.86 | 0.80 | 0.80 | 0.80 | 0.84 |
Sitagliptin_MPO | 0.34 | 0.33 | 0.38 | 0.33 | 0.39 | 0.32 | 0.39 |
Thiothixene_rediscovery | 0.41 | 0.41 | 0.56 | 0.45 | 0.48 | 0.48 | 0.58 |
Troglitazone_rediscovery | 0.31 | 0.31 | 0.47 | 0.34 | 0.35 | 0.46 | 0.52 |
Valsartan_smarts | 0.03 | 0.02 | 0.02 | 0.02 | 0.02 | 0.03 | 0.03 |
Zaleplon_MPO | 0.47 | 0.47 | 0.52 | 0.48 | 0.47 | 0.50 | 0.52 |
Total | 13.67 | 13.60 | 15.65 | 13.91 | 14.27 | 14.65 | 15.80 |
If you use ACEGEN in your work, please refer to this BibTeX entry to cite it:
@article{bou2024acegen,
title={ACEGEN: Reinforcement learning of generative chemical agents for drug discovery},
author={Bou, Albert and Thomas, Morgan and Dittert, Sebastian and Navarro Ramírez, Carles and Majewski, Maciej and Wang, Ye and Patel, Shivam and Tresadern, Gary and Ahmad, Mazen and Moens, Vincent and Sherman, Woody and Sciabola, Simone and De Fabritiis, Gianni},
eprint={2405.04657},
archivePrefix={arXiv},
year={2024}
}