Graph Neural Networks for predicting molecule activity

Details

Supervisors

Nijssen, Siegfried

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en science des données, orientation technologies de l'information, à finalité spécialisée

Abstract

The development of new medicines involves an expensive and long process. AI and in particular Deep Learning techniques have caused a technological and scientific leap in many domains. This thesis explores how deep learning models are used to predict molecular properties and assist drug development. Specifically, it focuses on the graph representation of molecules and graph neural networks. Traditional AI techniques are well ahead at dealing with common formats like images, text, but graph data structures remain a research question. To evaluate the interest of working with graph structures and neural networks in the cheminformatics domain, a comparison study based on a common benchmark dataset for HIV drug development was executed. Classic techniques, GNNs, and alternative approaches are compared. Also, the impact of transfer learning strategies is tested to assess its potential by evaluating the accuracy gains. There exist great flexibility and potential to work with graphs directly on neural networks. However, the results don’t show drastic performance gains, and they remain similar to classical cheminformatics approaches. The results suggest that the use of pretraining strategies like supervised and self-supervised transfer learning can improve accuracy.