Multi-Objective Molecular Design Enhanced by Conditional Normalizing Flows

Generative AI is transforming molecular discovery in drug design and materials science by exploring the vast and largely unexplored chemical space. However, current methods, including normalizing flows, struggle to balance multi-objective optimization and sampling speed, particularly when generating specific compound classes and more intricate scaffolds such as aromatic rings. This project aimed to create a generative model capable of efficiently sampling novel organo-phosphate molecules while optimizing drug-likeness, synthetic accessibility and chemical reactivity. To achieve this, we employed conditional normalizing flows with its base density conditioned on Quantitative Estimate of Drug-likeness (QED), Synthetic Accessibility (SA) scores and electronic density on the central phosphorus atom approximated by Hirschfeld charges calculated with density functional theory. To enhance scaffold diversity and simplify molecular representation, we utilized group Self-Referencing Embedded Strings (group-SELFIES) fragments from the ZINC-250k database. Our framework efficiently generated diverse organo-phosphates, demonstrating that combining conditional normalizing flows with group-SELFIES can address key limitations in the inverse molecular design. This paves the way towards targeted therapies and enables the optimization of complex molecular objectives.