This repository contains the code for the prediction of triplet energies and spin populations of organic molecules as described in this paper.
The models can also be used in a web application: entdecker.uni-muenster.de
For installation run
git clone --recurse-submodules https://github.com/le-schlo/EnTdecker.git
# option --recurse-submodules is required in order to clone the respective submodules
#go to EnTdecker directory and install all required dependencies
cd EnTdecker
pip install -r requirements.txt
#Additionally you need to go to the EasyChemML directory and install the necessary dependencies separately.
cd EasyChemML
pip install ./
The datasets can be found in Data/EnT_data.csv and Data/EnT_SD_data.csv, for the triplet energy
prediction and the spin population prediction, respectively. The triplet energy values in Data/EnT_data.csv are given in kcal/mol. Data/EnT_SD_data.csv contains the input and target strings, to which the spin population for each heavy atom is inserted. Those files can be directly used for training the models. The dataset for finetuning the model can be found in Data/Retrain_Disulfide.csv for the triplet energy prediction and in Data/Retrain_SD_Disulfide.csv for the spin population prediction. Data/Pretrain.csv contains the data
for pretraining the sequence-to-sequence model on the canonicalisation task.
The folder DataProcessing contains helper functions and scripts for obtaining the splits for the triplet
energy prediction (RandomSplitter.py, MurckoSplitter.py, SimilaritySplitter.py, OutOfSample.py) as
well as the data processing for the spin population prediction. The file SpinExtraction.py contains a
function that returns the input for the sequence-to-sequence model given a smiles and the
corresponding spin population. The file TargetRearrangement.py contains a function that returns
rearranged SMILES maintaing the correct order of binned spin population.
Sample scripts to train the models can be found in the Models directory and the respective sub-directories.
Scripts for training the models for predicting the triplet energies can be found in
Models/triplet_energy. For training an AttentiveFP-GNN model the file
Models/triplet_energy/AttFP/train.py can be run. This code was slightly adapted from the original
repository. For training a chemprop-D-MPNN model
the file Models/triplet_energy/chemprop/train.py can be run. Information about defining additional
hyperparameters can be found in the documentation of chemprop.
For training a transformer-CNN model the file
Models/triplet_energy/transformer-cnn/transformer-cnn.py can be run. All parameters for training the
model need to be specified in the Models/triplet_energy/transformer-cnn/config.cfg file. Please note
that running the code for the transformer-CNN code requires different dependencies than all other
code. A tutorial on how to set the environment up can be found in the
Models/triplet_energy/transformer-cnn/README.md file. For training a CatBoost-GDBT model with
the MFF fingerprint the file Models/triplet_energy/CatBoost/train.py can be run.
Scripts for training and evaluating the sequence-to-sequence model for predicting the spin population
can be found under Models/spin_population/train.py and Models/spin_population/eval.py.
The file Models/spin_population/settings.json contains parameters for running the training as well as for running the evaluation.
“dir_name”: provide the path to the directory were the code is executed. This ensures that
the model checkpoints and the results of the evaluation are saved in the correct
location
“file_path”: provide the path to the data file used for training when using running the
train.py or for evaluation when running the eval.py
“src_vocab_size”: specify the size of the vocabulary of the inputs. Default: 106
“src_len”: specify the maximum length of the input sequence. Default: 100
“epochs”: specify the number of epochs used for training the model
“batch_size”: specify the batch size during training and evaluation
“print_every”: specify the frequency an output is printed
“loss_fname”: specify the name of the file the loss will be saved in during training.
Pretrained models can be downloaded from Zenodo.
Sample scripts to use these for obtaining predictions can be found under Models/triplet_energy/chemprop/eval.py and Models/spin_population/eval.py.
@article{doi:10.1021/jacs.4c01352,
author = {Schlosser, Leon and Rana, Debanjan and Pflüger, Philipp and Katzenburg, Felix and Glorius, Frank},
title = {EnTdecker − A Machine Learning-Based Platform for Guiding Substrate Discovery in Energy Transfer Catalysis},
journal = {Journal of the American Chemical Society},
volume = {146},
number = {19},
pages = {13266-13275},
year = {2024},
doi = {10.1021/jacs.4c01352},
note ={PMID: 38695558},
URL = {https://doi.org/10.1021/jacs.4c01352},
eprint = {https://doi.org/10.1021/jacs.4c01352}
}