Distributed LocalRetro Implementation

This package extends AiZynthFinder to seamlessly integrate various single-step retrosynthesis models, as illustrated in our papers Models Matter: The Impact of Single-Step Models on Synthesis Prediction and Mind the Retrosynthesis Gap: Bridging the divide between Single-step and Multi-step Retrosynthesis Prediction.

Reimplementation of LocalRetro to preprocess data in a distributed fashion using slurm. Starting point is to create a preprocessing pipeline with the slurm/Makefile. Just copy the "/slurm" folder, change the relevant paths in the Makefile, and run "make create_experiment". Afterwards, follow the steps in the newly created make file.

Note: Whenever <Path> is present, it is necessary to adjust the paths to the correct absolute folder.

Funding

This study was partially funded by the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie Innovative Training Network European Industrial Doctorate grant agreement No. 956832 “Advanced machine learning for Innovative Drug Discovery”

Old LocalRetro Readme

Implementation of Retrosynthesis Prediction with LocalRetro developed by prof. Yousung Jung group at KAIST (contact: [email protected]).

Developer

Shuan Chen (contact: [email protected])

Requirements

• Python (version >= 3.6)
• Numpy (version >= 1.16.4)
• PyTorch (version >= 1.0.0)
• RDKit (version >= 2019)
• DGL (version >= 0.5.2)
• DGLLife (version >= 0.2.6)

Requirements

Create a virtual environment to run the code of LocalRetro.
Install pytorch with the cuda version that fits your device.

cd LocalRetro
conda create -c conda-forge -n rdenv python=3.7 -y
conda activate rdenv
conda install pytorch cudatoolkit=10.2 -c pytorch -y
conda install -c conda-forge rdkit -y
pip install dgl
pip install dgllife

Publication

Shuan Chen and Yousung Jung. Deep Retrosynthetic Reaction Prediction using Local Reactivity and Global Attention, JACS Au 2021.

Code reuploading announcement (2022.02.09)

We cleaned the code and made the template more simplied, which yields 658 local reaction templates for USPTO_50K dataset and 20,221 local reaction templates for USPTO_MIT dataset. Therefore we tested the top-k accuracy again and the results are updated at the bottom of README.md. The training takes around 100 minutes on NVIDIA GeForce RTX 3090

Code cleaning announcement (2021.09.16)

Currently, we are cleaning up the codes, and the codes will be uploaded back afterwards.

Usage

[1] Download the raw data of USPTO-50K or USPTO-MIT dataset

See the README in ./data to download the raw data files for training and testing the model.

[2] Data preprocessing

A two-step data preprocessing is needed to train the LocalRetro model.

1) Local reaction template derivation

First go to the data processing folder

cd preprocessing

and extract the reaction template with specified dataset name (default: USPTO_50K).

python Extract_from_train_data.py -d USPTO_50K

This will give you four files, including (1) atom_templates.csv (2) bond_templates.csv (3) template_infos.csv (4) template_rxnclass.csv (if train_class.csv exists in data folder)

2) Assign the derived templates to raw data

By running

python Run_preprocessing.py -d USPTO_50K

You can get four preprocessed files, including (1) preprocessed_train.csv (2) preprocessed_val.csv (3) preprocessed_test.csv (4) labeled_data.csv

[3] Train LocalRetro model

Go to the localretro folder

cd ../scripts

and run the following to train the model with specified dataset (default: USPTO_50K)

python Train.py -d USPTO_50K

The trained model will be saved at LocalRetro/models/LocalRetro_USPTO_50K.pth

[4] Test LocalRetro model

To use the model to test on test set, simply run

python Test.py -d USPTO_50K

to get the raw prediction file saved at LocalRetro/outputs/raw_prediction/LocalRetro_USPTO_50K.txt
Finally you can get the reactants of each prediciton by decoding the raw prediction file

python Decode_predictions.py -d USPTO_50K

The decoded reactants will be saved at LocalRetro/outputs/decoded_prediction/LocalRetro_USPTO_50K.txt
and LocalRetro/outputs/decoded_prediction_class/LocalRetro_USPTO_50K.txt

Exact match accuracy (%) on USPTO-50K dataset without given reaction class

*AT = Augmented Transformer

Method	Top-1	Top-3	Top-5	Top-10	Top-50
GLN	52.5	69.0	75.6	83.7	92.4
G2Gs	48.9	67.6	72.5	75.5	/
GraphRetro	53.7	68.3	72.2	75.5	/
AT	53.5	69.4	81.0	85.7	/
MEGAN	48.1	70.7	78.4	86.1	93.2
LocalRetro	53.4	77.5	85.6	92.4	98.4

Exact match accuracy (%) on USPTO-50K dataset with given reaction class

Method	Top-1	Top-3	Top-5	Top-10	Top-50
GLN	64.2	79.1	85.2	90.0	93.2
G2Gs	61.0	81.3	86.0	88.7	/
GraphRetro	63.9	81.5	85.2	88.1	/
MEGAN	60.7	82.0	87.5	91.6	95.3
LocalRetro	64.2	86.8	93.0	96.9	98.6