Potential Energy Prediction with Persistence Images

CHEM 420 final project: Owen Queen, Sai Thatigotla, and Henry Eigen

Installing Dependencies

Install the package ChemTDA by using the setup.py file with pip:

pip install -e .

This will allow you to access the ChemTDA package, which is what we use to make the persistence images. The next step is to install all dependencies:

pip install -r requirements.txt

XTB

To generate the conformations that we used, you will have to download XTB.

RDkit

This project also requires RdKit, which requires separate installation here.

How to run code

Code below is listed in order of appearance in the PowerPoint. Run all code from the PotentialEnergies directory unless instructed otherwise.

Generating Conformations

To generate the conformations, first change the path within the generate/md.sh to where you have the xtb executable downloaded.

Then, run the following:

cd generate
./md.sh <structure number>

"structure number" can be any integer between 1 and 40. This short shell script generates conformations with XTB and then parses and moves all files into the appropriate position in the data directory.

Visualize PI for Structure

To visualize a PI for a given input xyz file, run:

python3 show_PI.py <xyz path>

For example, you can shown a structure from base structure 5 with:

python3 show_PI.py data/xtb_data/base_5/STRUCTS/struct_3.xyz

TSNE Plot

To make the TSNE plot for molecules 1-5, run:

python3 models/tsne.py many

To make the TSNE plot just for base structure 1, with points highlighted by energy, run:

python3 models/tsne.py one

Distribution Plots

To make the plots of the distributions of energies, run

python3 models/energy_distribution.py 1
python3 models/energy_distribution.py 2

The command line argument specifies the base structure number for which to show the distribution.

Basic ML Models

To run the basic ML models, please run the Jupyter Notebook models/PI_basic_predictors.ipynb.

Convolutional Neural Networks

To run the CNN, please execute the following command:

python3 models/CNN.py <epochs>

Where the epochs are the number of training iterations for which you want to train the CNN. We recommend 100, but this will take about 5 minutes to train.

Note: Running the CNN is highly stochastic, so results may not match what was presented in the PowerPoint.

ChemRacer/PotentialEnergies