Metabolite Translator (MetaTrans)

Metabolite Translator (MetaTrans) is a deep learning, Transformer-based, architecture for predicting metabolites of small molecules in humans.

MetaTrans is developed using transfer learning: First, we trained a Transformer model for predcting the outcome of general chemical reactions. Subsequently, we fine-tuned it on human metabolic reactions. Finally, we constructed an ensemble model, combining multiple fine-tuned models. The methodology is described in detain in the paper Prediction of drug metabolites using neural machine translation

The specifications for pre-training the Transformer model were based on the Molecular Transformer for reaction outcome prediction.

The implementation of the Transformer model is based on the OpenNMT toolkit.

Processing of the chemical data is based on the RDKit software.

Installation

Create a conda environment:

conda create -n metatrans python=3.5
source activate metatrans
conda install rdkit -c rdkit
conda install future six tqdm pandas
conda install pytorch=1.1.0 torchvision -c pytorch
pip install torchtext==0.3.1
pip install -e .

Prediction of human metabolites for small molecules

Get Trained models

Step 1: Download trained models from this link and place thm inside the folder models.

Prepare data

Step 2: Prepare a csv with the molecules in SMILES notation (Sample input files are given in datasets/test/input.csv). The csv should contain the molecule id in the 1st column and the molecule SMILES in the 2nd column. Then prepare the data (canonicalise and tokenise SMILES) for translation:

python prepare_input_file.py -input_file ${infile} -output_file ${outfile}

input_file the name of the input csv file. output_file (optional) the name of the output txt file which will contain the processed data. Default: processed_data.txt

Translate

Step 3: Edit the bash file translate_molecules: define the source file (output of the previous step) and the beam size. The default beam size is 5 and the user can change it to get fewer or more predictions per molecule. A beam size of 5 approximatily gives top-10 predictions. A beam size of 2 gives top-5. A beam size of 10 gives top-20. The user can also define the min and max length of the predicted SMILES sequences. Then, translate the input molecules into metabolites:

./translate_molecules

Get predictions

Step 4: Obtain the predictions:

python process_predictions.py -input_file ${infile} -output_file ${outfile} -beam_size ${beam} -visualise_predictions ${bool}

input_file the name of the input csv file (same as the input file in step 2) output_file (optional) the name of the output csv file where the predictions will be saved. Default: predicted_metabolites.csv beam_size (optional) the beam size which has to be the same as the beam size used in step 3. Default: 5 visualise_predictions (optional) if True then the predicted metabolites will be visualised using RDKit. Default: False.

This will generate a csv file with the predicted metabolites.

Datasets

The datasets of human metabolic transformations we constructed for training, validating and testing MetaTrans are in the folder datasets.

Citation

@article{metatrans,
  author = {Litsa, Eleni E. and Das, Payel and Kavraki, Lydia E.},
  title = {Prediction of drug metabolites using neural machine translation},
  journal = {Chemical Science},
  year = {2020},
  month = sep,
  publisher = {The Royal Society of Chemistry},
  doi = {10.1039/D0SC02639E},
  url = {http://dx.doi.org/10.1039/D0SC02639E},
  issue = {11},
  pages = {12777-12788}
}