PhD defence by Peter Bjørn Jørgensen

Deep learning methods for screening of molecules and materials

Discovery of new drugs as well as new and better materials for solar cells and batteries relies on fast screening of new molecules and materials using computationally costly physics computations. Through these computations the properties of the molecules can be determined without synthesizing them in reality. However, the current physics based methods are very computational costly making it infeasible to screen large numbers of candidate molecules and/or molecules that consist of many atoms.

Machine learning methods can be trained on big databases of already computed molecules and materials and learn to predict the properties of molecules that have never been seen before. We can describe a molecule by the atoms that it consists of and the bonds between atoms, which mathematically is represented as a molecular graph, but most machine learning algorithms are not well suited for learning from graph data. In the thesis Deep learning methods for screening of molecules and materials we propose new methods for learning from graph data. The methods are based on deep neural networks and we show that the accuracy of the proposed method is significantly better than previous methods. The machine learning models are also able to predict the properties of molecules and materials without knowing the exact spatial structure, but can rely on just the atoms, bonds and symmetries. This is very useful for screening purposes, because often costly computations are required to determine the spatial structure.

Another problem in virtual screening of molecules and materials is to come up with new candidates that have the desired properties, e.g. come up with molecules for polymer solar cells that effectively absorbs the light from the sun. For this purpose we also investigate generative models, i.e. models that learn to generate new molecules that looks like a given set. The generator learns its own internal representation of the molecules and this representation is then used to optimise the molecules towards the desired properties. We demonstrate the application of this method to the discovery of new molecules for polymer solar cells and accelerate the rate of useful molecules by a factor of five.

Principal supervisor: Associate Professor Mikkel N.Schmidt
Co-supervisor: Associate Professor Morten Mørup

(Chairman)  Professor Lars Kai Hansen, DTU Compute
Associate Professor Anatole von Lilienfeld, University of Basel. Switzerland
Associate Professor Sune Darkner, Universety of Copenhagen

Chairperson at defence:
Professor Ole Winther



tir 07 maj 19
9:00 - 10:00


DTU Compute



The Technical University of Denmark
Richard Petersen's Plads
Building 324, Room 240