Whole-Graph Embedding and Adversarial Attacks for Life Sciences

Networks provide a suitable model for many scientific and technological problems that require the representation of complex entities and their relations. Life sciences applications include systems biology, where molecular components are represented in integrated systems in which the interactions among them provide richer information than single components taken separately, or neuroimaging, where brain networks allow representing the connectivity between different brain locations. In the examples we focus on, a set of networks is available, with each network representing an entity (e.g., a molecule, a macro molecule, or a patient) and links expressing their relation in the chemical/biological domain. The growing size and complexity of biomedical networks and the high computational complexity of graph analysis methods have lead to the investigation of the so-called whole-graph embedding techniques. Here, graphs are projected into lower dimensional vector spaces, while retaining their structural properties, allowing to reducing the data complexity at the same time keeping the topological and structural information. These techniques are showing very promising results in terms of their usability and potential. However, little research has focused on the analysis of their reliability and robustness. This need is strongly felt for real world applications, where corrupted data, either due to acquisition noise or to intentional attacks, could lead to misleading conclusions for the task at hand. Our objective here is to investigate on the adoption of adversarial attacks to whole-graph embedding methods for evaluating their robustness for classification in applications of interest for life sciences.