Reliability of Machine Learning Approaches Towards the Identification of Novel PI4KIIIβ Inhibitors.

INTRODUCTION: The type III Phosphatidylinositol 4-kinase (PI4KIIIβ) serves as a key host factor that mediates the replication of RNA viruses and is involved in essential membrane trafficking events. The development of PI4KIIIβ-specific inhibitors could give insights into the enzyme's function and contribute to the treatment of viral diseases. However, the molecular features that determine inhibitor activity and selectivity remain unclear. METHODS: PI4KIIIβ inhibitors with reported biological activities (pIC50) were systematically analyzed using machine-learning methods. Molecular characterization was performed using RDKit, Chemopy descriptors, and PubChem fingerprints. Multiple classification and regression models were developed and evaluated using 10-fold cross-validation to establish robust structureactivity relationships. RESULTS: It was suggested that the PubChem fingerprint-based linear regression, SMOreg, and GaussianProcesses models were more appropriate for the studied dataset to deliver the best predictive ability, which showed a correlation coefficient value > 0.7 in the 10-fold crossvalidation. PubChemFP389 and PubChemFP572, representing amidine-like fragments or guanidine- related motifs, and the aromatic amine fragment showed a negative correlation with respect to biological activity. DISCUSSION: The study describes predictive models and provides insights into essential structural features in chemical compounds relevant for inhibitory activity, which could assist in the design and optimization of novel PI4KIIIβ inhibitors for antiviral and anticancer applications. A similar approach may be extended to a larger dataset of molecules or to other related drug targets. CONCLUSIONS: The findings suggest that excessive basicity and certain polar functionalities may be detrimental to PI4KIIIβ inhibition, providing important insights into the structural features influencing ligand binding and activity.