Diffusion Model Generates Developable 3D Drug Molecules with High Affinity

Researchers have introduced conDitar-dev, a novel conditional diffusion-based framework designed to accelerate drug discovery by generating developable 3D molecules. This system addresses limitations in existing structure-based drug design (SBDD) methods, which often prioritize binding affinity over crucial developability properties like ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity). conDitar-dev integrates three key modules: a multi-scale pocket representation learning module (msPRL), a pocket-conditioned diffusion model (conDitar), and a generation-time optimization method for ligand developability (paOPT). The framework was rigorously tested on a new benchmark of human disease targets, where conDitar significantly outperformed state-of-the-art SBDD baselines, achieving an impressive average binding score. Crucially, conDitar-dev improved ADMET property performance by up to 73% over its base conDitar model, demonstrating its ability to generate more "drug-like" molecules. Further validation involved applying conDitar-dev to two known druggable targets, PD-L1 and CSF1R proteins. Experimentally synthesized molecules generated by conDitar-dev for PD-L1 showed strong binding affinities, and hit expansion based on its designs led to the identification of selective CSF1R inhibitors with very low IC50 values. This research highlights conDitar-dev's potential to not only discover new drug candidates but also to reposition existing ones, significantly streamlining the early stages of drug development.