Graph Neural Networks for Chemistry: Building Powerful Models Without Breaking the Laws of Physics

Johannes Klicpera
TU Munich

Recent advances in machine learning for Chemistry promise to speed up molecular property prediction, drug screening, inverse synthesis planning and other tasks by multiple orders of magnitude. Many of these recent advances are based on graph neural networks, which have established themselves as a powerful model for molecular and physical systems. By directly modeling the interactions between particles they originally promised to replace the sophisticated hand-crafted features prevalent in classical machine learning models (e.g. kernel ridge regression). However, while graph neural networks do indeed outperform classical methods on large benchmarks, they still struggle with generalization on tasks with limited available data.

In this talk I will first give a general overview of graph neural networks and their underlying principles and limitations. I will then focus on their application to molecular dynamics and present modern approaches that try to marry the best of both worlds by building physical knowledge directly into the model [1]. Doing so significantly improves both their overall performance and their generalization capabilities, and thus opens venues for research on molecules far away from the equilibrium [2].

References
[1] Johannes Klicpera, Janek Groß, Stephan Günnemann. Directional Message Passing for Molecular Graphs. ICLR 2020.
[2] Johannes Klicpera, Shankari Giri, Johannes T. Margraf, Stephan Günnemann. Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules. ML4Molecules workshop, NeurIPS 2020.

Speaker Bio
Johannes Klicpera is a PhD student in the Data Analytics and Machine Learning group lead by Stephan Günnemann at TU Munich. His research focuses on machine learning for relational data, from web-scale social networks to small molecules. Before starting his PhD he studied Computer Science and Physics at TU Munich and the University of Cambridge. More information on his research is available in his Google Scholar profile.