SiDGen: Structure-informed Diffusion for Generative modeling of Ligands for Proteins

SiDGen, introduced on November 13, 2025, represents a breakthrough in computational drug discovery by tackling the persistent challenge of designing ligands that are both chemically valid and structurally compatible with protein binding pockets. Traditional methods often overlook structural context or are hindered by high memory demands, limiting their scalability. SiDGen integrates masked SMILES generation with lightweight folding-derived features to enhance pocket awareness. It offers two conditioning pathways to balance expressivity and efficiency, ensuring that ligands generated are not only unique but also maintain high validity and desirable molecular properties. The framework's innovative coarse-stride folding mechanism mitigates memory costs, allowing for training on realistic sequence lengths. Furthermore, in-loop chemical validity checks and an invalidity penalty ensure learning stability. SiDGen's competitive performance in docking-based evaluations positions it as a promisi…