Iterative Interaction Fingerprints-Guided Multiobjective Molecular Generation

Structure-based molecular generation has made substantial progress in recent years, yet methods for multiobjective optimization remain lacking. Here, we introduce IFPGen, a framework for interaction fingerprints-guided multiobjective molecular generation. It integrates a conditional diffusion model for molecular generation guided by interaction fingerprints within an iterative optimization loop that refines the search for molecules with an optimal balance across multiple objectives. On test sets, IFPGen outperforms baseline models in generating molecules with interaction patterns that closely resemble those of reference ligands. IFPGen effectively achieves multiobjective optimization by dynamically updating reference ligands and interaction patterns during the optimization process. IFPGen was employed for lead optimization of an inhibitor targeting human glutaminyl cyclases, leading to the identification of a series of new inhibitors. The most potent inhibitor exhibited a substantial improvement in potency, likely due to the establishment of predefined hydrogen-bonding interactions.