Abstract
We present the Full-Spectrum Generative Lead Discovery (FSGLD), a deep learning-driven pipeline for efficient drug lead identification. FSGLD integrates generative modeling with molecular docking, molecular dynamics simulations, ligand-residue interaction profile, MM-PBSA, thermodynamic integration (TI), and experimental validation to bridge theoretical design and practical application. The core multiscale DRUG-GAN models enable de novo design for both drug-like and target-specific compounds across three scenarios: I. generation of random drug-like compounds, II. generation of target-specific compounds, III. generation of target-biased compound series featuring shared chemical structures. FSGLD significantly outperformed traditional computer-aided drug design methods in generating novel chemicals which specifically target the CB2 receptor. Additionally, a computational protocol for TI calculations was established to reduce computation time by 80-90% while maintaining accuracy. By integrating generative models with in silico and in vitro evaluation techniques, FSGLD reduces the cost of identifying novel yet viable lead compounds, offering remarkable benefits to both academic and industry.