The QM9 dataset [1] contains information about 134k organic molecules containing Hydrogen (H), Carbon (C), Nitrogen (N) and Fluorine (F). For each molecule, computational quantum mechanical modeling was used to find each atom's “positions” as well as a wide range of interesting and fundamental chemical properties, such as dipole moment, isotropic polarizability, enthalpy at 25ºC, etc.

The model presented in this example follows the GNN architecture used in [2], which consists of:

• Feed-forward neural network to build atom to atom messages using the hidden states along with edge information (atom to atom distance and bond type).
• Gated Recurrent Unit (GRU) to update atom's hidden states.
• Gated feed-forward neural network as readout to compute target properties.

By default we use as target the molecules' dipole moment, but the provided data contains all molecule properties in the original dataset to explore other options.

For this example you can find the directory data containing a very small subset of the dataset. In particular, 1000 molecules for training and 100 for validation. In addition, we have included the rest of framework files properly filled, and thus there are no prerequisites to execute it.

To train the corresponding QM9 GNN with the default settings, just run:

    python main.py

This command will create the GNN specified in model_description file, with the variables specified in the global_variables file. To learn more about the implementation details, refer to the framework documentation.

If you want to execute any other functionality that is not train and validate, simply change the main file, see the documentation page for more details.

Although minimal example data is provided, one can generate bigger datasets by executing the generate_dataset Python script. This downloads the QM9 dataset and processeses a subset of molecules to create the associated NetworkX graphs, writing them as arrays in JSON files.

Through changing the global variables in the top of the script one can change the amount of training/validation samples, and other properties.

Please note that the script requires specific dependencies to generate some atomical features which the model requires, see the script's docstring for more details.

1. Ramakrishnan, Raghunathan; Dral, Pavlo; Rupp, Matthias; Anatole von Lilienfeld, O. (2014): Quantum chemistry structures and properties of 134 kilo molecules. figshare. Collection. https://doi.org/10.6084/m9.figshare.c.978904.v5

2. Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. 2017. Neural message passing for Quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning - Volume 70 (ICML'17). JMLR.org, 1263–1272.