Chemoenzymatic Multistep Retrosynthesis by Disconnection Aware Triple Transformer Loops

This repository complements the Chemical Science article "Multistep retrosynthesis combining a disconnection aware triple transformer loop with a route penalty score guided tree search". The work has been extended to Enzymatic reactions as described in the ChemRxiv preprint "Chemoenzymatic Multistep Retrosynthesis with Transformer Loops".

The goal of this tool is to predict multistep retrosynthesis routes combining a tree search strategy and Transformer Disconnection-Aware Retrosynthesis models. The single-step retrosynthesis models are augmented by coupling a bias-free systematic tagging as well as a template-based tagging using a reaction center substructure identification from known reactions. The work was extended to biocatalysis by running two TTLs in parallel.

Setup Environment

Create the conda environment, clone this repo and install:

conda create -n MultiStepRetro_ENZ python=3.8.16 -y
conda activate MultiStepRetro_ENZ

git clone https://github.com/reymond-group/MultiStepRetrosynthesisTTL.git

cd MultiStepRetrosynthesisTTL

pip install .

Download USPTO Models

Auto-Tag, Disconnection-Aware Retrosynthesis (T1), Reagent Prediction (T2), and Forward Validation Transformer (T3) models are required and should be placed into your models folder and referenced in the config file under parameters: USPTO_AutoTag_path, USPTO_T1_path, USPTO_T2_path and USPTO_T3_path.

The USPTO models can be downloaded from Zenodo and placed in the models folder with the following commands:

wget https://zenodo.org/record/8160148/files/USPTO_STEREO_separated_T0_AutoTag_260000.pt?download=1 -O models/USPTO_STEREO_separated_T0_AutoTag_260000.pt
wget https://zenodo.org/record/8160148/files/USPTO_STEREO_separated_T1_Retro_255000.pt?download=1 -O models/USPTO_STEREO_separated_T1_Retro_255000.pt
wget https://zenodo.org/record/8160148/files/USPTO_STEREO_separated_T2_Reagent_Pred_225000.pt?download=1 -O models/USPTO_STEREO_separated_T2_Reagent_Pred_225000.pt
wget https://zenodo.org/record/8160148/files/USPTO_STEREO_separated_T3_Forward_255000.pt?download=1 -O models/USPTO_STEREO_separated_T3_Forward_255000.pt

ENZR Models

Models can not be made public as trained from data under Reaxys commercial subscription. If you have a Reaction Reaxys API licence, you can download the ENZR enzymatic reaction dataset using notebook/API_reaxys_ENZR_dataset.py with your API credentials. The list of reaction IDs used for this work can be found in notebook/ENZR_Rxn_IDs.txt. The raw data obtained needs to be converted to SMILES to train the models following the preprocessing described in the OpenNMT repository.

Commercial Building Blocks

The list of commercial compounds should be requested and downloaded from MolPort and/or from Enamine. SMILES should be canonicalized using the same environment and located as one SMILES per line in the stocks folder. The file should be referenced in the config file as "commercial_file_path".

Usage for Multistep Prediction

Edit the /configs/config_example.yaml configuration file, change the target compound (target_cpd) and tree search parameters as needed. Start the multistep search as follow:

conda activate MultiStepRetro_ENZ
retrosynthesis --config configs/config_example.yaml

Visualizing Results

Results are written into output/project_name/ as pickle files. Forward validated single step-reaction predictions are stored as output/project_name/[DateJob]__prediction.pkl, and full predicted route paths are stored as output/project_name/[DateJob]__tree.pkl, which refers to reaction indexes from [DateJob]__prediction.pkl. Routes could be sorted by scores to get the best ones. Temporary checkpoints are written in the output/project_name/ folder after each iteration to monitor the progress of retrosynthesis, but also to resume a job starting from checkpoints. If logs are enabled, those are written into output/project_name/.

To visualize predicted routes, check this notebook /notebooks/visualize_results.ipynb or the following example:

import pandas as pd
import ttlretro.view_routes as vr

project_name = 'config_example'
log_time_stamp = 'YYYY_MM_DD__HHMMSS'

predictions = pd.read_pickle('output/{}/{}__prediction.pkl'.format(project_name, log_time_stamp))
tree = pd.read_pickle('output/{}/{}__tree.pkl'.format(project_name, log_time_stamp))
tree = vr.get_advanced_scores(tree=tree, predictions=predictions)

bests = vr.get_best_first_branches(
    tree=tree, 
    predictions=predictions, 
    num_branches=10, 
    score_metric='Fwd_Conf_Score'
)

vr.display_branch(
    branch_tree_index_or_list_rxn_id=0, 
    tree=bests, 
    predictions=predictions, 
    forwarddirection=True
)

Citations

This repository makes use of existing projects:

OpenNMT-py

Publication: OpenNMT: Neural Machine Translation Toolkit

OpenNMT technical report

@inproceedings{klein-etal-2017-opennmt,
    title = "{O}pen{NMT}: Open-Source Toolkit for Neural Machine Translation",
    author = "Klein, Guillaume  and
      Kim, Yoon  and
      Deng, Yuntian  and
      Senellart, Jean  and
      Rush, Alexander",
    booktitle = "Proceedings of {ACL} 2017, System Demonstrations",
    month = jul,
    year = "2017",
    address = "Vancouver, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P17-4012",
    pages = "67--72",
}

SCScore

Publication: SCScore: Synthetic Complexity Learned from a Reaction Corpus

GitHub repository: SCScore

@article{coley_scscore_2018,
	title = {{SCScore}: {Synthetic} {Complexity} {Learned} from a {Reaction} {Corpus}},
	author = {Coley, Connor W. and Rogers, Luke and Green, William H. and Jensen, Klavs F.},
	volume = {58},
	issn = {1549-9596},
	shorttitle = {{SCScore}},
	url = {https://doi.org/10.1021/acs.jcim.7b00622},
	doi = {10.1021/acs.jcim.7b00622},
	number = {2},
	urldate = {2022-09-02},
	journal = {Journal of Chemical Information and Modeling},
	month = feb,
	year = {2018},
	note = {Publisher: American Chemical Society},
	pages = {252--261},
}

Cite this work

Publication of the original Triple Transformer Loop: Multistep retrosynthesis combining a disconnection aware triple transformer loop with a route penalty score guided tree search.

@article{kreutter_multistep_2023,
	title = {Multistep Retrosynthesis Combining a Disconnection Aware Triple Transformer Loop 
		with a Route Penalty Score Guided Tree Search},
	author = {Kreutter, David and Reymond, Jean-Louis},
	url = {https://pubs.rsc.org/en/content/articlelanding/2023/sc/d3sc01604h},
	doi = {10.1039/D3SC01604H},
	date = {2023-09-20},
	journaltitle = {Chemical Science},
	shortjournal = {Chem. Sci.},
	volume = {14},
	number = {36},
	pages = {9959--9969},
	publisher = {{The Royal Society of Chemistry}},
}

Preprint of the Enzymatic work: Chemoenzymatic Multistep Retrosynthesis with Transformer Loops.

@article{kreutterChemoenzymaticMultistepRetrosynthesis2024,
  title = {Chemoenzymatic {{Multistep Retrosynthesis}} with {{Transformer Loops}}},
  author = {Kreutter, David and Reymond, Jean-Louis},
  date = {2024-04-12},
  eprinttype = {ChemRxiv},
  doi = {10.26434/chemrxiv-2024-svr99},
  url = {https://chemrxiv.org/engage/chemrxiv/article-details/6617b48391aefa6ce157c2b4},
  pubstate = {preprint}
}