Multimodal Spectroscopic Dataset

This project is designed to work with Multimodal-Spectroscopic-Dataset. It requires a specific conda environment and data files to run properly.

Installation

Clone the repository or download the project files by running

git clone https://github.com/rxn4chemistry/multimodal-spectroscopic-dataset.git

Conda install

Open a terminal or command prompt and navigate to the project directory. Create a new conda environment named multispecdata with the required dependencies:

conda create -n multispecdata python=3.10

activate the created conda enviroment and insall the required dependencies

conda activate multispecdata
pip install -r requirements.txt

Data Download

The project requires specific data files to run. You can download the data from the following link: https://zenodo.org/records/11611178

Extract the downloaded ZIP file into the project directory in the data folder.

Replicating experiments

XGBoost and CNN

To replicate the results presented in the paper for XGBoost and CNNs use execute the following two scripts:

run_cnn.sh
run_xgboost.sh

Transformer models:

For the transformer models we present a set of different experiments including predicting the structure of a molecule from spectra, the spectra from the molecule. For all experiments the training data was generated using the following script:

python benchmark/generate_input.py 
  # Paths
  --analytical_data <Path to the folder containing the dataset> 
  --out_path <Path where the data will be saved>

  # Flags controlling input 
  --formula <Flag: Wether or not to include molecular formula in the input>
  --h_nmr <Flag: Wether or not to include 13C-NMR in the input>
  --c_nmr <Flag: Wether or not to include 1H-NMR in the input>
  --ir <Flag: Wether or not to include IR spectra in the input>
  --pos_msms <Flag: Wether or not to include positive MS/MS spectra in the input>
  --neg_msms <Flag: Wether or not to include negative MS/MS spectra in the input>

  # Flags controlling target
  --pred_spectra <Flag: By default the target is the molecular structure as SMILES. If this flag is used this is reversed with SMILES as input and spectra as output.>

  --seed <Default: 3245>

To start training the model use the following script:

python benchmark/start_training.py 
  --output_path <Path to the out_path specified when generating the data.>
  --template_path <Path to the OpenNMT config template to use. Default: ./benchmark/transformer_template.yaml>
  --seed <Seed. Default: 3245>

With a trained model inference can be run using the following OpenNMT command:

onmt_translate 
  -model <model_path> 
  -src <src_path> 
  -output <out_file> 
  -beam_size 10 
  -n_best 10 
  -min_length 5 
  -gpu 0

To analyse the results use the following script:

python benchmark/analyse_results.py 
  --pred_path <Path to the predictions made with onmt_translate>
  --test_path <Path to the ground truth test data>

We have provided a script run_transformer.sh to train models used in the structure prediction experiments for 1H-NMR, 13C-NMR and the combination of 1H-NMR and 13C-NMR.

Citation

Cite our work as followes:

@article{alberts2024unraveling,
  title={Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry},
  author={Alberts, Marvin and Schilter, Oliver and Zipoli, Federico and Hartrampf, Nina and Laino, Teodoro},
  year={2024},
  url={https://arxiv.org/abs/2407.17492}, 
}

Acknowledgements

This publication was created as part of NCCR Catalysis (grant number 180544), a National Centre of Competence in Research funded by the Swiss National Science Foundation.