ChemistryFeaturization.jl

Unified graph building and featurizing for Weave.jl, AtomicGraphNets.jl, and (maybe soon) more!

Features

Graph-building and featurization from CIF files

Build graphs (as SimpleWeightedGraphs) from CIF files using PyCall to pymatgen functions
Visualization using GraphPlot, check out the visualize_graph function in the graph_functions.jl file, you can make pretty pictures like these, whether the graph is simpler or more complicated (thickness of connections indicates weight of edge in graph (higher weights for nearer neighbors)):

Flexible featurization (currently onehot-style) and decoding: choose features to include, level of discretization, etc., and directly decode feature vectors to check values:

SMILES input

Sean to add...

Requirements

Julia 1.4+
packages listed in Project.toml
In addition, you will need your PyCall to have access to the pymatgen package, which can be added using Conda.jl as: Conda.add("pymatgen"; channel="conda-forge"), as well as the rdkit package (Conda.add("rdkit"; channel="conda-forge"))

Terminological notes

There are a lot of seemingly similar terms used for quantities in this package that refer to disparate things. Here is my best attempt to clarify them as I use them here:

Feature: A quality or quantity associated with an atom that we wish to encode, such as atomic mass, row in the periodic table, etc.
Feature vector: The (typically one-hot-style) encoding of the values of a set of features associated with a particular atom. For example, if we were encoding the atomic mass (across five possible bins) and periodic table block (s, p, d, or f) of hydrogen, the associated feature vector would be [1 0 0 0 0 1 0 0 0] (the first five slots corresponding to atomic mass and the last four to block)
Featurization: either the process of assigning feature vectors to chemical elements, or a description of a scheme for doing so, encoded as a Vector of AtomFeat objects
Feature matrix: A collection of feature vectors associated with the atoms in a structure. Its shape should be (# features, # nodes).

Future Plans:

v0.1

graph-building and featurizing for CrystalGraphConvNets.jl
featurizing for Weave.jl
featurization and inversion
some basic documentation/examples

later...

saving out featurization metadata (features used, bins, etc.) in consistent format
consistent handling of numerical vs. categorical features
"hybrid" featurizations using features from multiple paradigms if available
more input file formats? e.g. SELFIES

z-q-y/ChemistryFeaturization.jl