DRAGON-3D: an electron-density-aware generative model for dual-target drug molecule generation

Dual-target drug molecule generation is a significant computational challenge due to the structural constraints imposed by two protein binding pockets. Existing generative models typically rely on scarce dual-target drug datasets and struggle to capture three-dimensional (3D) structural and physicochemical interactions between protein pockets and ligands, limiting generalization to new target pairs. To address these limitations, we propose DRAGON-3D, a general dual-target drug molecule generation model based on latent-space fusion of 3D electron-density representations of diverse target pairs. Our model employs a novel bi-Net β-VAE architecture and an innovative dual-pretraining strategy, enabling dual-target molecule generation using only single-target data. By integrating 3D spatial electron-density with 2D molecular topology, DRAGON-3D enhances generalization and spatial compatibility across diverse dual-target protein pairs. Comprehensive benchmarks show that DRAGON-3D is a promising framework for the generation of candidate dual-target molecules, particularly in molecular diversity, drug-likeness (QED), deployability, generalizability, and interpretability. Overall, DRAGON-3D provides an efficient, interpretable, and electron-density-informed paradigm for de novo dual-target molecular generation.