从潜在的多元体到功能性探针:一个可解释的,基因组规模的生成机器学习框架,用于家庭向的基因酶抑制剂设计
bioRxiv : the preprint server for biology
|January 16, 2026
概括
生成性AI可以设计激酶抑制剂,但缺乏可解释性. 本研究介绍了SRC激酶抑制剂设计的框架,改善了支架转换,并揭示了复杂分子生成当前AI模型的局限性.
科学领域:
- 药用化学 医学化学
- 人工智能的人工智能
- 药物发现 药物发现 药物发现
背景情况:
- 设计选择性激酶抑制剂是具有挑战性的,因为保留了ATP结合部位.
- 生成性AI模型提供快速的化学空间探索,但往往缺乏可解释性.
研究的目的:
- 为*de novo* SRC激酶抑制剂设计开发一个模块化,可解释的生成框架.
- 解决人工智能驱动药物设计中的瓶问题,重点关注可解释性和化学逻辑.
主要方法:
- 整合了ChemVAE隐性空间建模,激酶抑制可能性 (KIL) 评分函数和贝叶斯优化.
- 利用集群指导的本地社区采样用于支架转换和化学空间的探索.
- 采用了一个模块化框架,将可解释的机器学习与生成的AI结合起来.
主要成果:
- 激酶抑制剂在潜空间中形成一个连贯的多重体,SRC作为支架转换的"枢纽".
- 成功将LCK抑制剂转化为新型SRC类化学型,其中LCK衍生分子占产出的40%左右.
- 发现了一个代表性差距:基于SMILES的生成与临床酶抑制剂至关重要的多环药相斗争.
结论:
- 结合可解释的ML和生成AI的混合方法对于透明和有效的药物设计至关重要.
- 需要具有拓意识的表示来克服复杂药理的当前生成模型的局限性.
- 在抑制器设计中,在导航化学空间时,无偏见的,集群引导的采样比积极偏见的优化更有效.
相关概念视频
02:54Protein Kinases and Phosphatases
4.4K
4.4K
02:54Protein Kinases and Phosphatases
15.0K
Proteins undergo chemical modifications that trigger changes in the charge, structure, and conformation of the proteins. Phosphorylation, acetylation, glycosylation, nitrosylation, ubiquitination, lipidation, methylation, and proteolysis are various protein modifications that regulate protein activity. Such modifications are usually enzyme-driven.
Protein kinases
Many proteins in the cell are regulated by phosphorylation, the addition of a phosphate group. A family of enzymes called kinases...
Protein kinases
Many proteins in the cell are regulated by phosphorylation, the addition of a phosphate group. A family of enzymes called kinases...
15.0K
01:27Transducer Mechanism: Enzyme-Linked Receptors
3.9K
Enzyme-linked receptors are cell-surface receptors acting as an enzyme or associating with an enzyme intracellularly. They make excellent drug targets. Drugs can bind to the extracellular ligand-binding domain or directly affect their enzymatic domain and alter their activity.
Major types that are helpful drug targets include:
Major types that are helpful drug targets include:
3.9K