TextReact

This repository contains the code for TextReact, a novel method that directly augments predictive chemistry with text retrieval.

@inproceedings{TextReact,
  author       = {Yujie Qian and
                  Zhening Li and
                  Zhengkai Tu and
                  Connor W. Coley and
                  Regina Barzilay},
  title        = {Predictive Chemistry Augmented with Text Retrieval},
  booktitle    = {Proceedings of the 2023 Conference on Empirical Methods in Natural
                  Language Processing, {EMNLP} 2023, Singapore, December 6-10, 2023},
  pages        = {12731--12745},
  publisher    = {Association for Computational Linguistics},
  year         = {2023},
  url          = {https://aclanthology.org/2023.emnlp-main.784}
}

Requirements

We implement the code with torch==1.11.0, pytorch-lightning==2.0.0, and transformers==4.27.3. To reproduce our experiments, we recommend creating a conda environment with the same dependencies:

conda env create -f environment.yml -n textreact
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113

Data

Run the following commands to download and unzip the preprocessed datasets:

git clone https://huggingface.co/datasets/yujieq/TextReact data
cd data
unzip '*'

Training Scripts

TextReact consists of two modules: SMILES-To-text retriever and text-augmented predictor. This repository only contains the code for training the predictor, while the code for the retriever is available in a separate repository: https://github.com/thomas0809/tevatron.

The training scripts are located under scripts:

• train_RCR.sh trains a model for reaction condition recommendation (RCR) on the random split of the USPTO dataset.
• train_RetroSyn_tf.sh trains a template-free model for retrosynthesis on the random split of the USPTO-50K dataset.
• train_RetroSyn_tb.sh trains a template-based model for retrosynthesis on the random split of the USPTO-50K dataset. In addition, train_RCR_TS.sh, train_RetroSyn_tf_TS.sh and train_RetroSyn_tb_TS.sh train the corresponding models on the time-based split of the dataset.

If you're working on a distributed file system, it is recommended to add to the script a --cache_path option specifying a local path to reduce network time.

To run the script scripts/train_MODEL.sh, run the following command at the root of the folder:

bash scripts/train_MODEL.sh

At the end of training, two dictionaries are printed with the top-k test accuracies. The first one corresponds to retrieving from the full corpus and the second one corresponds to retrieving from the gold-removed corpus.

Models and test predictions are stored under the path specified by the SAVE_PATH variable in the script.

• best.ckpt is the checkpoint with the highest validation accuracy so far, whereas
• last.ckpt is the last checkpoint.
• prediction_test_0.json contains the test predictions when retrieving from the full corpus.
• prediction_test_1.json contains the predictions when retrieving from the gold-removed corpus.