🌐 Code | 📃 Paper | 💻 Try it online

This repository includes the scripts used for the manuscript "Chemoenzymatic synthesis planning guided by synthetic potential score."

AceRetro (Asynchronous ChemoEnzymatic Retrosynthesis) is a computer-aided chemoenzymatic synthesis planning tool that uitilize Synthetic Potential to advance the synthesis planning algorithms.

Dependencies

• conda=22.9.0
• python=3.9.12
• numpy=1.21.5
• pytorch=1.12.1
• rdkit-pypi==2022.3.5
• map4==1.0 (MAP4 is not required if only ECFP4 embedding is used)

Operating systems

• Ubuntu 20.04.5 LTS

./sfscore/sfscore.py shows the basic usage. If you want to use SFScore elsewhere, follow the steps below.

Git clone this repository and cd to the project folder.

conda create -n sfscore_train python=3.9
conda activate sfscore_train
conda install -c tmap tmap
pip install -r ./process_reaction_database/requirements.txt

Create Conda environment though conda env create -f ./process_reaction_database/environment.yml

Activate Conda environment conda activate sfscore_train

sfscore_path = [PROJECT_DIRECTORY]/'sfscore' # replace PROJECT_DIRECTORY
import sys
sys.path.append(str(sfscore_path))
from sfscore import SFScore
sfscore_model = SFScore()
# load the model in 'process_reaction_database/saved_model/ecfp4_4096_3_layer_epoch10.pt' by default
sfscore_model.load() 
smiles = 'O=C(COP(=O)(O)O)[C@H](O)[C@H](O)CO'
sfscore = sfscore_model.score_from_smi(smiles)
print('SMILES:',smiles,', SFScore:',sfscore)
# SMILES: O=C(COP(=O)(O)O)[C@H](O)[C@H](O)CO , SFScore: [0.4322359 0.6146086]

The scripts and checkpoints of trained models are in ./process_reaction_database.

process_reaction_database/fingerprint_embedding.ipynb: Process USPTO and ECREACT database. Process ECFP4 and MAP4 fingerprint embedding. (~ 2h)

process_reaction_database/model_training.ipynb: SFScore model training and results processing. (~ 40 min)

evalueate_score/benchmark_zinc_one_step.ipynb: SFScore results and one-step retrosynthesis(RXN4Chemistry)/retrobiosynthesis(ASKCOS-Enzy) on 11K molecules from ZINC database (Cost ~ 43.5 hours) (Please refer to the next section for one-step retrosynthesis/retrobiosynthesis prediction).

evalueate_score/process_zinc_one_step.ipynb: Process one-step retrosynthesis results. (Fig. 2)

evalueate_score/benchmark_askcos_results.ipynb: SFScore results on pathways predicted by ASKCOS-hybrid (https://github.com/itai-levin/hybmind) in Nat Commun 13, 7747 (2022). (https://doi.org/10.1038/s41467-022-35422-y) (Fig. 3 and ASKCOS pathways in Fig. 5)

Installation of ASKCOS (Nat Commun, 13, 7747 (2022)):

• Download ASKCOS (https://github.com/xuanliugit/askcos-core/tree/aceretro) to ./pathway_search_standalone/askcos-core
• Download ASKCOS data (https://github.com/xuanliugit/askcos-data) to ./pathway_search_standalone/askcos-core/askcos/data

# Download ASKCOS
## Directory: ./pathway_search_standalone
git clone https://github.com/xuanliugit/askcos-core.git
## Change branch to aceretro
cd askcos-core
git switch aceretro

# Download ASKCOS data
## Directory: ./pathway_search_standalone/askcos-core/askcos/
git lfs clone https://github.com/xuanliugit/askcos-data.git
## Rename 'askcos-data' to 'data'
mv askcos-data data

Installation of RXN4Chemistry (Digital Discovery, 2023,2, 489-501):

• Download RXN4Chemistry (https://github.com/xuanliugit/rxn_cluster_token_prompt/tree/aceretro) to pathway_search_standalone/rxn_cluster_token_prompt
• RXN4Chemistry's model is the intellectual property of IBM, which is not open-source but available in https://rxn.res.ibm.com. To use the open-source models download them in pathway_search_standalone/rxn_cluster_token_prompt/models following this link.

# Download RXN4Chemistry
## Directory: ./pathway_search_standalone
git clone https://github.com/xuanliugit/rxn_cluster_token_prompt.git
## Change branch to aceretro
cd rxn_cluster_token_prompt
git switch aceretro

## Model Directory: ./pathway_search_standalone/rxn_cluster_token_prompt/models

# Create environmentt
conda create -n aceretro python=3.6.13
conda activate aceretro
pip install rdkit-pypi

# Install packages for ASKCOS
pip install -r ./pathway_search_standalone/askcos-core/requirements.txt

# Install packages for RXN4Chemistry
cd pathway_search_standalone/rxn_cluster_token_prompt
pip install -e .

# Install other packages
pip install pubchempy==1.0.4
pip install git+https://github.com/reymond-group/[email protected]
conda install -c tmap tmap

# ./pathway_search_standalone/
from scripts.search_utils import hybridSearch
hybridSearch = hybridSearch()
smiles = 'CC[C@@H](CO)NCCN[C@@H](CC)CO' # Target molecule's SMILES
# Conduct a 3 min asynchronous chemoenzymatic synthesis planning
explored_rxns, explored_nodes, start_node = hybridSearch.get_chemoenzy_path_async(smiles, max_depth=10, chem_topk=10, max_num_templates=250, max_branching=15, time_lim=180)

To process the search result:

# ./pathway_search_standalone/
from scripts.search_utils import build_graph_from_async
import networkx as nx
graph_json = build_graph_from_async(explored_rxns,explored_nodes,start_node)
g = nx.node_link_graph(graph_json) # reaction graph
# Check ./pathway_search_standalone/result_processing.ipynb to extract synthesis routes from reaction graph 

pathway_search_standalone/search_rxn4chemistry_askcos.ipynb: Search hybrid pathways to 1k molecules from ZINC (Fig. 5) (around 6.25 days) and case studies (Fig. 6,7,8).

pathway_search_standalone/result_processing.ipynb: Process the search results and translate graphs to readable pathways (less than 5 min)