Computational prediction of regioselectivity in the metabolism of xenobiotics

优化药物、化妆品和农用化学品的代谢特性的策略通常依赖于代谢酶的区域选择性的知识。计算方法在预测代谢不稳定的原子位置（代谢位点; SoMs）方面具有巨大的潜力，但仍处于发展的早期阶段，在适用性、准确性、可解释性和可用性方面具有深远的局限性。该项目的目的是系统地研究和开发新的计算方法，克服这些限制，并允许准确预测各种化合物，酶和物种的SoMs。一个新的和全面的高质量的数据集的底物代谢酶和他们的专业分配的机制SoMs将探索模型开发的第一次。一系列不同的机器学习分类器将在一组精心设计的有物理意义的原子描述符上进行训练。这将导致这些模型和方法的适用性和范围的实质性扩展，从细胞色素P450介导的代谢到1 + 2相代谢的完全覆盖，以及从药物样分子到广泛的外源性物质及其代谢物。对于大多数现有的SoM预测器观察到的有问题的假阳性预测率将通过不同类型的模型和进一步策略的整合和组合来抵消。将研究估计预测误差的新方法和适用范围。重要的是，还将开发用于将生物转化类型分配给SoMs和定性估计代谢物丰度的模型，从而提供关于可能代谢物的化学结构和相关性的有价值的额外信息。这些模型将接受严格的评价，包括由独立实验室进行的肝细胞测定的前瞻性验证。

Strategies for optimising the metabolic properties of drugs, cosmetics and agrochemicals often rely on the knowledge of the regioselectivity of metabolising enzymes. Computational methods have an enormous potential for predicting metabolically labile atom positions (sites of metabolism; SoMs) but are at an early state of development and have far-reaching limitations with respect to applicability, accuracy, interpretability and availability. The aim of this project is the systematic research and development of new computational methods that overcome these limitations and allow the accurate prediction of SoMs for a large variety of compounds, enzymes and species. A new and comprehensive high-quality dataset of substrates of metabolising enzymes and their expertly assigned mechanistic SoMs will be explored for model development for the first time. A range of different machine learning classifiers will be trained on an elaborate set of physically meaningful atomic descriptors. This will lead to a substantial expansion of the applicability and scope of these models and methods, from cytochrome P450-mediated metabolism toward full coverage of phase 1 + 2 metabolism, and from drug-like molecules to a broad range of xenobiotics and their metabolites. Problematic false positive prediction rates observed for most of the existing SoM predictors will be countered by the integration and combination of different types of models and further strategies. New methods for estimating prediction errors and the applicability domain will be investigated. Importantly, models for the assignment of biotransformation types to SoMs and qualitative estimation of metabolite abundance will also be developed, hence providing valuable additional information on the chemical structure and relevance of likely metabolites. The models will be subject to rigorous evaluation including the prospective validation in hepatocyte assays by an independent laboratory.