Exploring Chemical Reaction Space With Reaction Difference Fingerprints and Parametric t-SNE

Exploring Chemical Reaction Space With Reaction Difference Fingerprints and Parametric t-SNE
Mikhail G. Andronov,Maxim V. Fedorov,and Sergey Sosnin. Full text: https://pubs.acs.org/doi/full/10.1021/acsomega.1c04778

This is the repositary accompanying the paper. It contains a pytorch implementation of the parametric t-SNE algorithm and the pretrained models to visualize chemical reaction space.

Usage

Before using the code, create a new conda environment and install requred packages:

conda create -n <environment name> anaconda python=3.7
conda activate <environment name>
conda install -c conda-forge rdkit
conda install pytorch lightgbm

Training of new models

Use the script train.py to train a new model. This requires a train dataset in the form of a file with reaction SMILES and a config file. Rename the file example.config.yaml as config.yaml and set the instructions and parameters in it. The config file contains the following instructions and hyperparameters:

• "device": "cpu" or "cuda" - device to train a model on
• "seed": random seed for pytorch
• "save_model": a flag indicating whather to save the model or not
• "problem_settings":
  • "filename": path to the train dataset
  • "fp_method": "structural" or "difference" - what type of reaction fingerprints to use
  • "n_bits": the length of the reaction fingerprint
  • "fp_type": "MorganFP", "AtomPairFP" or "TopologicalTorsion" - three options for the type of fingerprints
  • "include_agents": whether to include agents in a reaction fingerprint or not
  • "agent_weight": agent weight in difference fingerprints if agents are included
  • "non_agent_weight": non-agent weight in difference fingerprints
  • "bit_ratio_agent": the ratio of agent bits in structural fingerprints
• "optimization":
  • "lr": learning rate for the Adam optimizer
• "training": section for training hyperparameters, which is used in model.fit_model function