We offer a 9-month internship. The project aims at combining graph neural networks and causal inference to address diverse problems in computational biology.
Graph neural networks and causal inferences represent two state-of-the-art approaches to connectionist and symbolic learning. Graph neural networks, based on knowledge encoded in graphs and neural networks, are powerful tools for knowledge injection in machine learning and for representation learning (see the previous post on using knowledge-graph for patient-level data). Causal inference, based on Directed Acyclic Graphs (DAGs) and statistical models, is a powerful to discern causal effects from correlations caused by confounding (see a tutorial here on GitHub).
Our research question is whether we can find a midway that combines the advantages of both approaches - powerful prediction on the one hand, interpretability and counterfactual queries on the other hand - to predict and to understand mechanism and safety profiles of drugs. The premise is that drugs exert their efficacy and invoke adverse effects by interacting with and modulating targets and off-targets, which are components of biological networks. By combining our limited knowledge of biological networks and observational data, we wish to gain a better understanding of how and why drugs induce certain phenotypes by leveraging both computational approaches.