通过注意力增强图形神经网络预测对化剂的毒性,并从基线毒性转移学习
Kunyang Zhang1,2, Philippe Schwaller3,4, Kathrin Fenner1,2
1Department of Environmental Chemistry, Eawag, 8600 Dübendorf, Switzerland.
Environmental science & technology
|February 27, 2025
概括
使用图形神经网络 (GNN) 的新机制引导转移学习策略有效预测化学物质对环境的影响. 这种方法识别了关键的有毒子结构,改进了化学风险评估和安全设计原则.
科学领域:
- 环境化学环境化学
- 计算毒理学计算毒理学
- 机器学习用于化学安全.
背景情况:
- 实验性生态毒性测试是耗时且昂贵的.
- 图形神经网络 (GNN) 为生态毒性提供了预测能力,但可能会在小型数据集上过度匹配.
- 机械洞察对于理解和减轻化学物质对环境的影响至关重要.
研究的目的:
- 为GNN开发一种高效的机制引导转移学习策略,以预测化学生态毒性.
- 为模型预训练利用基线毒性和化剂毒性之间的机械联系.
- 提高GNN的解释性,以识别关键的有毒子结构及其贡献.
主要方法:
- 在脂友性 (log P) 数据上预训练一个GNN,并在化剂毒性数据上进行微调.
- 使用一种机制引导的转移学习方法.
- 采用调整的多头注意力和调整的沙普利值方法,以提高GNN的解释性和高效的基层结构贡献量化.
主要成果:
- 实现了与在较大,不太相关的数据集上预先训练的GNN可比的预测性能.
- 成功确定和量化关键子结构与基线毒性和化剂中毒性作用的特定模式相关的贡献.
- 突出显示的子结构与已知的结构性毒性警报保持一致.
结论:
- 拟议的策略使化学生态毒性的有效和机械信息预测成为可能.
- 增强的GNN解释性有助于发现各种生态毒性终点的新结构性警报.
- 这种方法支持改进化学风险评估,并促进安全设计原则.
相关概念视频
01:24Drug Toxicity: Dose-Dependent Reactions
245
Drug toxicities can be stratified into pharmacological, pathological, or genotoxic based on their mechanisms. The incidence and severity of these toxicities generally increase with the drug's concentration in the body and exposure time.Pharmacological toxicity is evident when the therapeutic effects of drugs overshoot into adverse reactions in a predictable, dose-dependent manner. Central nervous system (CNS) depression from barbiturates is a classic example, with effects escalating from...
245
01:23Toxicity Testing in Animals
222
Toxicity tests in animals are grounded on two main assumptions: first, the effects observed in laboratory animals can be extrapolated to humans, especially when adjusted for body surface area; second, high-dose exposure in animals is essential to identify potential human hazards from lower doses. This is based on the quantal dose-response concept, which faces the challenge of extrapolating results from relatively few test animals to much larger human populations. For example, a 0.01% incidence...
222