Multitask Graph Convolutional Networks for Molecular Property Prediction

Overview Our project seeks to combine multitask learning with graph convolutional networks to build a network capable of predicting molecular properties with high speed and accuracy. It builds upon the Spektral library and uses Psi4 for as a baseline for speed comparisons.

Required Packages and Installation Instructions

1. We recommend the use of the conda package manager, available from Anaconda, for easy installation of packages. In this project, we used Python 3.7 and machine learning and scientific computing libraries Tensorflow, Scikit-Learn, and matplotlib. These may be installed with conda install tensorflow scikit-learn matplotlib.
2. Spektral - Spektral is a library built on top of Tensorflow and Keras including packages for graph learning and molecular datasets. The documetation is found at https://danielegrattarola.github.io/spektral/. To install on Ubuntu, install the required dependencies using sudo apt install python3-dev graphviz libgraphviz-dev libcgraph6 pkg-config and then run pip install spektral. The complete installation instructions are found at https://pypi.org/project/spektral/
3. Psi4 - Psi4 is a suite of program for chemical and molecular simulation and property prediction. Psi4 can be installed into Anaconda by running conda install psi4 psi4-rt -c psi4. The complete installation instructions are found at http://www.psicode.org/psi4manual/master/build_obtaining.html
4. RDKit - RDKit is a cheminformatics library that provides various chemical visualization and conversion tools. To install in Anaconda, run conda install rdkit -c rdkit. The complete installation instructions can be found at https://rdkit.org/docs/Install.html

Dataset We are learning from the QM9 dataset, which has chemical properties for 134k organic molecules with up to 9 heavy atoms (CONF). More information about this dataset and download instructions can be found at http://www.quantum-machine.org/datasets/. Note that graphical representations of the molecules come bundled with Spektral, but the true geometries are necessary for density-functional computation with Psi4.

Build and Run Instructions All of our files are interactive Python notetbooks and may be downloaded and run in Jupyter Notebook or Google Colab.

Paper Our final writeup is included as final_paper.pdf above.