This code accompanies the paper Learning the Language of NMR: Structure Elucidation from NMR spectra using Transformer Models. The repo contains scripts to simulate NMR spectra, prepare the data and train models.
Create a conda environment for the scripts:
conda create -n nmr python=3.9
conda activate nmr
Clone the repository and instal via:
pip install .
For development install via:
pip install -e .[dev]
This section explains how to simulate NMR spectra using MestreNova. A working version of MestreNova is required.
Running the MestreNova script requires a .csv file containing SMILES of the molecules for which NMR spectra will be generated. This .csv file requires one columns, Smiles. The index of the smiles is used to name the output files. An 'example.csv' is provided. To run the simulations run the following script:
run_simulation --smiles_csv example/smiles.csv --out_folder example/simulation_out/1H --sim_type 1H --mnova_path <Absolute Path to your MestreNova executable> --script_path <Absolute path to the folder containing the MestreNova scripts>
The MestreNova simulation scripts can be found here [link to folder]. This script uses the MestreNova scripting tool to run the simulations. As a result all MestreNova will be opened and you will be able to see how spectra are simulated. A file is saved for each molecule.
To run the simulations for 13C simply replace the --sim_type with 13C.
After running the script the files need to be compiled into one dataframe making handling easier. This can be accomplished via the gather_data script. To gather the data from the 1H simulation from the above example use:
gather_data --results_folder example/simulation_out/1H --out_path example/simulation_out/results.pkl --sim_type 1H
This compiles the data into a format readable by the scripts to prepare the data for training. In the above example the dataframe only contains 1H NMR spectra. To include 13C NMRs as well, use the optional argument --add_to_existing_df upon which the additional NMR information will be added to the same dataframe and saved according to the path set with --out_path. Additionally, set the --sim_type to 13C.
The following scripts will use the data generated in the previous steps, format the NMRs, tokenize the strings and split into a train, test and validation set.
This script can be used to prepare training data for a model that predicts the structure from the NMR. Using the data from above, training data can be prepared as such:
prepare_nmr_input --nmr_data example/simulation_out/results.pkl --out_path example/training/1H --mode hnmr
The above command will prepare training data for a model that predicts the structure solely from a 1H NMR. Further options are available as described in the paper.
The following script prepares training data for the second task of predicting the correct molecule from a set given the NMR spectrum. Additionally, a file containg reactions from which to build the molecule sets is required. The script expects this file to be a pickled dataframe of one column Use the following syntax to run the script:
prepare_nmr_rxn_input --nmr_data example/simulation_out/results.pkl --rxn_data <path to Reaction file> --out_path example/training/1H --mode hnmr
Run a training of the model using the run_training.py script. This script requires a directory with a subfolder called data which contains the training, validation and test data as generated by the two scripts above. Another requirement is a path to a OpenNMT configuration template. The template with which all trainings were performed in the template is provided in src/nmr_to_structure/training/transformer_template.yaml.
For the example try:
train_model --template_path src/nmr_to_structure/training/transformer_template.yaml --data_folder example/train
Note: The template expects a GPU to be present and to run a real training much more data is required.
For testing purposes, training a model for only few steps with a tiny model on CPU only:
train_model --template_path src/nmr_to_structure/training/transformer_template_tiny_cpu.yaml --data_folder example/training/1H
Run the following command to do inference:
onmt_translate -model <model_path> -src <src_path> -output <out_file> -beam_size 10 -n_best 10 -min_length 5 -gpu 0
For the example (dummy model, valid set, CPU only), this corresponds to the following:
onmt_translate -model example/training/1H/model_step_5.pt -src example/training/1H/data/src-val.txt -output example/training/1H/pred-val.txt -beam_size 10 -n_best 10 -min_length 5
Scoring can be done via the score.py script:
python scripts/score.py --tgt_path <path to tgt file> --inference_path <out file from inference>
For the example above:
score_model --tgt_path example/training/1H/data/tgt-val.txt --inference_path example/training/1H/pred-val.txt
Use --n_beams if the beam size is changed from 10 during inference.