Bayesian neural networks for molecular discovery

The resurgence of deep neural networks in the past decade has led to large advances in the machine learning field and deep learning has provided impressive results in a wide range of applications including computer vision and natural language processing. However, the success of modern deep learning has mostly been limited to simple data types such as regular grids and simple sequences, i.e., images and text. In contrast, the recently developed graph neural networks (GNNs) promise to expand the modern deep learning toolbox to complex domains with rich relational structures, i.e., domains where data can be represented as (relational) graphs.

In particular, GNNs have a large untapped potential in molecular data science and are expected to play an important role in materials science and drug discovery. Molecules are naturally represented as graphs and GNNs can, for example, be used to aid the design of new molecules or to predict molecular properties, replacing resource-intensive quantum mechanical simulations.

However, deep learning models, including GNNs, are inherently overconfident and exhibit poorly calibrated predictive uncertainties. In other words, deep learning models do not know, what they do not know and for this reason it is difficult to know when they should be trusted.

This project aims to improve the calibration of deep learning models, and in particular GNNs, by developing novel uncertainty quantification methods for deep learning. We will apply the developed methods to molecular data science problems, and envision that improved calibration will enable trustworthy molecular property predictions and a more efficient exploration of the chemical compound space during the design of new molecules.