This repository hosts DST (Differentiable Scaffolding Tree for Molecule Optimization) (Tianfan Fu*, Wenhao Gao*, Cao Xiao, Jacob Yasonik, Connor W. Coley, Jimeng Sun), which enables a gradient-based optimization on a chemical graph.
To install locally, we recommend to install from pip
and conda
. Please see conda.yml
for the package dependency.
conda create -n dst python=3.7
conda activate dst
pip install torch
pip install PyTDC
conda install -c rdkit rdkit
Activate conda environment.
conda activate dst
make directory
mkdir -p save_model result
In our setup, we restrict the number of oracle calls. In realistic discovery settings, the oracle acquisition cost is usually not negligible.
We use ZINC
database, which contains around 250K drug-like molecules and can be downloaded download ZINC
.
python src/download.py
- output
data/zinc.tab
: all the smiles in ZINC, around 250K.
Oracle is a property evaluator and is a function whose input is molecular structure, and output is the property. We consider following oracles:
JNK3
: biological activity to JNK3, ranging from 0 to 1.GSK3B
biological activity to GSK3B, ranging from 0 to 1.QED
: Quantitative Estimate of Drug-likeness, ranging from 0 to 1.SA
: Synthetic Accessibility, we normalize SA to (0,1).LogP
: solubility and synthetic accessibility of a compound. It ranges from negative infinity to positive infinity.
For all the property scores above, higher is more desirable.
There are two kinds of optimization tasks: single-objective and multi-objective optimization.
Multi-objective optimization contains jnkgsk
(JNK3 + GSK3B), qedsajnkgsk
(QED + SA + JNK3 + GSK3B).
In this project, the basic unit is substructure
, which can be atoms or single rings.
The vocabulary is the set of frequent substructures
.
python src/vocabulary.py
- input
data/zinc.tab
: all the smiles in ZINC, around 250K.
- output
data/substructure.txt
: including all the substructures in ZINC.data/vocabulary.txt
: vocabulary, frequent substructures.
We remove the molecules that contains substructure that is not in vocabulary.
python src/clean.py
- input
data/vocabulary.txt
: vocabularydata/zinc.tab
: all the smiles in ZINC
- output
data/zinc_clean.txt
We use oracle to evaluate molecule's properties to obtain the labels for training graph neural network.
python src/labelling.py
- input
data/zinc_clean.txt
: all the smiles in ZINC, around 250K.
- output
data/zinc_label.txt
: including 6 columns,smiles
,qed
,sa
,jnk
,gsk
,logp
. We only contains subset of zinc (10K).
In our setup, we restrict the number of oracle calls in both training GNN and de novo design.
It corresponds to Section 3.2 in the paper.
python src/train.py $prop $train_oracle
prop
represent the property to optimize, includingqed
,logp
,jnk
,gsk
,jnkgsk
,qedsajnkgsk
.train_oracle
is number of oracle calls in training GNN.- input
data/zinc_label.txt
: training data includes(SMILES,y)
pairs, whereSMILES
is the molecule,y
is the label.y = GNN(SMILES)
- output
save_model/model_epoch_*.ckpt
: saved GNN model.
- log
"loss/{$prop}.pkl"
save the valid loss. For example,
python src/train.py jnkgsk 5000
It corresponds to Section 3.3 and 3.4 in the paper.
python src/denovo.py $prop $denovo_oracle
prop
represent the property to optimize, includingqed
,logp
,jnk
,gsk
,jnkgsk
,qedsajnkgsk
.denovo_oracle
is number of oracle calls.- input
save_model/{$prop}_*.ckpt
: saved GNN model. * is number of iteration or epochs.
- output
result/{$prop}.pkl
: set of generated molecules.
For example,
python src/denovo.py jnkgsk 5000
python src/evaluate.py $prop
- input
result/{$prop}.pkl
- output
diversity
,novelty
,average property
of top-100 molecules with highest property.
For example,
python src/evaluate.py jnkgsk
python src/multiobjective.py
Please contact futianfan@gmail.com or gaowh19@gmail.com for help or submit an issue.
If you found this package useful, please cite our paper:
@article{fu2020differentiable,
title={Differentiable Scaffolding Tree for Molecule Optimization},
author={Tianfan Fu*, Wenhao Gao*, Cao Xiao, Jacob Yasonik, Connor W. Coley, Jimeng Sun},
journal={International Conference on Learning Representation (ICLR)},
year={2022}
}