Awesome

Magicmol

Implemention of paper : Magicmol - A light-weighted pipeline for drug-like molecule evolution and quick chemical space exploration

A copy of this work is available at https://zenodo.org/record/7776906#.ZCsbYcpBwUE

Magicmol provides a light-weighted pipeline for de novo novel molecule generation and evolution (combine with STONED-SELFIES), and can be utilized for either positive sampling or negative sampling. The process can be visulized in the following picture.

Data cleaning and preparation

Training Data

Original ChEMBL30 (https://www.ebi.ac.uk/chembl/) dataset, cleaned dataset are available at : https://drive.google.com/drive/folders/1ULI0ZxBk26EiCw_p0LnLB1ckgstWK0Hq?usp=sharing

190W: Original Chembl-30 small molecule dataset

database.cleaned.smi / database.smi : pure SMILES data / SMILES data cleaned by clean.py

database_smiles_0.25.pkl / database_smiles_0.5.pkl : random choosed 25% / 50% of the cleaned molecule data

Data Cleaning

python clean.py --in_path database.smi --out_path database_cleaned.smi --vocab_path ../vocab/chembl_selfies_vocab.yaml

Model Weight

The weight file is available at : https://drive.google.com/drive/folders/10LDOoQQzuL0XeDRdsWidhVsoN8W1wxFU?usp=sharing

Update: Model trained with SMILES is provided also in the above link.

Update: SMILES / DeepSMILES corpus and processing code are available now (see vocab). Temporarily our Model adopting DeepSMILES reaches a relatively lower performance (with approximately 70% validity for generating molecules, I may tried to fix it later).

However, we recommend you train the backbone model by yourself and leave it a cup of tea time (approximately 15 min on a single RTX3090) to finish this (XD) and experience how fast the process could be finished. The hardware for training and sampling can be altered as long as CUDA is available, and 6G video memory will also be affluent for training or sampling.

But how to run ?

Model Training

By default the training process is done by adopting SELFIES, more super parameters can be customized based on your requirements.

python train.py

For example, training with SMILES:

python train.py --num_embeddings 101 --vocab_path ./vocab/chembl_regex_vocab.yaml --which_vocab regex

SA Optimization

In our work, optimization is done by varying the SA score judged by SYBA from sampled molecules derived from the well-trained backbone model.

For positive optimization:

python synthetic_accessibility_modification.py --model_weight_path ../model_parameters/trained_model.pth --use_syba True --optim_task posi

For negative optimization:

python synthetic_accessibility_modification.py --model_weight_path ../model_parameters/trained_model.pth --use_syba True --optim_task nega

Sampling

Sampling with SELFIES corpus, for SMILES or DeepSMILES, change them to correspond to your local directory

python sample.py --num_samples 1024 --num_batches 100 --vocab_path vocab/chembl_selfies_vocab.yaml --model_weight_path model_parameters/trained_model.pth --which_vocab selfies

Testing

We provided a simple way for computing some metrics. See after_generate.py.

Acknowledgement

STONED-SELFIES is available at: https://github.com/aspuru-guzik-group/stoned-selfies

SYBA is available at https://github.com/lich-uct/syba

If you think our research is interesting, consider citing this paper? 👇

Chen, L., Shen, Q., & Lou, J. (2023). Magicmol: a light-weighted pipeline for drug-like molecule evolution and quick chemical space exploration. BMC bioinformatics, 24(1), 173. https://doi.org/10.1186/s12859-023-05286-0

Attention

Noted that our work was done by SELIFES 1.x, the above version may caused encoding problems.

Pickle needs python 3.8 or higher version for using pickle.HIGHEST_PROTOCOL.