Graph Neural Networks for Advanced Molecular Data Analysis

Niccolò Pancino's Ph.D. Thesis, Ph.D. program in Smart Computing, Cycle XXXV, discussed on 29/5/2023. Graphs are data structures composed of collections of nodes and edges, which can be used to represent objects, or patterns, along with their relationships. Deep Learning techniques, and in particular Deep Neural Networks, have recently known a great development and have been employed in solving tasks of increasing complexity and variety. In particular, Graph Neural Networks (GNNs) have been extensively studied in the last decade, with many theoretical and practical innovations. Indeed, their main feature is the capability of processing graphs with minimal loss of structural information, which has caused GNNs to be applied to an increasing number of problems of different nature, leading to the development of new theories, models, and techniques. In particular, biological data proved to be a very suitable application field for GNNs, with metabolic networks, molecules, and proteins representing just few examples of data that are naturally encoded as graphs. In this thesis, a software framework for implementing GNN models was developed and discussed. Furthermore, some applications of GNNs to molecular data, relevant both from the point of view of Deep Learning and Bioinformatics, are discussed. The main focus of the work is on the drug side–effect prediction problem. This is a challenging task, since predictions can be based on homogeneous as well as heterogeneous and complex data, with graphs collecting nodes and edges representing different entities and relationships. On the other side, protein–protein interfaces can be detected by identifying the maximum clique in a correspondence graph of protein secondary structures, a problem which can be solved with Layered Graph Neural Networks (LGNNs). Promising experimental outcomes offer valuable insights and permit drawing interesting conclusions about the abilities of GNNs in analyzing molecular data. Given the growing interest of the AI research community on graph–based models, these applications, inspired by real–world problems, constitute a very good testing ground for evaluating GNN computational capabilities, in order to improve and evolve the actual models and extend them to more complex tasks, particularly in the biological field.