Pocket-Based Generative Diffusion Model Accelerates Potent Influenza A Hemagglutinin Inhibitor Discovery

The deep generative model has recently advanced 3D chemical space exploration but overlooked the balance between target affinity and structural rationality, limiting their effectiveness in drug discovery. Herein, we established a novel dual conditional diffusion model (DCDM) that leveraged ligand-protein interaction features to refine 3D target-based molecular generation. DCDM exhibited superiority in enhancing predicted binding affinity while maintaining high structural rationality and diversity. Subsequently, we applied DCDM to optimize penindolone (PND), a marine-derived lead compound from our laboratory, targeting influenza A hemagglutinin (HA). Efficiently, a promising candidate (compound C2e) was successfully obtained from eight synthesized derivatives inspired by the DCDM-generated molecules, with a 26-fold higher affinity for HA. Notably, C2e exhibited a 10-fold decrease in IC50 compared with the parent compound PND. Further in vivo assessments demonstrated its potent antiviral activity and safety. All results indicate that DCDM is a valuable generative model, capable of accelerating drug development in real-world applications.深度生成模型（Deep Generative Model）近年来推动了三维化学空间探索的发展，但却忽视了靶点亲和力与结构合理性之间的平衡，这限制了其在药物发现领域的应用效能。据此，我们构建了一种新型双条件扩散模型（dual conditional diffusion model，DCDM），该模型借助配体-蛋白质相互作用特征，优化基于三维靶点的分子生成任务。DCDM在提升预测结合亲和力的同时，能够维持较高的结构合理性与分子多样性，展现出显著优势。随后，我们将DCDM应用于优化喷尼多酮（Penindolone，PND）——本实验室开发的一种海洋来源先导化合物，其靶点为甲型流感血凝素（Influenza A Hemagglutinin，HA）。依托DCDM生成的分子设计思路，我们成功合成8种衍生物并高效筛选得到一株极具开发潜力的候选化合物C2e，其对HA的亲和力较母体化合物提升26倍。值得注意的是，与母体化合物PND相比，C2e的半数抑制浓度（IC50）降低了10倍。进一步的体内评价实验证实，C2e具备强效的抗病毒活性与良好的安全性。综上所有实验结果表明，DCDM是一款极具应用价值的生成模型，能够切实加速实际场景中的药物研发进程。