Building A Graph Convolutional Network for Molecular Property Prediction

Artificial intelligence has taken the world by storm. Every week, new models, tools, and applications emerge that promise to push the boundaries of human endeavor. The availability of open-source tools that enable users to train and employ complex machine learning models in a modest number of lines of code have truly democratized AI; at the same time, while many of these off-the-shelf models may provide excellent predictive capabilities, their usage as black box models may deprive inquisitive students of AI of a deeper understanding of how they work and why they were developed in the first place. This understanding is particularly important in the natural sciences, where knowing that a model is accurate is not enough — it is also essential to know its connection to other physical theories, its limitations, and its generalizability to other systems. In this article, we will explore the basics of one particular ML model — a graph convolutional network — through the lens of chemistry. This is not meant to be a mathematically rigorous exploration; instead, we will try to compare features of the network with traditional models in the natural sciences and think about why it works as well as it does. <a href="https://towardsdatascience.com/building-a-graph-convolutional-network-for-molecular-property-prediction-978b0ae10ec4">Click Here</a>