Zhou Yao, Sang Zhipei, Xu Chao, Cao Ze, Xiao Kaixiang, Jia Qian, He Yutao, Luo Haibin, Huang Shuheng
Key Laboratory of Tropical Biological Resources of Ministry of Education and Hainan Engineering Research Center for Drug Screening and Evaluation, School of Pharmaceutical Sciences, Hainan University, Haikou, 570228, China.
Mol Divers. 2025 Sep 4. doi: 10.1007/s11030-025-11342-z.
Lead optimization is a crucial step in drug design. Generative AI-driven molecular modification has emerged as a powerful strategy to accelerate lead optimization by efficiently exploring chemical space and enhancing key drug-like properties. However, current AI tools primarily focus on de novo scaffold design rather than targeted modifications of validated lead compounds, limiting their practical utility in medicinal chemistry. Herein, we developed MolMod ( http://software.tdd-lab.com/molmod ), a web-based platform that enables site-specific molecular modifications through fragment-based optimization. MolMod employs a transformer model trained on 8.3 million ZINC20 compounds and fine-tuned with ~30,000 medicinal chemistry fragments from ChEMBL. Users mark specific modification sites on their molecules, and the model generates property-optimized fragments for these positions. The platform achieves high scaffold retention while maintaining a ≥99.99% fragment assembly success rate across extensive validation tests. Single-property optimization achieved >93% success rates, while multi-property constraints maintained 95% accuracy. Experimental validation confirmed the platform's accuracy: optimization of α-mangostin increased aqueous solubility from <5 μg/mL to 789 μg/mL through single-site modification, closely matching computational predictions (LogS: -6.128 to -3.829). MolMod provides ADMET profiles for all generated molecules and enables real-time visualization of structural modifications. By focusing on site-specific modifications rather than de novo generation, MolMod aligns with medicinal chemistry workflows and provides a practical tool for both computational and experimental scientists.
先导化合物优化是药物设计中的关键步骤。生成式人工智能驱动的分子修饰已成为一种强大的策略,可通过有效探索化学空间和增强关键的类药性质来加速先导化合物优化。然而,当前的人工智能工具主要侧重于从头开始的骨架设计,而非对已验证的先导化合物进行靶向修饰,这限制了它们在药物化学中的实际应用。在此,我们开发了MolMod(http://software.tdd-lab.com/molmod),这是一个基于网络的平台,可通过基于片段的优化实现位点特异性分子修饰。MolMod采用了一个在830万个ZINC20化合物上训练的变压器模型,并使用来自ChEMBL的约30000个药物化学片段进行了微调。用户在其分子上标记特定的修饰位点,该模型会为这些位置生成性质优化的片段。在广泛的验证测试中,该平台在保持≥99.99%的片段组装成功率的同时,实现了高骨架保留率。单性质优化的成功率超过93%,而多性质约束的准确率保持在95%。实验验证证实了该平台的准确性:通过单点修饰,α-山竹黄酮的水溶性从<5μg/mL提高到789μg/mL,与计算预测结果(LogS:-6.128至-3.829)密切匹配。MolMod为所有生成的分子提供ADMET概况,并能够实时可视化结构修饰。通过专注于位点特异性修饰而非从头生成,MolMod与药物化学工作流程相一致,为计算科学家和实验科学家提供了一个实用工具。