Drug side effects (DSEs), or adverse drug reactions (ADRs), constitute an important health risk, given the approximately 197,000 annual DSE deaths in Europe alone. Therefore, during the drug development process, DSE detection is of utmost importance, and the occurrence of ADRs prevents many candidate molecules from going through clinical trials. Thus, early prediction of DSEs has the potential to massively reduce drug development times and costs. In this work, data are represented in a non-euclidean manner, in the form of a graph-of-graphs domain. In such a domain, structures of molecule are represented by molecular graphs, each of which becomes a node in the higher-level graph. In the latter, nodes stand for drugs and genes, and arcs represent their relationships. This relational nature represents an important novelty for the DSE prediction task, and it is directly used during the prediction. For this purpose, the MolecularGNN model is proposed. This new classifier is based on graph neural networks, a connectionist model capable of processing data in the form of graphs. The approach represents an improvement over a previous method, called DruGNN, as it is also capable of extracting information from the graph-based molecular structures, producing a task-based neural fingerprint (NF) of the molecule which is adapted to the specific task. The architecture has been compared with other GNN models in terms of performance, showing that the proposed approach is very promising.

Lensless cameras disregard the conventional design that imaging should mimic the human eye. This is done by replacing the lens with a thin mask, and moving image formation to the digital post-processing. State-of-the-art lensless imaging techniques use learned approaches that combine physical modeling and neural networks. However, these approaches make simplifying modeling assumptions for ease of calibration and computation. Moreover, the generalizability of learned approaches to lensless measurements of new masks has not been studied. To this end, we utilize a modular learned reconstruction in which a key component is a pre-processor prior to image recovery. We theoretically demonstrate the pre-processor's necessity for standard image recovery techniques (Wiener filtering and iterative algorithms), and through extensive experiments show its effectiveness for multiple lensless imaging approaches and across datasets of different mask types (amplitude and phase). We also perform the first generalization benchmark across mask types to evaluate how well reconstructions trained with one system generalize to others. Our modular reconstruction enables us to use pre-trained components and transfer learning on new systems to cut down weeks of tedious measurements and training. As part of our work, we open-source four datasets, and software for measuring datasets and for training our modular reconstruction.