So you have networks … now what?

When & Where:
3 Nov 6 PM — Phys Room 539 (Blackett)

Speaker profile:
Guadalupe González is a PhD candidate in Computing at Imperial College London, supervised by Michael Bronstein and Kirill Veselkov. Her research focuses on the application of graph machine learning to biomedical problems, in the context of drug discovery and development.

Abstract:

Networks are powerful representations of data at many levels in biology and medicine: from protein structure networks to protein-protein interaction networks and even patient-healthcare interaction networks.

If you’re not familiar with networks, network-structured data looks pretty much like any other dataset: a table with features for each sample or node to which you can apply simple linear models or even deep learning approaches to solve predictive tasks. If you’ve wondered, however, if it’s possible to process this data in a more principled way to obtain more expressive, powerful representations and higher predictive performance. The answer is yes.

In this session, we will define what networks are, and will present machine learning paradigms, from graph-theoretic techniques to graph neural networks, that work directly on them to solve predictive tasks. We will have a detailed look at 3 application works that have used these paradigms in biology, pushing the boundaries of scientific discovery: drug repositioning using a multiscale interactome, prediction of anticancer molecules within foods, and prediction of antibiotics for antibiotic-resistant bacteria. By the end of the session, you will know how to work with network-structured data, which algorithms to use for a given predictive task and the state-of-the-art from a research perspective.