• graphs
• generative models
• bioinformatics
• cheminformatics
• structured data
• deep learning

The application of Deep Learning models to complex biological problems has recently revolutionized the field of the life sciences, leading to advancements that could potentially change the quality and length of human life for the better. A significant contribution to this success comes from the ability of deep neural networks to well approximate non-linear biological functions and their flexibility to learn from structured biological data directly. This thesis presents two relevant applications where Deep Learning is used on biological graphs to learn complex tasks. The first concerns the prediction of dynamical properties of chemical reactions at the cellular level represented as Petri graphs. Our contribution is a Deep Learning model that can learn the task solely relying on the structure of such graphs, which is orders of magnitude faster than running expensive simulations. The second application is about accelerating the drug design process by discovering novel drug candidates. We present a deep and generative framework in which novel graphs, corresponding to molecules with desired characteristics, can be obtained by combining chemically meaningful molecular fragments. Our work suggests that coupling the power of Deep Learning with the ability to handle structured data remains one of the preferred avenues to pursue towards solving well-known biological problems, as well as an effective methodology to tackle new ones.