Depriving cancer cells of sufficient NAD levels, mainly through interfering with their NAD-producing capacity, has been conceived as a promising anti-cancer strategy. Numerous inhibitors of the NAD-producing enzyme, nicotinamide phosphoribosyltransferase (NAMPT), have been developed over the past two decades. However, their limited anti-cancer activity in clinical trials raised the possibility that cancer cells may also exploit alternative NAD-producing enzymes. Recent studies show the relevance of nicotinic acid phosphoribosyltransferase (NAPRT), the rate-limiting enzyme of the Preiss–Handler NAD-production pathway for a large group of human cancers. We demonstrated that the NAPRT inhibitor 2-hydroxynicotinic acid (2-HNA) cooperates with the NAMPT inhibitor FK866 in killing NAPRT-proficient cancer cells that were otherwise insensitive to FK866 alone. Despite this emerging relevance of NAPRT as a potential target in cancer therapy, very few NAPRT inhibitors exist. Starting from a high-throughput virtual screening approach, we were able to identify and annotate two additional chemical scaffolds that function as NAPRT inhibitors. These compounds show comparable anti-cancer activity to 2-HNA and improved predicted aqueous solubility, in addition to demonstrating favorable drug-like profiles.

The development of anticancer therapies has increasingly relied on advanced 3D in vitro models, which more accurately mimic the tumor microenvironment compared to traditional 2D cultures. This review describes the evolution of these 3D models, highlighting significant advancements and their impact on cancer research. We discuss the integration of machine learning (ML) and artificial intelligence (AI) in enhancing the predictive power and efficiency of these models, potentially reducing the dependence on animal testing. ML and AI offer innovative approaches for analyzing complex data, optimizing experimental conditions, and predicting therapeutic outcomes with higher accuracy. By leveraging these technologies, the next generation of 3D in vitro models could revolutionize anticancer drug development, offering effective alternatives to animal experiments.

Beatrice Costa was not included as a first co-author in the original publication

Protein–protein interactions (PPIs) are increasingly recognised as therapeutically actionable targets, yet the discovery of small-molecule PPI inhibitors remains challenging because PPI interfaces are often large, shallow, dynamic and structurally heterogeneous. Machine learning (ML) approaches have shown promise for PPI inhibitor prediction, but most models rely on ligand-centred descriptors and may therefore capture scaffold similarity rather than compatibility with a specific protein–protein interface. In this study, we evaluated a structure-aware ML framework for the prediction of small-molecule PPI inhibitors by comparing ligand-only Extended Connectivity Fingerprints (ECFPs) with Protein–Ligand Extended Connectivity (PLEC) fingerprints. ECFP-based models showed high sensitivity on the external labelled prediction set, while PLEC-based models behaved more conservatively, recovering fewer positives but showing stronger rejection of ChEMBL-derived negatives. Family-specific ECFP models further demonstrated that PPI inhibitors could often be assigned to the correct PPI family, particularly when additional negative examples were included during training. Leave-one-family-out analyses revealed that extrapolation to unseen PPI families was strongly family-dependent: ECFP descriptors provided broader generalisation across several families, whereas PLEC descriptors improved recovery in selected cases. Overall, these results indicate that ligand-only and structure-aware descriptors provide complementary information for PPI inhibitor prediction.