directmultistep 1.1.0

No project description provided

DirectMultiStep: Direct Route Generation for Multi-Step Retrosynthesis

Overview

This work has been published in J. Chem. Inf. Model. The preprint for this work was posted on arXiv.

You can use DMS models without installation through our web interface at models.batistalab.com. Or, if you want, you can install the package from pypi pip install directmultistep. Check out dms.batistalab.com for full documentation.

How to use

Here's a quick example to generate a retrosynthesis route (you can get relevant checkpoints by running bash download_files.sh).

from directmultistep.generate import generate_routes
from pathlib import Path

data_path = Path(__file__).resolve().parents[1] / "data"
ckpt_path = data_path / "checkpoints"
fig_path = data_path / "figures"
config_path = data_path / "configs" / "dms_dictionary.yaml"

# Generate a route for a target molecule
target = "CNCc1cc(-c2ccccc2F)n(S(=O)(=O)c2cccnc2)c1"
starting_material = "CN"

# Find routes with different models:
# Using flash model with starting material
paths = generate_routes(
    target, 
    n_steps=2, 
    starting_material=starting_material, 
    model="flash", beam_size=5,
    config_path=config_path, ckpt_dir=ckpt_path
)

# Or use explorer model to automatically determine steps
paths = generate_routes(
    target,
    starting_material=starting_material,
    model="explorer",
    beam_size=5,
    config_path=config_path, ckpt_dir=ckpt_path
)

See use-examples/generate-route.py to see more examples with other models. Other example scripts include:

• train-model.py: Train a new model with customizable configuration for local or cluster environments
• eval-subset.py: Evaluate a trained model on a subset of data
• paper-figures.py: Reproduce figures from the paper
• visualize-train-curves.py: Plot training curves and metrics

Citing

If you use DirectMultiStep in an academic project, please consider citing our publication in J. Chem. Inf. Model:

@article{directmultistep,
    author = {Shee, Yu and Morgunov, Anton and Li, Haote and Batista, Victor S.},
    title = {DirectMultiStep: Direct Route Generation for Multistep Retrosynthesis},
    journal = {Journal of Chemical Information and Modeling},
    volume = {0},
    number = {0},
    pages = {null},
    year = {0},
    doi = {10.1021/acs.jcim.4c01982},
    URL = {https://doi.org/10.1021/acs.jcim.4c01982},
    eprint = {https://doi.org/10.1021/acs.jcim.4c01982}
}

Extra Materials

Through download_files.sh you can download canonicalized versions of eMols (23M SMILES), Buyables (329k SMILES), ChEMBL-5000 (5k SMILES), and USPTO-190 (190 SMILES). Using pre-canonicalized version saves you roughly a day of cpu time. If you happen to use these canonicalized versions, consider citing the repo from figshare:

@misc{shee2025figshare,
    author = {Yu Shee and Anton Morgunov},
    title = {Data for ``DirectMultiStep: Direct Route Generation for Multistep Retrosynthesis''},
    year = {2025},
    month = {3},
    howpublished = {\url{https://figshare.com/articles/dataset/Data_for_DirectMultiStep_Direct_Route_Generation_for_Multistep_Retrosynthesis_/28629470}},
    doi = {"10.6084/m9.figshare.28629470.v1"},
    note = {Accessed: 20xx-xx-xx}
}

Also check out the HigherLev Retro repo which is the source of the Buyables stock set. route-distances is the source of ChEMBL-5000. Retro* is the source of the eMols stock set and USPTO-190.

Licenses

All code is licensed under MIT License. The content of the pre-print on arXiv is licensed under CC-BY 4.0.