taoshen99/AutoMolDesigner

The official open-source repository for AutoMolDesigner, an easy-to-use Python application dedicated to automated molecular design.

AutoMolDesigner V1.1

Introduction

AutoMolDesigner is a computer-aided drug design software that enables automated design and screening of drug-like molecules through combination of chemical language model and automated machine learning (AutoML) framework. This open-source repository holds its command-line interface (CLI) version. It was developed on Linux (Ubuntu 22.04 LTS) with Python3，and it has also been tested on Windows and MacOS. Moreover, we provide the graphical user interface (GUI) version of this software that has been tested on Windows and MacOS. The ready-to-use software packages are freely available at Zenodo, wherein the user manual along with all datasets and trained models is also available.

Science News

Here are two pieces of Chinese news promoting our work:

1. ComputArt (WeChat Channel): AutoMolDesigner - Leveraging AI for automated small-molelcule antibiotic discovery (BaiduNetDisk provided for Chinese users)
2. ISTIS Shanghai: Deep application of AI in biomedicine

Requirements

• This tool supports CUDA-enabled GPU and can also run on CPU with limited performance.
• Basic requirements are Python>=3.8, pytorch>=1.13.1 and autogluon>=0.8.2, other packages including rdkit, molvs, descriptastorus, dimorphite-dl, openpyxl and so on.

Installation

Hereinafter, the command lines are bash scripts.

0. Install Anaconda and Git.
1. Clone this repository and navigate to it:

$ git clone https://github.com/taoshen99/AutoMolDesigner.git
$ cd AutoMolDesigner

2. Create the conda environment called automoldesigner:

$ conda env create -f environment.yml

3. Activate the environment:

$ conda activate automoldesigner

Usage

This software is composed of two sub-modules, i.e., deep molecular generation and molecular property prediction. The first module aims at establishing focused virtual libraries with generative deep learning, and the second module employs AutoML-based predictive model to prioritize the most promising molecules from virtual libraries. The combination of two sub-modules can facilitate experimental chemists to fast conceptualize novel candidates for new projects. However, they can also take advantage of the independent module to train customized models to satisfy various demands.

Deep molecular generation

This software implements deep molecular generation with chemical language model based on recurrent neural networks. The complete workflow includes Pretrain, Finetune and Sample. To build the model and sample molecules from scratch, one should execute the following commands in sequence. However, one can also run each module independently as long as the corresponding requirements are met.

Pretrain

Model pretraining aims at learning the basic grammar of chemical language, and moreover, excavating the intrinsic features of drug-like molecules. We have provided the pretrained model (trained with millions of drug-like small molecules from ChEMBL32). To pretrain your own model, you should have the drug-like molecules (hundreds of thousands to millions):

1. Deposit and curate the molecules for pretraining (datasets/pretrain-chembl-mini.smi raw molecules for pretraining, one SMILES sequence per row):

$ python prep_smi.py -i datasets/pretrain-chembl-mini.smi -o datasets/pretrain-chembl-mini_preped.smi -a 5 -u

Curation includes:

• Standardizing molecules, removing duplicates, salts and stereo information.
• Using -a to declare data augmentation with SMILES Enumeration (herein, augment five fold).
• Using -u to declare neutralization of molecules.

2. Compile the config for pretraining: template_pretrain.json under template : rename prj_name with your project name, change pretrain_data to your path of pretraining data datasets/pretrain-chembl-mini_preped.smi and leave other parameters as default (recommended).

3. Do pretraining with compiled config:

$ python pretrain.py

After each iteration of the whole data, the model will be deposited at projects/{prj_name}/ if the validation loss decreases. The pretraining process will be terminated if no progress of validation loss is made during five consecutive epochs. The best model is deposited as CLM-pretrain_best-{epoch}-{validation loss}.pkl

Finetune

Model finetuning is an effective strategy to navigate molecular generation to the focused chemical space of interest. To finetune your own model, you should have both the pretrained model and molecules of interest (tens to thousands). Herein, we take the generation of virtual library for antibacterial small molecules against Escherichia coli (E. coli) as an guiding example:

1. Deposit and curate the molecules for finetuning (In this case, we have prepared the dataset for finetuning; see manuscript for detailed description):

2. Compile the config for finetuning: template_finetune.json under template: rename prj_name with your project name, change finetune_data to your path of finetuning data and change pretrain_model to your path of pretrain_model. You may adjust the batch_size according to your data size. Leave other parameters as default (recommended).

3. Do finetuning with compiled config:

$ python finetune.py

After each iteration of the whole data, the model will be deposited at projects/{prj_name}/ if the validation loss decreases. The finetuning process will be terminated if no progress of validation loss is made during five consecutive epochs. The best model is deposited as CLM-finetune_best-{epoch}-{validation loss}.pkl

Sample

Model sampling is to generate the desired number of molecules represented by SMILES. To sample molecules, you should have the trained model (either pretrained one or finetuned one):

1. Compile the config for sampling: template_sample.json under template: rename prj_name with your project name and change trained_model to your path of trained model. You may adjust the sampling_num to obatin specific number of molecules. The parameter sampling_temp affects the validity and novelty of sampled SMILES, detailed illustration can be referred to Gupta et al..

2. Do sampling with compiled config:

$ python sample.py

The sampled SMILESs are available at projects/{prj_name}/sampled_SMILES.smi.

Postprocess

This additional module provides curations for raw sampled molecules in order to remove invalid and duplicated SMILES: Check Validity and Uniqueness:

$ python post_process.py -i raw_sampled.smi -o curated_sampled.smi

To check Novelty additionally, you should provide the molecules for training with -k:

$ python post_process.py -i raw_sampled.smi -o curated_sampled.smi -k known_training.smi

The definitions of Validity, Uniqueness and Novelty can be referred to Polykovskiy et al..

Molecular property prediction

This software implements molecular property prediction with AutoGluon, an open-source AutoML framework developed by Amazon Co. Ltd.,. It takes molecular descriptors as input and can be applied for both binary classification tasks and regression tasks. Two modes are provided, i.e., retrospective benchmarking and prospective prediction. Herein, we provide a case study on screening active antibacterial small molecules against E. coli as a basic guideline.

Data curation

The raw active data was retrieved from ChEMBL32, and further curated to make corresponding Maximal Unbiased Benchmarking Datasets (MUBD) with MUBD-DecoyMaker 2.0 (see manuscript for detailed description). The curated datasets were provided as MIC_train.xlsx. User can still take prep_smi.py to curate raw data (this script also supports .csv and .xlsx file, see template) and -p is additionally provided to protonate molecules at a certain range of pH.

Model training

The training config is provided as template_ag_train.json . To customize your training, follow the instructions below:

argument	description
prj_name	name of the project that will be created under the directory of projects
train_data	directory of tabular data file for training, it should contain two columns, the first column is named "SMILES" containing prepared molecules and the second is named "label" containing binary label (0, 1) or continuous numerical value; support file suffix with csv and xlsx
test_size	[0, 1), fraction of test data split from input train data, 0 means no data will be held out for testing
features	molecular descriptors, one of {'ecfp4', 'fcfp6', 'maccs', 'rdkit_2d', 'rdkit_2d_norm'}
eval_metric	metrics used to evaluate model performance during training, one of {'f1', 'roc_auc', 'mean_absolute_error', 'root_mean_squared_error'}
quality	presets of AutoGluon model quality, better quality consumes more resources, one of {'medium quality', 'good_quality', 'high_quality', 'best_quality'}
time_limit	maximum amount of time (in seconds) for model training
opt_deploy	whether to perform deployment optimization

$ python ag_train.py

The trained AutoGluon models will be saved as projects/{prj_name}/ag_models. If test_size is not zero, the split train and test data will be saved as ag_train.csv and ag_test.csv.

Model prediction

The predicting config is provided as template_ag_predict.json . To predict new data or perform benchmarking, follow the instructions below:

argument	description
prj_name	name of the project that will be created under the directory of projects
smiles_dir	directory of new data for prediction or test data for benchmarking, if new data prediction required, the data file should be plain text with one SMILES sequence per row, and end with .smi; if benchmark required, the data file should be test data generated during model training, for example, ag_test.csv (also support .xlsx)
model_dir	directory of AutoGluon models, for example, projects/{prj_name}/ag_models
features	molecular descriptors used in model training, one of {'ecfp4', 'fcfp6', 'maccs', 'rdkit_2d', 'rdkit_2d_norm'}

$ python ag_predict.py 

If the model is used for new data prediction, the prediction results will be saved under the directory of projects/{prj_name} ending with .csv and .sdf. Otherwise, if the model is used for benchmark, the model performance will be saved as projects/{prj_name}/metrics.csv.

Acknowledgements

The AutoML framework in this work adopted AutoGluon, see its documentation for advanced usage. Please cite their reported work AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data if molecular property prediction module in this application is used.

We appreciate the developers of Dimorphite-DL. We directly use that computational tool to protonate raw actives. It is highly recommended to cite their publication Dimorphite-DL: an open-source program for enumerating the ionization states of drug-like small molecules, if you use this module in your work.

The tokenization method for SMILES and Temperature Sampling technique referred to LSTM_Chem that was developed by Gupta et al., who contributed to the early establishment of chemical language model.

Citation

If you use AutoMolDesigner or related materials, please cite:

Shen, T.; Guo, J.; Han, Z.; Zhang, G.; Liu, Q.; Si, X.; Wang, D.; Wu, S.; Xia, J., AutoMolDesigner for Antibiotic Discovery: An AI-Based Open-Source Software for Automated Design of Small-Molecule Antibiotics. J. Chem. Inf. Model. 2024, 64, 575-583.

or BibTex:

@article{doi:10.1021/acs.jcim.3c01562,
author = {Shen, Tao and Guo, Jiale and Han, Zunsheng and Zhang, Gao and Liu, Qingxin and Si, Xinxin and Wang, Dongmei and Wu, Song and Xia, Jie},
title = {AutoMolDesigner for Antibiotic Discovery: An AI-Based Open-Source Software for Automated Design of Small-Molecule Antibiotics},
journal = {Journal of Chemical Information and Modeling},
volume = {64},
number = {3},
pages = {575-583},
year = {2024},
doi = {10.1021/acs.jcim.3c01562},
note = {PMID: 38265916},
URL = {https://doi.org/10.1021/acs.jcim.3c01562},
eprint = {https://doi.org/10.1021/acs.jcim.3c01562}
}