Thermal barriers of azobenzene derivatives

This repository contains code for computing the thermal barriers of azobenzene derivatives. The code base uses a neural network force field (NFF) to compute energies and gradients. These are in turn used to optimize reactants, products, transition states, and singlet-triplet crossing points for intersystem crossing.

If you use this repository, please cite our paper:

@article{axelrod2023thermal,
  title={Thermal half-lives of azobenzene derivatives: virtual screening based on intersystem crossing using a machine learning potential},
  author={Axelrod, Simon and Shakhnovich, Eugene and Gomez-Bombarelli, Rafael},
  journal={ACS Central Science},
  volume={9},
  number={2},
  pages={166-176},
  year={2023},
  publisher={ACS Publications}
}

Conda environment

We recommend creating a conda environment to run the code. You can learn more about managing conda environments by reading this page.

To create the envionrment, please do the following:

1. Run ./setup.sh to create the environment.
2. Run source ~/.bashrc (linux) or source ~/.bash_profile (Mac).

You're now ready to use the code base! Watch out for an SSH permission error when cloning NeuralForceField. If this arises, please authenticate your SSH keys, and then run cd .. && git clone [email protected]:learningmatter-mit/NeuralForceField.git.

If something else goes wrong, please see this file.

Tutorials

Jupyter notebook tutorials show how to load and interpret our published barrier data. The data can be found here. The tutorials also show how to load any data you may generate yourself. To learn how to generate your own data with pre-made scripts using our models, see the Examples section in this document.

Examples

An example calculation can be found in the examples folder. To test it out, run the following code on the command line:

cd examples
./run.sh

To run this for your own molecules, simply supply their SMILES strings in the file examples/job_info.json:

"smiles_list": [...]

The script should produce a series of calculations for two different molecules. The calculations include:

• Initial 3D structure generation through RDKit
• 4 relaxed scans per molecule to generate 4 possible transition states (TSs) of different mechanisms
• Metadynamics-based conformer generation to generate reactant, product, and TS conformers
• Eigenvector following to optimize the TSs
• Hessian calculations on the optimized reactants and products
• Singlet-triplet minimum energy crossing point search
• Intrinsic reaction coordinate generation for the optimized TSs
• Single point $S_0/S_1$ gap calculations on the optimized cis and trans geometries

Note that you only need to provide one cis or trans SMILES per molecule. Optionally, can also change the directory of your singlet neural network model (weightpath), the directory of your triplet model (triplet_weightpath), the directory of your $S_0/S_1$ gap model (s0_s1_weightpath), the device you want to use (cpu if you have no GPUs, or the index of the GPU you want to use), and the number of parallel jobs to run at once for each of the configs (num_parallel).

The final results are stored in examples/summary.pickle. Tutorials show how to load, visualize, and interpret the results. They also go into some detail about other parameters you can specify in job_info.json

Pre-trained models

A set of pre-trained models can be found in models.