| 引用本文: | 陆桂文,刘宏德,罗坤.基于靶点结构和分子片段的药物分子[]设计神经网络模型[J].生物信息学,2026,24(2):114-126. |
| LU Guiwen,LIU Hongde,LUO Kun.Neural network modeling for drug molecular design based on target structures and molecular fragments[J].Chinese Journal of Bioinformatics,2026,24(2):114-126. |
|
| 摘要: |
| 针对蛋白质靶点设计适配分子是一项具有挑战性的任务。近年,深度生成模型成功应用于新药分子设计。从原子等级构建分子的模型容易生成不符合化学规律的结构；开发容易合成、亲和力高的分子设计模型依然是该领域的重要研究内容。本文开发了一种面向靶点口袋的药物分子设计模型,BFMOL。模型由两个编码器和一个门控自注意力解码器构成；编码器用于提取靶点和配体分子的特征,解码器实现两种特征的关联并解码为分子片段概率,进而生成配体分子。模型通过BRICS方法对分子进行片段化,并在原子、残基和残基间三个层次对靶点口袋进行特征化。结果表明,在Crossdocked数据集上,BFMOL设计的分子,其平均合成可及性得分达到0.769,平均类药性得分达到0.607,高于PMDM模型(0.611和0.594)。在针对激酶类靶点的分子设计中,模型生成的分子具备更多双六元环结构,具有激酶抑制剂的典型特征。在子结构分析中,模型生成的分子具有与真实分子更接近的环结构和键角分布。总之,本文开发了一种更加合理的面向靶点口袋的潜在药物分子设计的神经网络模型。 |
| 关键词: 深度学习 药物设计 分子生成 |
| DOI：10.12113/202501013 |
| 分类号:Q811.4 |
| 文献标识码:A |
| 基金项目:“天山英才”项目(No.2023TSYCCJ0030). |
|
| Neural network modeling for drug molecular design based on target structures and molecular fragments |
|
LU Guiwen1, LIU Hongde1,LUO Kun2
|
|
(1.School of Biological Science and Medical Engineering, Southeast University, Nanjing 211106, China;2.Department of Neurosurgery, The Second Affiliated Hospital of Xinjiang Medical University, Urumqi 830063, China)[HJ1.5mm]
|
| Abstract: |
| Designing adaptor molecules against protein targets is a challenging task. In recent years, deep generative models have been successfully applied to the molecular design of new drugs. Models that construct molecules from the atomic level are prone to generate structures that do not conform to the chemical laws; the development of molecular design models that are easy to synthesize and have high affinity remains an important part of research in this field. In this paper, we develop a target protein-oriented drug molecular design model, BFMOL. The model consists of two encoders and a gated self-attentive decoder; the encoder is used to extract the features of the target and ligand molecules, and the decoder realizes the association of the two features and decodes them into the probability of the molecular fragments, which then generates the ligand molecules. The model fragments the molecule by BRICS method and characterizes the target at three levels: atom, residue and inter-residue. The results show that on the Crossdocked dataset, the molecules designed by BFMOL have an average synthetic accessibility score of 0.769 and an average druggability score of 0.607, which are higher than the baseline model PMDM (0.611 and 0.594). In the design of molecules against kinase-like targets, the model-generated molecules possessed more double six-membered ring systems with typical features of kinase inhibitors. In substructure analysis, the model-generated molecules have the closest ring structure and bond angle distribution to the real molecules. In conclusion, a more rational neural network model for potential drug molecule design toward target proteins was developed in this work. |
| Key words: Deep learning Drug design Molecular generation |
