GLN

Implementation of Retrosynthesis Prediction with Conditional Graph Logic Network

https://arxiv.org/abs/2001.01408

Setup

• Install package

This package requires the rdkit and pytorch.

For rdkit, we tested with 2019_03_3. Please build the package from source and setup the PYTHONPATH and LD_LIBRARY_PATH properly, as this package requires the dynamic lib built from rdkit.

For pytorch, we tested from 1.2.0 to 1.3.1. Please don't use versions older than that, as this package contains c++ customized ops which relies on specific toolchains from pytorch.

After the above preparation, simply navigate to the project root folder and install:

cd GLN
pip install -e .

Note that by default the cuda ops will not be enabled on Mac OSX.

• Dropbox

We provide the raw datasets, the cooked (data after preprocessing) datasets, and also the trained model dumps in a dropbox folder.

https://www.dropbox.com/sh/6ideflxcakrak10/AADTbFBC0F8ax55-z-EDgrIza?dl=0

The cooked dataset is pretty large. You can also simply download the raw datasets only, and use the script provided in this repo for preprocessing. You don't have to have a dropbox to download, and the result folder doesn't have to be in your dropbox. The only thing needed is to create a symbolic link named dropbox and put it in the right place.

Finally the folder structure will look like this:

GLN
|___gln  # source code
|   |___common # common implementations
|   |___...
|
|___setup.py 
|
|___dropbox  # data, trained model dumps and cooked data, this can be a symbolic link
    |___schneider50k # raw data
    |___|__raw_train.csv
    |   |__...
    |
    |___cooked_schneider50k # cooked data
    |___schneider50k.ckpt # trained model dump
...

• Remark: full USPTO dataset

We also released our cleaned USPTO dataset used in the paper via the above dropbox link (see uspto_multi folder under the dropbox folder). Meanwhile, the script for cleaning and de-duplication can be found under gln/data_process/clean_uspto.py. The version of USPTO is 1976_Sep2016_USPTOgrants_smiles.rsmi (which can also be found via above dropbox link). If you run the clean_uspto.py on this raw rsmi file, you are expected to get the same data split as we used in the paper.

Preprocessing

If you download the cooked data in the previous step, you can simply skip this step.

Our paper mainly focused on schneider50k dataset. The raw data and data split is the same as https://github.com/connorcoley/retrosim/blob/master/retrosim/data/get_data.py

Note that for both the reaction type unknown and type given experiments, we go throught the same steps as below. The only difference is the dataset name: schneider50k (type unknown) v.s. typed_schneider50k (type given)

First go to the data processing folder:

cd gln/data_process

Then run the script in the following order:

• Get canonical smiles

Please specify the dataset name accordingly

./step0.0_run_get_cano_smiles.sh

• Extract raw templates

Please specify the dataset name, as well as #cpu threads (the more the better).

./step0.1_run_raw_template_extract.sh

1. Filter template

One can filter out uncommon templates using this script. We didn't filter out any template in our paper, though a careful selection of templates may further improve the performance.

Please specify the dataset name, template name (arbitrary one is fine), and possibly the minimum number of occurance one template needs to have in order to be included (default=1)

./step1_filter_template.sh

2. Get subgraph SMARTS

Please specify the dataset name, template name (same as step 1.)

./step2_run_get_cano_smarts.sh

3. Get feasible centers

Please specify the dataset name, template name (same as step 1.), # cpus (the more the better)

./step3_run_find_centers.sh

4. Get the support for the graphical model

Please specify the dataset name, template name (same as step 1.), # cpus (the more the better). A 40-core machine would get it done in 15min.

./step4_run_find_all_reactions.sh

5. Get graph feature dumps

Please specify the dataset name, template name (same as step 1.)

./step5_run_dump_graphs.sh

Training

To train the model from scratch, first navigate to the training script folder

cd gln/training/scripts

Then run the default script run_mf.sh with the dataset name.

• To run type unknown model, use ./run_mf.sh schneider50k
• To run type conditional model, use ./run_mf.sh typed_schneider50k

Usually ~10 x 3000 iterations would be able to get reasonable results that match the numbers in the paper. You are also welcome to tune any hyper-parameters or configurations in the script.

Test

First navigate to the test folder:

cd gln/test

1. Reproducing results in the paper

To test the existing model dump in the dropbox, use the following commands:

• To test pretrained type unknown model, use ./test_single.sh schneider50k
• To test pretrained type conditional model, use ./test_single.sh typed_schneider50k

You can also test whatever single model you want, by changing the -model_for_test argument to the model dump you want.

2. Pick models that trained from scratch with best validation

The best model is picked with best validation loss. To do so, first get the performance of all model dumps:

• ./test_all.sh schneider50k YOUR_MODEL_DUMP_ROOT
• python report_test_stats.py YOUR_MODEL_DUMP_ROOT

Reference

If you find our paper/code is useful, please consider citing our paper:

@inproceedings{dai2019retrosynthesis,
  title={Retrosynthesis Prediction with Conditional Graph Logic Network},
  author={Dai, Hanjun and Li, Chengtao and Coley, Connor and Dai, Bo and Song, Le},
  booktitle={Advances in Neural Information Processing Systems},
  pages={8870--8880},
  year={2019}
}

The orignal version of rdchiral comes from https://github.com/connorcoley/rdchiral, and this repo has made a wrapper over that. Please also cite the corresponding paper if possible:

@article{coley2019rdchiral,
  title={RDChiral: An RDKit Wrapper for Handling Stereochemistry in Retrosynthetic Template Extraction and Application},
  author={Coley, Connor W and Green, William H and Jensen, Klavs F},
  journal={Journal of chemical information and modeling},
  publisher={ACS Publications}
}