T-Rex Model

The code for the paper: T-Rex: Text-assisted Retrosynthesis Prediction.

Stage 0: Necessary Package Installation

• pip install torchdrug
• pip install wandb
• pip install torch
• pip install transformers
• pip install openai
• pip install rdkit

• cd ChemicalReaction
  

Stage 1: Candidate Ranking

• First modify the "YOUR_WANDB_KEY" in scripts/train/candidate_generation.sh
• Then run by the following code (you can change the parameters of fold and device):

  • bash scripts/train/candidate_ranking.sh
    

  • or the complete code(change FOLD, DEVICE NAME to the actual fold, GPU number and project name):

  • bash scripts/train/candidate_ranking.sh FOLD DEVICE NAME
    

• Test by the following code

  • bash scripts/train/cr_test.sh
    

  • samely the complete code is: (due to the bug in torchdrug, the evaluation is shown with "Evaluate on train" but actually it is testing on some file)

  • bash scripts/train/cr_test.sh FOLD DEVICE NAME
    

Mid-Stage: Candidate Preparation

• You should have the openai key for generating the text description of the retrosynthesis pair.
• First generate candidate pairs by:

  • bash scripts/chatgpt/candidate_generator.sh
    

• Then generate the text descriptions using chatgpt("gpt-3.5-turbo-0301")

  • bash scripts/chatgpt/chatgpt_generator.sh
    

Stage 2: Re-ranking

• Then train re-rank model by:

  • bash scripts/train/reranking_training.sh
    

  • or the complete code(change FOLD, DEVICE NAME to the actual fold, GPU number and project name):

  • bash scripts/train/reranking_training.sh FOLD DEVICE NAME
    

• And test by:

  • bash scripts/train/reranking_test.sh
    

  • or the complete code(change FOLD, DEVICE NAME to the actual fold, GPU number and checkpoint name, which is a little different from the NAME above, it is in logs-ckpt folder):

  • bash scripts/train/reranking_test.sh FOLD DEVICE NAME