Google Research, Brain Team

 Molecules, the building blocks of the material world, come in different sizes (e.g., from small amino acids to large proteins) and with different symmetries (e.g., periodic for polymers). Observational data on molecules and their properties opens the doors to 1) property prediction and 2) inverse design, i.e. searching the molecular space and optimizing for a property of interest. But how do we convert molecules into a digital representation that is searchable, optimizable, and editable, while taking into account the vast diversity of molecular scales and symmetries? And once that is done, what are the suitable machine learning techniques for modeling molecules? In this primer, we will go over several examples of property prediction, drug design, and considerations for making discoveries in a small-data regime. While not comprehensive, we hope this primer serves as a starting point for different ways of thinking about molecules and modeling them.

MIA Talks Search