deepsa 0.1.2

A Deep-learning Driven Predictor of Compound Synthesis Accessibility

With the continuous development of artificial intelligence technology, more and more deep-generation models are used for molecule generation. However, most new molecules generated by the generation models often face great challenges in terms of synthetic accessibility.

DeepSA is a deep learning-based tool for predicting the synthetic accessibility of compounds, helping users evaluate the synthesis difficulty of molecules to select more easily synthesizable molecules for drug discovery and development. DeepSA has a much higher early enrichment rate in discriminating molecules that are difficult to synthesize. This helps users to select less expensive molecules for synthesis, thus reducing the time for drug discovery and development.

Installation

Requirements

Dependencies can be installed using the following command:

conda create -n DeepSA python=3.12
conda activate DeepSA
# for gpu version
pip3 install autogluon==1.2
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
pip3 install rdkit

Install via pip

You can install DeepSA directly using pip:

pip install deepsa

News

• 2024-12, because AutoGluon stopped supporting python version 3.8 starting in October 2024. Therefore, we have updated DeepSA to use Python version 3.12 and updated the training and inference scripts to adapt to the latest version of AutoGluon, thanks for your interest in DeepSA!

• 2023-7, DeepSA_v1.0 has been released, welcome to provide feedback on the issue!

Data

The expand training and test datasets could be easily downloaded at https://drive.google.com/drive/folders/1iup6T3Bqyy-uvpdFyP0Of_WQqn-9l62h?usp=sharing

Usage

Python Package Usage

1. Predict synthetic accessibility of a single SMILES

from deepsa import predict_sa

# Predict a single SMILES
result = predict_sa("CCO")  # Ethanol
print(f"Synthetic accessibility score: {result['SA_score']:.4f}")
print(f"Heavy atom count: {result['HA_num']}")
print(f"Ring count: {result['Ring_num']}")
print(f"Ring system count: {result['RingSystem_num']}")
print(f"Rule of five compliance: {result['rule_of_five']}")

2. Predict synthetic accessibility of multiple SMILES

import pandas as pd
from deepsa import predict_sa_from_file

# Create DataFrame containing SMILES
smiles_list = ["CCO", "c1ccccc1", "CC(=O)OC1=CC=CC=C1C(=O)O"]
df = pd.DataFrame({"smiles": smiles_list})

# Predict and save results
results = predict_sa_from_file(df, output_path="results.csv")
print(results[["smiles", "easy", "hard"]])

3. Predict from CSV file

from deepsa import predict_sa_from_file

# Predict from CSV file (file must contain smiles column)
results = predict_sa_from_file("compounds.csv")

4. Command line usage

# Predict a single SMILES
deepsa-predict "CCO"

# Predict SMILES from a CSV file
deepsa-predict compounds.csv --output results.csv

Usage For Researchers

If you want to train your own model, you can run it from the command line,

running:

python DeepSA_training.py <dataset.csv/training.csv:test.csv> DeepSA_model ./data/test_set.list

If you want to use the model we proposed,

running:

python DeepSA_predict.py <input_data.csv> DeepSA_model

Online Server

We deployed a pre-trained model on a dedicated server, which is publicly available at https://bailab.siais.shanghaitech.edu.cn/deepsa, to make it easy for biomedical researcher users to utilize DeepSA in their research activity.

Users can upload their SMILES or csv files to the server, and then they can quickly obtain the predicted results.

Citation

If you find this repository useful in your research, please consider citing our paper:

Wang, S., Wang, L., Li, F. et al. DeepSA: a deep-learning driven predictor of compound synthesis accessibility. J Cheminform 15, 103 (2023). https://doi.org/10.1186/s13321-023-00771-3

Contact

If you have any questions, please feel free to contact Shihang Wang (Email: [e-mail]) or Lin Wang (Email: [e-mail]).

Pull requests are highly welcomed!

Acknowledgments

We are grateful for the support from HPC Platform of ShanghaiTech University.

Thank you all for your attention to this work.