SMILES-to-SMILES transformer

The repository contains an implementation of the Molecular Transformer in Pytorch Lightning.

Speculative decoding

The repository also contains the code to accompany the manuscript
"Accelerating the inference of string generation-based chemical reaction models for industrial applications".

The idea of the manuscript is that reaction product prediction and single-step retrosynthesis with the Molecular Transformer
can be accelerated by ~3 times on inference without losing accuracy.

Installation

Create a new isolated environment with Python 3.10 and install the necessary packages:

pip install lightning
pip install jsonargparse[signatures]
pip install rdkit
pip install tensorboard
pip install gdown
pip install -e .

Data

For reaction prediction, we used USPTO_MIT mixed from the Molecular Transformer paper. For single-step retrosynthesis, we used USPTO50k as prepared in the RSMILES paper.

Download USPTO MIT mixed:

gdown https://drive.google.com/drive/folders/1fJ7Hm55IDevIi5Apna7v-rQBQStTH7Yg -O data/MIT_mixed --folder
python3 src/detokenize.py --data_dir data/MIT_mixed

Download USPTO 50K and augment it using RSMILES augmentation.
Clone the RSMILES repository to some path in your system.

pip install pandas textdistance

THIS_REPO_PATH=$(pwd) # The full path to this repository 
RSMILES_PATH=../retrosynthesis  # as an example; the path to the RSMILES repository

gdown https://drive.google.com/drive/folders/1la4OgBKgm2K-IRwuV-GHUNjN3bcCrl6v -O ${RSMILES_PATH}/dataset/USPTO_50K --folder
cd ${RSMILES_PATH}
AUGMENTATIONS=20
PROCESSES=8
python3 preprocessing/generate_PtoR_data.py -augmentation ${AUGMENTATIONS} -processes ${PROCESSES} -test_except
python3 preprocessing/generate_PtoR_data.py -augmentation 1 -processes ${PROCESSES} -test_only -canonical
mv dataset/USPTO_50K_PtoR_aug${AUGMENTATIONS} ${THIS_REPO_PATH}/data # The augmented dataset is now in this repository
mv dataset/USPTO_50K_PtoR_aug1 ${THIS_REPO_PATH}/data
cd $THIS_REPO_PATH
python3 src/detokenize.py --data_dir data/USPTO_50K_PtoR_aug1/test
python3 src/detokenize.py --data_dir data/USPTO_50K_PtoR_aug20/train
python3 src/detokenize.py --data_dir data/USPTO_50K_PtoR_aug20/val

Models

The file main.py calls a Pytorch Lightning model for either training or inference. The directory scripts contains bash scripts for reaction product prediction and single-step retrosynthesis with speculative decoding.

Checkpoints

Trained checkpoints and config files are available at Google Drive. Download reaction prediction checkpoints:

mkdir checkpoints
mkdir checkpoints/reaction_prediction
gdown https://drive.google.com/drive/folders/1sBiVgFZyD4F42nVqR835-0Tl90LkQvU9 -O checkpoints/reaction_prediction --folder

Download single-step retrosynthesis checkpoints

mkdir checkpoints/single_step_retrosynthesis
gdown https://drive.google.com/drive/folders/1v4pKYWlE0qNA-ksa7yX55i7qMeesURON -O checkpoints/single_step_retrosynthesis --folder

Inference

The scripts scripts/product_prediction.sh and scripts/single_step_retrosynthesis.sh run the models for reaction product prediction and single-step retrosynthesis, respectively.
The forward pass of the transformer is implemented in src/model/lightning_model.py.
The sampling of sequences from the logits predicted by the transformer is implemented in src/decoding/standard_decoding.py and src/decoding/speculative_decoding.py.
The generated sequences are written to CSV files using the PredictionWriter callback implemented in src/callbacks.py. For example, the script scripts/product_prediction.sh loads a checkpoint of the transformer and runs the model on the USPTO MIT mixed test dataset.
The generated sequences are then saved to, e.g., results_product_greedy_speculative/MIT_mixed_greedy_speculative_batched_bs_1_draftlen_10.csv.

Evaluating performance

The script src/score_predictions.py scores the predictions saved in CSV files. It calculates the top-N accuracy and the percentage of invalid SMILES.

Example usage:

python3 src/score_predictions.py -f results_product_greedy_speculative/MIT_mixed_greedy_speculative_batched_bs_1_draftlen_10.csv

The accuracy is printed to the terminal.