Molecular Set Transformer

A deep learning model, namely Molecular Set Transformer, was designed for enabling high-throughput co-crystal screening for any type of molecular pairs. The model is able to provide score and uncertainty for any given molecular pair based on its probability to form a multicomponent crystal.

Installation

We recommend installing the package by following the instructions below.

conda create --name cocrystals python=3.7.13
conda activate cocrystals
git clone https://github.com/lrcfmd/MolecularSetTransformer.git
cd MolecularSetTransformer
pip install -r requirements.txt

Or cloning the environment from the cocrystals.yaml file

conda create --name myclone --clone cocrystals.yml

Extract co-crystals from the Cambridge Structural Database (CSD)

You need the CSD licence (https://www.ccdc.cam.ac.uk/) to use this script. You can extract multicomponent crystals after screening the whole > 1 million CSD database. If you do not want to include the solvents , you can use the option --no_solvents

python extract_cocrystals/extract_cocrystals.py --no_solvents

Our extracted csd co-crystals dataset using CSD 2020 can be found in csd_data/csd_cocrystals2020.csv The training co-crystals dataset was created after removing the duplicate pairs, i.e., pairs found both in the csd_cocrystals2020 and in the benchmark data and is referred as csd_data/csd_training_set.csv

python extract_cocrystals/drop_duplicates.py

The code for generating the co-crystals tree map can be found in the TMAP-cocrystals folder. The PPD code for calculating the distances between the co-crystals can be found in https://github.com/dwiddo/average-minimum-distance

Benchmarks (Publically available co-crystal screening data)

All the is-silico co-crystal screening datasets gathered from literature can be found on the validation_database folder and are described below in chronological order:

	Dataset reference	Number or data	computational methods tested on these datasets
1	Karki et al, CrystEngComm, 2010, 12, 4038–4041	75 (73 negatives + 2 positives)
2	Grecu et al, Cryst. Growth Des., 2014, 14, 165–171	432 (300 negatives + 132 positives)	MEPS, COSMO-RS
3	Wicker et al, CrystEngComm, 2017, 19, 5336–5340	680 (408 negatives + 272 positives
4	Mapp et al, Cryst. Growth Des., 2017, 17, 163–174	108 (90 negatives + 18 positives)	MC
5	Przybyłek et al, Cryst. Growth Des., 2018, 18, 3524–3534	226 (58 negatives + 168 positives)
6	Przybyłek et al, Cryst. Growth Des., 2019, 19, 3876–3887	712 (104 negatives + 608 positives)
7	L. Roca-Paixão et al, CrystEngComm, 2019, 21, 6991–7001	21 (8 negative and 13 positives)
8	Aakeröy et al, Cryst. Growth Des., 2019, 9, 432–441	82 (17 negatives + 65 positives)
9	M. Khalaji et al, Cryst. Growth Des., 2021, 21, 2301–2314	19 (9 negatives and 10 positives)
10	Vriza et al, Chem. Sci., 2021, 12, 1702–1719	6 (4 negatives + 2 positives)
11	Devogelaer et al, Cryst. Growth Des., 2021, 21, 3428–3437	30 (18 negatives + 12 positives)
12	Wu et al, Cryst. Growth Des., 2021, 21, 4531–4546	63 (22 negatives + 41 positives)	COSMO-RS, MC
13	J. Yuan et al, CrystEngComm, 2021, 23, 6039–6044	16 (9 negatives + 7 positives)	COSMO-RS

Train your own models based on the prefered molecular representation

python train.py --model <model> --training_data <data> --save_dir <save_dir> -n_epochs <epochs> -lr <learning_rate>

• model The molecular embeddings to use. Choose among:

  • gnn
  • morgan
  • chemberta
  • mordred

• training_data The location of the .csv file containing the molecular pairs for training

• save_dir The directory to save the trained model

• n_epochs The number of epochs to use for training

• lr Learning rate

• use wandb We use Weight&Biases to monitor the hyperparameters during the training. This not a strict requirement but it is suggested. If you want to use it, you need to use the 'use_wandb' flag and to create an account here and install the Python package:

  pip install wandb
  

Example

python train.py --model gnn --training_data csd_data/csd_cocrystals2020.csv --save_dir pretrained_models --model_name gnn_trained -n_epochs 10 -lr 0.001 --use_wandb

Load a model after training

from load_pretrained_models import *
gnn_model = load_model('gnn', 'pretrained_models/model.pth')

Notebooks

• cocrystal_statistics.ipynb Contains the code to reproduce the plots regarding the statistical analysis of the current co-crystal related research

Comparing with other co-crystal screening methods

• ccgnet_reports Contains the outcome from the tests using the CCGnet model(github repo) -MEPS and COSMO-RS data

Use our pretrained models to get quick predictions for any given SMILES pair

Click on the link: Co-crystal screeening GUI

The full code and dependencies can be found to the following repository: https://github.com/katerinavr/streamlit