template-splits

Code to complement our paper: An exploration of dataset importance in single-step retrosynthesis prediction. preprint

This repository contains the code for cleaning and splitting USPTO and Pistachio datasets based on templates, as well as the configuration files used to train models used in the paper.

1. Installation/Setup

1. Clone the repository: git clone https://github.com/duartegroup/template-splits.git
2. Install the required Python environment using the provided yaml file:
   conda env create -f envs/environment.yml
3. Activate environment: conda activate template-splits.
4. Add the directory of this repo to your Python path:
   export PYTHONPATH=${PYTHONPATH}:$(pwd) (from inside the directory this README lives in).

2. Preprocessing

Note that we take advantage of high memory availability to load and process each dataset in full. Use the --uspto-only flag on each script if not using the Pistachio dataset.

2.1 Initialisation

As Pistachio is a proprietary database owned by NextMove Software, we do not provide the data here. Instead it must be manually imported into preprocessing/data/raw as pistachio.smi before running the following script.

1_initialisation.py: Initialises USPTO and Pistachio datasets for further preprocessing. The Pistachio dataset is filtered to remove US patents and remapped using RXNMapper (GPU recommended).

2.2 Cleaning

2_cleaning.py: Cleans and filters atom-mapped reaction strings for the initialised USPTO and Pistachio datasets according to the steps detailed in S1.

2.3 Template extraction

3_template_extracting: Extracts templates from USPTO and Pistachio via the LocalTemplate algorithm. The module for this algorithm was obtained from the original respository (https://github.com/kaist-amsg/LocalRetro) in February 2024 when it was licensed under Apache 2.0; subsequent removal by original authors does not affect this version’s licence. We have redistributed the necessary scripts for reproducibility here under the directory preprocessing/LocalTemplate.

2.4 Template filtering

4_template_filtering.py: Filters USPTO and Pistachio by the number of template occurrences, removing all templates with less than 5 or 20 reactions respectively.

2.5 Template splitting

5_template_splitting.py: Splits the USPTO dataset via the narrow and broad template splitting strategies, as detailed in the paper. Splits the Pistachio dataset into two 10k reaction test sets, where the ID and OOD test sets contains templates present or not present in USPTO respectively.

3. Training

All configuration files used to train the models in the paper are provided. Models were trained according to their respective packages:

• LocalRetro: https://github.com/kaist-amsg/LocalRetro (since removed by the authors)
• MEGAN: https://github.com/molecule-one/megan
• RootAligned: https://github.com/otori-bird/retrosynthesis